Raven-1/tests/test_agents_code_flow_exec.py

807 lines
33 KiB
Python
Raw Permalink Normal View History

feat(docker): run raven in a docker container and make a first run configurable from the page ## Summary Two things arrive together because the second is what made the first usable. **Run raven from a container.** A Dockerfile builds the page, the Python environment and the engine into one image; `cd docker && docker compose up` serves it. nginx serves the built page and proxies the RPC socket to the engine over the loopback they share -- the engine binds 127.0.0.1 only, so a proxy in a second container would have no address to dial. Compose publishes on 127.0.0.1, and the entrypoint prints a one-time sign-in link on boot. **A first run is no longer a dead end.** The gateway used to refuse to start until a provider was configured, while being the thing that serves the page where a provider gets configured. It now starts, says what is missing, and picks up a key saved in Settings on the next turn without a restart. The onboarding overlay no longer takes the page either: New chat, the model chip and Send send the reader to Settings > Models instead, and an empty model list is marked on the control that fills it. Also carried: the provider catalogue registry (capabilities, modalities and context windows per model), the reworked Settings > Models pane, and four fixes found in review of that work -- a sibling catalogue answering 200 with a non-catalogue body no longer crashes the whole fetch; adding a tag no longer erases the label and description already stored for that model; `authenticated` no longer demands a bearer token from a local provider reached by address; and registry rows are keyed by the spelling every lookup produces, which recovers 593 rows that no reader could reach. ## Type - [ ] Fix - [x] Feature - [x] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification Commands run on the branch head, with their results: - `uv run pytest tests -q -k "provider or registry or catalog or model or rpc or gateway or docker or i18n"` -- 4389 passed, 46 skipped - `npx vitest run` in `ui-web` -- 1573 passed (101 files) - `npx vitest run` in `ui-tui` -- 2037 passed (142 files) - `uv run ruff check raven scripts tests` and `ruff format --check` -- clean - `npx tsc --noEmit` in `ui-web` and `npm run type-check` in `ui-tui` -- clean - `npm run lint:rpc` in `ui-tui` -- generated.ts in sync - `make check-source-language`, `make check-large-files` -- clean Each of the four review fixes was checked by reverting it and watching a named test go red, then restoring it. The container image itself has not been built by CI or by me -- no Docker daemon was available on the machine this was written on. It has been built and run by the author of the branch. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [x] User-facing docs or screenshots are updated when needed ## Risk Two deliberate security trade-offs in the container deployment, both documented in `docker/README.md`: - nginx rewrites the `Origin` header, because the engine accepts only an Origin naming its own loopback port and no browser reaching a published port can send one. This drops a defence-in-depth layer. The session cookie is SameSite=Strict, so an unauthenticated socket is still refused -- but SameSite ignores ports, so another page on localhost counts as same-site. - `RAVEN_AUTO_LOGIN=1` is the default: nginx attaches the session cookie to every page load, so anyone who can reach the published port is signed in. It is contained by Compose publishing on 127.0.0.1 only. Set it to 0 for any deployment reachable from elsewhere. These are two settings that must agree, and nothing warns when they do not. Behaviour changes outside the container: the gateway now starts without a configured provider rather than exiting 1, and the first-run onboarding overlay no longer opens by default (`?onboard=1` still forces it). Rollback is by revert; nothing here migrates state or changes an on-disk format. - [x] Security impact considered - [x] Backward compatibility considered - [x] Rollback path is clear for risky changes ## Related Issues N/A
2026-09-11 03:27:52 +00:00
"""The code flow's exec: trunk's shell tool with the fork's fixes.
feat(acp): a prompt is held open while the agent sleeps on an armed wake, and takes a steer ## Summary An on-call agent submits a job, arms a wake and ends its turn saying so; the wake runs a later turn on the session on its own clock. Over ACP that turn's `end_turn` read as "finished" to the caller: a DAG judging its node spent every continuation on "still watching, wake at 13:13" while the job ran fine (ar-r4, 2026-09-08), and the main agent re-dispatched the work by hand. ACP has no stop reason for "waiting". Replaces the closed !573: **no DAG code changes**, no new node status. The node is `running` because a turn is genuinely in flight. - **The signal.** The loop-hooks paper gains one key the ACP layer reads for itself: `acp_meta["raven.holdTurn"]` (`{"untilMs", "why"}`). oncall-flow files it at the send fire when the window's campaign is claimed, not concluded, and has one pending wake -- read the same way the owner's answer pulls that wake forward. No campaign, a wake that stood down, or a concluded campaign files nothing and the turn ends as before. - **The hold** (`raven/acp/methods.py`). A turn that ended `end_turn` with the key filed is not answered. One `tool_call` stands for the wait (its title names the campaign and the next look), with a `tool_call_update` heartbeat every `HOLD_HEARTBEAT_S` (60 s) so a client that clocks liveness by frames sees the agent alive while its model sleeps. The turn slot is reopened with no id to correlate on, so whichever turn ends next on the session -- the wake's, a steer's -- settles it, and the handler decides again from that turn's stash. A wake that arms again keeps the hold; a turn ending with nothing pending answers `end_turn`; a cancel answers `cancelled` as before. The key itself is not passed to the client. - **Steer.** `_raven/session/steer` is served and advertised under `agentCapabilities._meta` as `raven.steer` (the names raven's own client already uses). Accepted while a prompt is open on the session, held or running: the text is shown on the wire as the person's own line, then sent with `turn.send`'s new `busy: "inject"`. Answers `injected` / `no_turn`. - **Wire.** `turn.send` gains the optional `busy` param (`const "inject"`): an ordinary send on an idle lane, `BusyPolicy.INJECT` into the running turn otherwise -- the merge a channel's mid-turn message already gets (`gateway_commands.py`). No slot is bound for an injected message; the running turn keeps the lane's. `openrpc.json` and both generated clients moved together. What raven's client side already does, unchanged: the DAG stall watcher resets on any frame (the heartbeat keeps it quiet, and a stopped heartbeat raises the existing stall notice after ten minutes); a sub-agent process that dies fails the request and the node takes its ordinary path; the pool's `_responsive` treats a prompt in flight as busy. ## Review round 1 (2026-09-09, on !600, superseded by this MR) - **`busy: inject` reads the scheduler's lane state too.** `is_turn_active` knows only the turns `turn.send` started; an armed wake is submitted to the scheduler directly (`make_on_session_wake`) and never binds that map, so a steer during a running wake queued a second turn and answered `injected`. `turn.send` now asks `Scheduler.has_inflight(lane)` as well, and merges (`BusyPolicy.INJECT`) when either record says the lane is running. Pinned: `test_busy_inject_merges_into_a_wake_the_real_scheduler_is_running` -- the reviewer's reproduction on the real `Scheduler`, a turn submitted the way `make_on_session_wake` submits one; with the lane check bypassed it reads `pending=1, inject_mailbox=0`, with it `(1, 0)` in the mailbox and the running turn drains the text. - Branch renamed to snake_case and both commits carry the `Co-authored-by: Claude (<model-id>)` trailer, per AGENTS.md §2–3; that is why this MR replaces !600 rather than continuing it. ## Review round 2 (2026-09-10) - **A wake re-armed during its own run survives the writeback.** `CronService._writeback_locked` reloaded the store after the run and, for a one-shot with `delete_after_run`, overwrote the id with the finished run's state and removed it -- but the turn the wake ran had just called `schedule_wake` under the same key, so the id now named the *next* look. It vanished, and the held prompt had nothing left to settle it. A stored job under the fired id whose `at` schedule the run never had is treated as the new wake it is: claim cleared, run recorded, schedule kept, not deleted. Pinned on the real service: `test_a_wake_re_armed_during_its_own_run_survives_the_writeback` (reads `[]` without the fix, the reviewer's state) and the unchanged one-shot deletion beside it. Cron, proactive, sentinel and on-call suites: 968 passed. ## Review round 3 (2026-09-10) - **An undrained inject stays visible and cancellable.** `busy: inject` dropped the scheduler's handle; when the host turn ended past its last drain, the spine's fallback ran the text as a turn nothing owned -- `turn.cancel` a no-op, no `message.start`. The handle is now kept per lane under an id minted for the text; `turn.cancel` / `session.interrupt` reach it when the active slot is empty (mailbox, queued fallback, or running); and when the fallback turn starts, the RPC sink (`on_turn_start`, wired from bootstrap) binds the lane to it and opens it on the wire with the person's words. `turn.send` answers the inject's own id. Pinned through `build_rpc_spine` with the reviewer's shape (a loop that drains once, an inject after it): `test_an_undrained_inject_runs_as_a_turn_the_surface_can_see_and_cancel` and `test_a_pending_inject_can_be_cancelled_before_it_merges_or_falls_back`; both read the reviewer's state (`turn_ids` empty, cancel finding nothing) with the fix bypassed. - **Line ceiling.** The contract comment is five lines; `raven/contracts` is 2,519 against 2,520, one line of headroom on the current `main` (2,514). - Non-blocking items taken: `SchedulerDrainingError` on the inject path exits as `turn_failed`; the title is announced before the hold; the steer constants sit below the cancel constant's docstring; three edges pinned (a held prompt settled by a wake turn's uncorrelated ending through `send_frame`, the hold key never in the client's `_meta`, `raven.toolName` on the wait row). - Focused suite (reviewer's list plus the touched files): 948 passed on Python 3.12. ## Review round 4 (2026-09-10) - **Every pending inject is kept, per lane and per id.** One replaceable record per lane kept only the newest steer; two steers after the host's last drain fall back in turn, and the first started unowned while a cancel took out the second. The record is now `lane -> {inject id: (handle, text)}` in arrival order: each fallback promotes its own handle, and a cancel with no active turn reaches the oldest pending one -- the one the spine runs next. Pinned: `test_two_undrained_injects_each_get_their_own_turn_promoted_and_cancelled` (both promoted with their own ids and `message.start`, cancelled one after the other; runner saw all three turns). Focused suite: 949 passed. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification ``` uv run pytest tests/test_acp_methods.py tests/test_acp_protocol.py tests/test_rpc_turn_send.py \ tests/test_rpc_schema_match.py tests/test_acp_updates.py tests/test_acp_server.py -q # 766 passed uv run pytest tests/test_agents_oncall_flow_hooks.py tests/test_agents_oncall_flow_actions.py \ tests/test_agents_oncall_flow_budget.py tests/test_agents_oncall_flow_readings.py \ tests/test_agents_oncall_flow_escalation_tools.py -q # 99 passed uv run ruff check raven tests agents && uv run ruff format --check raven tests agents # clean uv run lint-imports # 9 kept, 0 broken cd ui-tui && npm run lint:rpc && npx tsc --noEmit; cd ui-web && npm run gen:check && npx tsc --noEmit # in sync, clean ``` New tests: `TestHeldTurn` (held across an armed wake, held again on a re-arm, heartbeat on the wait call, cancel while held, a non-`end_turn` stop is not held; every frame validated against the ACP schema), `TestSteer` (advertised, `no_turn` when idle, a held prompt takes the words as a turn with `busy: inject`, refusals), `busy: inject` on a busy and an idle lane in `test_rpc_turn_send.py`, and the plugin's hold stamping through the real `CronService` grant (filed with a pending wake; not filed without one, after the wake stands down, or once the campaign concluded). Not done: no end-to-end run of a real on-call wake through a held prompt yet -- unit tests only. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [ ] User-facing docs or screenshots are updated when needed (the hold is invisible to a person except as a "waiting for ..." tool row) ## Risk - [x] Backward compatibility considered: a turn that files no `raven.holdTurn` is answered exactly as before; `busy` is optional and absent from every existing caller; the steer method and capability are additive. - [x] Rollback path is clear: revert the two commits. A held prompt holds one `max_concurrent_subagents` slot (default 8) for the wait's duration on the caller's side; raised in config if it bites. ## Related Replaces !573 (closed). Follows !518 (stall notice) and !520 (unprompted-turn route; a wake that runs under a held prompt no longer takes it for DAG nodes).
2026-09-11 02:37:55 +00:00
A long timeout is clamped instead of refused; a timed-out command keeps the
feat(docker): run raven in a docker container and make a first run configurable from the page ## Summary Two things arrive together because the second is what made the first usable. **Run raven from a container.** A Dockerfile builds the page, the Python environment and the engine into one image; `cd docker && docker compose up` serves it. nginx serves the built page and proxies the RPC socket to the engine over the loopback they share -- the engine binds 127.0.0.1 only, so a proxy in a second container would have no address to dial. Compose publishes on 127.0.0.1, and the entrypoint prints a one-time sign-in link on boot. **A first run is no longer a dead end.** The gateway used to refuse to start until a provider was configured, while being the thing that serves the page where a provider gets configured. It now starts, says what is missing, and picks up a key saved in Settings on the next turn without a restart. The onboarding overlay no longer takes the page either: New chat, the model chip and Send send the reader to Settings > Models instead, and an empty model list is marked on the control that fills it. Also carried: the provider catalogue registry (capabilities, modalities and context windows per model), the reworked Settings > Models pane, and four fixes found in review of that work -- a sibling catalogue answering 200 with a non-catalogue body no longer crashes the whole fetch; adding a tag no longer erases the label and description already stored for that model; `authenticated` no longer demands a bearer token from a local provider reached by address; and registry rows are keyed by the spelling every lookup produces, which recovers 593 rows that no reader could reach. ## Type - [ ] Fix - [x] Feature - [x] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification Commands run on the branch head, with their results: - `uv run pytest tests -q -k "provider or registry or catalog or model or rpc or gateway or docker or i18n"` -- 4389 passed, 46 skipped - `npx vitest run` in `ui-web` -- 1573 passed (101 files) - `npx vitest run` in `ui-tui` -- 2037 passed (142 files) - `uv run ruff check raven scripts tests` and `ruff format --check` -- clean - `npx tsc --noEmit` in `ui-web` and `npm run type-check` in `ui-tui` -- clean - `npm run lint:rpc` in `ui-tui` -- generated.ts in sync - `make check-source-language`, `make check-large-files` -- clean Each of the four review fixes was checked by reverting it and watching a named test go red, then restoring it. The container image itself has not been built by CI or by me -- no Docker daemon was available on the machine this was written on. It has been built and run by the author of the branch. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [x] User-facing docs or screenshots are updated when needed ## Risk Two deliberate security trade-offs in the container deployment, both documented in `docker/README.md`: - nginx rewrites the `Origin` header, because the engine accepts only an Origin naming its own loopback port and no browser reaching a published port can send one. This drops a defence-in-depth layer. The session cookie is SameSite=Strict, so an unauthenticated socket is still refused -- but SameSite ignores ports, so another page on localhost counts as same-site. - `RAVEN_AUTO_LOGIN=1` is the default: nginx attaches the session cookie to every page load, so anyone who can reach the published port is signed in. It is contained by Compose publishing on 127.0.0.1 only. Set it to 0 for any deployment reachable from elsewhere. These are two settings that must agree, and nothing warns when they do not. Behaviour changes outside the container: the gateway now starts without a configured provider rather than exiting 1, and the first-run onboarding overlay no longer opens by default (`?onboard=1` still forces it). Rollback is by revert; nothing here migrates state or changes an on-disk format. - [x] Security impact considered - [x] Backward compatibility considered - [x] Rollback path is clear for risky changes ## Related Issues N/A
2026-09-11 03:27:52 +00:00
output it produced and says so in as many words; an oversized output is saved
whole under Agent home with the path in the result. Keeping the output is no
longer a difference -- trunk does that itself now -- so what this adds on that
axis is the explicit marker. And the sandbox discipline: the factory serves this
feat(acp): a prompt is held open while the agent sleeps on an armed wake, and takes a steer ## Summary An on-call agent submits a job, arms a wake and ends its turn saying so; the wake runs a later turn on the session on its own clock. Over ACP that turn's `end_turn` read as "finished" to the caller: a DAG judging its node spent every continuation on "still watching, wake at 13:13" while the job ran fine (ar-r4, 2026-09-08), and the main agent re-dispatched the work by hand. ACP has no stop reason for "waiting". Replaces the closed !573: **no DAG code changes**, no new node status. The node is `running` because a turn is genuinely in flight. - **The signal.** The loop-hooks paper gains one key the ACP layer reads for itself: `acp_meta["raven.holdTurn"]` (`{"untilMs", "why"}`). oncall-flow files it at the send fire when the window's campaign is claimed, not concluded, and has one pending wake -- read the same way the owner's answer pulls that wake forward. No campaign, a wake that stood down, or a concluded campaign files nothing and the turn ends as before. - **The hold** (`raven/acp/methods.py`). A turn that ended `end_turn` with the key filed is not answered. One `tool_call` stands for the wait (its title names the campaign and the next look), with a `tool_call_update` heartbeat every `HOLD_HEARTBEAT_S` (60 s) so a client that clocks liveness by frames sees the agent alive while its model sleeps. The turn slot is reopened with no id to correlate on, so whichever turn ends next on the session -- the wake's, a steer's -- settles it, and the handler decides again from that turn's stash. A wake that arms again keeps the hold; a turn ending with nothing pending answers `end_turn`; a cancel answers `cancelled` as before. The key itself is not passed to the client. - **Steer.** `_raven/session/steer` is served and advertised under `agentCapabilities._meta` as `raven.steer` (the names raven's own client already uses). Accepted while a prompt is open on the session, held or running: the text is shown on the wire as the person's own line, then sent with `turn.send`'s new `busy: "inject"`. Answers `injected` / `no_turn`. - **Wire.** `turn.send` gains the optional `busy` param (`const "inject"`): an ordinary send on an idle lane, `BusyPolicy.INJECT` into the running turn otherwise -- the merge a channel's mid-turn message already gets (`gateway_commands.py`). No slot is bound for an injected message; the running turn keeps the lane's. `openrpc.json` and both generated clients moved together. What raven's client side already does, unchanged: the DAG stall watcher resets on any frame (the heartbeat keeps it quiet, and a stopped heartbeat raises the existing stall notice after ten minutes); a sub-agent process that dies fails the request and the node takes its ordinary path; the pool's `_responsive` treats a prompt in flight as busy. ## Review round 1 (2026-09-09, on !600, superseded by this MR) - **`busy: inject` reads the scheduler's lane state too.** `is_turn_active` knows only the turns `turn.send` started; an armed wake is submitted to the scheduler directly (`make_on_session_wake`) and never binds that map, so a steer during a running wake queued a second turn and answered `injected`. `turn.send` now asks `Scheduler.has_inflight(lane)` as well, and merges (`BusyPolicy.INJECT`) when either record says the lane is running. Pinned: `test_busy_inject_merges_into_a_wake_the_real_scheduler_is_running` -- the reviewer's reproduction on the real `Scheduler`, a turn submitted the way `make_on_session_wake` submits one; with the lane check bypassed it reads `pending=1, inject_mailbox=0`, with it `(1, 0)` in the mailbox and the running turn drains the text. - Branch renamed to snake_case and both commits carry the `Co-authored-by: Claude (<model-id>)` trailer, per AGENTS.md §2–3; that is why this MR replaces !600 rather than continuing it. ## Review round 2 (2026-09-10) - **A wake re-armed during its own run survives the writeback.** `CronService._writeback_locked` reloaded the store after the run and, for a one-shot with `delete_after_run`, overwrote the id with the finished run's state and removed it -- but the turn the wake ran had just called `schedule_wake` under the same key, so the id now named the *next* look. It vanished, and the held prompt had nothing left to settle it. A stored job under the fired id whose `at` schedule the run never had is treated as the new wake it is: claim cleared, run recorded, schedule kept, not deleted. Pinned on the real service: `test_a_wake_re_armed_during_its_own_run_survives_the_writeback` (reads `[]` without the fix, the reviewer's state) and the unchanged one-shot deletion beside it. Cron, proactive, sentinel and on-call suites: 968 passed. ## Review round 3 (2026-09-10) - **An undrained inject stays visible and cancellable.** `busy: inject` dropped the scheduler's handle; when the host turn ended past its last drain, the spine's fallback ran the text as a turn nothing owned -- `turn.cancel` a no-op, no `message.start`. The handle is now kept per lane under an id minted for the text; `turn.cancel` / `session.interrupt` reach it when the active slot is empty (mailbox, queued fallback, or running); and when the fallback turn starts, the RPC sink (`on_turn_start`, wired from bootstrap) binds the lane to it and opens it on the wire with the person's words. `turn.send` answers the inject's own id. Pinned through `build_rpc_spine` with the reviewer's shape (a loop that drains once, an inject after it): `test_an_undrained_inject_runs_as_a_turn_the_surface_can_see_and_cancel` and `test_a_pending_inject_can_be_cancelled_before_it_merges_or_falls_back`; both read the reviewer's state (`turn_ids` empty, cancel finding nothing) with the fix bypassed. - **Line ceiling.** The contract comment is five lines; `raven/contracts` is 2,519 against 2,520, one line of headroom on the current `main` (2,514). - Non-blocking items taken: `SchedulerDrainingError` on the inject path exits as `turn_failed`; the title is announced before the hold; the steer constants sit below the cancel constant's docstring; three edges pinned (a held prompt settled by a wake turn's uncorrelated ending through `send_frame`, the hold key never in the client's `_meta`, `raven.toolName` on the wait row). - Focused suite (reviewer's list plus the touched files): 948 passed on Python 3.12. ## Review round 4 (2026-09-10) - **Every pending inject is kept, per lane and per id.** One replaceable record per lane kept only the newest steer; two steers after the host's last drain fall back in turn, and the first started unowned while a cancel took out the second. The record is now `lane -> {inject id: (handle, text)}` in arrival order: each fallback promotes its own handle, and a cancel with no active turn reaches the oldest pending one -- the one the spine runs next. Pinned: `test_two_undrained_injects_each_get_their_own_turn_promoted_and_cancelled` (both promoted with their own ids and `message.start`, cancelled one after the other; runner saw all three turns). Focused suite: 949 passed. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification ``` uv run pytest tests/test_acp_methods.py tests/test_acp_protocol.py tests/test_rpc_turn_send.py \ tests/test_rpc_schema_match.py tests/test_acp_updates.py tests/test_acp_server.py -q # 766 passed uv run pytest tests/test_agents_oncall_flow_hooks.py tests/test_agents_oncall_flow_actions.py \ tests/test_agents_oncall_flow_budget.py tests/test_agents_oncall_flow_readings.py \ tests/test_agents_oncall_flow_escalation_tools.py -q # 99 passed uv run ruff check raven tests agents && uv run ruff format --check raven tests agents # clean uv run lint-imports # 9 kept, 0 broken cd ui-tui && npm run lint:rpc && npx tsc --noEmit; cd ui-web && npm run gen:check && npx tsc --noEmit # in sync, clean ``` New tests: `TestHeldTurn` (held across an armed wake, held again on a re-arm, heartbeat on the wait call, cancel while held, a non-`end_turn` stop is not held; every frame validated against the ACP schema), `TestSteer` (advertised, `no_turn` when idle, a held prompt takes the words as a turn with `busy: inject`, refusals), `busy: inject` on a busy and an idle lane in `test_rpc_turn_send.py`, and the plugin's hold stamping through the real `CronService` grant (filed with a pending wake; not filed without one, after the wake stands down, or once the campaign concluded). Not done: no end-to-end run of a real on-call wake through a held prompt yet -- unit tests only. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [ ] User-facing docs or screenshots are updated when needed (the hold is invisible to a person except as a "waiting for ..." tool row) ## Risk - [x] Backward compatibility considered: a turn that files no `raven.holdTurn` is answered exactly as before; `busy` is optional and absent from every existing caller; the steer method and capability are additive. - [x] Rollback path is clear: revert the two commits. A held prompt holds one `max_concurrent_subagents` slot (default 8) for the wait's duration on the caller's side; raised in config if it bites. ## Related Replaces !573 (closed). Follows !518 (stall notice) and !520 (unprompted-turn route; a wake that runs under a held prompt no longer takes it for DAG nodes).
2026-09-11 02:37:55 +00:00
host-running replacement only where the host's own exec would run on the host
too, and declines under any configured sandbox backend.
