## Summary
Reported as "deliver_files is broken". It is not: **the tool never ran.**
Every `deliver_files` call on one install was refused before dispatch -- five in a row in a single session -- and nothing was ever delivered:
```
Error: Invalid parameters for tool 'deliver_files': files should be array
```
What the model sent was the array, spelled as a string:
```json
{"files": "[{\"path\": \"...poster.png\", \"title\": \"...\"}]", "message": "..."}
```
`cast_params` already forgives that spelling for every scalar type -- `"3"` for an integer, `"true"` for a boolean -- and did not for arrays or objects, which are exactly the types a model is most likely to stringify.
## Type
- [x] Fix
- [ ] Feature
## The edge of the forgiving is the design
When the value is a string, try `json.loads`, and take the result **only if it is the type the schema asked for**. A string that is not JSON, or is JSON of the wrong shape, is returned unchanged so the validator still names the real problem -- the point is to widen what is accepted, not to swallow a type error.
**Nothing here can raise.** Turning "your argument has the wrong type" into a traceback would be worse than the refusal it replaces, so the parse is wrapped and failure falls through.
**A `type: "string"` field is never parsed.** JSON is ordinary text in a message body or a file's content, and replacing it would hand the tool something it never asked for. Two independent things prevent it today (the string branch returns before the parse is reached, and the parse only wins on a type match), which is why the case pinning it is a tripwire rather than a load-bearing assertion -- said so in its own docstring rather than left to look like coverage.
## Not just deliver_files
Six built-in tools declare an array or object parameter and were all exposed:
| tool | parameter |
|---|---|
| `deliver_files` | `files` |
| `ask_user` | `questions` |
| `image_generate` | `images` |
| `message` | `media` |
| `create_playbook` | `skills` |
| `video_generate` | `params` |
The same failure was already seen from the other side: `acp_client/acp_agent.py` records `update_task_state` and `operations should be array` from a real dispatch.
## Verification
```
make lint-python ruff check + ruff format --check, 1850 files
make lint-imports 9 contracts kept, 0 broken
pytest tests/ 19,699 passed, 7 failed (all pre-existing, see below)
```
- [x] Relevant tests pass locally
- [x] Relevant lint / type checks pass locally
- [ ] User-facing docs or screenshots -- no user-facing string
**End to end, through the real tool**, with the exact payload that failed five times in production:
```
files sent as: str -> [{"path": ".../poster.png", "title": "..."}]
returned: Delivered 1 file(s): poster.png (72 B).
still "should be array": False
registered in the store: True
```
Twelve cases, five mutations, each killing at least one:
| mutation | red |
|---|---|
| drop the whole branch (the old behaviour) | 3 |
| raise instead of falling through on a parse failure | 5 |
| accept the parse whatever type it yielded | 1 |
| do not re-cast the parsed value, so items are not cast | 1 |
| widen the tuple to include `string` | **survived -- unreachable** |
The last is worth stating rather than hiding: that mutation cannot change behaviour, because the `type == "string"` branch returns earlier and the type check would refuse the result anyway. I also tried moving the new block above that branch, alone and combined with widening the tuple, and neither reddens anything either. The invariant holds by construction; the case that documents it says so.
**Pre-existing failures, checked against the base.** 7 failed in the full run, all in `tests/test_security_network.py` (`test_a_redirect_to_a_dangerous_address_is_refused` and siblings). Running that file alone on this same tree gives 4 -- the count is order-dependent -- and the change touches only `raven/agent/tools/params.py` plus a new test file, neither of which `raven/security/network.py` imports. The failures are the box's resolver: it returns poisoned answers roughly half the time, which I measured separately (`en.wikipedia.org` came back as `2001::1` on 3 of 6 lookups). Not introduced here, and worth someone's attention on its own.
## Risk
- [x] Security impact considered -- widens what reaches a tool's own validation, not what bypasses it. A parsed value is still validated against the same schema, and a parse that yields the wrong type is refused exactly as before
- [x] Backward compatibility considered -- a call that worked before takes the identical path; the new branch is reached only by a `str` where an array or object is declared, which previously always failed
- [x] Rollback path is clear -- single-commit revert
## Related Issues
N/A
## Swept
`mr-review-patterns` pre-submit sweep, over `git diff origin/main...HEAD`.
- **1 (a write location moved)** -- nothing moved; a branch was added to an existing dispatch on `target_type`. Grepped every caller of `cast_params`: the registry is the only one, and it runs `validate_params` right after, which is the reader this affects.
- **2 (a value with several entry points)** -- the cast is the single funnel every tool argument passes through, which is why the fix belongs here and not in `deliver_files`.