"""
from __future__ import annotations
import asyncio
import json
import os
feat(*): add bounded Python diagnostics after code edits (#967) ## Summary Raven-Code reports Python syntax errors after file edits but misses unresolved names and equivalent feedback after shell edits. Add advisory Pyflakes F821/F822/F823 checks to the existing tool result; optional F401 notices remain off by default. No additional model call or code execution is involved. ```text write_file / edit_file -> inspect final source, compare with a baseline that can be read and analyzed -> retain syntax warnings; suppress existing name findings exec -> require checkpoint visibility for the working directory -> read the changed files from the host's write record for the command -> ask visibility in batches of 256, select at most 8, check up to 4 concurrently -> append ordered diagnostics, preserving exit status and file metadata ``` Existing syntax failures remain visible after unrelated edits and identical rewrites. A baseline that cannot be read, parsed or analyzed is marked unknown; overwriting non-UTF-8 source still checks the known replacement. Source nested too deeply for the parser or checker, and a changed file that cannot be read, report diagnostics unavailable for that file rather than passing as clean. Every check reads a file as the bytes Python runs: text a write or edit stores is checked as its UTF-8 bytes, the content it replaced included, and a file from disk as its bytes, each decoded as Python decodes source, by its coding declaration or as UTF-8. Bytes that do not decode are reported as a syntax error at the line where decoding stops, so a declaration the written text cannot satisfy is reported however the file was changed. Noqa comments are read as Ruff reads them, including after other comment directives and never inside string literals: a file-level `ruff: noqa` / `flake8: noqa` counts only on a line of its own, and a marker counts only when codes follow its colon or, with no colon, whitespace, another comment or the end of the comment follows it. Unlike Ruff, which takes the first directive in a comment, every directive in it counts. The seeded tool guide gives the syntax note's own advice. With no checkpoint or no visibility verdict, exec diagnostics do not inspect changed file contents; direct file checks remain available. Pyflakes is imported when the tool face loads, not by whichever check first needs it: imported lazily, its cold import fell inside that check's one-second wait, and on a slow Windows runner the one test that paid it reported diagnostics unavailable for a plain syntax error. Private workers bound reads and checks to four active jobs per process. Saturated capacity rejects new work without queuing, and timeout/cancellation retains each slot until the actual work finishes. Name checks are capped at 256 KiB of source and the syntax check at 2 MiB, and a check skipped for size says so in one line. Findings are capped at eight per file, and exec diagnostic text at 6,000 characters plus omission notices. Runtime changes stay inside the Raven-Code code-flow plugin, except for one fix to `CheckpointService.trackable`, the visibility verdict that exec diagnostics and the host's exec write record both rely on. On macOS, git precomposes a file name it reads off the disk before matching an ignore rule, but takes the paths `trackable` sends to `git check-ignore --stdin` as written. A file stored with a decomposed (NFD) name under a `.gitignore` rule written precomposed (NFC) is therefore left out of the checkpoint, yet main vouches for it, so the write record's diff carries its text, and exec diagnostics would quote its source lines to the model. Nor is git's precomposition NFC: it composes only neighbouring characters, by Unicode 3.2's table, and reorders nothing, so a name whose combining marks are out of canonical order comes out as neither the name as given nor its NFC. `trackable` now asks about each name as given, as NFC and in git's macOS spelling, and withholds it when any spelling is ignored. Other shared file/shell/snapshot implementations are unchanged. Raven-Code has no separate dependency manifest, so Pyflakes uses the existing root dependencies. It is the only added package; uv normalized existing lockfile markers without changing other package versions. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification - **Before:** on main, Raven-Code's file tools report Python syntax errors only, and its exec reports nothing about the Python files a command changed. On macOS, main's `CheckpointService.trackable` vouches for a file stored as `secret_cafe` + U+0301 + `.py` under the rule `secret_caf` + U+00E9 + `.py`: the user's own git ignores it and the checkpoint's staged tree leaves it out, yet its text reaches the exec write record's stored diff. - **After:** on macOS 26.6.2 / arm64 with Apple Git 2.54.0, the command below passes 364 tests on Python 3.12.13 and on 3.14.6. The full default suite on 3.12.13 has 32283 passed, 169 skipped, 20 failed and 3 errors over 22 test ids: 21 of them also fail on the exact base in the same environment, and the other, a remote process-tree cancel test that missed its deadline under the suite's load, passes when run alone. Tests cover actual worker overlap, retained capacity after timeout/cancellation, host-tool isolation, hidden-file quotas, visibility batching through the host's write-record wiring, baseline handling, parser-limit failures, the size bounds and their skip notes, undecodable bytes placed where decoding stops, coding declarations in written and edited text, the baseline read staying on a worker, the checker loaded with its module, the file each unavailable note names, comment directives and their marker boundary, the tool guide's agreement with the syntax note, portable commands in paths with spaces/non-ASCII characters, and the checkpoint's verdict on decomposed names and on names git spells as neither given nor NFC, held against its own staged tree. Removing any of the three spellings from the query fails the cases that pin it, as does removing the marker boundary, the parser-limit report, the unknown-baseline handling, the syntax check past the name-check bound and the decoding report, checking written text as its stored bytes, the worker around the baseline read, loading the checker with its module, or the file name in an unavailable note; restoring the guide's earlier advice fails its agreement test. git's macOS spelling is what libiconv's UTF-8-MAC conversion returns, which git precomposes with: composing neighbours by Unicode 3.2's table reproduced it for all 281,928 names built from every decomposable character's decomposition (reordered, partly composed, or with a mark added), and across 784,220 further mixed names and pairs no name was spelled outside the three asked; the same held under Python 3.13 and 3.14. The decomposed-name check asserts the macOS outcome only where git precomposes names, so it also passes with `core.precomposeunicode = false` set through `GIT_CONFIG_GLOBAL`. Across 102 noqa comment variants the reported lines match Ruff 0.15.12, except the three that hold several directives in one comment. Through the real `AgentLoop`, code-flow tool face, `CodeExecutor` and checkpoint, with the working directory under `/tmp`, under `$TMPDIR` and in a path with spaces and CJK characters, exec diagnostics report created, modified and decomposed-name files, follow `working_dir`, report an undecodable file at the line where decoding stops, a syntax error past 256 KiB and a skip past 2 MiB, and withhold `.gitignore`d files, including decomposed names and names git spells as neither given nor NFC; across repeated runs on both interpreters, a file near the 256 KiB cap is checked in 52-135 ms (117-128 ms against a baseline), the syntax of a file near 2 MiB in 318-371 ms, and 600 ignored files plus one visible file take 73-90 ms with the event loop held at most 20 ms. At this head CI passes the unit suite on Linux/Python 3.12 (32619 tests over four shards), and its two Windows steps pass the diagnostic files on native Windows/Python 3.13: 122 file/worker tests, and 27 exec diagnostic tests with 17 pre-existing executor tests deselected; the symlink fixture skips only where the host cannot create links. ```sh env -u SERPER_API_KEY -u RAVEN_MIRRORED_SERPER_API_KEY uv run --no-sync pytest tests/test_agents_code_diagnostic_workers.py tests/test_agents_code_tools_plugin.py tests/test_agents_code_flow_exec.py tests/test_agents_code_flow_plugin.py tests/test_agents_code_flow_read_state.py tests/test_agents_code_launcher.py tests/test_workdir_snapshot.py tests/test_shell_command_writes.py tests/test_runtime_checkpoint.py tests/test_trajectory_upload.py::test_blocked_upload_hits_the_deadline -o addopts= -q --tb=short --maxfail=3 uv run --no-sync ruff check raven/agent/loop/checkpoint.py agents/raven-code/plugins/code-flow/code_flow/tools agents/raven-code/plugins/code-flow/code_flow/config.py tests/test_runtime_checkpoint.py tests/test_agents_code_flow_exec.py tests/test_agents_code_tools_plugin.py tests/test_agents_code_diagnostic_workers.py uv run --no-sync ty check raven/agent/loop/checkpoint.py agents/raven-code/plugins/code-flow/code_flow/tools agents/raven-code/plugins/code-flow/code_flow/config.py uv run --no-sync lint-imports uv run --no-sync ruff format --check raven/agent/loop/checkpoint.py agents/raven-code/plugins/code-flow/code_flow/tools agents/raven-code/plugins/code-flow/code_flow/config.py tests/test_runtime_checkpoint.py tests/test_agents_code_flow_exec.py tests/test_agents_code_tools_plugin.py ``` Native Windows CI commands: ```sh uv run --frozen --python 3.13 --no-sync pytest tests/test_agents_code_diagnostic_workers.py tests/test_agents_code_tools_plugin.py -q -n 0 --tb=short uv run --frozen --python 3.13 --no-sync pytest tests/test_agents_code_flow_exec.py -k test_exec_ -q -n 0 --tb=short ``` Ruff, format and type checks passed; all 10 import contracts kept; the commit-message, large-file and source-language checks pass over the branch. The launcher tests unset inherited search API keys to keep the expected tool list deterministic. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [x] User-facing docs or screenshots are updated when needed ## Risk Scope: Raven-Code's enabled code-flow tool face; installing Pyflakes does not activate checks in other agents. The change is reversible and never fixes/removes code or rolls back edits. Static diagnostics can need judgment for dynamic names, target Python versions, and project lint configuration; only explicit noqa directives are honored. Passing diagnostics do not replace tests. The `trackable` fix reaches every exec write record, not only Raven-Code: a file whose name, as given, as NFC or in git's macOS spelling, matches an ignore rule keeps its counts and loses its diff and removal body. Where git matches another of those spellings, as off macOS for a decomposed name under a precomposed rule, the checkpoint stores the file but it is still withheld; its counts remain. Exec diagnostics now require an active checkpoint. Noninteractive runs using the default interactive checkpoint policy must explicitly enable checkpoints to receive this feedback. Visibility is checked in batches of up to 256 files and stops once eight visible files are found; each batch holds the event loop while the host resolves its paths, about 130 ms for a full batch on native Windows. Each file-check wait is one second, and post-command diagnostics have a three-second wait budget; past it, or when they fail as a whole, they are dropped and the command's result is kept. These are caller budgets, not hard I/O deadlines: running OS reads and inherited synchronous host callbacks cannot be forcibly interrupted. Private worker capacity remains bounded. Between 256 KiB and 2 MiB only syntax is checked: near 2 MiB that took 0.3-0.4 s and about 190 MiB here, up to four checks run at once, and on slower hardware such a check can outlast the one-second wait and report diagnostics unavailable. Python's GIL means concurrency does not promise a CPU speedup. Windows CI exercises the diagnostics layer with the real native DirectExecutor. The existing CodeExecutor environment probe uses POSIX quoting and already fails on Windows at the base; this PR neither changes that probe nor claims end-to-end compatibility for it. Symlink setup is skipped only if the host cannot create the fixture. Disable `pythonLint` for syntax-only feedback, `execPythonDiagnostics` for no post-command checks, or `pythonSyntaxNote` for no diagnostics, under `plugins.config["code-flow"].tools`. The PR can also be reverted as a unit. - [x] Security impact considered - [x] Backward compatibility considered - [x] Rollback path is clear for risky changes ## Related Issues N/A --------- Co-authored-by: Codex <noreply@openai.com> Co-authored-by: Dizhan Xue <xuedizhan17@mails.ucas.ac.cn> Co-authored-by: Claude (claude-opus-5-5) <noreply@anthropic.com>
2026-10-18 20:21:04 +08:00
import shlex
import subprocess
feat(acp): a prompt is held open while the agent sleeps on an armed wake, and takes a steer ## Summary An on-call agent submits a job, arms a wake and ends its turn saying so; the wake runs a later turn on the session on its own clock. Over ACP that turn's `end_turn` read as "finished" to the caller: a DAG judging its node spent every continuation on "still watching, wake at 13:13" while the job ran fine (ar-r4, 2026-09-08), and the main agent re-dispatched the work by hand. ACP has no stop reason for "waiting". Replaces the closed !573: **no DAG code changes**, no new node status. The node is `running` because a turn is genuinely in flight. - **The signal.** The loop-hooks paper gains one key the ACP layer reads for itself: `acp_meta["raven.holdTurn"]` (`{"untilMs", "why"}`). oncall-flow files it at the send fire when the window's campaign is claimed, not concluded, and has one pending wake -- read the same way the owner's answer pulls that wake forward. No campaign, a wake that stood down, or a concluded campaign files nothing and the turn ends as before. - **The hold** (`raven/acp/methods.py`). A turn that ended `end_turn` with the key filed is not answered. One `tool_call` stands for the wait (its title names the campaign and the next look), with a `tool_call_update` heartbeat every `HOLD_HEARTBEAT_S` (60 s) so a client that clocks liveness by frames sees the agent alive while its model sleeps. The turn slot is reopened with no id to correlate on, so whichever turn ends next on the session -- the wake's, a steer's -- settles it, and the handler decides again from that turn's stash. A wake that arms again keeps the hold; a turn ending with nothing pending answers `end_turn`; a cancel answers `cancelled` as before. The key itself is not passed to the client. - **Steer.** `_raven/session/steer` is served and advertised under `agentCapabilities._meta` as `raven.steer` (the names raven's own client already uses). Accepted while a prompt is open on the session, held or running: the text is shown on the wire as the person's own line, then sent with `turn.send`'s new `busy: "inject"`. Answers `injected` / `no_turn`. - **Wire.** `turn.send` gains the optional `busy` param (`const "inject"`): an ordinary send on an idle lane, `BusyPolicy.INJECT` into the running turn otherwise -- the merge a channel's mid-turn message already gets (`gateway_commands.py`). No slot is bound for an injected message; the running turn keeps the lane's. `openrpc.json` and both generated clients moved together. What raven's client side already does, unchanged: the DAG stall watcher resets on any frame (the heartbeat keeps it quiet, and a stopped heartbeat raises the existing stall notice after ten minutes); a sub-agent process that dies fails the request and the node takes its ordinary path; the pool's `_responsive` treats a prompt in flight as busy. ## Review round 1 (2026-09-09, on !600, superseded by this MR) - **`busy: inject` reads the scheduler's lane state too.** `is_turn_active` knows only the turns `turn.send` started; an armed wake is submitted to the scheduler directly (`make_on_session_wake`) and never binds that map, so a steer during a running wake queued a second turn and answered `injected`. `turn.send` now asks `Scheduler.has_inflight(lane)` as well, and merges (`BusyPolicy.INJECT`) when either record says the lane is running. Pinned: `test_busy_inject_merges_into_a_wake_the_real_scheduler_is_running` -- the reviewer's reproduction on the real `Scheduler`, a turn submitted the way `make_on_session_wake` submits one; with the lane check bypassed it reads `pending=1, inject_mailbox=0`, with it `(1, 0)` in the mailbox and the running turn drains the text. - Branch renamed to snake_case and both commits carry the `Co-authored-by: Claude (<model-id>)` trailer, per AGENTS.md §2–3; that is why this MR replaces !600 rather than continuing it. ## Review round 2 (2026-09-10) - **A wake re-armed during its own run survives the writeback.** `CronService._writeback_locked` reloaded the store after the run and, for a one-shot with `delete_after_run`, overwrote the id with the finished run's state and removed it -- but the turn the wake ran had just called `schedule_wake` under the same key, so the id now named the *next* look. It vanished, and the held prompt had nothing left to settle it. A stored job under the fired id whose `at` schedule the run never had is treated as the new wake it is: claim cleared, run recorded, schedule kept, not deleted. Pinned on the real service: `test_a_wake_re_armed_during_its_own_run_survives_the_writeback` (reads `[]` without the fix, the reviewer's state) and the unchanged one-shot deletion beside it. Cron, proactive, sentinel and on-call suites: 968 passed. ## Review round 3 (2026-09-10) - **An undrained inject stays visible and cancellable.** `busy: inject` dropped the scheduler's handle; when the host turn ended past its last drain, the spine's fallback ran the text as a turn nothing owned -- `turn.cancel` a no-op, no `message.start`. The handle is now kept per lane under an id minted for the text; `turn.cancel` / `session.interrupt` reach it when the active slot is empty (mailbox, queued fallback, or running); and when the fallback turn starts, the RPC sink (`on_turn_start`, wired from bootstrap) binds the lane to it and opens it on the wire with the person's words. `turn.send` answers the inject's own id. Pinned through `build_rpc_spine` with the reviewer's shape (a loop that drains once, an inject after it): `test_an_undrained_inject_runs_as_a_turn_the_surface_can_see_and_cancel` and `test_a_pending_inject_can_be_cancelled_before_it_merges_or_falls_back`; both read the reviewer's state (`turn_ids` empty, cancel finding nothing) with the fix bypassed. - **Line ceiling.** The contract comment is five lines; `raven/contracts` is 2,519 against 2,520, one line of headroom on the current `main` (2,514). - Non-blocking items taken: `SchedulerDrainingError` on the inject path exits as `turn_failed`; the title is announced before the hold; the steer constants sit below the cancel constant's docstring; three edges pinned (a held prompt settled by a wake turn's uncorrelated ending through `send_frame`, the hold key never in the client's `_meta`, `raven.toolName` on the wait row). - Focused suite (reviewer's list plus the touched files): 948 passed on Python 3.12. ## Review round 4 (2026-09-10) - **Every pending inject is kept, per lane and per id.** One replaceable record per lane kept only the newest steer; two steers after the host's last drain fall back in turn, and the first started unowned while a cancel took out the second. The record is now `lane -> {inject id: (handle, text)}` in arrival order: each fallback promotes its own handle, and a cancel with no active turn reaches the oldest pending one -- the one the spine runs next. Pinned: `test_two_undrained_injects_each_get_their_own_turn_promoted_and_cancelled` (both promoted with their own ids and `message.start`, cancelled one after the other; runner saw all three turns). Focused suite: 949 passed. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification ``` uv run pytest tests/test_acp_methods.py tests/test_acp_protocol.py tests/test_rpc_turn_send.py \ tests/test_rpc_schema_match.py tests/test_acp_updates.py tests/test_acp_server.py -q # 766 passed uv run pytest tests/test_agents_oncall_flow_hooks.py tests/test_agents_oncall_flow_actions.py \ tests/test_agents_oncall_flow_budget.py tests/test_agents_oncall_flow_readings.py \ tests/test_agents_oncall_flow_escalation_tools.py -q # 99 passed uv run ruff check raven tests agents && uv run ruff format --check raven tests agents # clean uv run lint-imports # 9 kept, 0 broken cd ui-tui && npm run lint:rpc && npx tsc --noEmit; cd ui-web && npm run gen:check && npx tsc --noEmit # in sync, clean ``` New tests: `TestHeldTurn` (held across an armed wake, held again on a re-arm, heartbeat on the wait call, cancel while held, a non-`end_turn` stop is not held; every frame validated against the ACP schema), `TestSteer` (advertised, `no_turn` when idle, a held prompt takes the words as a turn with `busy: inject`, refusals), `busy: inject` on a busy and an idle lane in `test_rpc_turn_send.py`, and the plugin's hold stamping through the real `CronService` grant (filed with a pending wake; not filed without one, after the wake stands down, or once the campaign concluded). Not done: no end-to-end run of a real on-call wake through a held prompt yet -- unit tests only. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [ ] User-facing docs or screenshots are updated when needed (the hold is invisible to a person except as a "waiting for ..." tool row) ## Risk - [x] Backward compatibility considered: a turn that files no `raven.holdTurn` is answered exactly as before; `busy` is optional and absent from every existing caller; the steer method and capability are additive. - [x] Rollback path is clear: revert the two commits. A held prompt holds one `max_concurrent_subagents` slot (default 8) for the wait's duration on the caller's side; raised in config if it bites. ## Related Replaces !573 (closed). Follows !518 (stall notice) and !520 (unprompted-turn route; a wake that runs under a held prompt no longer takes it for DAG nodes).
2026-09-11 02:37:55 +00:00
import sys
feat(*): add bounded Python diagnostics after code edits (#967) ## Summary Raven-Code reports Python syntax errors after file edits but misses unresolved names and equivalent feedback after shell edits. Add advisory Pyflakes F821/F822/F823 checks to the existing tool result; optional F401 notices remain off by default. No additional model call or code execution is involved. ```text write_file / edit_file -> inspect final source, compare with a baseline that can be read and analyzed -> retain syntax warnings; suppress existing name findings exec -> require checkpoint visibility for the working directory -> read the changed files from the host's write record for the command -> ask visibility in batches of 256, select at most 8, check up to 4 concurrently -> append ordered diagnostics, preserving exit status and file metadata ``` Existing syntax failures remain visible after unrelated edits and identical rewrites. A baseline that cannot be read, parsed or analyzed is marked unknown; overwriting non-UTF-8 source still checks the known replacement. Source nested too deeply for the parser or checker, and a changed file that cannot be read, report diagnostics unavailable for that file rather than passing as clean. Every check reads a file as the bytes Python runs: text a write or edit stores is checked as its UTF-8 bytes, the content it replaced included, and a file from disk as its bytes, each decoded as Python decodes source, by its coding declaration or as UTF-8. Bytes that do not decode are reported as a syntax error at the line where decoding stops, so a declaration the written text cannot satisfy is reported however the file was changed. Noqa comments are read as Ruff reads them, including after other comment directives and never inside string literals: a file-level `ruff: noqa` / `flake8: noqa` counts only on a line of its own, and a marker counts only when codes follow its colon or, with no colon, whitespace, another comment or the end of the comment follows it. Unlike Ruff, which takes the first directive in a comment, every directive in it counts. The seeded tool guide gives the syntax note's own advice. With no checkpoint or no visibility verdict, exec diagnostics do not inspect changed file contents; direct file checks remain available. Pyflakes is imported when the tool face loads, not by whichever check first needs it: imported lazily, its cold import fell inside that check's one-second wait, and on a slow Windows runner the one test that paid it reported diagnostics unavailable for a plain syntax error. Private workers bound reads and checks to four active jobs per process. Saturated capacity rejects new work without queuing, and timeout/cancellation retains each slot until the actual work finishes. Name checks are capped at 256 KiB of source and the syntax check at 2 MiB, and a check skipped for size says so in one line. Findings are capped at eight per file, and exec diagnostic text at 6,000 characters plus omission notices. Runtime changes stay inside the Raven-Code code-flow plugin, except for one fix to `CheckpointService.trackable`, the visibility verdict that exec diagnostics and the host's exec write record both rely on. On macOS, git precomposes a file name it reads off the disk before matching an ignore rule, but takes the paths `trackable` sends to `git check-ignore --stdin` as written. A file stored with a decomposed (NFD) name under a `.gitignore` rule written precomposed (NFC) is therefore left out of the checkpoint, yet main vouches for it, so the write record's diff carries its text, and exec diagnostics would quote its source lines to the model. Nor is git's precomposition NFC: it composes only neighbouring characters, by Unicode 3.2's table, and reorders nothing, so a name whose combining marks are out of canonical order comes out as neither the name as given nor its NFC. `trackable` now asks about each name as given, as NFC and in git's macOS spelling, and withholds it when any spelling is ignored. Other shared file/shell/snapshot implementations are unchanged. Raven-Code has no separate dependency manifest, so Pyflakes uses the existing root dependencies. It is the only added package; uv normalized existing lockfile markers without changing other package versions. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification - **Before:** on main, Raven-Code's file tools report Python syntax errors only, and its exec reports nothing about the Python files a command changed. On macOS, main's `CheckpointService.trackable` vouches for a file stored as `secret_cafe` + U+0301 + `.py` under the rule `secret_caf` + U+00E9 + `.py`: the user's own git ignores it and the checkpoint's staged tree leaves it out, yet its text reaches the exec write record's stored diff. - **After:** on macOS 26.6.2 / arm64 with Apple Git 2.54.0, the command below passes 364 tests on Python 3.12.13 and on 3.14.6. The full default suite on 3.12.13 has 32283 passed, 169 skipped, 20 failed and 3 errors over 22 test ids: 21 of them also fail on the exact base in the same environment, and the other, a remote process-tree cancel test that missed its deadline under the suite's load, passes when run alone. Tests cover actual worker overlap, retained capacity after timeout/cancellation, host-tool isolation, hidden-file quotas, visibility batching through the host's write-record wiring, baseline handling, parser-limit failures, the size bounds and their skip notes, undecodable bytes placed where decoding stops, coding declarations in written and edited text, the baseline read staying on a worker, the checker loaded with its module, the file each unavailable note names, comment directives and their marker boundary, the tool guide's agreement with the syntax note, portable commands in paths with spaces/non-ASCII characters, and the checkpoint's verdict on decomposed names and on names git spells as neither given nor NFC, held against its own staged tree. Removing any of the three spellings from the query fails the cases that pin it, as does removing the marker boundary, the parser-limit report, the unknown-baseline handling, the syntax check past the name-check bound and the decoding report, checking written text as its stored bytes, the worker around the baseline read, loading the checker with its module, or the file name in an unavailable note; restoring the guide's earlier advice fails its agreement test. git's macOS spelling is what libiconv's UTF-8-MAC conversion returns, which git precomposes with: composing neighbours by Unicode 3.2's table reproduced it for all 281,928 names built from every decomposable character's decomposition (reordered, partly composed, or with a mark added), and across 784,220 further mixed names and pairs no name was spelled outside the three asked; the same held under Python 3.13 and 3.14. The decomposed-name check asserts the macOS outcome only where git precomposes names, so it also passes with `core.precomposeunicode = false` set through `GIT_CONFIG_GLOBAL`. Across 102 noqa comment variants the reported lines match Ruff 0.15.12, except the three that hold several directives in one comment. Through the real `AgentLoop`, code-flow tool face, `CodeExecutor` and checkpoint, with the working directory under `/tmp`, under `$TMPDIR` and in a path with spaces and CJK characters, exec diagnostics report created, modified and decomposed-name files, follow `working_dir`, report an undecodable file at the line where decoding stops, a syntax error past 256 KiB and a skip past 2 MiB, and withhold `.gitignore`d files, including decomposed names and names git spells as neither given nor NFC; across repeated runs on both interpreters, a file near the 256 KiB cap is checked in 52-135 ms (117-128 ms against a baseline), the syntax of a file near 2 MiB in 318-371 ms, and 600 ignored files plus one visible file take 73-90 ms with the event loop held at most 20 ms. At this head CI passes the unit suite on Linux/Python 3.12 (32619 tests over four shards), and its two Windows steps pass the diagnostic files on native Windows/Python 3.13: 122 file/worker tests, and 27 exec diagnostic tests with 17 pre-existing executor tests deselected; the symlink fixture skips only where the host cannot create links. ```sh env -u SERPER_API_KEY -u RAVEN_MIRRORED_SERPER_API_KEY uv run --no-sync pytest tests/test_agents_code_diagnostic_workers.py tests/test_agents_code_tools_plugin.py tests/test_agents_code_flow_exec.py tests/test_agents_code_flow_plugin.py tests/test_agents_code_flow_read_state.py tests/test_agents_code_launcher.py tests/test_workdir_snapshot.py tests/test_shell_command_writes.py tests/test_runtime_checkpoint.py tests/test_trajectory_upload.py::test_blocked_upload_hits_the_deadline -o addopts= -q --tb=short --maxfail=3 uv run --no-sync ruff check raven/agent/loop/checkpoint.py agents/raven-code/plugins/code-flow/code_flow/tools agents/raven-code/plugins/code-flow/code_flow/config.py tests/test_runtime_checkpoint.py tests/test_agents_code_flow_exec.py tests/test_agents_code_tools_plugin.py tests/test_agents_code_diagnostic_workers.py uv run --no-sync ty check raven/agent/loop/checkpoint.py agents/raven-code/plugins/code-flow/code_flow/tools agents/raven-code/plugins/code-flow/code_flow/config.py uv run --no-sync lint-imports uv run --no-sync ruff format --check raven/agent/loop/checkpoint.py agents/raven-code/plugins/code-flow/code_flow/tools agents/raven-code/plugins/code-flow/code_flow/config.py tests/test_runtime_checkpoint.py tests/test_agents_code_flow_exec.py tests/test_agents_code_tools_plugin.py ``` Native Windows CI commands: ```sh uv run --frozen --python 3.13 --no-sync pytest tests/test_agents_code_diagnostic_workers.py tests/test_agents_code_tools_plugin.py -q -n 0 --tb=short uv run --frozen --python 3.13 --no-sync pytest tests/test_agents_code_flow_exec.py -k test_exec_ -q -n 0 --tb=short ``` Ruff, format and type checks passed; all 10 import contracts kept; the commit-message, large-file and source-language checks pass over the branch. The launcher tests unset inherited search API keys to keep the expected tool list deterministic. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [x] User-facing docs or screenshots are updated when needed ## Risk Scope: Raven-Code's enabled code-flow tool face; installing Pyflakes does not activate checks in other agents. The change is reversible and never fixes/removes code or rolls back edits. Static diagnostics can need judgment for dynamic names, target Python versions, and project lint configuration; only explicit noqa directives are honored. Passing diagnostics do not replace tests. The `trackable` fix reaches every exec write record, not only Raven-Code: a file whose name, as given, as NFC or in git's macOS spelling, matches an ignore rule keeps its counts and loses its diff and removal body. Where git matches another of those spellings, as off macOS for a decomposed name under a precomposed rule, the checkpoint stores the file but it is still withheld; its counts remain. Exec diagnostics now require an active checkpoint. Noninteractive runs using the default interactive checkpoint policy must explicitly enable checkpoints to receive this feedback. Visibility is checked in batches of up to 256 files and stops once eight visible files are found; each batch holds the event loop while the host resolves its paths, about 130 ms for a full batch on native Windows. Each file-check wait is one second, and post-command diagnostics have a three-second wait budget; past it, or when they fail as a whole, they are dropped and the command's result is kept. These are caller budgets, not hard I/O deadlines: running OS reads and inherited synchronous host callbacks cannot be forcibly interrupted. Private worker capacity remains bounded. Between 256 KiB and 2 MiB only syntax is checked: near 2 MiB that took 0.3-0.4 s and about 190 MiB here, up to four checks run at once, and on slower hardware such a check can outlast the one-second wait and report diagnostics unavailable. Python's GIL means concurrency does not promise a CPU speedup. Windows CI exercises the diagnostics layer with the real native DirectExecutor. The existing CodeExecutor environment probe uses POSIX quoting and already fails on Windows at the base; this PR neither changes that probe nor claims end-to-end compatibility for it. Symlink setup is skipped only if the host cannot create the fixture. Disable `pythonLint` for syntax-only feedback, `execPythonDiagnostics` for no post-command checks, or `pythonSyntaxNote` for no diagnostics, under `plugins.config["code-flow"].tools`. The PR can also be reverted as a unit. - [x] Security impact considered - [x] Backward compatibility considered - [x] Rollback path is clear for risky changes ## Related Issues N/A --------- Co-authored-by: Codex <noreply@openai.com> Co-authored-by: Dizhan Xue <xuedizhan17@mails.ucas.ac.cn> Co-authored-by: Claude (claude-opus-5-5) <noreply@anthropic.com>
2026-10-18 20:21:04 +08:00
import tempfile
import threading
import time
feat(acp): a prompt is held open while the agent sleeps on an armed wake, and takes a steer ## Summary An on-call agent submits a job, arms a wake and ends its turn saying so; the wake runs a later turn on the session on its own clock. Over ACP that turn's `end_turn` read as "finished" to the caller: a DAG judging its node spent every continuation on "still watching, wake at 13:13" while the job ran fine (ar-r4, 2026-09-08), and the main agent re-dispatched the work by hand. ACP has no stop reason for "waiting". Replaces the closed !573: **no DAG code changes**, no new node status. The node is `running` because a turn is genuinely in flight. - **The signal.** The loop-hooks paper gains one key the ACP layer reads for itself: `acp_meta["raven.holdTurn"]` (`{"untilMs", "why"}`). oncall-flow files it at the send fire when the window's campaign is claimed, not concluded, and has one pending wake -- read the same way the owner's answer pulls that wake forward. No campaign, a wake that stood down, or a concluded campaign files nothing and the turn ends as before. - **The hold** (`raven/acp/methods.py`). A turn that ended `end_turn` with the key filed is not answered. One `tool_call` stands for the wait (its title names the campaign and the next look), with a `tool_call_update` heartbeat every `HOLD_HEARTBEAT_S` (60 s) so a client that clocks liveness by frames sees the agent alive while its model sleeps. The turn slot is reopened with no id to correlate on, so whichever turn ends next on the session -- the wake's, a steer's -- settles it, and the handler decides again from that turn's stash. A wake that arms again keeps the hold; a turn ending with nothing pending answers `end_turn`; a cancel answers `cancelled` as before. The key itself is not passed to the client. - **Steer.** `_raven/session/steer` is served and advertised under `agentCapabilities._meta` as `raven.steer` (the names raven's own client already uses). Accepted while a prompt is open on the session, held or running: the text is shown on the wire as the person's own line, then sent with `turn.send`'s new `busy: "inject"`. Answers `injected` / `no_turn`. - **Wire.** `turn.send` gains the optional `busy` param (`const "inject"`): an ordinary send on an idle lane, `BusyPolicy.INJECT` into the running turn otherwise -- the merge a channel's mid-turn message already gets (`gateway_commands.py`). No slot is bound for an injected message; the running turn keeps the lane's. `openrpc.json` and both generated clients moved together. What raven's client side already does, unchanged: the DAG stall watcher resets on any frame (the heartbeat keeps it quiet, and a stopped heartbeat raises the existing stall notice after ten minutes); a sub-agent process that dies fails the request and the node takes its ordinary path; the pool's `_responsive` treats a prompt in flight as busy. ## Review round 1 (2026-09-09, on !600, superseded by this MR) - **`busy: inject` reads the scheduler's lane state too.** `is_turn_active` knows only the turns `turn.send` started; an armed wake is submitted to the scheduler directly (`make_on_session_wake`) and never binds that map, so a steer during a running wake queued a second turn and answered `injected`. `turn.send` now asks `Scheduler.has_inflight(lane)` as well, and merges (`BusyPolicy.INJECT`) when either record says the lane is running. Pinned: `test_busy_inject_merges_into_a_wake_the_real_scheduler_is_running` -- the reviewer's reproduction on the real `Scheduler`, a turn submitted the way `make_on_session_wake` submits one; with the lane check bypassed it reads `pending=1, inject_mailbox=0`, with it `(1, 0)` in the mailbox and the running turn drains the text. - Branch renamed to snake_case and both commits carry the `Co-authored-by: Claude (<model-id>)` trailer, per AGENTS.md §2–3; that is why this MR replaces !600 rather than continuing it. ## Review round 2 (2026-09-10) - **A wake re-armed during its own run survives the writeback.** `CronService._writeback_locked` reloaded the store after the run and, for a one-shot with `delete_after_run`, overwrote the id with the finished run's state and removed it -- but the turn the wake ran had just called `schedule_wake` under the same key, so the id now named the *next* look. It vanished, and the held prompt had nothing left to settle it. A stored job under the fired id whose `at` schedule the run never had is treated as the new wake it is: claim cleared, run recorded, schedule kept, not deleted. Pinned on the real service: `test_a_wake_re_armed_during_its_own_run_survives_the_writeback` (reads `[]` without the fix, the reviewer's state) and the unchanged one-shot deletion beside it. Cron, proactive, sentinel and on-call suites: 968 passed. ## Review round 3 (2026-09-10) - **An undrained inject stays visible and cancellable.** `busy: inject` dropped the scheduler's handle; when the host turn ended past its last drain, the spine's fallback ran the text as a turn nothing owned -- `turn.cancel` a no-op, no `message.start`. The handle is now kept per lane under an id minted for the text; `turn.cancel` / `session.interrupt` reach it when the active slot is empty (mailbox, queued fallback, or running); and when the fallback turn starts, the RPC sink (`on_turn_start`, wired from bootstrap) binds the lane to it and opens it on the wire with the person's words. `turn.send` answers the inject's own id. Pinned through `build_rpc_spine` with the reviewer's shape (a loop that drains once, an inject after it): `test_an_undrained_inject_runs_as_a_turn_the_surface_can_see_and_cancel` and `test_a_pending_inject_can_be_cancelled_before_it_merges_or_falls_back`; both read the reviewer's state (`turn_ids` empty, cancel finding nothing) with the fix bypassed. - **Line ceiling.** The contract comment is five lines; `raven/contracts` is 2,519 against 2,520, one line of headroom on the current `main` (2,514). - Non-blocking items taken: `SchedulerDrainingError` on the inject path exits as `turn_failed`; the title is announced before the hold; the steer constants sit below the cancel constant's docstring; three edges pinned (a held prompt settled by a wake turn's uncorrelated ending through `send_frame`, the hold key never in the client's `_meta`, `raven.toolName` on the wait row). - Focused suite (reviewer's list plus the touched files): 948 passed on Python 3.12. ## Review round 4 (2026-09-10) - **Every pending inject is kept, per lane and per id.** One replaceable record per lane kept only the newest steer; two steers after the host's last drain fall back in turn, and the first started unowned while a cancel took out the second. The record is now `lane -> {inject id: (handle, text)}` in arrival order: each fallback promotes its own handle, and a cancel with no active turn reaches the oldest pending one -- the one the spine runs next. Pinned: `test_two_undrained_injects_each_get_their_own_turn_promoted_and_cancelled` (both promoted with their own ids and `message.start`, cancelled one after the other; runner saw all three turns). Focused suite: 949 passed. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification ``` uv run pytest tests/test_acp_methods.py tests/test_acp_protocol.py tests/test_rpc_turn_send.py \ tests/test_rpc_schema_match.py tests/test_acp_updates.py tests/test_acp_server.py -q # 766 passed uv run pytest tests/test_agents_oncall_flow_hooks.py tests/test_agents_oncall_flow_actions.py \ tests/test_agents_oncall_flow_budget.py tests/test_agents_oncall_flow_readings.py \ tests/test_agents_oncall_flow_escalation_tools.py -q # 99 passed uv run ruff check raven tests agents && uv run ruff format --check raven tests agents # clean uv run lint-imports # 9 kept, 0 broken cd ui-tui && npm run lint:rpc && npx tsc --noEmit; cd ui-web && npm run gen:check && npx tsc --noEmit # in sync, clean ``` New tests: `TestHeldTurn` (held across an armed wake, held again on a re-arm, heartbeat on the wait call, cancel while held, a non-`end_turn` stop is not held; every frame validated against the ACP schema), `TestSteer` (advertised, `no_turn` when idle, a held prompt takes the words as a turn with `busy: inject`, refusals), `busy: inject` on a busy and an idle lane in `test_rpc_turn_send.py`, and the plugin's hold stamping through the real `CronService` grant (filed with a pending wake; not filed without one, after the wake stands down, or once the campaign concluded). Not done: no end-to-end run of a real on-call wake through a held prompt yet -- unit tests only. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [ ] User-facing docs or screenshots are updated when needed (the hold is invisible to a person except as a "waiting for ..." tool row) ## Risk - [x] Backward compatibility considered: a turn that files no `raven.holdTurn` is answered exactly as before; `busy` is optional and absent from every existing caller; the steer method and capability are additive. - [x] Rollback path is clear: revert the two commits. A held prompt holds one `max_concurrent_subagents` slot (default 8) for the wait's duration on the caller's side; raised in config if it bites. ## Related Replaces !573 (closed). Follows !518 (stall notice) and !520 (unprompted-turn route; a wake that runs under a held prompt no longer takes it for DAG nodes).
2026-09-11 02:37:55 +00:00
from pathlib import Path
feat(*): add bounded Python diagnostics after code edits (#967) ## Summary Raven-Code reports Python syntax errors after file edits but misses unresolved names and equivalent feedback after shell edits. Add advisory Pyflakes F821/F822/F823 checks to the existing tool result; optional F401 notices remain off by default. No additional model call or code execution is involved. ```text write_file / edit_file -> inspect final source, compare with a baseline that can be read and analyzed -> retain syntax warnings; suppress existing name findings exec -> require checkpoint visibility for the working directory -> read the changed files from the host's write record for the command -> ask visibility in batches of 256, select at most 8, check up to 4 concurrently -> append ordered diagnostics, preserving exit status and file metadata ``` Existing syntax failures remain visible after unrelated edits and identical rewrites. A baseline that cannot be read, parsed or analyzed is marked unknown; overwriting non-UTF-8 source still checks the known replacement. Source nested too deeply for the parser or checker, and a changed file that cannot be read, report diagnostics unavailable for that file rather than passing as clean. Every check reads a file as the bytes Python runs: text a write or edit stores is checked as its UTF-8 bytes, the content it replaced included, and a file from disk as its bytes, each decoded as Python decodes source, by its coding declaration or as UTF-8. Bytes that do not decode are reported as a syntax error at the line where decoding stops, so a declaration the written text cannot satisfy is reported however the file was changed. Noqa comments are read as Ruff reads them, including after other comment directives and never inside string literals: a file-level `ruff: noqa` / `flake8: noqa` counts only on a line of its own, and a marker counts only when codes follow its colon or, with no colon, whitespace, another comment or the end of the comment follows it. Unlike Ruff, which takes the first directive in a comment, every directive in it counts. The seeded tool guide gives the syntax note's own advice. With no checkpoint or no visibility verdict, exec diagnostics do not inspect changed file contents; direct file checks remain available. Pyflakes is imported when the tool face loads, not by whichever check first needs it: imported lazily, its cold import fell inside that check's one-second wait, and on a slow Windows runner the one test that paid it reported diagnostics unavailable for a plain syntax error. Private workers bound reads and checks to four active jobs per process. Saturated capacity rejects new work without queuing, and timeout/cancellation retains each slot until the actual work finishes. Name checks are capped at 256 KiB of source and the syntax check at 2 MiB, and a check skipped for size says so in one line. Findings are capped at eight per file, and exec diagnostic text at 6,000 characters plus omission notices. Runtime changes stay inside the Raven-Code code-flow plugin, except for one fix to `CheckpointService.trackable`, the visibility verdict that exec diagnostics and the host's exec write record both rely on. On macOS, git precomposes a file name it reads off the disk before matching an ignore rule, but takes the paths `trackable` sends to `git check-ignore --stdin` as written. A file stored with a decomposed (NFD) name under a `.gitignore` rule written precomposed (NFC) is therefore left out of the checkpoint, yet main vouches for it, so the write record's diff carries its text, and exec diagnostics would quote its source lines to the model. Nor is git's precomposition NFC: it composes only neighbouring characters, by Unicode 3.2's table, and reorders nothing, so a name whose combining marks are out of canonical order comes out as neither the name as given nor its NFC. `trackable` now asks about each name as given, as NFC and in git's macOS spelling, and withholds it when any spelling is ignored. Other shared file/shell/snapshot implementations are unchanged. Raven-Code has no separate dependency manifest, so Pyflakes uses the existing root dependencies. It is the only added package; uv normalized existing lockfile markers without changing other package versions. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification - **Before:** on main, Raven-Code's file tools report Python syntax errors only, and its exec reports nothing about the Python files a command changed. On macOS, main's `CheckpointService.trackable` vouches for a file stored as `secret_cafe` + U+0301 + `.py` under the rule `secret_caf` + U+00E9 + `.py`: the user's own git ignores it and the checkpoint's staged tree leaves it out, yet its text reaches the exec write record's stored diff. - **After:** on macOS 26.6.2 / arm64 with Apple Git 2.54.0, the command below passes 364 tests on Python 3.12.13 and on 3.14.6. The full default suite on 3.12.13 has 32283 passed, 169 skipped, 20 failed and 3 errors over 22 test ids: 21 of them also fail on the exact base in the same environment, and the other, a remote process-tree cancel test that missed its deadline under the suite's load, passes when run alone. Tests cover actual worker overlap, retained capacity after timeout/cancellation, host-tool isolation, hidden-file quotas, visibility batching through the host's write-record wiring, baseline handling, parser-limit failures, the size bounds and their skip notes, undecodable bytes placed where decoding stops, coding declarations in written and edited text, the baseline read staying on a worker, the checker loaded with its module, the file each unavailable note names, comment directives and their marker boundary, the tool guide's agreement with the syntax note, portable commands in paths with spaces/non-ASCII characters, and the checkpoint's verdict on decomposed names and on names git spells as neither given nor NFC, held against its own staged tree. Removing any of the three spellings from the query fails the cases that pin it, as does removing the marker boundary, the parser-limit report, the unknown-baseline handling, the syntax check past the name-check bound and the decoding report, checking written text as its stored bytes, the worker around the baseline read, loading the checker with its module, or the file name in an unavailable note; restoring the guide's earlier advice fails its agreement test. git's macOS spelling is what libiconv's UTF-8-MAC conversion returns, which git precomposes with: composing neighbours by Unicode 3.2's table reproduced it for all 281,928 names built from every decomposable character's decomposition (reordered, partly composed, or with a mark added), and across 784,220 further mixed names and pairs no name was spelled outside the three asked; the same held under Python 3.13 and 3.14. The decomposed-name check asserts the macOS outcome only where git precomposes names, so it also passes with `core.precomposeunicode = false` set through `GIT_CONFIG_GLOBAL`. Across 102 noqa comment variants the reported lines match Ruff 0.15.12, except the three that hold several directives in one comment. Through the real `AgentLoop`, code-flow tool face, `CodeExecutor` and checkpoint, with the working directory under `/tmp`, under `$TMPDIR` and in a path with spaces and CJK characters, exec diagnostics report created, modified and decomposed-name files, follow `working_dir`, report an undecodable file at the line where decoding stops, a syntax error past 256 KiB and a skip past 2 MiB, and withhold `.gitignore`d files, including decomposed names and names git spells as neither given nor NFC; across repeated runs on both interpreters, a file near the 256 KiB cap is checked in 52-135 ms (117-128 ms against a baseline), the syntax of a file near 2 MiB in 318-371 ms, and 600 ignored files plus one visible file take 73-90 ms with the event loop held at most 20 ms. At this head CI passes the unit suite on Linux/Python 3.12 (32619 tests over four shards), and its two Windows steps pass the diagnostic files on native Windows/Python 3.13: 122 file/worker tests, and 27 exec diagnostic tests with 17 pre-existing executor tests deselected; the symlink fixture skips only where the host cannot create links. ```sh env -u SERPER_API_KEY -u RAVEN_MIRRORED_SERPER_API_KEY uv run --no-sync pytest tests/test_agents_code_diagnostic_workers.py tests/test_agents_code_tools_plugin.py tests/test_agents_code_flow_exec.py tests/test_agents_code_flow_plugin.py tests/test_agents_code_flow_read_state.py tests/test_agents_code_launcher.py tests/test_workdir_snapshot.py tests/test_shell_command_writes.py tests/test_runtime_checkpoint.py tests/test_trajectory_upload.py::test_blocked_upload_hits_the_deadline -o addopts= -q --tb=short --maxfail=3 uv run --no-sync ruff check raven/agent/loop/checkpoint.py agents/raven-code/plugins/code-flow/code_flow/tools agents/raven-code/plugins/code-flow/code_flow/config.py tests/test_runtime_checkpoint.py tests/test_agents_code_flow_exec.py tests/test_agents_code_tools_plugin.py tests/test_agents_code_diagnostic_workers.py uv run --no-sync ty check raven/agent/loop/checkpoint.py agents/raven-code/plugins/code-flow/code_flow/tools agents/raven-code/plugins/code-flow/code_flow/config.py uv run --no-sync lint-imports uv run --no-sync ruff format --check raven/agent/loop/checkpoint.py agents/raven-code/plugins/code-flow/code_flow/tools agents/raven-code/plugins/code-flow/code_flow/config.py tests/test_runtime_checkpoint.py tests/test_agents_code_flow_exec.py tests/test_agents_code_tools_plugin.py ``` Native Windows CI commands: ```sh uv run --frozen --python 3.13 --no-sync pytest tests/test_agents_code_diagnostic_workers.py tests/test_agents_code_tools_plugin.py -q -n 0 --tb=short uv run --frozen --python 3.13 --no-sync pytest tests/test_agents_code_flow_exec.py -k test_exec_ -q -n 0 --tb=short ``` Ruff, format and type checks passed; all 10 import contracts kept; the commit-message, large-file and source-language checks pass over the branch. The launcher tests unset inherited search API keys to keep the expected tool list deterministic. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [x] User-facing docs or screenshots are updated when needed ## Risk Scope: Raven-Code's enabled code-flow tool face; installing Pyflakes does not activate checks in other agents. The change is reversible and never fixes/removes code or rolls back edits. Static diagnostics can need judgment for dynamic names, target Python versions, and project lint configuration; only explicit noqa directives are honored. Passing diagnostics do not replace tests. The `trackable` fix reaches every exec write record, not only Raven-Code: a file whose name, as given, as NFC or in git's macOS spelling, matches an ignore rule keeps its counts and loses its diff and removal body. Where git matches another of those spellings, as off macOS for a decomposed name under a precomposed rule, the checkpoint stores the file but it is still withheld; its counts remain. Exec diagnostics now require an active checkpoint. Noninteractive runs using the default interactive checkpoint policy must explicitly enable checkpoints to receive this feedback. Visibility is checked in batches of up to 256 files and stops once eight visible files are found; each batch holds the event loop while the host resolves its paths, about 130 ms for a full batch on native Windows. Each file-check wait is one second, and post-command diagnostics have a three-second wait budget; past it, or when they fail as a whole, they are dropped and the command's result is kept. These are caller budgets, not hard I/O deadlines: running OS reads and inherited synchronous host callbacks cannot be forcibly interrupted. Private worker capacity remains bounded. Between 256 KiB and 2 MiB only syntax is checked: near 2 MiB that took 0.3-0.4 s and about 190 MiB here, up to four checks run at once, and on slower hardware such a check can outlast the one-second wait and report diagnostics unavailable. Python's GIL means concurrency does not promise a CPU speedup. Windows CI exercises the diagnostics layer with the real native DirectExecutor. The existing CodeExecutor environment probe uses POSIX quoting and already fails on Windows at the base; this PR neither changes that probe nor claims end-to-end compatibility for it. Symlink setup is skipped only if the host cannot create the fixture. Disable `pythonLint` for syntax-only feedback, `execPythonDiagnostics` for no post-command checks, or `pythonSyntaxNote` for no diagnostics, under `plugins.config["code-flow"].tools`. The PR can also be reverted as a unit. - [x] Security impact considered - [x] Backward compatibility considered - [x] Rollback path is clear for risky changes ## Related Issues N/A --------- Co-authored-by: Codex <noreply@openai.com> Co-authored-by: Dizhan Xue <xuedizhan17@mails.ucas.ac.cn> Co-authored-by: Claude (claude-opus-5-5) <noreply@anthropic.com>
2026-10-18 20:21:04 +08:00
import pytest
from raven.agent.loop.checkpoint import CheckpointService
feat(acp): a prompt is held open while the agent sleeps on an armed wake, and takes a steer ## Summary An on-call agent submits a job, arms a wake and ends its turn saying so; the wake runs a later turn on the session on its own clock. Over ACP that turn's `end_turn` read as "finished" to the caller: a DAG judging its node spent every continuation on "still watching, wake at 13:13" while the job ran fine (ar-r4, 2026-09-08), and the main agent re-dispatched the work by hand. ACP has no stop reason for "waiting". Replaces the closed !573: **no DAG code changes**, no new node status. The node is `running` because a turn is genuinely in flight. - **The signal.** The loop-hooks paper gains one key the ACP layer reads for itself: `acp_meta["raven.holdTurn"]` (`{"untilMs", "why"}`). oncall-flow files it at the send fire when the window's campaign is claimed, not concluded, and has one pending wake -- read the same way the owner's answer pulls that wake forward. No campaign, a wake that stood down, or a concluded campaign files nothing and the turn ends as before. - **The hold** (`raven/acp/methods.py`). A turn that ended `end_turn` with the key filed is not answered. One `tool_call` stands for the wait (its title names the campaign and the next look), with a `tool_call_update` heartbeat every `HOLD_HEARTBEAT_S` (60 s) so a client that clocks liveness by frames sees the agent alive while its model sleeps. The turn slot is reopened with no id to correlate on, so whichever turn ends next on the session -- the wake's, a steer's -- settles it, and the handler decides again from that turn's stash. A wake that arms again keeps the hold; a turn ending with nothing pending answers `end_turn`; a cancel answers `cancelled` as before. The key itself is not passed to the client. - **Steer.** `_raven/session/steer` is served and advertised under `agentCapabilities._meta` as `raven.steer` (the names raven's own client already uses). Accepted while a prompt is open on the session, held or running: the text is shown on the wire as the person's own line, then sent with `turn.send`'s new `busy: "inject"`. Answers `injected` / `no_turn`. - **Wire.** `turn.send` gains the optional `busy` param (`const "inject"`): an ordinary send on an idle lane, `BusyPolicy.INJECT` into the running turn otherwise -- the merge a channel's mid-turn message already gets (`gateway_commands.py`). No slot is bound for an injected message; the running turn keeps the lane's. `openrpc.json` and both generated clients moved together. What raven's client side already does, unchanged: the DAG stall watcher resets on any frame (the heartbeat keeps it quiet, and a stopped heartbeat raises the existing stall notice after ten minutes); a sub-agent process that dies fails the request and the node takes its ordinary path; the pool's `_responsive` treats a prompt in flight as busy. ## Review round 1 (2026-09-09, on !600, superseded by this MR) - **`busy: inject` reads the scheduler's lane state too.** `is_turn_active` knows only the turns `turn.send` started; an armed wake is submitted to the scheduler directly (`make_on_session_wake`) and never binds that map, so a steer during a running wake queued a second turn and answered `injected`. `turn.send` now asks `Scheduler.has_inflight(lane)` as well, and merges (`BusyPolicy.INJECT`) when either record says the lane is running. Pinned: `test_busy_inject_merges_into_a_wake_the_real_scheduler_is_running` -- the reviewer's reproduction on the real `Scheduler`, a turn submitted the way `make_on_session_wake` submits one; with the lane check bypassed it reads `pending=1, inject_mailbox=0`, with it `(1, 0)` in the mailbox and the running turn drains the text. - Branch renamed to snake_case and both commits carry the `Co-authored-by: Claude (<model-id>)` trailer, per AGENTS.md §2–3; that is why this MR replaces !600 rather than continuing it. ## Review round 2 (2026-09-10) - **A wake re-armed during its own run survives the writeback.** `CronService._writeback_locked` reloaded the store after the run and, for a one-shot with `delete_after_run`, overwrote the id with the finished run's state and removed it -- but the turn the wake ran had just called `schedule_wake` under the same key, so the id now named the *next* look. It vanished, and the held prompt had nothing left to settle it. A stored job under the fired id whose `at` schedule the run never had is treated as the new wake it is: claim cleared, run recorded, schedule kept, not deleted. Pinned on the real service: `test_a_wake_re_armed_during_its_own_run_survives_the_writeback` (reads `[]` without the fix, the reviewer's state) and the unchanged one-shot deletion beside it. Cron, proactive, sentinel and on-call suites: 968 passed. ## Review round 3 (2026-09-10) - **An undrained inject stays visible and cancellable.** `busy: inject` dropped the scheduler's handle; when the host turn ended past its last drain, the spine's fallback ran the text as a turn nothing owned -- `turn.cancel` a no-op, no `message.start`. The handle is now kept per lane under an id minted for the text; `turn.cancel` / `session.interrupt` reach it when the active slot is empty (mailbox, queued fallback, or running); and when the fallback turn starts, the RPC sink (`on_turn_start`, wired from bootstrap) binds the lane to it and opens it on the wire with the person's words. `turn.send` answers the inject's own id. Pinned through `build_rpc_spine` with the reviewer's shape (a loop that drains once, an inject after it): `test_an_undrained_inject_runs_as_a_turn_the_surface_can_see_and_cancel` and `test_a_pending_inject_can_be_cancelled_before_it_merges_or_falls_back`; both read the reviewer's state (`turn_ids` empty, cancel finding nothing) with the fix bypassed. - **Line ceiling.** The contract comment is five lines; `raven/contracts` is 2,519 against 2,520, one line of headroom on the current `main` (2,514). - Non-blocking items taken: `SchedulerDrainingError` on the inject path exits as `turn_failed`; the title is announced before the hold; the steer constants sit below the cancel constant's docstring; three edges pinned (a held prompt settled by a wake turn's uncorrelated ending through `send_frame`, the hold key never in the client's `_meta`, `raven.toolName` on the wait row). - Focused suite (reviewer's list plus the touched files): 948 passed on Python 3.12. ## Review round 4 (2026-09-10) - **Every pending inject is kept, per lane and per id.** One replaceable record per lane kept only the newest steer; two steers after the host's last drain fall back in turn, and the first started unowned while a cancel took out the second. The record is now `lane -> {inject id: (handle, text)}` in arrival order: each fallback promotes its own handle, and a cancel with no active turn reaches the oldest pending one -- the one the spine runs next. Pinned: `test_two_undrained_injects_each_get_their_own_turn_promoted_and_cancelled` (both promoted with their own ids and `message.start`, cancelled one after the other; runner saw all three turns). Focused suite: 949 passed. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification ``` uv run pytest tests/test_acp_methods.py tests/test_acp_protocol.py tests/test_rpc_turn_send.py \ tests/test_rpc_schema_match.py tests/test_acp_updates.py tests/test_acp_server.py -q # 766 passed uv run pytest tests/test_agents_oncall_flow_hooks.py tests/test_agents_oncall_flow_actions.py \ tests/test_agents_oncall_flow_budget.py tests/test_agents_oncall_flow_readings.py \ tests/test_agents_oncall_flow_escalation_tools.py -q # 99 passed uv run ruff check raven tests agents && uv run ruff format --check raven tests agents # clean uv run lint-imports # 9 kept, 0 broken cd ui-tui && npm run lint:rpc && npx tsc --noEmit; cd ui-web && npm run gen:check && npx tsc --noEmit # in sync, clean ``` New tests: `TestHeldTurn` (held across an armed wake, held again on a re-arm, heartbeat on the wait call, cancel while held, a non-`end_turn` stop is not held; every frame validated against the ACP schema), `TestSteer` (advertised, `no_turn` when idle, a held prompt takes the words as a turn with `busy: inject`, refusals), `busy: inject` on a busy and an idle lane in `test_rpc_turn_send.py`, and the plugin's hold stamping through the real `CronService` grant (filed with a pending wake; not filed without one, after the wake stands down, or once the campaign concluded). Not done: no end-to-end run of a real on-call wake through a held prompt yet -- unit tests only. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [ ] User-facing docs or screenshots are updated when needed (the hold is invisible to a person except as a "waiting for ..." tool row) ## Risk - [x] Backward compatibility considered: a turn that files no `raven.holdTurn` is answered exactly as before; `busy` is optional and absent from every existing caller; the steer method and capability are additive. - [x] Rollback path is clear: revert the two commits. A held prompt holds one `max_concurrent_subagents` slot (default 8) for the wait's duration on the caller's side; raised in config if it bites. ## Related Replaces !573 (closed). Follows !518 (stall notice) and !520 (unprompted-turn route; a wake that runs under a held prompt no longer takes it for DAG nodes).
2026-09-11 02:37:55 +00:00
from raven.agent.tools.registry import ToolRegistry
from raven.agent.tools.shell import ExecTool as TrunkExecTool
from raven.plugins.context import PluginContext, ServiceLocator
feat(*): add bounded Python diagnostics after code edits (#967) ## Summary Raven-Code reports Python syntax errors after file edits but misses unresolved names and equivalent feedback after shell edits. Add advisory Pyflakes F821/F822/F823 checks to the existing tool result; optional F401 notices remain off by default. No additional model call or code execution is involved. ```text write_file / edit_file -> inspect final source, compare with a baseline that can be read and analyzed -> retain syntax warnings; suppress existing name findings exec -> require checkpoint visibility for the working directory -> read the changed files from the host's write record for the command -> ask visibility in batches of 256, select at most 8, check up to 4 concurrently -> append ordered diagnostics, preserving exit status and file metadata ``` Existing syntax failures remain visible after unrelated edits and identical rewrites. A baseline that cannot be read, parsed or analyzed is marked unknown; overwriting non-UTF-8 source still checks the known replacement. Source nested too deeply for the parser or checker, and a changed file that cannot be read, report diagnostics unavailable for that file rather than passing as clean. Every check reads a file as the bytes Python runs: text a write or edit stores is checked as its UTF-8 bytes, the content it replaced included, and a file from disk as its bytes, each decoded as Python decodes source, by its coding declaration or as UTF-8. Bytes that do not decode are reported as a syntax error at the line where decoding stops, so a declaration the written text cannot satisfy is reported however the file was changed. Noqa comments are read as Ruff reads them, including after other comment directives and never inside string literals: a file-level `ruff: noqa` / `flake8: noqa` counts only on a line of its own, and a marker counts only when codes follow its colon or, with no colon, whitespace, another comment or the end of the comment follows it. Unlike Ruff, which takes the first directive in a comment, every directive in it counts. The seeded tool guide gives the syntax note's own advice. With no checkpoint or no visibility verdict, exec diagnostics do not inspect changed file contents; direct file checks remain available. Pyflakes is imported when the tool face loads, not by whichever check first needs it: imported lazily, its cold import fell inside that check's one-second wait, and on a slow Windows runner the one test that paid it reported diagnostics unavailable for a plain syntax error. Private workers bound reads and checks to four active jobs per process. Saturated capacity rejects new work without queuing, and timeout/cancellation retains each slot until the actual work finishes. Name checks are capped at 256 KiB of source and the syntax check at 2 MiB, and a check skipped for size says so in one line. Findings are capped at eight per file, and exec diagnostic text at 6,000 characters plus omission notices. Runtime changes stay inside the Raven-Code code-flow plugin, except for one fix to `CheckpointService.trackable`, the visibility verdict that exec diagnostics and the host's exec write record both rely on. On macOS, git precomposes a file name it reads off the disk before matching an ignore rule, but takes the paths `trackable` sends to `git check-ignore --stdin` as written. A file stored with a decomposed (NFD) name under a `.gitignore` rule written precomposed (NFC) is therefore left out of the checkpoint, yet main vouches for it, so the write record's diff carries its text, and exec diagnostics would quote its source lines to the model. Nor is git's precomposition NFC: it composes only neighbouring characters, by Unicode 3.2's table, and reorders nothing, so a name whose combining marks are out of canonical order comes out as neither the name as given nor its NFC. `trackable` now asks about each name as given, as NFC and in git's macOS spelling, and withholds it when any spelling is ignored. Other shared file/shell/snapshot implementations are unchanged. Raven-Code has no separate dependency manifest, so Pyflakes uses the existing root dependencies. It is the only added package; uv normalized existing lockfile markers without changing other package versions. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification - **Before:** on main, Raven-Code's file tools report Python syntax errors only, and its exec reports nothing about the Python files a command changed. On macOS, main's `CheckpointService.trackable` vouches for a file stored as `secret_cafe` + U+0301 + `.py` under the rule `secret_caf` + U+00E9 + `.py`: the user's own git ignores it and the checkpoint's staged tree leaves it out, yet its text reaches the exec write record's stored diff. - **After:** on macOS 26.6.2 / arm64 with Apple Git 2.54.0, the command below passes 364 tests on Python 3.12.13 and on 3.14.6. The full default suite on 3.12.13 has 32283 passed, 169 skipped, 20 failed and 3 errors over 22 test ids: 21 of them also fail on the exact base in the same environment, and the other, a remote process-tree cancel test that missed its deadline under the suite's load, passes when run alone. Tests cover actual worker overlap, retained capacity after timeout/cancellation, host-tool isolation, hidden-file quotas, visibility batching through the host's write-record wiring, baseline handling, parser-limit failures, the size bounds and their skip notes, undecodable bytes placed where decoding stops, coding declarations in written and edited text, the baseline read staying on a worker, the checker loaded with its module, the file each unavailable note names, comment directives and their marker boundary, the tool guide's agreement with the syntax note, portable commands in paths with spaces/non-ASCII characters, and the checkpoint's verdict on decomposed names and on names git spells as neither given nor NFC, held against its own staged tree. Removing any of the three spellings from the query fails the cases that pin it, as does removing the marker boundary, the parser-limit report, the unknown-baseline handling, the syntax check past the name-check bound and the decoding report, checking written text as its stored bytes, the worker around the baseline read, loading the checker with its module, or the file name in an unavailable note; restoring the guide's earlier advice fails its agreement test. git's macOS spelling is what libiconv's UTF-8-MAC conversion returns, which git precomposes with: composing neighbours by Unicode 3.2's table reproduced it for all 281,928 names built from every decomposable character's decomposition (reordered, partly composed, or with a mark added), and across 784,220 further mixed names and pairs no name was spelled outside the three asked; the same held under Python 3.13 and 3.14. The decomposed-name check asserts the macOS outcome only where git precomposes names, so it also passes with `core.precomposeunicode = false` set through `GIT_CONFIG_GLOBAL`. Across 102 noqa comment variants the reported lines match Ruff 0.15.12, except the three that hold several directives in one comment. Through the real `AgentLoop`, code-flow tool face, `CodeExecutor` and checkpoint, with the working directory under `/tmp`, under `$TMPDIR` and in a path with spaces and CJK characters, exec diagnostics report created, modified and decomposed-name files, follow `working_dir`, report an undecodable file at the line where decoding stops, a syntax error past 256 KiB and a skip past 2 MiB, and withhold `.gitignore`d files, including decomposed names and names git spells as neither given nor NFC; across repeated runs on both interpreters, a file near the 256 KiB cap is checked in 52-135 ms (117-128 ms against a baseline), the syntax of a file near 2 MiB in 318-371 ms, and 600 ignored files plus one visible file take 73-90 ms with the event loop held at most 20 ms. At this head CI passes the unit suite on Linux/Python 3.12 (32619 tests over four shards), and its two Windows steps pass the diagnostic files on native Windows/Python 3.13: 122 file/worker tests, and 27 exec diagnostic tests with 17 pre-existing executor tests deselected; the symlink fixture skips only where the host cannot create links. ```sh env -u SERPER_API_KEY -u RAVEN_MIRRORED_SERPER_API_KEY uv run --no-sync pytest tests/test_agents_code_diagnostic_workers.py tests/test_agents_code_tools_plugin.py tests/test_agents_code_flow_exec.py tests/test_agents_code_flow_plugin.py tests/test_agents_code_flow_read_state.py tests/test_agents_code_launcher.py tests/test_workdir_snapshot.py tests/test_shell_command_writes.py tests/test_runtime_checkpoint.py tests/test_trajectory_upload.py::test_blocked_upload_hits_the_deadline -o addopts= -q --tb=short --maxfail=3 uv run --no-sync ruff check raven/agent/loop/checkpoint.py agents/raven-code/plugins/code-flow/code_flow/tools agents/raven-code/plugins/code-flow/code_flow/config.py tests/test_runtime_checkpoint.py tests/test_agents_code_flow_exec.py tests/test_agents_code_tools_plugin.py tests/test_agents_code_diagnostic_workers.py uv run --no-sync ty check raven/agent/loop/checkpoint.py agents/raven-code/plugins/code-flow/code_flow/tools agents/raven-code/plugins/code-flow/code_flow/config.py uv run --no-sync lint-imports uv run --no-sync ruff format --check raven/agent/loop/checkpoint.py agents/raven-code/plugins/code-flow/code_flow/tools agents/raven-code/plugins/code-flow/code_flow/config.py tests/test_runtime_checkpoint.py tests/test_agents_code_flow_exec.py tests/test_agents_code_tools_plugin.py ``` Native Windows CI commands: ```sh uv run --frozen --python 3.13 --no-sync pytest tests/test_agents_code_diagnostic_workers.py tests/test_agents_code_tools_plugin.py -q -n 0 --tb=short uv run --frozen --python 3.13 --no-sync pytest tests/test_agents_code_flow_exec.py -k test_exec_ -q -n 0 --tb=short ``` Ruff, format and type checks passed; all 10 import contracts kept; the commit-message, large-file and source-language checks pass over the branch. The launcher tests unset inherited search API keys to keep the expected tool list deterministic. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [x] User-facing docs or screenshots are updated when needed ## Risk Scope: Raven-Code's enabled code-flow tool face; installing Pyflakes does not activate checks in other agents. The change is reversible and never fixes/removes code or rolls back edits. Static diagnostics can need judgment for dynamic names, target Python versions, and project lint configuration; only explicit noqa directives are honored. Passing diagnostics do not replace tests. The `trackable` fix reaches every exec write record, not only Raven-Code: a file whose name, as given, as NFC or in git's macOS spelling, matches an ignore rule keeps its counts and loses its diff and removal body. Where git matches another of those spellings, as off macOS for a decomposed name under a precomposed rule, the checkpoint stores the file but it is still withheld; its counts remain. Exec diagnostics now require an active checkpoint. Noninteractive runs using the default interactive checkpoint policy must explicitly enable checkpoints to receive this feedback. Visibility is checked in batches of up to 256 files and stops once eight visible files are found; each batch holds the event loop while the host resolves its paths, about 130 ms for a full batch on native Windows. Each file-check wait is one second, and post-command diagnostics have a three-second wait budget; past it, or when they fail as a whole, they are dropped and the command's result is kept. These are caller budgets, not hard I/O deadlines: running OS reads and inherited synchronous host callbacks cannot be forcibly interrupted. Private worker capacity remains bounded. Between 256 KiB and 2 MiB only syntax is checked: near 2 MiB that took 0.3-0.4 s and about 190 MiB here, up to four checks run at once, and on slower hardware such a check can outlast the one-second wait and report diagnostics unavailable. Python's GIL means concurrency does not promise a CPU speedup. Windows CI exercises the diagnostics layer with the real native DirectExecutor. The existing CodeExecutor environment probe uses POSIX quoting and already fails on Windows at the base; this PR neither changes that probe nor claims end-to-end compatibility for it. Symlink setup is skipped only if the host cannot create the fixture. Disable `pythonLint` for syntax-only feedback, `execPythonDiagnostics` for no post-command checks, or `pythonSyntaxNote` for no diagnostics, under `plugins.config["code-flow"].tools`. The PR can also be reverted as a unit. - [x] Security impact considered - [x] Backward compatibility considered - [x] Rollback path is clear for risky changes ## Related Issues N/A --------- Co-authored-by: Codex <noreply@openai.com> Co-authored-by: Dizhan Xue <xuedizhan17@mails.ucas.ac.cn> Co-authored-by: Claude (claude-opus-5-5) <noreply@anthropic.com>
2026-10-18 20:21:04 +08:00
from raven.sandbox.direct_executor import DirectExecutor
feat(acp): a prompt is held open while the agent sleeps on an armed wake, and takes a steer ## Summary An on-call agent submits a job, arms a wake and ends its turn saying so; the wake runs a later turn on the session on its own clock. Over ACP that turn's `end_turn` read as "finished" to the caller: a DAG judging its node spent every continuation on "still watching, wake at 13:13" while the job ran fine (ar-r4, 2026-09-08), and the main agent re-dispatched the work by hand. ACP has no stop reason for "waiting". Replaces the closed !573: **no DAG code changes**, no new node status. The node is `running` because a turn is genuinely in flight. - **The signal.** The loop-hooks paper gains one key the ACP layer reads for itself: `acp_meta["raven.holdTurn"]` (`{"untilMs", "why"}`). oncall-flow files it at the send fire when the window's campaign is claimed, not concluded, and has one pending wake -- read the same way the owner's answer pulls that wake forward. No campaign, a wake that stood down, or a concluded campaign files nothing and the turn ends as before. - **The hold** (`raven/acp/methods.py`). A turn that ended `end_turn` with the key filed is not answered. One `tool_call` stands for the wait (its title names the campaign and the next look), with a `tool_call_update` heartbeat every `HOLD_HEARTBEAT_S` (60 s) so a client that clocks liveness by frames sees the agent alive while its model sleeps. The turn slot is reopened with no id to correlate on, so whichever turn ends next on the session -- the wake's, a steer's -- settles it, and the handler decides again from that turn's stash. A wake that arms again keeps the hold; a turn ending with nothing pending answers `end_turn`; a cancel answers `cancelled` as before. The key itself is not passed to the client. - **Steer.** `_raven/session/steer` is served and advertised under `agentCapabilities._meta` as `raven.steer` (the names raven's own client already uses). Accepted while a prompt is open on the session, held or running: the text is shown on the wire as the person's own line, then sent with `turn.send`'s new `busy: "inject"`. Answers `injected` / `no_turn`. - **Wire.** `turn.send` gains the optional `busy` param (`const "inject"`): an ordinary send on an idle lane, `BusyPolicy.INJECT` into the running turn otherwise -- the merge a channel's mid-turn message already gets (`gateway_commands.py`). No slot is bound for an injected message; the running turn keeps the lane's. `openrpc.json` and both generated clients moved together. What raven's client side already does, unchanged: the DAG stall watcher resets on any frame (the heartbeat keeps it quiet, and a stopped heartbeat raises the existing stall notice after ten minutes); a sub-agent process that dies fails the request and the node takes its ordinary path; the pool's `_responsive` treats a prompt in flight as busy. ## Review round 1 (2026-09-09, on !600, superseded by this MR) - **`busy: inject` reads the scheduler's lane state too.** `is_turn_active` knows only the turns `turn.send` started; an armed wake is submitted to the scheduler directly (`make_on_session_wake`) and never binds that map, so a steer during a running wake queued a second turn and answered `injected`. `turn.send` now asks `Scheduler.has_inflight(lane)` as well, and merges (`BusyPolicy.INJECT`) when either record says the lane is running. Pinned: `test_busy_inject_merges_into_a_wake_the_real_scheduler_is_running` -- the reviewer's reproduction on the real `Scheduler`, a turn submitted the way `make_on_session_wake` submits one; with the lane check bypassed it reads `pending=1, inject_mailbox=0`, with it `(1, 0)` in the mailbox and the running turn drains the text. - Branch renamed to snake_case and both commits carry the `Co-authored-by: Claude (<model-id>)` trailer, per AGENTS.md §2–3; that is why this MR replaces !600 rather than continuing it. ## Review round 2 (2026-09-10) - **A wake re-armed during its own run survives the writeback.** `CronService._writeback_locked` reloaded the store after the run and, for a one-shot with `delete_after_run`, overwrote the id with the finished run's state and removed it -- but the turn the wake ran had just called `schedule_wake` under the same key, so the id now named the *next* look. It vanished, and the held prompt had nothing left to settle it. A stored job under the fired id whose `at` schedule the run never had is treated as the new wake it is: claim cleared, run recorded, schedule kept, not deleted. Pinned on the real service: `test_a_wake_re_armed_during_its_own_run_survives_the_writeback` (reads `[]` without the fix, the reviewer's state) and the unchanged one-shot deletion beside it. Cron, proactive, sentinel and on-call suites: 968 passed. ## Review round 3 (2026-09-10) - **An undrained inject stays visible and cancellable.** `busy: inject` dropped the scheduler's handle; when the host turn ended past its last drain, the spine's fallback ran the text as a turn nothing owned -- `turn.cancel` a no-op, no `message.start`. The handle is now kept per lane under an id minted for the text; `turn.cancel` / `session.interrupt` reach it when the active slot is empty (mailbox, queued fallback, or running); and when the fallback turn starts, the RPC sink (`on_turn_start`, wired from bootstrap) binds the lane to it and opens it on the wire with the person's words. `turn.send` answers the inject's own id. Pinned through `build_rpc_spine` with the reviewer's shape (a loop that drains once, an inject after it): `test_an_undrained_inject_runs_as_a_turn_the_surface_can_see_and_cancel` and `test_a_pending_inject_can_be_cancelled_before_it_merges_or_falls_back`; both read the reviewer's state (`turn_ids` empty, cancel finding nothing) with the fix bypassed. - **Line ceiling.** The contract comment is five lines; `raven/contracts` is 2,519 against 2,520, one line of headroom on the current `main` (2,514). - Non-blocking items taken: `SchedulerDrainingError` on the inject path exits as `turn_failed`; the title is announced before the hold; the steer constants sit below the cancel constant's docstring; three edges pinned (a held prompt settled by a wake turn's uncorrelated ending through `send_frame`, the hold key never in the client's `_meta`, `raven.toolName` on the wait row). - Focused suite (reviewer's list plus the touched files): 948 passed on Python 3.12. ## Review round 4 (2026-09-10) - **Every pending inject is kept, per lane and per id.** One replaceable record per lane kept only the newest steer; two steers after the host's last drain fall back in turn, and the first started unowned while a cancel took out the second. The record is now `lane -> {inject id: (handle, text)}` in arrival order: each fallback promotes its own handle, and a cancel with no active turn reaches the oldest pending one -- the one the spine runs next. Pinned: `test_two_undrained_injects_each_get_their_own_turn_promoted_and_cancelled` (both promoted with their own ids and `message.start`, cancelled one after the other; runner saw all three turns). Focused suite: 949 passed. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification ``` uv run pytest tests/test_acp_methods.py tests/test_acp_protocol.py tests/test_rpc_turn_send.py \ tests/test_rpc_schema_match.py tests/test_acp_updates.py tests/test_acp_server.py -q # 766 passed uv run pytest tests/test_agents_oncall_flow_hooks.py tests/test_agents_oncall_flow_actions.py \ tests/test_agents_oncall_flow_budget.py tests/test_agents_oncall_flow_readings.py \ tests/test_agents_oncall_flow_escalation_tools.py -q # 99 passed uv run ruff check raven tests agents && uv run ruff format --check raven tests agents # clean uv run lint-imports # 9 kept, 0 broken cd ui-tui && npm run lint:rpc && npx tsc --noEmit; cd ui-web && npm run gen:check && npx tsc --noEmit # in sync, clean ``` New tests: `TestHeldTurn` (held across an armed wake, held again on a re-arm, heartbeat on the wait call, cancel while held, a non-`end_turn` stop is not held; every frame validated against the ACP schema), `TestSteer` (advertised, `no_turn` when idle, a held prompt takes the words as a turn with `busy: inject`, refusals), `busy: inject` on a busy and an idle lane in `test_rpc_turn_send.py`, and the plugin's hold stamping through the real `CronService` grant (filed with a pending wake; not filed without one, after the wake stands down, or once the campaign concluded). Not done: no end-to-end run of a real on-call wake through a held prompt yet -- unit tests only. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [ ] User-facing docs or screenshots are updated when needed (the hold is invisible to a person except as a "waiting for ..." tool row) ## Risk - [x] Backward compatibility considered: a turn that files no `raven.holdTurn` is answered exactly as before; `busy` is optional and absent from every existing caller; the steer method and capability are additive. - [x] Rollback path is clear: revert the two commits. A held prompt holds one `max_concurrent_subagents` slot (default 8) for the wait's duration on the caller's side; raised in config if it bites. ## Related Replaces !573 (closed). Follows !518 (stall notice) and !520 (unprompted-turn route; a wake that runs under a held prompt no longer takes it for DAG nodes).
2026-09-11 02:37:55 +00:00
REPO = Path(__file__).resolve().parent.parent
PLUGIN_DIR = REPO / "agents" / "raven-code" / "plugins" / "code-flow"
sys.path.insert(0, str(PLUGIN_DIR))
from code_flow.tools import plugin as factories # noqa: E402
from code_flow.tools.exec import ( # noqa: E402
MAX_OUTPUT,
SPILL_SUBDIR,
TIMED_OUT_NOTE,
CodeExecResult,
CodeExecTool,
CodeExecutor,
)
def _run(coro):
return asyncio.run(coro)
def _ctx(tmp_path: Path, tools: dict | None = None) -> PluginContext:
return PluginContext(
config={"enabled": True, "tools": {"enabled": True, **(tools or {})}},
services=ServiceLocator(workspace=tmp_path / "home", user_id="u", agent_id="a"),
)
def _tool(tmp_path: Path, max_timeout: int = 1200) -> CodeExecTool:
home = tmp_path / "home"
home.mkdir(exist_ok=True)
return CodeExecTool(
timeout=60,
working_dir=str(home),
executor=CodeExecutor(max_timeout=max_timeout, spill_dir=home / SPILL_SUBDIR),
extra_allowed_dirs=(home,),
max_timeout=max_timeout,
)
feat(*): add bounded Python diagnostics after code edits (#967) ## Summary Raven-Code reports Python syntax errors after file edits but misses unresolved names and equivalent feedback after shell edits. Add advisory Pyflakes F821/F822/F823 checks to the existing tool result; optional F401 notices remain off by default. No additional model call or code execution is involved. ```text write_file / edit_file -> inspect final source, compare with a baseline that can be read and analyzed -> retain syntax warnings; suppress existing name findings exec -> require checkpoint visibility for the working directory -> read the changed files from the host's write record for the command -> ask visibility in batches of 256, select at most 8, check up to 4 concurrently -> append ordered diagnostics, preserving exit status and file metadata ``` Existing syntax failures remain visible after unrelated edits and identical rewrites. A baseline that cannot be read, parsed or analyzed is marked unknown; overwriting non-UTF-8 source still checks the known replacement. Source nested too deeply for the parser or checker, and a changed file that cannot be read, report diagnostics unavailable for that file rather than passing as clean. Every check reads a file as the bytes Python runs: text a write or edit stores is checked as its UTF-8 bytes, the content it replaced included, and a file from disk as its bytes, each decoded as Python decodes source, by its coding declaration or as UTF-8. Bytes that do not decode are reported as a syntax error at the line where decoding stops, so a declaration the written text cannot satisfy is reported however the file was changed. Noqa comments are read as Ruff reads them, including after other comment directives and never inside string literals: a file-level `ruff: noqa` / `flake8: noqa` counts only on a line of its own, and a marker counts only when codes follow its colon or, with no colon, whitespace, another comment or the end of the comment follows it. Unlike Ruff, which takes the first directive in a comment, every directive in it counts. The seeded tool guide gives the syntax note's own advice. With no checkpoint or no visibility verdict, exec diagnostics do not inspect changed file contents; direct file checks remain available. Pyflakes is imported when the tool face loads, not by whichever check first needs it: imported lazily, its cold import fell inside that check's one-second wait, and on a slow Windows runner the one test that paid it reported diagnostics unavailable for a plain syntax error. Private workers bound reads and checks to four active jobs per process. Saturated capacity rejects new work without queuing, and timeout/cancellation retains each slot until the actual work finishes. Name checks are capped at 256 KiB of source and the syntax check at 2 MiB, and a check skipped for size says so in one line. Findings are capped at eight per file, and exec diagnostic text at 6,000 characters plus omission notices. Runtime changes stay inside the Raven-Code code-flow plugin, except for one fix to `CheckpointService.trackable`, the visibility verdict that exec diagnostics and the host's exec write record both rely on. On macOS, git precomposes a file name it reads off the disk before matching an ignore rule, but takes the paths `trackable` sends to `git check-ignore --stdin` as written. A file stored with a decomposed (NFD) name under a `.gitignore` rule written precomposed (NFC) is therefore left out of the checkpoint, yet main vouches for it, so the write record's diff carries its text, and exec diagnostics would quote its source lines to the model. Nor is git's precomposition NFC: it composes only neighbouring characters, by Unicode 3.2's table, and reorders nothing, so a name whose combining marks are out of canonical order comes out as neither the name as given nor its NFC. `trackable` now asks about each name as given, as NFC and in git's macOS spelling, and withholds it when any spelling is ignored. Other shared file/shell/snapshot implementations are unchanged. Raven-Code has no separate dependency manifest, so Pyflakes uses the existing root dependencies. It is the only added package; uv normalized existing lockfile markers without changing other package versions. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification - **Before:** on main, Raven-Code's file tools report Python syntax errors only, and its exec reports nothing about the Python files a command changed. On macOS, main's `CheckpointService.trackable` vouches for a file stored as `secret_cafe` + U+0301 + `.py` under the rule `secret_caf` + U+00E9 + `.py`: the user's own git ignores it and the checkpoint's staged tree leaves it out, yet its text reaches the exec write record's stored diff. - **After:** on macOS 26.6.2 / arm64 with Apple Git 2.54.0, the command below passes 364 tests on Python 3.12.13 and on 3.14.6. The full default suite on 3.12.13 has 32283 passed, 169 skipped, 20 failed and 3 errors over 22 test ids: 21 of them also fail on the exact base in the same environment, and the other, a remote process-tree cancel test that missed its deadline under the suite's load, passes when run alone. Tests cover actual worker overlap, retained capacity after timeout/cancellation, host-tool isolation, hidden-file quotas, visibility batching through the host's write-record wiring, baseline handling, parser-limit failures, the size bounds and their skip notes, undecodable bytes placed where decoding stops, coding declarations in written and edited text, the baseline read staying on a worker, the checker loaded with its module, the file each unavailable note names, comment directives and their marker boundary, the tool guide's agreement with the syntax note, portable commands in paths with spaces/non-ASCII characters, and the checkpoint's verdict on decomposed names and on names git spells as neither given nor NFC, held against its own staged tree. Removing any of the three spellings from the query fails the cases that pin it, as does removing the marker boundary, the parser-limit report, the unknown-baseline handling, the syntax check past the name-check bound and the decoding report, checking written text as its stored bytes, the worker around the baseline read, loading the checker with its module, or the file name in an unavailable note; restoring the guide's earlier advice fails its agreement test. git's macOS spelling is what libiconv's UTF-8-MAC conversion returns, which git precomposes with: composing neighbours by Unicode 3.2's table reproduced it for all 281,928 names built from every decomposable character's decomposition (reordered, partly composed, or with a mark added), and across 784,220 further mixed names and pairs no name was spelled outside the three asked; the same held under Python 3.13 and 3.14. The decomposed-name check asserts the macOS outcome only where git precomposes names, so it also passes with `core.precomposeunicode = false` set through `GIT_CONFIG_GLOBAL`. Across 102 noqa comment variants the reported lines match Ruff 0.15.12, except the three that hold several directives in one comment. Through the real `AgentLoop`, code-flow tool face, `CodeExecutor` and checkpoint, with the working directory under `/tmp`, under `$TMPDIR` and in a path with spaces and CJK characters, exec diagnostics report created, modified and decomposed-name files, follow `working_dir`, report an undecodable file at the line where decoding stops, a syntax error past 256 KiB and a skip past 2 MiB, and withhold `.gitignore`d files, including decomposed names and names git spells as neither given nor NFC; across repeated runs on both interpreters, a file near the 256 KiB cap is checked in 52-135 ms (117-128 ms against a baseline), the syntax of a file near 2 MiB in 318-371 ms, and 600 ignored files plus one visible file take 73-90 ms with the event loop held at most 20 ms. At this head CI passes the unit suite on Linux/Python 3.12 (32619 tests over four shards), and its two Windows steps pass the diagnostic files on native Windows/Python 3.13: 122 file/worker tests, and 27 exec diagnostic tests with 17 pre-existing executor tests deselected; the symlink fixture skips only where the host cannot create links. ```sh env -u SERPER_API_KEY -u RAVEN_MIRRORED_SERPER_API_KEY uv run --no-sync pytest tests/test_agents_code_diagnostic_workers.py tests/test_agents_code_tools_plugin.py tests/test_agents_code_flow_exec.py tests/test_agents_code_flow_plugin.py tests/test_agents_code_flow_read_state.py tests/test_agents_code_launcher.py tests/test_workdir_snapshot.py tests/test_shell_command_writes.py tests/test_runtime_checkpoint.py tests/test_trajectory_upload.py::test_blocked_upload_hits_the_deadline -o addopts= -q --tb=short --maxfail=3 uv run --no-sync ruff check raven/agent/loop/checkpoint.py agents/raven-code/plugins/code-flow/code_flow/tools agents/raven-code/plugins/code-flow/code_flow/config.py tests/test_runtime_checkpoint.py tests/test_agents_code_flow_exec.py tests/test_agents_code_tools_plugin.py tests/test_agents_code_diagnostic_workers.py uv run --no-sync ty check raven/agent/loop/checkpoint.py agents/raven-code/plugins/code-flow/code_flow/tools agents/raven-code/plugins/code-flow/code_flow/config.py uv run --no-sync lint-imports uv run --no-sync ruff format --check raven/agent/loop/checkpoint.py agents/raven-code/plugins/code-flow/code_flow/tools agents/raven-code/plugins/code-flow/code_flow/config.py tests/test_runtime_checkpoint.py tests/test_agents_code_flow_exec.py tests/test_agents_code_tools_plugin.py ``` Native Windows CI commands: ```sh uv run --frozen --python 3.13 --no-sync pytest tests/test_agents_code_diagnostic_workers.py tests/test_agents_code_tools_plugin.py -q -n 0 --tb=short uv run --frozen --python 3.13 --no-sync pytest tests/test_agents_code_flow_exec.py -k test_exec_ -q -n 0 --tb=short ``` Ruff, format and type checks passed; all 10 import contracts kept; the commit-message, large-file and source-language checks pass over the branch. The launcher tests unset inherited search API keys to keep the expected tool list deterministic. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [x] User-facing docs or screenshots are updated when needed ## Risk Scope: Raven-Code's enabled code-flow tool face; installing Pyflakes does not activate checks in other agents. The change is reversible and never fixes/removes code or rolls back edits. Static diagnostics can need judgment for dynamic names, target Python versions, and project lint configuration; only explicit noqa directives are honored. Passing diagnostics do not replace tests. The `trackable` fix reaches every exec write record, not only Raven-Code: a file whose name, as given, as NFC or in git's macOS spelling, matches an ignore rule keeps its counts and loses its diff and removal body. Where git matches another of those spellings, as off macOS for a decomposed name under a precomposed rule, the checkpoint stores the file but it is still withheld; its counts remain. Exec diagnostics now require an active checkpoint. Noninteractive runs using the default interactive checkpoint policy must explicitly enable checkpoints to receive this feedback. Visibility is checked in batches of up to 256 files and stops once eight visible files are found; each batch holds the event loop while the host resolves its paths, about 130 ms for a full batch on native Windows. Each file-check wait is one second, and post-command diagnostics have a three-second wait budget; past it, or when they fail as a whole, they are dropped and the command's result is kept. These are caller budgets, not hard I/O deadlines: running OS reads and inherited synchronous host callbacks cannot be forcibly interrupted. Private worker capacity remains bounded. Between 256 KiB and 2 MiB only syntax is checked: near 2 MiB that took 0.3-0.4 s and about 190 MiB here, up to four checks run at once, and on slower hardware such a check can outlast the one-second wait and report diagnostics unavailable. Python's GIL means concurrency does not promise a CPU speedup. Windows CI exercises the diagnostics layer with the real native DirectExecutor. The existing CodeExecutor environment probe uses POSIX quoting and already fails on Windows at the base; this PR neither changes that probe nor claims end-to-end compatibility for it. Symlink setup is skipped only if the host cannot create the fixture. Disable `pythonLint` for syntax-only feedback, `execPythonDiagnostics` for no post-command checks, or `pythonSyntaxNote` for no diagnostics, under `plugins.config["code-flow"].tools`. The PR can also be reverted as a unit. - [x] Security impact considered - [x] Backward compatibility considered - [x] Rollback path is clear for risky changes ## Related Issues N/A --------- Co-authored-by: Codex <noreply@openai.com> Co-authored-by: Dizhan Xue <xuedizhan17@mails.ucas.ac.cn> Co-authored-by: Claude (claude-opus-5-5) <noreply@anthropic.com>
2026-10-18 20:21:04 +08:00
def _diagnostic_tool(tmp_path: Path, tools: dict | None = None) -> CodeExecTool:
(tmp_path / "home").mkdir(exist_ok=True)
tool = factories.make_exec(_ctx(tmp_path, tools))
assert tool is not None
# The existing CodeExecutor probe uses POSIX quoting. Exercise the new
# diagnostic layer with the real native host executor on Windows.
if os.name == "nt":
tool._executor = DirectExecutor()
repos = {}
def shadow(root):
if root not in repos:
repos[root] = CheckpointService(root)
return repos[root]
# As the loop wires a same-name exec: the write record and the visibility
# hook arrive together, through the one door production uses.
tool.measure_writes(shadow)
return tool
def _python_command(tmp_path: Path, source: str) -> str:
with tempfile.NamedTemporaryFile(mode="w", encoding="utf-8", dir=tmp_path, suffix=".py", delete=False) as script:
script.write(source)
argv = [sys.executable, script.name]
return subprocess.list2cmdline(argv) if os.name == "nt" else shlex.join(argv)
def _write_command(tmp_path: Path, files: dict[str, str], *, exit_code: int = 0) -> str:
source = "from pathlib import Path\n"
for name, text in files.items():
source += f"Path({name!r}).parent.mkdir(parents=True, exist_ok=True)\n"
source += f"Path({name!r}).write_text({text!r}, encoding='utf-8')\n"
return _python_command(tmp_path, source + f"raise SystemExit({exit_code})\n")
def test_exec_created_python_gets_name_diagnostics_in_the_command_result(tmp_path):
tool = _diagnostic_tool(tmp_path)
out = _run(tool.execute(command=_write_command(tmp_path, {"created.py": "result = missing_value\n"})))
assert "Exit code: 0" in str(out) and "F821" in str(out) and "missing_value" in str(out)
assert getattr(out, "ok", None) is True
assert (tmp_path / "home" / "created.py").read_text() == "result = missing_value\n"
def test_exec_diagnostics_work_in_paths_with_spaces_and_non_ascii_characters(tmp_path):
root = tmp_path / "space path \u00e9"
root.mkdir()
tool = _diagnostic_tool(root)
out = _run(tool.execute(command=_write_command(root, {"created.py": "result = missing_value\n"})))
assert out.ok is True and "F821" in str(out)
assert (root / "home" / "created.py").read_text() == "result = missing_value\n"
def test_exec_reuses_the_hosts_write_record_and_keeps_it_in_the_result(tmp_path):
tool = _diagnostic_tool(tmp_path)
out = _run(tool.execute(command=_write_command(tmp_path, {"recorded.py": "result = missing_value\n"})))
assert "F821" in str(out)
assert len(out.written) == 1 and out.written[0].path.endswith("recorded.py")
def test_exec_modified_python_labels_unknown_baseline_and_ignores_untouched_files(tmp_path):
tool = _diagnostic_tool(tmp_path)
home = tmp_path / "home"
(home / "untouched.py").write_text("old = untouched_missing\n")
(home / "changed.py").write_text("value = 1\n")
out = str(_run(tool.execute(command=_write_command(tmp_path, {"changed.py": "value = newly_missing\n"}))))
assert "newly_missing" in out and "baseline is unknown" in out
assert "untouched_missing" not in out
assert "Python diagnostics" not in str(
_run(
tool.execute(
command=_python_command(tmp_path, "from pathlib import Path; print(Path('changed.py').read_text())")
)
)
)
def test_exec_syntax_feedback_keeps_command_failure_and_does_not_rollback(tmp_path):
tool = _diagnostic_tool(tmp_path)
out = _run(tool.execute(command=_write_command(tmp_path, {"broken.py": "def f(:\n"}, exit_code=7)))
assert "Exit code: 7" in str(out) and "syntax error" in str(out)
assert getattr(out, "ok", None) is False
assert (tmp_path / "home" / "broken.py").read_text() == "def f(:\n"
def test_exec_diagnostics_follow_explicit_working_directory_and_have_a_switch(tmp_path):
home = tmp_path / "home"
home.mkdir()
elsewhere = home / "elsewhere"
elsewhere.mkdir()
tool = _diagnostic_tool(tmp_path)
out = str(
_run(tool.execute(command=_write_command(tmp_path, {"other.py": "x = missing\n"}), working_dir=str(elsewhere)))
)
assert os.path.join("elsewhere", "other.py") in out and "F821" in out
quiet = _diagnostic_tool(tmp_path, {"execPythonDiagnostics": False})
out = str(_run(quiet.execute(command=_write_command(tmp_path, {"quiet.py": "def f(:\n"}))))
assert "Python diagnostics" not in out
def test_exec_notes_preserve_file_metadata_and_control_flags():
from code_flow.tools.exec import _prefixed
from raven.contracts.tool import Continuation, FileRemoval, FileWrite, ToolOutput, ToolResult
for cls in (ToolOutput, ToolResult):
original = cls(
"command failed",
ok=False,
retryable=False,
blocks_call=True,
continuation=Continuation.CONTINUE,
written=(FileWrite(path="a.py", created=True, size=1),),
removed=(FileRemoval(path="gone.py"),),
)
annotated = _prefixed(" diagnostic", original, append=True)
assert annotated.written == original.written and annotated.removed == original.removed
assert annotated.ok is False and annotated.retryable is False and annotated.blocks_call is True
def test_exec_diagnostics_do_not_read_files_withheld_by_the_host(tmp_path):
class Shadow:
async def stage_tree(self):
return None
async def trackable(self, paths):
return {path for path in paths if Path(path).name != "private.py"}
tool = _diagnostic_tool(tmp_path)
tool.measure_writes(lambda root: Shadow())
out = _run(
tool.execute(
command=_write_command(
tmp_path,
{"private.py": "private_value = private_missing\n", "public.py": "public_value = public_missing\n"},
)
)
)
assert "public_missing" in str(out)
assert "private_missing" not in str(out)
assert out.ok is True and len(out.written) == 2
def test_exec_without_a_write_record_reads_nothing(tmp_path, monkeypatch):
from code_flow.tools import diagnostics
from raven.contracts.tool import FileWrite, ToolOutput
tool = _diagnostic_tool(tmp_path)
target = tmp_path / "home" / "recorded.py"
reads, records = [], []
monkeypatch.setattr(diagnostics, "read_source", lambda path: reads.append(path) or b"")
async def host(*args, **kwargs):
# With no record, as when the host's listing gave up: the output alone.
return ToolOutput("Exit code: 0", ok=True, written=tuple(records))
monkeypatch.setattr(TrunkExecTool, "execute", host)
out = _run(tool.execute(command="test command"))
assert out.ok is True and str(out) == "Exit code: 0" and reads == []
# The control: the same tool reads the file once the host records it.
records.append(FileWrite(path=str(target), created=True, size=24))
_run(tool.execute(command="test command"))
assert reads == [target]
@pytest.mark.parametrize("kind", ["timeout", "exception"])
def test_exec_post_command_checker_failure_keeps_original_exit_and_output(tmp_path, monkeypatch, kind):
from code_flow.tools import exec as module
tool = _diagnostic_tool(tmp_path)
async def broken(*args):
if kind == "timeout":
await asyncio.Event().wait()
raise RuntimeError("diagnostic failed")
monkeypatch.setattr(tool, "_diagnose_writes", broken)
monkeypatch.setattr(module, "_DIAGNOSTIC_WAIT_SECONDS", 0.01)
out = _run(tool.execute(command=_python_command(tmp_path, "print('evidence'); raise SystemExit(9)")))
assert "evidence" in str(out) and "Exit code: 9" in str(out)
assert out.ok is False
def test_exec_refused_command_takes_no_diagnostics(tmp_path, monkeypatch):
tool = _diagnostic_tool(tmp_path)
calls = []
async def diagnose(*args):
calls.append(args)
return "unexpected"
monkeypatch.setattr(tool, "_diagnose_writes", diagnose)
monkeypatch.setattr(tool, "_guard_command", lambda *args: "refused for this test")
out = _run(tool.execute(command="echo refused"))
assert not calls and out.blocks_call
def test_exec_diagnostics_limit_changed_files_and_output(tmp_path):
tool = _diagnostic_tool(tmp_path)
files = {f"file{i}.py": "result = missing\n" for i in range(10)}
files.update({f"notes{i}.txt": "not python\n" for i in range(3)})
out = str(_run(tool.execute(command=_write_command(tmp_path, files))))
assert out.count("Python diagnostics for") == 8
assert "2 more were skipped" in out and "notes" not in out
assert len(out) < 6500
def test_exec_checks_four_files_at_a_time_in_stable_order_with_an_eight_file_cap(tmp_path, monkeypatch):
from code_flow.tools import diagnostics
tool = _diagnostic_tool(tmp_path)
barrier = threading.Barrier(4, timeout=2)
lock = threading.Lock()
active = peak = 0
checked = []
def note(self, path, text, **kwargs):
nonlocal active, peak
with lock:
active += 1
peak = max(peak, active)
checked.append(path.name)
try:
barrier.wait()
return f"checked:{path.name}"
finally:
with lock:
active -= 1
monkeypatch.setattr(diagnostics.PythonDiagnostics, "note", note)
out = str(_run(tool.execute(command=_write_command(tmp_path, {f"file{i}.py": "x = 1\n" for i in range(10)}))))
assert peak == 4 and len(checked) == 8
assert sorted(checked) == [f"file{i}.py" for i in range(8)]
positions = [out.index(f"checked:file{i}.py") for i in range(8)]
assert positions == sorted(positions)
assert "2 more were skipped" in out
@pytest.mark.parametrize("visible_count", [1, 10])
def test_exec_withheld_files_do_not_consume_the_diagnostic_file_limit(tmp_path, monkeypatch, visible_count):
from raven.contracts.tool import FileWrite, ToolOutput
batches = []
class Shadow:
async def trackable(self, paths):
batches.append(list(paths))
return {path for path in paths if Path(path).name.startswith("visible")}
tool = _diagnostic_tool(tmp_path)
tool.measure_writes(lambda root: Shadow())
paths = [tmp_path / "home" / f"private{i}.py" for i in range(8)] + [
tmp_path / "home" / f"visible{i}.py" for i in range(visible_count)
]
for path in paths:
path.write_text("value = missing_binding\n")
async def host(*args, **kwargs):
return ToolOutput(
"Exit code: 0", ok=True, written=tuple(FileWrite(path=str(p), created=True, size=24) for p in paths)
)
monkeypatch.setattr(TrunkExecTool, "execute", host)
out = str(_run(tool.execute(command="test command")))
assert "visible0.py" in out and "F821" in out
assert all(f"private{i}.py" not in out for i in range(8))
assert [len(batch) for batch in batches] == [8 + visible_count]
assert out.count("Python diagnostics for") == min(8, visible_count)
if visible_count > 8:
assert "2 more were skipped" in out and "visible8.py" not in out
else:
assert "were skipped" not in out
def test_exec_ignored_files_are_asked_about_in_wide_batches(tmp_path, monkeypatch):
from raven.contracts.tool import FileWrite, ToolOutput
queries = []
class Shadow:
async def trackable(self, paths):
queries.append(len(paths))
return {path for path in paths if Path(path).name == "visible.py"}
tool = _diagnostic_tool(tmp_path)
tool.measure_writes(lambda root: Shadow())
home = tmp_path / "home"
ignored = [home / "build" / f"mod{i}.py" for i in range(600)]
visible = home / "visible.py"
visible.write_text("shown = visible_missing\n")
async def host(*args, **kwargs):
# Ordered the way the host sorts its record: the ignored tree first.
records = [FileWrite(path=str(p), created=True, size=1) for p in ignored]
records.append(FileWrite(path=str(visible), created=True, size=24))
return ToolOutput("Exit code: 0", ok=True, written=tuple(records))
monkeypatch.setattr(TrunkExecTool, "execute", host)
out = str(_run(tool.execute(command="test command")))
assert queries == [256, 256, 89]
assert "visible_missing" in out and "mod0.py" not in out
def test_exec_visibility_queries_stop_once_the_file_limit_is_visible(tmp_path, monkeypatch):
from raven.contracts.tool import FileWrite, ToolOutput
queries = []
class Shadow:
async def trackable(self, paths):
queries.append(len(paths))
return set(paths)
tool = _diagnostic_tool(tmp_path)
tool.measure_writes(lambda root: Shadow())
paths = [tmp_path / "home" / f"file{i:03}.py" for i in range(300)]
for path in paths[:8]:
path.write_text("value = missing_binding\n")
async def host(*args, **kwargs):
return ToolOutput(
"Exit code: 0", ok=True, written=tuple(FileWrite(path=str(p), created=True, size=24) for p in paths)
)
monkeypatch.setattr(TrunkExecTool, "execute", host)
out = str(_run(tool.execute(command="test command")))
assert queries == [256]
assert out.count("Python diagnostics for") == 8
assert "248 more were skipped" in out
assert "44 additional changed files were not checked" in out
def test_exec_visibility_batches_yield_to_the_post_command_deadline(tmp_path, monkeypatch):
from code_flow.tools import exec as module
from raven.contracts.tool import FileWrite, ToolOutput
calls = []
class Shadow:
async def trackable(self, paths):
# Stands for the host resolving every path on the loop: no await.
calls.append(len(paths))
time.sleep(0.06)
return set()
tool = _diagnostic_tool(tmp_path)
tool.measure_writes(lambda root: Shadow())
home = tmp_path / "home"
records = tuple(FileWrite(path=str(home / f"mod{i}.py"), created=True, size=1) for i in range(3000))
async def host(*args, **kwargs):
return ToolOutput("Exit code: 0", ok=True, written=records)
monkeypatch.setattr(TrunkExecTool, "execute", host)
monkeypatch.setattr(module, "_DIAGNOSTIC_WAIT_SECONDS", 0.05)
out = _run(tool.execute(command="test command"))
assert str(out) == "Exit code: 0" and out.ok is True
# Twelve batches in all; the deadline stops them at a yield in between.
assert 1 <= len(calls) < 12
@pytest.mark.asyncio
async def test_exec_diagnostics_do_not_stat_walk_or_open_files_on_the_event_loop(tmp_path, monkeypatch):
"""The plugin's own selection, with the host's visibility hook faked out:
every stat, walk and open it needs happens on a diagnostic worker."""
import io
from raven.contracts.tool import FileWrite, ToolOutput
loop_thread = threading.get_ident()
on_loop, walked = [], []
active = False
def watch(module, name):
original = getattr(module, name)
def watched(*args, **kwargs):
if active and threading.get_ident() == loop_thread:
on_loop.append(name)
if active and name == "scandir":
walked.append(args)
return original(*args, **kwargs)
monkeypatch.setattr(module, name, watched)
for module, name in ((os, "stat"), (os, "lstat"), (os, "scandir"), (os.path, "realpath"), (io, "open")):
watch(module, name)
class Shadow:
async def trackable(self, paths):
return {path for path in paths if Path(path).name == "visible.py"}
tool = _diagnostic_tool(tmp_path)
tool.measure_writes(lambda root: Shadow())
home = tmp_path / "home"
visible = home / "visible.py"
visible.write_text("shown = visible_missing\n")
records = [FileWrite(path=str(home / "build" / f"mod{i}.py"), created=True, size=1) for i in range(3000)]
records.append(FileWrite(path=str(visible), created=True, size=24))
active = True
try:
note = await tool._diagnose_writes(home, ToolOutput("Exit code: 0", ok=True, written=tuple(records)))
finally:
active = False
assert "visible_missing" in note
assert on_loop == [] and walked == []
@pytest.mark.asyncio
async def test_exec_checks_a_changed_file_as_a_written_one_is_checked(tmp_path):
"""Syntax past the name checks' bound, a line for each check that did not
run, and bytes Python cannot decode reported where decoding stops."""
from code_flow.tools.diagnostics import MAX_SOURCE_BYTES, MAX_SYNTAX_BYTES
from raven.contracts.tool import FileWrite, ToolOutput
class Shadow:
async def trackable(self, paths):
return set(paths)
tool = _diagnostic_tool(tmp_path)
tool.measure_writes(lambda root: Shadow())
home = tmp_path / "home"
rows = MAX_SOURCE_BYTES // len("value = 1\n") + 1
files = {
"large_broken.py": ("value = 1\n" * rows + "def broken(:\n").encode(),
"large_clean.py": ("value = 1\n" * rows).encode(),
"huge.py": b"#" * (MAX_SYNTAX_BYTES + 1),
"legacy.py": "# coding: gbk\nlabel = '\u4e2d'\n".encode(),
}
for name, raw in files.items():
(home / name).write_bytes(raw)
records = tuple(FileWrite(path=str(home / name), created=True, size=len(raw)) for name, raw in files.items())
note = await tool._diagnose_writes(home, ToolOutput("Exit code: 0", ok=True, written=records))
assert f"WARNING: line {rows + 1}, column 12: Python syntax error" in note
assert f"Python name checks skipped for {home / 'large_clean.py'}" in note
assert f"Python diagnostics skipped for {home / 'huge.py'}: the source is over 2 MiB" in note
assert "line 2, column 12: Python syntax error: 'gbk' codec can't decode byte 0xad" in note
def test_exec_diagnostics_skip_symlinks_cache_and_non_python_files(tmp_path):
tool = _diagnostic_tool(tmp_path)
home = tmp_path / "home"
outside = tmp_path / "outside.py"
outside.write_text("value = external_missing\n")
try:
(tmp_path / "probe_link.py").symlink_to(outside)
except (OSError, NotImplementedError):
pytest.skip("This host cannot create the symlink fixture")
command = _python_command(
tmp_path,
"import os\n"
"from pathlib import Path\n"
f"os.symlink({str(outside)!r}, 'link.py')\n"
"Path('.venv').mkdir()\n"
"Path('.venv/cached.py').write_text('x = cached_missing\\n')\n"
"Path('text.txt').write_text('def f(:\\n')\n"
"Path('shown.py').write_text('y = shown_missing\\n')\n",
)
out = str(_run(tool.execute(command=command)))
assert (home / "link.py").is_symlink()
assert "shown_missing" in out
assert "external_missing" not in out and "cached_missing" not in out and "text.txt" not in out
@pytest.mark.parametrize("params", [{"machine": "elsewhere"}, {"run_in_background": True}])
def test_exec_nonlocal_or_background_commands_take_no_local_diagnostics(tmp_path, monkeypatch, params):
from raven.contracts.tool import ToolOutput
tool = _diagnostic_tool(tmp_path)
calls = []
async def diagnose(*args):
calls.append(args)
return "unexpected"
async def host(*args, **kwargs):
return ToolOutput("host response", ok=True)
monkeypatch.setattr(tool, "_diagnose_writes", diagnose)
monkeypatch.setattr(TrunkExecTool, "execute", host)
assert str(_run(tool.execute(command="test command", **params))) == "host response"
assert calls == []
def test_exec_missing_visibility_verdict_never_falls_back_to_reading_the_file(tmp_path, monkeypatch):
from raven.contracts.tool import FileWrite, ToolOutput
class Shadow:
async def trackable(self, paths):
raise RuntimeError("visibility unavailable")
tool = _diagnostic_tool(tmp_path)
target = tmp_path / "home" / "withheld.py"
target.write_text("value = confidential_name\n")
tool.measure_writes(lambda root: Shadow())
async def host(*args, **kwargs):
return ToolOutput("Exit code: 0", ok=True, written=(FileWrite(path=str(target), created=True, size=1),))
monkeypatch.setattr(TrunkExecTool, "execute", host)
out = _run(tool.execute(command="test command"))
assert out.ok is True and len(out.written) == 1
assert "confidential_name" not in str(out)
@pytest.mark.parametrize("hook", [None, lambda root: None])
def test_exec_without_checkpoint_never_reads_changed_file_contents(tmp_path, monkeypatch, hook):
from code_flow.tools import diagnostics
from raven.contracts.tool import FileWrite, ToolOutput
tool = _tool(tmp_path)
tool.measure_writes(hook)
target = tmp_path / "home" / "local_secrets.py"
target.write_text("token = confidential_name\n")
(target.parent / ".gitignore").write_text("local_secrets.py\n")
reads = []
def read(path):
reads.append(path)
return path.read_bytes()
async def host(*args, **kwargs):
return ToolOutput("Exit code: 0", ok=True, written=(FileWrite(path=str(target), created=True, size=26),))
monkeypatch.setattr(diagnostics, "read_source", read)
monkeypatch.setattr(TrunkExecTool, "execute", host)
out = _run(tool.execute(command="test command"))
assert reads == [] and str(out) == "Exit code: 0"
assert out.ok is True and len(out.written) == 1
feat(acp): a prompt is held open while the agent sleeps on an armed wake, and takes a steer ## Summary An on-call agent submits a job, arms a wake and ends its turn saying so; the wake runs a later turn on the session on its own clock. Over ACP that turn's `end_turn` read as "finished" to the caller: a DAG judging its node spent every continuation on "still watching, wake at 13:13" while the job ran fine (ar-r4, 2026-09-08), and the main agent re-dispatched the work by hand. ACP has no stop reason for "waiting". Replaces the closed !573: **no DAG code changes**, no new node status. The node is `running` because a turn is genuinely in flight. - **The signal.** The loop-hooks paper gains one key the ACP layer reads for itself: `acp_meta["raven.holdTurn"]` (`{"untilMs", "why"}`). oncall-flow files it at the send fire when the window's campaign is claimed, not concluded, and has one pending wake -- read the same way the owner's answer pulls that wake forward. No campaign, a wake that stood down, or a concluded campaign files nothing and the turn ends as before. - **The hold** (`raven/acp/methods.py`). A turn that ended `end_turn` with the key filed is not answered. One `tool_call` stands for the wait (its title names the campaign and the next look), with a `tool_call_update` heartbeat every `HOLD_HEARTBEAT_S` (60 s) so a client that clocks liveness by frames sees the agent alive while its model sleeps. The turn slot is reopened with no id to correlate on, so whichever turn ends next on the session -- the wake's, a steer's -- settles it, and the handler decides again from that turn's stash. A wake that arms again keeps the hold; a turn ending with nothing pending answers `end_turn`; a cancel answers `cancelled` as before. The key itself is not passed to the client. - **Steer.** `_raven/session/steer` is served and advertised under `agentCapabilities._meta` as `raven.steer` (the names raven's own client already uses). Accepted while a prompt is open on the session, held or running: the text is shown on the wire as the person's own line, then sent with `turn.send`'s new `busy: "inject"`. Answers `injected` / `no_turn`. - **Wire.** `turn.send` gains the optional `busy` param (`const "inject"`): an ordinary send on an idle lane, `BusyPolicy.INJECT` into the running turn otherwise -- the merge a channel's mid-turn message already gets (`gateway_commands.py`). No slot is bound for an injected message; the running turn keeps the lane's. `openrpc.json` and both generated clients moved together. What raven's client side already does, unchanged: the DAG stall watcher resets on any frame (the heartbeat keeps it quiet, and a stopped heartbeat raises the existing stall notice after ten minutes); a sub-agent process that dies fails the request and the node takes its ordinary path; the pool's `_responsive` treats a prompt in flight as busy. ## Review round 1 (2026-09-09, on !600, superseded by this MR) - **`busy: inject` reads the scheduler's lane state too.** `is_turn_active` knows only the turns `turn.send` started; an armed wake is submitted to the scheduler directly (`make_on_session_wake`) and never binds that map, so a steer during a running wake queued a second turn and answered `injected`. `turn.send` now asks `Scheduler.has_inflight(lane)` as well, and merges (`BusyPolicy.INJECT`) when either record says the lane is running. Pinned: `test_busy_inject_merges_into_a_wake_the_real_scheduler_is_running` -- the reviewer's reproduction on the real `Scheduler`, a turn submitted the way `make_on_session_wake` submits one; with the lane check bypassed it reads `pending=1, inject_mailbox=0`, with it `(1, 0)` in the mailbox and the running turn drains the text. - Branch renamed to snake_case and both commits carry the `Co-authored-by: Claude (<model-id>)` trailer, per AGENTS.md §2–3; that is why this MR replaces !600 rather than continuing it. ## Review round 2 (2026-09-10) - **A wake re-armed during its own run survives the writeback.** `CronService._writeback_locked` reloaded the store after the run and, for a one-shot with `delete_after_run`, overwrote the id with the finished run's state and removed it -- but the turn the wake ran had just called `schedule_wake` under the same key, so the id now named the *next* look. It vanished, and the held prompt had nothing left to settle it. A stored job under the fired id whose `at` schedule the run never had is treated as the new wake it is: claim cleared, run recorded, schedule kept, not deleted. Pinned on the real service: `test_a_wake_re_armed_during_its_own_run_survives_the_writeback` (reads `[]` without the fix, the reviewer's state) and the unchanged one-shot deletion beside it. Cron, proactive, sentinel and on-call suites: 968 passed. ## Review round 3 (2026-09-10) - **An undrained inject stays visible and cancellable.** `busy: inject` dropped the scheduler's handle; when the host turn ended past its last drain, the spine's fallback ran the text as a turn nothing owned -- `turn.cancel` a no-op, no `message.start`. The handle is now kept per lane under an id minted for the text; `turn.cancel` / `session.interrupt` reach it when the active slot is empty (mailbox, queued fallback, or running); and when the fallback turn starts, the RPC sink (`on_turn_start`, wired from bootstrap) binds the lane to it and opens it on the wire with the person's words. `turn.send` answers the inject's own id. Pinned through `build_rpc_spine` with the reviewer's shape (a loop that drains once, an inject after it): `test_an_undrained_inject_runs_as_a_turn_the_surface_can_see_and_cancel` and `test_a_pending_inject_can_be_cancelled_before_it_merges_or_falls_back`; both read the reviewer's state (`turn_ids` empty, cancel finding nothing) with the fix bypassed. - **Line ceiling.** The contract comment is five lines; `raven/contracts` is 2,519 against 2,520, one line of headroom on the current `main` (2,514). - Non-blocking items taken: `SchedulerDrainingError` on the inject path exits as `turn_failed`; the title is announced before the hold; the steer constants sit below the cancel constant's docstring; three edges pinned (a held prompt settled by a wake turn's uncorrelated ending through `send_frame`, the hold key never in the client's `_meta`, `raven.toolName` on the wait row). - Focused suite (reviewer's list plus the touched files): 948 passed on Python 3.12. ## Review round 4 (2026-09-10) - **Every pending inject is kept, per lane and per id.** One replaceable record per lane kept only the newest steer; two steers after the host's last drain fall back in turn, and the first started unowned while a cancel took out the second. The record is now `lane -> {inject id: (handle, text)}` in arrival order: each fallback promotes its own handle, and a cancel with no active turn reaches the oldest pending one -- the one the spine runs next. Pinned: `test_two_undrained_injects_each_get_their_own_turn_promoted_and_cancelled` (both promoted with their own ids and `message.start`, cancelled one after the other; runner saw all three turns). Focused suite: 949 passed. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification ``` uv run pytest tests/test_acp_methods.py tests/test_acp_protocol.py tests/test_rpc_turn_send.py \ tests/test_rpc_schema_match.py tests/test_acp_updates.py tests/test_acp_server.py -q # 766 passed uv run pytest tests/test_agents_oncall_flow_hooks.py tests/test_agents_oncall_flow_actions.py \ tests/test_agents_oncall_flow_budget.py tests/test_agents_oncall_flow_readings.py \ tests/test_agents_oncall_flow_escalation_tools.py -q # 99 passed uv run ruff check raven tests agents && uv run ruff format --check raven tests agents # clean uv run lint-imports # 9 kept, 0 broken cd ui-tui && npm run lint:rpc && npx tsc --noEmit; cd ui-web && npm run gen:check && npx tsc --noEmit # in sync, clean ``` New tests: `TestHeldTurn` (held across an armed wake, held again on a re-arm, heartbeat on the wait call, cancel while held, a non-`end_turn` stop is not held; every frame validated against the ACP schema), `TestSteer` (advertised, `no_turn` when idle, a held prompt takes the words as a turn with `busy: inject`, refusals), `busy: inject` on a busy and an idle lane in `test_rpc_turn_send.py`, and the plugin's hold stamping through the real `CronService` grant (filed with a pending wake; not filed without one, after the wake stands down, or once the campaign concluded). Not done: no end-to-end run of a real on-call wake through a held prompt yet -- unit tests only. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [ ] User-facing docs or screenshots are updated when needed (the hold is invisible to a person except as a "waiting for ..." tool row) ## Risk - [x] Backward compatibility considered: a turn that files no `raven.holdTurn` is answered exactly as before; `busy` is optional and absent from every existing caller; the steer method and capability are additive. - [x] Rollback path is clear: revert the two commits. A held prompt holds one `max_concurrent_subagents` slot (default 8) for the wait's duration on the caller's side; raised in config if it bites. ## Related Replaces !573 (closed). Follows !518 (stall notice) and !520 (unprompted-turn route; a wake that runs under a held prompt no longer takes it for DAG nodes).
2026-09-11 02:37:55 +00:00
# --- the schema and the clamp -------------------------------------------------------
def test_the_timeout_has_a_ceiling_but_no_schema_maximum(tmp_path):
"""Trunk refuses ``timeout > 600`` at validation; here the schema carries no
maximum, so the call runs and the clamp is reported instead."""
tool = _tool(tmp_path)
timeout = tool.parameters["properties"]["timeout"]
assert "maximum" not in timeout
assert "clamped" in timeout["description"] and "1200" in timeout["description"]
trunk = TrunkExecTool().parameters["properties"]["timeout"]
assert trunk["maximum"] == 600, "the trunk shape this replacement departs from"
assert tool.timeout_seconds == 1260.0
def test_a_timeout_above_the_ceiling_is_clamped_and_said_so(tmp_path):
tool = _tool(tmp_path)
out = _run(tool.execute(command="echo clamped-run", timeout=99_999))
text = str(out)
assert text.startswith("[note: timeout 99999s clamped to the 1200s ceiling")
assert "clamped-run" in text and "Exit code: 0" in text
assert getattr(out, "ok", None) is True, "the note does not change the verdict"
def test_a_timeout_within_the_ceiling_runs_untouched(tmp_path):
tool = _tool(tmp_path)
out = str(_run(tool.execute(command="echo fine", timeout=900)))
assert not out.startswith("[note") and "fine" in out
def test_the_registry_runs_a_long_timeout_instead_of_refusing_it(tmp_path):
"""Through the registry, which validates against the schema: trunk's tool
would answer a validation error, this one runs."""
reg = ToolRegistry()
reg.register(_tool(tmp_path))
out = str(_run(reg.execute("exec", {"command": "echo via-registry", "timeout": 5000})))
assert "via-registry" in out and "clamped to the 1200s ceiling" in out
# --- partial output on timeout -----------------------------------------------------------
def test_a_timed_out_command_keeps_the_output_it_produced(tmp_path):
tool = _tool(tmp_path)
out = _run(tool.execute(command="echo started; echo more >&2; sleep 30", timeout=1))
text = str(out)
assert "started" in text, "what the command printed before the kill is kept"
assert "STDERR:\nmore" in text
assert "Timed out after 1s" in text and TIMED_OUT_NOTE in text
assert "Exit code: -1" in text
assert getattr(out, "ok", None) is False
feat(docker): run raven in a docker container and make a first run configurable from the page ## Summary Two things arrive together because the second is what made the first usable. **Run raven from a container.** A Dockerfile builds the page, the Python environment and the engine into one image; `cd docker && docker compose up` serves it. nginx serves the built page and proxies the RPC socket to the engine over the loopback they share -- the engine binds 127.0.0.1 only, so a proxy in a second container would have no address to dial. Compose publishes on 127.0.0.1, and the entrypoint prints a one-time sign-in link on boot. **A first run is no longer a dead end.** The gateway used to refuse to start until a provider was configured, while being the thing that serves the page where a provider gets configured. It now starts, says what is missing, and picks up a key saved in Settings on the next turn without a restart. The onboarding overlay no longer takes the page either: New chat, the model chip and Send send the reader to Settings > Models instead, and an empty model list is marked on the control that fills it. Also carried: the provider catalogue registry (capabilities, modalities and context windows per model), the reworked Settings > Models pane, and four fixes found in review of that work -- a sibling catalogue answering 200 with a non-catalogue body no longer crashes the whole fetch; adding a tag no longer erases the label and description already stored for that model; `authenticated` no longer demands a bearer token from a local provider reached by address; and registry rows are keyed by the spelling every lookup produces, which recovers 593 rows that no reader could reach. ## Type - [ ] Fix - [x] Feature - [x] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification Commands run on the branch head, with their results: - `uv run pytest tests -q -k "provider or registry or catalog or model or rpc or gateway or docker or i18n"` -- 4389 passed, 46 skipped - `npx vitest run` in `ui-web` -- 1573 passed (101 files) - `npx vitest run` in `ui-tui` -- 2037 passed (142 files) - `uv run ruff check raven scripts tests` and `ruff format --check` -- clean - `npx tsc --noEmit` in `ui-web` and `npm run type-check` in `ui-tui` -- clean - `npm run lint:rpc` in `ui-tui` -- generated.ts in sync - `make check-source-language`, `make check-large-files` -- clean Each of the four review fixes was checked by reverting it and watching a named test go red, then restoring it. The container image itself has not been built by CI or by me -- no Docker daemon was available on the machine this was written on. It has been built and run by the author of the branch. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [x] User-facing docs or screenshots are updated when needed ## Risk Two deliberate security trade-offs in the container deployment, both documented in `docker/README.md`: - nginx rewrites the `Origin` header, because the engine accepts only an Origin naming its own loopback port and no browser reaching a published port can send one. This drops a defence-in-depth layer. The session cookie is SameSite=Strict, so an unauthenticated socket is still refused -- but SameSite ignores ports, so another page on localhost counts as same-site. - `RAVEN_AUTO_LOGIN=1` is the default: nginx attaches the session cookie to every page load, so anyone who can reach the published port is signed in. It is contained by Compose publishing on 127.0.0.1 only. Set it to 0 for any deployment reachable from elsewhere. These are two settings that must agree, and nothing warns when they do not. Behaviour changes outside the container: the gateway now starts without a configured provider rather than exiting 1, and the first-run onboarding overlay no longer opens by default (`?onboard=1` still forces it). Rollback is by revert; nothing here migrates state or changes an on-disk format. - [x] Security impact considered - [x] Backward compatibility considered - [x] Rollback path is clear for risky changes ## Related Issues N/A
2026-09-11 03:27:52 +00:00
def test_trunks_executor_preserves_output_on_timeout(tmp_path):
"""The shared executor preserves output produced before the timeout.
This is the retired half of the reason the fork replaced it: trunk used to
wait on ``communicate()`` and lose the buffered output when that raised,
which is why this file's own partial-output fix was written. Trunk drains
the pipes itself now, so that axis no longer tells the two apart.
What is still this fork's on the same axis is the explicit marker. Trunk
reports the kill through the stderr line and exit ``-1`` and nothing more,
while ``CodeExecResult`` also flags ``timed_out`` and renders
``TIMED_OUT_NOTE``, which tells the model its output is a fragment. The
last assertion is what keeps the two from quietly converging.
"""
feat(acp): a prompt is held open while the agent sleeps on an armed wake, and takes a steer ## Summary An on-call agent submits a job, arms a wake and ends its turn saying so; the wake runs a later turn on the session on its own clock. Over ACP that turn's `end_turn` read as "finished" to the caller: a DAG judging its node spent every continuation on "still watching, wake at 13:13" while the job ran fine (ar-r4, 2026-09-08), and the main agent re-dispatched the work by hand. ACP has no stop reason for "waiting". Replaces the closed !573: **no DAG code changes**, no new node status. The node is `running` because a turn is genuinely in flight. - **The signal.** The loop-hooks paper gains one key the ACP layer reads for itself: `acp_meta["raven.holdTurn"]` (`{"untilMs", "why"}`). oncall-flow files it at the send fire when the window's campaign is claimed, not concluded, and has one pending wake -- read the same way the owner's answer pulls that wake forward. No campaign, a wake that stood down, or a concluded campaign files nothing and the turn ends as before. - **The hold** (`raven/acp/methods.py`). A turn that ended `end_turn` with the key filed is not answered. One `tool_call` stands for the wait (its title names the campaign and the next look), with a `tool_call_update` heartbeat every `HOLD_HEARTBEAT_S` (60 s) so a client that clocks liveness by frames sees the agent alive while its model sleeps. The turn slot is reopened with no id to correlate on, so whichever turn ends next on the session -- the wake's, a steer's -- settles it, and the handler decides again from that turn's stash. A wake that arms again keeps the hold; a turn ending with nothing pending answers `end_turn`; a cancel answers `cancelled` as before. The key itself is not passed to the client. - **Steer.** `_raven/session/steer` is served and advertised under `agentCapabilities._meta` as `raven.steer` (the names raven's own client already uses). Accepted while a prompt is open on the session, held or running: the text is shown on the wire as the person's own line, then sent with `turn.send`'s new `busy: "inject"`. Answers `injected` / `no_turn`. - **Wire.** `turn.send` gains the optional `busy` param (`const "inject"`): an ordinary send on an idle lane, `BusyPolicy.INJECT` into the running turn otherwise -- the merge a channel's mid-turn message already gets (`gateway_commands.py`). No slot is bound for an injected message; the running turn keeps the lane's. `openrpc.json` and both generated clients moved together. What raven's client side already does, unchanged: the DAG stall watcher resets on any frame (the heartbeat keeps it quiet, and a stopped heartbeat raises the existing stall notice after ten minutes); a sub-agent process that dies fails the request and the node takes its ordinary path; the pool's `_responsive` treats a prompt in flight as busy. ## Review round 1 (2026-09-09, on !600, superseded by this MR) - **`busy: inject` reads the scheduler's lane state too.** `is_turn_active` knows only the turns `turn.send` started; an armed wake is submitted to the scheduler directly (`make_on_session_wake`) and never binds that map, so a steer during a running wake queued a second turn and answered `injected`. `turn.send` now asks `Scheduler.has_inflight(lane)` as well, and merges (`BusyPolicy.INJECT`) when either record says the lane is running. Pinned: `test_busy_inject_merges_into_a_wake_the_real_scheduler_is_running` -- the reviewer's reproduction on the real `Scheduler`, a turn submitted the way `make_on_session_wake` submits one; with the lane check bypassed it reads `pending=1, inject_mailbox=0`, with it `(1, 0)` in the mailbox and the running turn drains the text. - Branch renamed to snake_case and both commits carry the `Co-authored-by: Claude (<model-id>)` trailer, per AGENTS.md §2–3; that is why this MR replaces !600 rather than continuing it. ## Review round 2 (2026-09-10) - **A wake re-armed during its own run survives the writeback.** `CronService._writeback_locked` reloaded the store after the run and, for a one-shot with `delete_after_run`, overwrote the id with the finished run's state and removed it -- but the turn the wake ran had just called `schedule_wake` under the same key, so the id now named the *next* look. It vanished, and the held prompt had nothing left to settle it. A stored job under the fired id whose `at` schedule the run never had is treated as the new wake it is: claim cleared, run recorded, schedule kept, not deleted. Pinned on the real service: `test_a_wake_re_armed_during_its_own_run_survives_the_writeback` (reads `[]` without the fix, the reviewer's state) and the unchanged one-shot deletion beside it. Cron, proactive, sentinel and on-call suites: 968 passed. ## Review round 3 (2026-09-10) - **An undrained inject stays visible and cancellable.** `busy: inject` dropped the scheduler's handle; when the host turn ended past its last drain, the spine's fallback ran the text as a turn nothing owned -- `turn.cancel` a no-op, no `message.start`. The handle is now kept per lane under an id minted for the text; `turn.cancel` / `session.interrupt` reach it when the active slot is empty (mailbox, queued fallback, or running); and when the fallback turn starts, the RPC sink (`on_turn_start`, wired from bootstrap) binds the lane to it and opens it on the wire with the person's words. `turn.send` answers the inject's own id. Pinned through `build_rpc_spine` with the reviewer's shape (a loop that drains once, an inject after it): `test_an_undrained_inject_runs_as_a_turn_the_surface_can_see_and_cancel` and `test_a_pending_inject_can_be_cancelled_before_it_merges_or_falls_back`; both read the reviewer's state (`turn_ids` empty, cancel finding nothing) with the fix bypassed. - **Line ceiling.** The contract comment is five lines; `raven/contracts` is 2,519 against 2,520, one line of headroom on the current `main` (2,514). - Non-blocking items taken: `SchedulerDrainingError` on the inject path exits as `turn_failed`; the title is announced before the hold; the steer constants sit below the cancel constant's docstring; three edges pinned (a held prompt settled by a wake turn's uncorrelated ending through `send_frame`, the hold key never in the client's `_meta`, `raven.toolName` on the wait row). - Focused suite (reviewer's list plus the touched files): 948 passed on Python 3.12. ## Review round 4 (2026-09-10) - **Every pending inject is kept, per lane and per id.** One replaceable record per lane kept only the newest steer; two steers after the host's last drain fall back in turn, and the first started unowned while a cancel took out the second. The record is now `lane -> {inject id: (handle, text)}` in arrival order: each fallback promotes its own handle, and a cancel with no active turn reaches the oldest pending one -- the one the spine runs next. Pinned: `test_two_undrained_injects_each_get_their_own_turn_promoted_and_cancelled` (both promoted with their own ids and `message.start`, cancelled one after the other; runner saw all three turns). Focused suite: 949 passed. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification ``` uv run pytest tests/test_acp_methods.py tests/test_acp_protocol.py tests/test_rpc_turn_send.py \ tests/test_rpc_schema_match.py tests/test_acp_updates.py tests/test_acp_server.py -q # 766 passed uv run pytest tests/test_agents_oncall_flow_hooks.py tests/test_agents_oncall_flow_actions.py \ tests/test_agents_oncall_flow_budget.py tests/test_agents_oncall_flow_readings.py \ tests/test_agents_oncall_flow_escalation_tools.py -q # 99 passed uv run ruff check raven tests agents && uv run ruff format --check raven tests agents # clean uv run lint-imports # 9 kept, 0 broken cd ui-tui && npm run lint:rpc && npx tsc --noEmit; cd ui-web && npm run gen:check && npx tsc --noEmit # in sync, clean ``` New tests: `TestHeldTurn` (held across an armed wake, held again on a re-arm, heartbeat on the wait call, cancel while held, a non-`end_turn` stop is not held; every frame validated against the ACP schema), `TestSteer` (advertised, `no_turn` when idle, a held prompt takes the words as a turn with `busy: inject`, refusals), `busy: inject` on a busy and an idle lane in `test_rpc_turn_send.py`, and the plugin's hold stamping through the real `CronService` grant (filed with a pending wake; not filed without one, after the wake stands down, or once the campaign concluded). Not done: no end-to-end run of a real on-call wake through a held prompt yet -- unit tests only. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [ ] User-facing docs or screenshots are updated when needed (the hold is invisible to a person except as a "waiting for ..." tool row) ## Risk - [x] Backward compatibility considered: a turn that files no `raven.holdTurn` is answered exactly as before; `busy` is optional and absent from every existing caller; the steer method and capability are additive. - [x] Rollback path is clear: revert the two commits. A held prompt holds one `max_concurrent_subagents` slot (default 8) for the wait's duration on the caller's side; raised in config if it bites. ## Related Replaces !573 (closed). Follows !518 (stall notice) and !520 (unprompted-turn route; a wake that runs under a held prompt no longer takes it for DAG nodes).
2026-09-11 02:37:55 +00:00
from raven.sandbox.direct_executor import DirectExecutor
feat(docker): run raven in a docker container and make a first run configurable from the page ## Summary Two things arrive together because the second is what made the first usable. **Run raven from a container.** A Dockerfile builds the page, the Python environment and the engine into one image; `cd docker && docker compose up` serves it. nginx serves the built page and proxies the RPC socket to the engine over the loopback they share -- the engine binds 127.0.0.1 only, so a proxy in a second container would have no address to dial. Compose publishes on 127.0.0.1, and the entrypoint prints a one-time sign-in link on boot. **A first run is no longer a dead end.** The gateway used to refuse to start until a provider was configured, while being the thing that serves the page where a provider gets configured. It now starts, says what is missing, and picks up a key saved in Settings on the next turn without a restart. The onboarding overlay no longer takes the page either: New chat, the model chip and Send send the reader to Settings > Models instead, and an empty model list is marked on the control that fills it. Also carried: the provider catalogue registry (capabilities, modalities and context windows per model), the reworked Settings > Models pane, and four fixes found in review of that work -- a sibling catalogue answering 200 with a non-catalogue body no longer crashes the whole fetch; adding a tag no longer erases the label and description already stored for that model; `authenticated` no longer demands a bearer token from a local provider reached by address; and registry rows are keyed by the spelling every lookup produces, which recovers 593 rows that no reader could reach. ## Type - [ ] Fix - [x] Feature - [x] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification Commands run on the branch head, with their results: - `uv run pytest tests -q -k "provider or registry or catalog or model or rpc or gateway or docker or i18n"` -- 4389 passed, 46 skipped - `npx vitest run` in `ui-web` -- 1573 passed (101 files) - `npx vitest run` in `ui-tui` -- 2037 passed (142 files) - `uv run ruff check raven scripts tests` and `ruff format --check` -- clean - `npx tsc --noEmit` in `ui-web` and `npm run type-check` in `ui-tui` -- clean - `npm run lint:rpc` in `ui-tui` -- generated.ts in sync - `make check-source-language`, `make check-large-files` -- clean Each of the four review fixes was checked by reverting it and watching a named test go red, then restoring it. The container image itself has not been built by CI or by me -- no Docker daemon was available on the machine this was written on. It has been built and run by the author of the branch. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [x] User-facing docs or screenshots are updated when needed ## Risk Two deliberate security trade-offs in the container deployment, both documented in `docker/README.md`: - nginx rewrites the `Origin` header, because the engine accepts only an Origin naming its own loopback port and no browser reaching a published port can send one. This drops a defence-in-depth layer. The session cookie is SameSite=Strict, so an unauthenticated socket is still refused -- but SameSite ignores ports, so another page on localhost counts as same-site. - `RAVEN_AUTO_LOGIN=1` is the default: nginx attaches the session cookie to every page load, so anyone who can reach the published port is signed in. It is contained by Compose publishing on 127.0.0.1 only. Set it to 0 for any deployment reachable from elsewhere. These are two settings that must agree, and nothing warns when they do not. Behaviour changes outside the container: the gateway now starts without a configured provider rather than exiting 1, and the first-run onboarding overlay no longer opens by default (`?onboard=1` still forces it). Rollback is by revert; nothing here migrates state or changes an on-disk format. - [x] Security impact considered - [x] Backward compatibility considered - [x] Rollback path is clear for risky changes ## Related Issues N/A
2026-09-11 03:27:52 +00:00
command = "echo started; sleep 30"
ci(*): make the idle ceiling a gate and take the last waits out (#428) ## Summary Follow-up to #413, from a measured look at what was left: a CI shard now spends 72 s on fixed cost (22 s environment, 48 s collection and worker start) against 82 s of tests, the four shards are within 15 s of each other, and 70 unmarked tests were still over the 2 s wall-clock ceiling on CI against 42 on a laptop because CPU-bound work runs three to four times slower there. This PR makes the ceiling a gate that is portable across machines, takes the remaining waits out, precompiles bytecode so the workers stop doing it four times, and fixes two product waits the same investigation surfaced. - `test(*)`: the ceiling judges idle time instead of wall clock, where idle is the wall clock minus every CPU second spent on the test's behalf, the reaped children's included (`os.times`). A test waiting on a production sleep or a process that answers nothing waits the same seconds on every machine; a test that computes for 2 s here and 7 s on CI, or drives a browser or eight spawned workers, is not the problem. The numbers ride on the teardown report so the xdist controller can judge them; options are `--idle-ceiling` / `--idle-ceiling-strict`. The fixture that shortens production ladders takes in the ACP cancel settle (5 s) and the gateway delivery grace (2 s); the CLI agent tests stub the one-shot command's unconditional 2 s hand-off grace; the rpc subagent tests no longer spawn a login shell or probe a real endpoint over the network; three literal timeouts shrink where the stub never answers anyway. What is left over the ceiling carries a marker: seven real LibreOffice renders and two tests that wait on a process answering nothing carry `slow`, two whose assertion is the wait itself carry `production_timing`. - `ci(*)`: every shard runs with `--idle-ceiling-strict` (on the make line, not in `addopts`: the Windows job runs `--noconftest`). Bytecode precompilation was tried and dropped, the numbers are below; the workflow is otherwise unchanged from #413. - `test(agent)`: a heartbeat test asked for five beats of the fifteen its intervals should produce in a 0.6 s window, and a loaded runner delivered three. It reads the two gaps instead, which the load cannot invert: the wait before the first beat is one interval, the wait after it is the other. Older than the gate that surfaced it; nothing here touches the registry. - `perf(ppt-engine)`: `MeasuredWidth.width` remembers measured widths (bounded, oldest key out); a 15-slide template's QA gates made 1848 freetype calls over 162 distinct keys. Importing `text_metrics` first raised on a circular import through `template.capacity`; capacity now imports the module and resolves the name at call time. - `perf(*)`: `raven gateway` and `raven web` start the litellm import on a daemon thread at boot, so saving a key from the settings page no longer waits three to five seconds on it. - `fix(providers)`: one thread at a time initialises litellm. It raises three logger levels for the duration of the import and puts the captured values back after, which holds only while one thread is doing it; a second caller arriving inside that window captures WARNING as the level to restore, and litellm's DEBUG and INFO stop reaching the file sink until a restart. The window was narrow while every import sat on a startup path and is ordinary now that a server warms it in the background, so the two callers share a module lock. The strict-mode warm-up also runs before the fixtures that redirect the home, and publishing the OAuth token directories created them, so it warms under a temporary home. ## Type - [ ] Fix - [ ] Feature - [ ] Docs - [x] CI / tooling - [ ] Refactor - [ ] Other ## Verification Whole suite, `uv run --frozen --all-extras pytest -q` (macOS, 4 workers), before and after the waits came out and with the ceiling strict: | run | result | idle list | |---|---|---| | main after #413 | 22052 passed, 43 skipped in 219.41 s | 39 over 2 s wall clock | | this branch, `--idle-ceiling` warning | 22057 passed, 43 skipped in 210.20 s | 14 over 2 s idle, all real render or subprocess, now marked | | this branch, `--idle-ceiling-strict` | 22058 passed, 43 skipped in 204.89 s, exit 0 | none | The ceiling is three seconds, placed from the suite's own distribution rather than guessed. Unmarked tests over one second of idle, measured across the whole suite: | idle | tests | what they are | |---|---|---| | 1.00-1.19 s | 19 | a test that sets a one second timeout and lets it expire | | 1.2-1.3 s | 3 | a subprocess handshake | | 1.5-1.6 s | 3 | an ACP turn interrupted after 0.8 s, and two like it | | above 1.6 s | 0 | the renders above this line all carry a marker | A runner stretches that 1.6 s to about 2.2 s, which is how a two second line failed two shards over tests waiting exactly as long as they meant to. Three seconds sits in the gap the population leaves, and stays under every ladder this guards against: the shortest constant the audit found was a three second grace, the rest run 5, 15, 30 and 60 seconds. What the metric separates, measured on this suite: a test that sleeps 5 s reports `4.96s idle of 5.02s`; a test that drives a headless browser for 13 s reports under 0.8 s; six `raven doctor` subprocesses report 1.2 s where wall clock said 4.2 s; eight spawned processes contending for a lock report under 0.8 s where wall clock said 3.6 s. Files touched by the product changes and the new markers (`uv run --frozen --all-extras pytest -q tests/test_litellm_setup.py tests/test_cli_gateway_commands.py tests/test_cli_serve_commands.py tests/test_ppt_engine_assets_fonts.py tests/test_ppt_engine_template_capacity.py tests/test_cli_doctor_commands.py tests/test_ppt_engine_end_to_end.py tests/test_ppt_engine_render_service.py tests/test_rpc_model.py`): `434 passed, 2 skipped in 33.39s`, none over 1 s idle. `python -c "import raven_ppt.services.assets.text_metrics"` in a fresh interpreter: ImportError before, imports after. Test-side speedups, measured per test: the ten CLI agent tests 2.0-4.4 s to 0.05-0.36 s; the three rpc subagent tests 1.4-2.6 s to 0.01-0.02 s; the rpc save-key test 2.2-2.8 s to 0.03 s; the two ACP probe tests 2.1 s to 0.2 s. Bytecode precompilation, seconds from the start of the pytest step to the first test, per shard, over three runs of this workflow: | what was compiled | shard 1 | shard 2 | shard 3 | shard 4 | mean | install step | |---|---|---|---|---|---|---| | nothing (#413) | 52.7 | 46.5 | 45.9 | 46.0 | 47.8 | 3-4 s | | site-packages and checkout | 39.8 | 39.3 | 35.6 | 45.5 | 40.1 | 12-17 s | | checkout only | 55.5 | 50.1 | 44.4 | 53.8 | 51.0 | 3-4 s | Compiling site-packages is the half that moved the number and it cost more in the install step than it returned. Compiling the checkout alone is inside the noise of doing nothing: the four workers share one `.pyc` cache, so the first to import a module writes it for the rest, and most of what a laptop measures as cold compilation is pytest rewriting test files through a cache `compileall` cannot fill. Both are gone; the comment in the workflow records it. The reworked heartbeat assertion was checked against the bug it guards by making the registry keep its first interval for every beat: one beat arrives in the window and the assertion fails on it. The diff gate is reproduced locally against `origin/main` after a full `make coverage`: `Diff coverage: 100.00% (12/12 executable changed lines)`. The two warm-up calls in the serving boot coroutines carry a no-cover pragma naming the test that pins them, because those coroutines bind a port and serve until they are killed. The lock is checked against the race it closes: with it taken away, a test that holds the first caller inside the import and starts a second fails on `[30, 30, 30] == [20, 20, 20]`. A strict session under a clean HOME now leaves no `.raven` behind. `make lint-python` clean. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [ ] User-facing docs or screenshots are updated when needed (none) ## Risk - The ceiling is now a gate on CI. A new test that waits more than 2 s without `slow` or `production_timing` turns its shard red; the summary names the test and the marker to use. A local run stays warning-only. - A subprocess that does real work is charged to the test that reaped it, so only a real wait trips the gate. A strict session imports litellm up front, because its 1.7 s of file opens would otherwise land on whichever test in the shard imports it first. - The litellm warm-up runs on a daemon thread at server boot only; unit tests never reach either boot path (their source is pinned instead), and a failed import is logged, not raised. - The width memo returns the same floats; the cap evicts the oldest key. The capacity import change resolves the same name one step later. - Rollback: revert the PR; every change is additive or a marker. - [x] Security impact considered (no security surface touched) - [x] Backward compatibility considered - [x] Rollback path is clear for risky changes ## Related Issues N/A --------- Co-authored-by: Claude (claude-fable-5-1) <noreply@anthropic.com>
2026-09-15 23:51:50 -07:00
result = _run(DirectExecutor().exec(command, timeout=0.3))
feat(acp): a prompt is held open while the agent sleeps on an armed wake, and takes a steer ## Summary An on-call agent submits a job, arms a wake and ends its turn saying so; the wake runs a later turn on the session on its own clock. Over ACP that turn's `end_turn` read as "finished" to the caller: a DAG judging its node spent every continuation on "still watching, wake at 13:13" while the job ran fine (ar-r4, 2026-09-08), and the main agent re-dispatched the work by hand. ACP has no stop reason for "waiting". Replaces the closed !573: **no DAG code changes**, no new node status. The node is `running` because a turn is genuinely in flight. - **The signal.** The loop-hooks paper gains one key the ACP layer reads for itself: `acp_meta["raven.holdTurn"]` (`{"untilMs", "why"}`). oncall-flow files it at the send fire when the window's campaign is claimed, not concluded, and has one pending wake -- read the same way the owner's answer pulls that wake forward. No campaign, a wake that stood down, or a concluded campaign files nothing and the turn ends as before. - **The hold** (`raven/acp/methods.py`). A turn that ended `end_turn` with the key filed is not answered. One `tool_call` stands for the wait (its title names the campaign and the next look), with a `tool_call_update` heartbeat every `HOLD_HEARTBEAT_S` (60 s) so a client that clocks liveness by frames sees the agent alive while its model sleeps. The turn slot is reopened with no id to correlate on, so whichever turn ends next on the session -- the wake's, a steer's -- settles it, and the handler decides again from that turn's stash. A wake that arms again keeps the hold; a turn ending with nothing pending answers `end_turn`; a cancel answers `cancelled` as before. The key itself is not passed to the client. - **Steer.** `_raven/session/steer` is served and advertised under `agentCapabilities._meta` as `raven.steer` (the names raven's own client already uses). Accepted while a prompt is open on the session, held or running: the text is shown on the wire as the person's own line, then sent with `turn.send`'s new `busy: "inject"`. Answers `injected` / `no_turn`. - **Wire.** `turn.send` gains the optional `busy` param (`const "inject"`): an ordinary send on an idle lane, `BusyPolicy.INJECT` into the running turn otherwise -- the merge a channel's mid-turn message already gets (`gateway_commands.py`). No slot is bound for an injected message; the running turn keeps the lane's. `openrpc.json` and both generated clients moved together. What raven's client side already does, unchanged: the DAG stall watcher resets on any frame (the heartbeat keeps it quiet, and a stopped heartbeat raises the existing stall notice after ten minutes); a sub-agent process that dies fails the request and the node takes its ordinary path; the pool's `_responsive` treats a prompt in flight as busy. ## Review round 1 (2026-09-09, on !600, superseded by this MR) - **`busy: inject` reads the scheduler's lane state too.** `is_turn_active` knows only the turns `turn.send` started; an armed wake is submitted to the scheduler directly (`make_on_session_wake`) and never binds that map, so a steer during a running wake queued a second turn and answered `injected`. `turn.send` now asks `Scheduler.has_inflight(lane)` as well, and merges (`BusyPolicy.INJECT`) when either record says the lane is running. Pinned: `test_busy_inject_merges_into_a_wake_the_real_scheduler_is_running` -- the reviewer's reproduction on the real `Scheduler`, a turn submitted the way `make_on_session_wake` submits one; with the lane check bypassed it reads `pending=1, inject_mailbox=0`, with it `(1, 0)` in the mailbox and the running turn drains the text. - Branch renamed to snake_case and both commits carry the `Co-authored-by: Claude (<model-id>)` trailer, per AGENTS.md §2–3; that is why this MR replaces !600 rather than continuing it. ## Review round 2 (2026-09-10) - **A wake re-armed during its own run survives the writeback.** `CronService._writeback_locked` reloaded the store after the run and, for a one-shot with `delete_after_run`, overwrote the id with the finished run's state and removed it -- but the turn the wake ran had just called `schedule_wake` under the same key, so the id now named the *next* look. It vanished, and the held prompt had nothing left to settle it. A stored job under the fired id whose `at` schedule the run never had is treated as the new wake it is: claim cleared, run recorded, schedule kept, not deleted. Pinned on the real service: `test_a_wake_re_armed_during_its_own_run_survives_the_writeback` (reads `[]` without the fix, the reviewer's state) and the unchanged one-shot deletion beside it. Cron, proactive, sentinel and on-call suites: 968 passed. ## Review round 3 (2026-09-10) - **An undrained inject stays visible and cancellable.** `busy: inject` dropped the scheduler's handle; when the host turn ended past its last drain, the spine's fallback ran the text as a turn nothing owned -- `turn.cancel` a no-op, no `message.start`. The handle is now kept per lane under an id minted for the text; `turn.cancel` / `session.interrupt` reach it when the active slot is empty (mailbox, queued fallback, or running); and when the fallback turn starts, the RPC sink (`on_turn_start`, wired from bootstrap) binds the lane to it and opens it on the wire with the person's words. `turn.send` answers the inject's own id. Pinned through `build_rpc_spine` with the reviewer's shape (a loop that drains once, an inject after it): `test_an_undrained_inject_runs_as_a_turn_the_surface_can_see_and_cancel` and `test_a_pending_inject_can_be_cancelled_before_it_merges_or_falls_back`; both read the reviewer's state (`turn_ids` empty, cancel finding nothing) with the fix bypassed. - **Line ceiling.** The contract comment is five lines; `raven/contracts` is 2,519 against 2,520, one line of headroom on the current `main` (2,514). - Non-blocking items taken: `SchedulerDrainingError` on the inject path exits as `turn_failed`; the title is announced before the hold; the steer constants sit below the cancel constant's docstring; three edges pinned (a held prompt settled by a wake turn's uncorrelated ending through `send_frame`, the hold key never in the client's `_meta`, `raven.toolName` on the wait row). - Focused suite (reviewer's list plus the touched files): 948 passed on Python 3.12. ## Review round 4 (2026-09-10) - **Every pending inject is kept, per lane and per id.** One replaceable record per lane kept only the newest steer; two steers after the host's last drain fall back in turn, and the first started unowned while a cancel took out the second. The record is now `lane -> {inject id: (handle, text)}` in arrival order: each fallback promotes its own handle, and a cancel with no active turn reaches the oldest pending one -- the one the spine runs next. Pinned: `test_two_undrained_injects_each_get_their_own_turn_promoted_and_cancelled` (both promoted with their own ids and `message.start`, cancelled one after the other; runner saw all three turns). Focused suite: 949 passed. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification ``` uv run pytest tests/test_acp_methods.py tests/test_acp_protocol.py tests/test_rpc_turn_send.py \ tests/test_rpc_schema_match.py tests/test_acp_updates.py tests/test_acp_server.py -q # 766 passed uv run pytest tests/test_agents_oncall_flow_hooks.py tests/test_agents_oncall_flow_actions.py \ tests/test_agents_oncall_flow_budget.py tests/test_agents_oncall_flow_readings.py \ tests/test_agents_oncall_flow_escalation_tools.py -q # 99 passed uv run ruff check raven tests agents && uv run ruff format --check raven tests agents # clean uv run lint-imports # 9 kept, 0 broken cd ui-tui && npm run lint:rpc && npx tsc --noEmit; cd ui-web && npm run gen:check && npx tsc --noEmit # in sync, clean ``` New tests: `TestHeldTurn` (held across an armed wake, held again on a re-arm, heartbeat on the wait call, cancel while held, a non-`end_turn` stop is not held; every frame validated against the ACP schema), `TestSteer` (advertised, `no_turn` when idle, a held prompt takes the words as a turn with `busy: inject`, refusals), `busy: inject` on a busy and an idle lane in `test_rpc_turn_send.py`, and the plugin's hold stamping through the real `CronService` grant (filed with a pending wake; not filed without one, after the wake stands down, or once the campaign concluded). Not done: no end-to-end run of a real on-call wake through a held prompt yet -- unit tests only. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [ ] User-facing docs or screenshots are updated when needed (the hold is invisible to a person except as a "waiting for ..." tool row) ## Risk - [x] Backward compatibility considered: a turn that files no `raven.holdTurn` is answered exactly as before; `busy` is optional and absent from every existing caller; the steer method and capability are additive. - [x] Rollback path is clear: revert the two commits. A held prompt holds one `max_concurrent_subagents` slot (default 8) for the wait's duration on the caller's side; raised in config if it bites. ## Related Replaces !573 (closed). Follows !518 (stall notice) and !520 (unprompted-turn route; a wake that runs under a held prompt no longer takes it for DAG nodes).
2026-09-11 02:37:55 +00:00
assert result.stdout == "started\n"
assert "Timed out" in result.stderr
assert result.exit_code == -1
feat(docker): run raven in a docker container and make a first run configurable from the page ## Summary Two things arrive together because the second is what made the first usable. **Run raven from a container.** A Dockerfile builds the page, the Python environment and the engine into one image; `cd docker && docker compose up` serves it. nginx serves the built page and proxies the RPC socket to the engine over the loopback they share -- the engine binds 127.0.0.1 only, so a proxy in a second container would have no address to dial. Compose publishes on 127.0.0.1, and the entrypoint prints a one-time sign-in link on boot. **A first run is no longer a dead end.** The gateway used to refuse to start until a provider was configured, while being the thing that serves the page where a provider gets configured. It now starts, says what is missing, and picks up a key saved in Settings on the next turn without a restart. The onboarding overlay no longer takes the page either: New chat, the model chip and Send send the reader to Settings > Models instead, and an empty model list is marked on the control that fills it. Also carried: the provider catalogue registry (capabilities, modalities and context windows per model), the reworked Settings > Models pane, and four fixes found in review of that work -- a sibling catalogue answering 200 with a non-catalogue body no longer crashes the whole fetch; adding a tag no longer erases the label and description already stored for that model; `authenticated` no longer demands a bearer token from a local provider reached by address; and registry rows are keyed by the spelling every lookup produces, which recovers 593 rows that no reader could reach. ## Type - [ ] Fix - [x] Feature - [x] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification Commands run on the branch head, with their results: - `uv run pytest tests -q -k "provider or registry or catalog or model or rpc or gateway or docker or i18n"` -- 4389 passed, 46 skipped - `npx vitest run` in `ui-web` -- 1573 passed (101 files) - `npx vitest run` in `ui-tui` -- 2037 passed (142 files) - `uv run ruff check raven scripts tests` and `ruff format --check` -- clean - `npx tsc --noEmit` in `ui-web` and `npm run type-check` in `ui-tui` -- clean - `npm run lint:rpc` in `ui-tui` -- generated.ts in sync - `make check-source-language`, `make check-large-files` -- clean Each of the four review fixes was checked by reverting it and watching a named test go red, then restoring it. The container image itself has not been built by CI or by me -- no Docker daemon was available on the machine this was written on. It has been built and run by the author of the branch. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [x] User-facing docs or screenshots are updated when needed ## Risk Two deliberate security trade-offs in the container deployment, both documented in `docker/README.md`: - nginx rewrites the `Origin` header, because the engine accepts only an Origin naming its own loopback port and no browser reaching a published port can send one. This drops a defence-in-depth layer. The session cookie is SameSite=Strict, so an unauthenticated socket is still refused -- but SameSite ignores ports, so another page on localhost counts as same-site. - `RAVEN_AUTO_LOGIN=1` is the default: nginx attaches the session cookie to every page load, so anyone who can reach the published port is signed in. It is contained by Compose publishing on 127.0.0.1 only. Set it to 0 for any deployment reachable from elsewhere. These are two settings that must agree, and nothing warns when they do not. Behaviour changes outside the container: the gateway now starts without a configured provider rather than exiting 1, and the first-run onboarding overlay no longer opens by default (`?onboard=1` still forces it). Rollback is by revert; nothing here migrates state or changes an on-disk format. - [x] Security impact considered - [x] Backward compatibility considered - [x] Rollback path is clear for risky changes ## Related Issues N/A
2026-09-11 03:27:52 +00:00
# Both sides measured in one run, because the absence on its own is the
# vacuous kind of assertion: it also holds for a fork that stopped marking
# anything. The pair says the difference exists and which way round it is.
ci(*): make the idle ceiling a gate and take the last waits out (#428) ## Summary Follow-up to #413, from a measured look at what was left: a CI shard now spends 72 s on fixed cost (22 s environment, 48 s collection and worker start) against 82 s of tests, the four shards are within 15 s of each other, and 70 unmarked tests were still over the 2 s wall-clock ceiling on CI against 42 on a laptop because CPU-bound work runs three to four times slower there. This PR makes the ceiling a gate that is portable across machines, takes the remaining waits out, precompiles bytecode so the workers stop doing it four times, and fixes two product waits the same investigation surfaced. - `test(*)`: the ceiling judges idle time instead of wall clock, where idle is the wall clock minus every CPU second spent on the test's behalf, the reaped children's included (`os.times`). A test waiting on a production sleep or a process that answers nothing waits the same seconds on every machine; a test that computes for 2 s here and 7 s on CI, or drives a browser or eight spawned workers, is not the problem. The numbers ride on the teardown report so the xdist controller can judge them; options are `--idle-ceiling` / `--idle-ceiling-strict`. The fixture that shortens production ladders takes in the ACP cancel settle (5 s) and the gateway delivery grace (2 s); the CLI agent tests stub the one-shot command's unconditional 2 s hand-off grace; the rpc subagent tests no longer spawn a login shell or probe a real endpoint over the network; three literal timeouts shrink where the stub never answers anyway. What is left over the ceiling carries a marker: seven real LibreOffice renders and two tests that wait on a process answering nothing carry `slow`, two whose assertion is the wait itself carry `production_timing`. - `ci(*)`: every shard runs with `--idle-ceiling-strict` (on the make line, not in `addopts`: the Windows job runs `--noconftest`). Bytecode precompilation was tried and dropped, the numbers are below; the workflow is otherwise unchanged from #413. - `test(agent)`: a heartbeat test asked for five beats of the fifteen its intervals should produce in a 0.6 s window, and a loaded runner delivered three. It reads the two gaps instead, which the load cannot invert: the wait before the first beat is one interval, the wait after it is the other. Older than the gate that surfaced it; nothing here touches the registry. - `perf(ppt-engine)`: `MeasuredWidth.width` remembers measured widths (bounded, oldest key out); a 15-slide template's QA gates made 1848 freetype calls over 162 distinct keys. Importing `text_metrics` first raised on a circular import through `template.capacity`; capacity now imports the module and resolves the name at call time. - `perf(*)`: `raven gateway` and `raven web` start the litellm import on a daemon thread at boot, so saving a key from the settings page no longer waits three to five seconds on it. - `fix(providers)`: one thread at a time initialises litellm. It raises three logger levels for the duration of the import and puts the captured values back after, which holds only while one thread is doing it; a second caller arriving inside that window captures WARNING as the level to restore, and litellm's DEBUG and INFO stop reaching the file sink until a restart. The window was narrow while every import sat on a startup path and is ordinary now that a server warms it in the background, so the two callers share a module lock. The strict-mode warm-up also runs before the fixtures that redirect the home, and publishing the OAuth token directories created them, so it warms under a temporary home. ## Type - [ ] Fix - [ ] Feature - [ ] Docs - [x] CI / tooling - [ ] Refactor - [ ] Other ## Verification Whole suite, `uv run --frozen --all-extras pytest -q` (macOS, 4 workers), before and after the waits came out and with the ceiling strict: | run | result | idle list | |---|---|---| | main after #413 | 22052 passed, 43 skipped in 219.41 s | 39 over 2 s wall clock | | this branch, `--idle-ceiling` warning | 22057 passed, 43 skipped in 210.20 s | 14 over 2 s idle, all real render or subprocess, now marked | | this branch, `--idle-ceiling-strict` | 22058 passed, 43 skipped in 204.89 s, exit 0 | none | The ceiling is three seconds, placed from the suite's own distribution rather than guessed. Unmarked tests over one second of idle, measured across the whole suite: | idle | tests | what they are | |---|---|---| | 1.00-1.19 s | 19 | a test that sets a one second timeout and lets it expire | | 1.2-1.3 s | 3 | a subprocess handshake | | 1.5-1.6 s | 3 | an ACP turn interrupted after 0.8 s, and two like it | | above 1.6 s | 0 | the renders above this line all carry a marker | A runner stretches that 1.6 s to about 2.2 s, which is how a two second line failed two shards over tests waiting exactly as long as they meant to. Three seconds sits in the gap the population leaves, and stays under every ladder this guards against: the shortest constant the audit found was a three second grace, the rest run 5, 15, 30 and 60 seconds. What the metric separates, measured on this suite: a test that sleeps 5 s reports `4.96s idle of 5.02s`; a test that drives a headless browser for 13 s reports under 0.8 s; six `raven doctor` subprocesses report 1.2 s where wall clock said 4.2 s; eight spawned processes contending for a lock report under 0.8 s where wall clock said 3.6 s. Files touched by the product changes and the new markers (`uv run --frozen --all-extras pytest -q tests/test_litellm_setup.py tests/test_cli_gateway_commands.py tests/test_cli_serve_commands.py tests/test_ppt_engine_assets_fonts.py tests/test_ppt_engine_template_capacity.py tests/test_cli_doctor_commands.py tests/test_ppt_engine_end_to_end.py tests/test_ppt_engine_render_service.py tests/test_rpc_model.py`): `434 passed, 2 skipped in 33.39s`, none over 1 s idle. `python -c "import raven_ppt.services.assets.text_metrics"` in a fresh interpreter: ImportError before, imports after. Test-side speedups, measured per test: the ten CLI agent tests 2.0-4.4 s to 0.05-0.36 s; the three rpc subagent tests 1.4-2.6 s to 0.01-0.02 s; the rpc save-key test 2.2-2.8 s to 0.03 s; the two ACP probe tests 2.1 s to 0.2 s. Bytecode precompilation, seconds from the start of the pytest step to the first test, per shard, over three runs of this workflow: | what was compiled | shard 1 | shard 2 | shard 3 | shard 4 | mean | install step | |---|---|---|---|---|---|---| | nothing (#413) | 52.7 | 46.5 | 45.9 | 46.0 | 47.8 | 3-4 s | | site-packages and checkout | 39.8 | 39.3 | 35.6 | 45.5 | 40.1 | 12-17 s | | checkout only | 55.5 | 50.1 | 44.4 | 53.8 | 51.0 | 3-4 s | Compiling site-packages is the half that moved the number and it cost more in the install step than it returned. Compiling the checkout alone is inside the noise of doing nothing: the four workers share one `.pyc` cache, so the first to import a module writes it for the rest, and most of what a laptop measures as cold compilation is pytest rewriting test files through a cache `compileall` cannot fill. Both are gone; the comment in the workflow records it. The reworked heartbeat assertion was checked against the bug it guards by making the registry keep its first interval for every beat: one beat arrives in the window and the assertion fails on it. The diff gate is reproduced locally against `origin/main` after a full `make coverage`: `Diff coverage: 100.00% (12/12 executable changed lines)`. The two warm-up calls in the serving boot coroutines carry a no-cover pragma naming the test that pins them, because those coroutines bind a port and serve until they are killed. The lock is checked against the race it closes: with it taken away, a test that holds the first caller inside the import and starts a second fails on `[30, 30, 30] == [20, 20, 20]`. A strict session under a clean HOME now leaves no `.raven` behind. `make lint-python` clean. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [ ] User-facing docs or screenshots are updated when needed (none) ## Risk - The ceiling is now a gate on CI. A new test that waits more than 2 s without `slow` or `production_timing` turns its shard red; the summary names the test and the marker to use. A local run stays warning-only. - A subprocess that does real work is charged to the test that reaped it, so only a real wait trips the gate. A strict session imports litellm up front, because its 1.7 s of file opens would otherwise land on whichever test in the shard imports it first. - The litellm warm-up runs on a daemon thread at server boot only; unit tests never reach either boot path (their source is pinned instead), and a failed import is logged, not raised. - The width memo returns the same floats; the cap evicts the oldest key. The capacity import change resolves the same name one step later. - Rollback: revert the PR; every change is additive or a marker. - [x] Security impact considered (no security surface touched) - [x] Backward compatibility considered - [x] Rollback path is clear for risky changes ## Related Issues N/A --------- Co-authored-by: Claude (claude-fable-5-1) <noreply@anthropic.com>
2026-09-15 23:51:50 -07:00
forked = str(_run(_tool(tmp_path).execute(command=command, timeout=0.3)))
feat(docker): run raven in a docker container and make a first run configurable from the page ## Summary Two things arrive together because the second is what made the first usable. **Run raven from a container.** A Dockerfile builds the page, the Python environment and the engine into one image; `cd docker && docker compose up` serves it. nginx serves the built page and proxies the RPC socket to the engine over the loopback they share -- the engine binds 127.0.0.1 only, so a proxy in a second container would have no address to dial. Compose publishes on 127.0.0.1, and the entrypoint prints a one-time sign-in link on boot. **A first run is no longer a dead end.** The gateway used to refuse to start until a provider was configured, while being the thing that serves the page where a provider gets configured. It now starts, says what is missing, and picks up a key saved in Settings on the next turn without a restart. The onboarding overlay no longer takes the page either: New chat, the model chip and Send send the reader to Settings > Models instead, and an empty model list is marked on the control that fills it. Also carried: the provider catalogue registry (capabilities, modalities and context windows per model), the reworked Settings > Models pane, and four fixes found in review of that work -- a sibling catalogue answering 200 with a non-catalogue body no longer crashes the whole fetch; adding a tag no longer erases the label and description already stored for that model; `authenticated` no longer demands a bearer token from a local provider reached by address; and registry rows are keyed by the spelling every lookup produces, which recovers 593 rows that no reader could reach. ## Type - [ ] Fix - [x] Feature - [x] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification Commands run on the branch head, with their results: - `uv run pytest tests -q -k "provider or registry or catalog or model or rpc or gateway or docker or i18n"` -- 4389 passed, 46 skipped - `npx vitest run` in `ui-web` -- 1573 passed (101 files) - `npx vitest run` in `ui-tui` -- 2037 passed (142 files) - `uv run ruff check raven scripts tests` and `ruff format --check` -- clean - `npx tsc --noEmit` in `ui-web` and `npm run type-check` in `ui-tui` -- clean - `npm run lint:rpc` in `ui-tui` -- generated.ts in sync - `make check-source-language`, `make check-large-files` -- clean Each of the four review fixes was checked by reverting it and watching a named test go red, then restoring it. The container image itself has not been built by CI or by me -- no Docker daemon was available on the machine this was written on. It has been built and run by the author of the branch. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [x] User-facing docs or screenshots are updated when needed ## Risk Two deliberate security trade-offs in the container deployment, both documented in `docker/README.md`: - nginx rewrites the `Origin` header, because the engine accepts only an Origin naming its own loopback port and no browser reaching a published port can send one. This drops a defence-in-depth layer. The session cookie is SameSite=Strict, so an unauthenticated socket is still refused -- but SameSite ignores ports, so another page on localhost counts as same-site. - `RAVEN_AUTO_LOGIN=1` is the default: nginx attaches the session cookie to every page load, so anyone who can reach the published port is signed in. It is contained by Compose publishing on 127.0.0.1 only. Set it to 0 for any deployment reachable from elsewhere. These are two settings that must agree, and nothing warns when they do not. Behaviour changes outside the container: the gateway now starts without a configured provider rather than exiting 1, and the first-run onboarding overlay no longer opens by default (`?onboard=1` still forces it). Rollback is by revert; nothing here migrates state or changes an on-disk format. - [x] Security impact considered - [x] Backward compatibility considered - [x] Rollback path is clear for risky changes ## Related Issues N/A
2026-09-11 03:27:52 +00:00
assert TIMED_OUT_NOTE in forked, "the fork tells the model its output is a fragment"
assert TIMED_OUT_NOTE not in result.as_text(), "trunk does not, and that is what is left"
feat(acp): a prompt is held open while the agent sleeps on an armed wake, and takes a steer ## Summary An on-call agent submits a job, arms a wake and ends its turn saying so; the wake runs a later turn on the session on its own clock. Over ACP that turn's `end_turn` read as "finished" to the caller: a DAG judging its node spent every continuation on "still watching, wake at 13:13" while the job ran fine (ar-r4, 2026-09-08), and the main agent re-dispatched the work by hand. ACP has no stop reason for "waiting". Replaces the closed !573: **no DAG code changes**, no new node status. The node is `running` because a turn is genuinely in flight. - **The signal.** The loop-hooks paper gains one key the ACP layer reads for itself: `acp_meta["raven.holdTurn"]` (`{"untilMs", "why"}`). oncall-flow files it at the send fire when the window's campaign is claimed, not concluded, and has one pending wake -- read the same way the owner's answer pulls that wake forward. No campaign, a wake that stood down, or a concluded campaign files nothing and the turn ends as before. - **The hold** (`raven/acp/methods.py`). A turn that ended `end_turn` with the key filed is not answered. One `tool_call` stands for the wait (its title names the campaign and the next look), with a `tool_call_update` heartbeat every `HOLD_HEARTBEAT_S` (60 s) so a client that clocks liveness by frames sees the agent alive while its model sleeps. The turn slot is reopened with no id to correlate on, so whichever turn ends next on the session -- the wake's, a steer's -- settles it, and the handler decides again from that turn's stash. A wake that arms again keeps the hold; a turn ending with nothing pending answers `end_turn`; a cancel answers `cancelled` as before. The key itself is not passed to the client. - **Steer.** `_raven/session/steer` is served and advertised under `agentCapabilities._meta` as `raven.steer` (the names raven's own client already uses). Accepted while a prompt is open on the session, held or running: the text is shown on the wire as the person's own line, then sent with `turn.send`'s new `busy: "inject"`. Answers `injected` / `no_turn`. - **Wire.** `turn.send` gains the optional `busy` param (`const "inject"`): an ordinary send on an idle lane, `BusyPolicy.INJECT` into the running turn otherwise -- the merge a channel's mid-turn message already gets (`gateway_commands.py`). No slot is bound for an injected message; the running turn keeps the lane's. `openrpc.json` and both generated clients moved together. What raven's client side already does, unchanged: the DAG stall watcher resets on any frame (the heartbeat keeps it quiet, and a stopped heartbeat raises the existing stall notice after ten minutes); a sub-agent process that dies fails the request and the node takes its ordinary path; the pool's `_responsive` treats a prompt in flight as busy. ## Review round 1 (2026-09-09, on !600, superseded by this MR) - **`busy: inject` reads the scheduler's lane state too.** `is_turn_active` knows only the turns `turn.send` started; an armed wake is submitted to the scheduler directly (`make_on_session_wake`) and never binds that map, so a steer during a running wake queued a second turn and answered `injected`. `turn.send` now asks `Scheduler.has_inflight(lane)` as well, and merges (`BusyPolicy.INJECT`) when either record says the lane is running. Pinned: `test_busy_inject_merges_into_a_wake_the_real_scheduler_is_running` -- the reviewer's reproduction on the real `Scheduler`, a turn submitted the way `make_on_session_wake` submits one; with the lane check bypassed it reads `pending=1, inject_mailbox=0`, with it `(1, 0)` in the mailbox and the running turn drains the text. - Branch renamed to snake_case and both commits carry the `Co-authored-by: Claude (<model-id>)` trailer, per AGENTS.md §2–3; that is why this MR replaces !600 rather than continuing it. ## Review round 2 (2026-09-10) - **A wake re-armed during its own run survives the writeback.** `CronService._writeback_locked` reloaded the store after the run and, for a one-shot with `delete_after_run`, overwrote the id with the finished run's state and removed it -- but the turn the wake ran had just called `schedule_wake` under the same key, so the id now named the *next* look. It vanished, and the held prompt had nothing left to settle it. A stored job under the fired id whose `at` schedule the run never had is treated as the new wake it is: claim cleared, run recorded, schedule kept, not deleted. Pinned on the real service: `test_a_wake_re_armed_during_its_own_run_survives_the_writeback` (reads `[]` without the fix, the reviewer's state) and the unchanged one-shot deletion beside it. Cron, proactive, sentinel and on-call suites: 968 passed. ## Review round 3 (2026-09-10) - **An undrained inject stays visible and cancellable.** `busy: inject` dropped the scheduler's handle; when the host turn ended past its last drain, the spine's fallback ran the text as a turn nothing owned -- `turn.cancel` a no-op, no `message.start`. The handle is now kept per lane under an id minted for the text; `turn.cancel` / `session.interrupt` reach it when the active slot is empty (mailbox, queued fallback, or running); and when the fallback turn starts, the RPC sink (`on_turn_start`, wired from bootstrap) binds the lane to it and opens it on the wire with the person's words. `turn.send` answers the inject's own id. Pinned through `build_rpc_spine` with the reviewer's shape (a loop that drains once, an inject after it): `test_an_undrained_inject_runs_as_a_turn_the_surface_can_see_and_cancel` and `test_a_pending_inject_can_be_cancelled_before_it_merges_or_falls_back`; both read the reviewer's state (`turn_ids` empty, cancel finding nothing) with the fix bypassed. - **Line ceiling.** The contract comment is five lines; `raven/contracts` is 2,519 against 2,520, one line of headroom on the current `main` (2,514). - Non-blocking items taken: `SchedulerDrainingError` on the inject path exits as `turn_failed`; the title is announced before the hold; the steer constants sit below the cancel constant's docstring; three edges pinned (a held prompt settled by a wake turn's uncorrelated ending through `send_frame`, the hold key never in the client's `_meta`, `raven.toolName` on the wait row). - Focused suite (reviewer's list plus the touched files): 948 passed on Python 3.12. ## Review round 4 (2026-09-10) - **Every pending inject is kept, per lane and per id.** One replaceable record per lane kept only the newest steer; two steers after the host's last drain fall back in turn, and the first started unowned while a cancel took out the second. The record is now `lane -> {inject id: (handle, text)}` in arrival order: each fallback promotes its own handle, and a cancel with no active turn reaches the oldest pending one -- the one the spine runs next. Pinned: `test_two_undrained_injects_each_get_their_own_turn_promoted_and_cancelled` (both promoted with their own ids and `message.start`, cancelled one after the other; runner saw all three turns). Focused suite: 949 passed. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification ``` uv run pytest tests/test_acp_methods.py tests/test_acp_protocol.py tests/test_rpc_turn_send.py \ tests/test_rpc_schema_match.py tests/test_acp_updates.py tests/test_acp_server.py -q # 766 passed uv run pytest tests/test_agents_oncall_flow_hooks.py tests/test_agents_oncall_flow_actions.py \ tests/test_agents_oncall_flow_budget.py tests/test_agents_oncall_flow_readings.py \ tests/test_agents_oncall_flow_escalation_tools.py -q # 99 passed uv run ruff check raven tests agents && uv run ruff format --check raven tests agents # clean uv run lint-imports # 9 kept, 0 broken cd ui-tui && npm run lint:rpc && npx tsc --noEmit; cd ui-web && npm run gen:check && npx tsc --noEmit # in sync, clean ``` New tests: `TestHeldTurn` (held across an armed wake, held again on a re-arm, heartbeat on the wait call, cancel while held, a non-`end_turn` stop is not held; every frame validated against the ACP schema), `TestSteer` (advertised, `no_turn` when idle, a held prompt takes the words as a turn with `busy: inject`, refusals), `busy: inject` on a busy and an idle lane in `test_rpc_turn_send.py`, and the plugin's hold stamping through the real `CronService` grant (filed with a pending wake; not filed without one, after the wake stands down, or once the campaign concluded). Not done: no end-to-end run of a real on-call wake through a held prompt yet -- unit tests only. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [ ] User-facing docs or screenshots are updated when needed (the hold is invisible to a person except as a "waiting for ..." tool row) ## Risk - [x] Backward compatibility considered: a turn that files no `raven.holdTurn` is answered exactly as before; `busy` is optional and absent from every existing caller; the steer method and capability are additive. - [x] Rollback path is clear: revert the two commits. A held prompt holds one `max_concurrent_subagents` slot (default 8) for the wait's duration on the caller's side; raised in config if it bites. ## Related Replaces !573 (closed). Follows !518 (stall notice) and !520 (unprompted-turn route; a wake that runs under a held prompt no longer takes it for DAG nodes).
2026-09-11 02:37:55 +00:00
def test_a_finished_command_reports_its_exit_code_and_both_streams(tmp_path):
tool = _tool(tmp_path)
text = str(_run(tool.execute(command="echo out; echo err >&2; exit 3")))
assert "out" in text and "STDERR:\nerr" in text and "Exit code: 3" in text
assert TIMED_OUT_NOTE not in text
# --- the output budget and the spill ---------------------------------------------------
def test_an_oversized_output_is_cut_at_the_budget_and_saved_whole(tmp_path):
tool = _tool(tmp_path)
text = str(_run(tool.execute(command="i=0; while [ $i -lt 4000 ]; do echo line-$i-0123456789; i=$((i+1)); done")))
assert len(text) < MAX_OUTPUT + 600
assert "chars truncated; full output" in text
assert "full output saved to " in text
saved = Path(text.split("full output saved to ", 1)[1].split(";", 1)[0])
assert saved.parent == tmp_path / "home" / SPILL_SUBDIR, "under Agent home, never in the repository"
body = saved.read_text()
assert body.count("line-") == 4000 and "line-3999-0123456789" in body
assert "line-0-0123456789" in text and "line-3999-0123456789" in text, "head and tail stay in view"
def test_the_budget_is_thirty_thousand_not_trunks_ten(tmp_path):
result = CodeExecResult(stdout="x" * 20_000, stderr="", exit_code=0, spill_dir=tmp_path / "spill")
assert "truncated" not in result.as_text(), "20k fits the new budget"
assert not (tmp_path / "spill").exists(), "nothing is saved while nothing was cut"
assert "truncated" in result.as_text(10_000), "and would not have fit trunk's"
assert (tmp_path / "spill").exists(), "cut at trunk's budget, the whole output is saved"
def test_a_spill_that_cannot_be_written_still_returns_the_head_and_tail(tmp_path):
blocked = tmp_path / "blocked"
blocked.write_text("a file where the directory should be")
result = CodeExecResult(stdout="y" * 40_000, stderr="", exit_code=0, spill_dir=blocked)
text = result.as_text()
assert "truncated" in text and "re-run redirecting to a file" in text
# --- the environment discipline --------------------------------------------------------
def _child_env_keys(text: str) -> set[str]:
body = text.split("\nExit code:", 1)[0]
return set(json.loads(body))
def test_the_child_sees_exactly_the_host_executors_baseline(tmp_path, monkeypatch):
"""No private import of the trunk allowlist (a cargo redline), yet no drift
from it either: the keys a child gets here equal the keys a child gets from
the host's own executor, and a credential in the process environment
reaches neither."""
from raven.sandbox.direct_executor import DirectExecutor
monkeypatch.setenv("CODE_FLOW_TEST_SECRET", "do-not-leak")
probe = f"{sys.executable} -c 'import json,os;print(json.dumps(sorted(os.environ)))'"
theirs = set(json.loads(_run(DirectExecutor().exec(probe, timeout=15)).stdout))
ours = _child_env_keys(str(_run(_tool(tmp_path).execute(command=probe))))
shell_added = {"PWD", "OLDPWD", "SHLVL", "_"}
assert ours - shell_added == theirs - shell_added
assert "CODE_FLOW_TEST_SECRET" not in ours and "CODE_FLOW_TEST_SECRET" not in theirs
assert "PATH" in ours
def test_the_baseline_is_learned_once_and_read_fresh(tmp_path, monkeypatch):
tool = _tool(tmp_path)
executor = tool._executor
first = _run(executor.allowed_keys())
assert first is _run(executor.allowed_keys()), "one probe per executor"
monkeypatch.setenv("PATH", os.environ["PATH"] + os.pathsep + str(tmp_path / "fresh-bin"))
text = str(_run(tool.execute(command="echo $PATH")))
assert str(tmp_path / "fresh-bin") in text, "values come from the process environment at spawn time"
def test_a_failed_probe_leaves_exec_erroring_rather_than_guessing(tmp_path, monkeypatch):
tool = _tool(tmp_path)
monkeypatch.setattr(sys, "executable", str(tmp_path / "no-such-python"))
out = str(_run(tool.execute(command="echo never")))
assert out.startswith("Error executing command:") and "environment baseline" in out
# --- the factory and the sandbox discipline ---------------------------------------------
def test_the_factory_serves_the_replacement_only_on_the_host_backend(tmp_path):
served = factories.make_exec(
_ctx(tmp_path, {"exec": {"sandboxBackend": "none", "maxTimeout": 900, "timeout": 120}})
)
assert isinstance(served, CodeExecTool)
assert served._MAX_TIMEOUT == 900 and served.timeout == 120
assert isinstance(served._executor, CodeExecutor) and served._executor.is_sandboxed is False
for backend in ("boxlite", "auto", "docker"):
assert factories.make_exec(_ctx(tmp_path, {"exec": {"sandboxBackend": backend}})) is None, backend
def test_the_factory_declines_with_the_rest_of_the_face(tmp_path):
off = PluginContext(
config={"enabled": True, "tools": {"enabled": False}},
services=ServiceLocator(workspace=tmp_path / "home", user_id="u", agent_id="a"),
)
assert factories.make_exec(off) is None
def test_the_same_name_replacement_answers_the_call(tmp_path):
reg = ToolRegistry()
reg.register(TrunkExecTool(working_dir=str(tmp_path)))
reg.register(factories.make_exec(_ctx(tmp_path)))
served = reg.get("exec")
assert type(served) is CodeExecTool
assert "clamped" in reg.get("exec").parameters["properties"]["timeout"]["description"]
def test_the_published_config_carries_the_ceiling_in_the_slice_not_in_a_dead_host_key():
"""Trunk's ``tools.exec`` knows ``timeout`` and ``pathAppend`` only; a
``maxTimeout`` there was silently ignored. It lives in this plugin's slice."""
published = json.loads((REPO / "agents" / "raven-code" / "config.json").read_text())
assert "maxTimeout" not in published["tools"]["exec"]
assert published["plugins"]["config"]["code-flow"]["tools"]["exec"] == {"maxTimeout": 1200}