- **3 (a guard)** -- checked both directions: what it wrongly lets through (JSON of the wrong type, non-JSON text -- both refused, with cases) and what it wrongly refuses (a real list, unaffected, with a case).
- **4 (tests)** -- five mutations above, one honestly reported as unreachable. Failing-ID sets diffed against the base rather than counts.
- **6 (stale text)** -- the module docstring enumerated what is coerced ("a string that should be an integer"); it now names arrays and objects too. Grepped the tracked tree for other statements of the rule: none.
- **9 (error handling)** -- the new path adds one way to end and it is caught: `json.loads` raises `ValueError` on bad text and `TypeError` on a non-string, and both fall through to the untouched value.
- **11** -- `git ls-files | grep -c node_modules` is 0; the commit-message gate was run, not eyeballed.
Not applicable, with reasons: **3b, 3c, 3d, 3e** (no tool schema produced, no state moved, one lane, no assertion lost its symbol); **4b, 4c, 4d** (no platform branch, no shell, no rendered number); **5, 5b** (no dependency; `json` is stdlib and already imported across this package); **6b, 6c** (no backup, no generated artifact); **7, 7b** (nothing on disk); **8** (no asset, no secret); **10** (no prompt text).
## Deliberately not fixed
The error text still ends with `[Analyze the error above and try a different approach.]`, and the model plainly did not -- it sent the same shape five times. A message naming the expected spelling (`files must be a JSON array, not a string containing one`) would give it something to act on. That is a change to the registry's error surface, affects every tool, and deserves its own argument.
## Moved from !594
Opened on a compliant source branch. AGENTS.md 2.1 asks for a 3-5 word snake_case description and the previous branch had seven; GitLab cannot change an existing merge request's source branch, so the change moves rather than the name.
Same commits, rebased onto `c2a9768f`. Both code findings from the review on !594 are already in them -- `RecursionError` caught with its regression case, and the implementation comment reduced to the two non-obvious reasons. Nothing else changed, so that review carries.
304 lines
10 KiB
Python
304 lines
10 KiB
Python
"""Per-trial cheap metadata extraction for downstream stratification.
|
|
|
|
Reads a legacy-runner trial dir and emits a :class:`ProxyFeatures` dataclass per
|
|
trial. The features are cheap to compute (parse one session.jsonl + one
|
|
result.json, no replay, no LLM, no container) and stable enough to feed
|
|
into the cold-start bandit's K-means sub-strata on the ``stable_fail`` 0/3
|
|
bucket — where ~70% of tasks land under v7 and where the bandit needs to
|
|
slice the population for cohort selection.
|
|
|
|
Feature set (all per-trial, all O(session.jsonl)):
|
|
|
|
- ``turn_count`` — assistant turns with at least one tool call
|
|
- ``final_exit_status`` — categorical (:class:`ExitStatus`)
|
|
- ``has_tool_calls_ever`` — bool: did the agent ever invoke a tool
|
|
- ``assistant_text_length_avg``— mean length (chars) of assistant.content across all assistant messages
|
|
- ``docker_error_count`` — count of docker-error patterns across tool responses + exception traceback
|
|
|
|
The docker-error count is a noise indicator (container-side issues that
|
|
surface as docker daemon errors during exec). It is intentionally loose
|
|
since per-trial occurrences should be rare; spikes indicate infra issues
|
|
worth filtering out before bandit stratification.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import re
|
|
from dataclasses import dataclass
|
|
from enum import Enum
|
|
from pathlib import Path
|
|
|
|
__all__ = [
|
|
"ExitStatus",
|
|
"ProxyFeatures",
|
|
"extract_features",
|
|
"extract_trial_dir",
|
|
]
|
|
|
|
|
|
class ExitStatus(str, Enum):
|
|
"""Outcome of a trial as observed at the harness level.
|
|
|
|
``PASSED`` / ``FAILED_VERIFIER`` are the normal exit modes; the rest
|
|
encode different failure modes the bandit may want to filter out
|
|
(e.g. wall-cap-bound tasks shouldn't be in the same K-means cluster
|
|
as agent-decision-bound tasks).
|
|
"""
|
|
|
|
PASSED = "passed"
|
|
FAILED_VERIFIER = "failed_verifier"
|
|
AGENT_TIMEOUT = "agent_timeout"
|
|
VERIFIER_TIMEOUT = "verifier_timeout"
|
|
REWARD_FILE_NOT_FOUND = "reward_file_not_found"
|
|
RUNTIME_ERROR = "runtime_error"
|
|
NO_SESSION = "no_session"
|
|
OTHER = "other"
|
|
|
|
|
|
@dataclass(frozen=True)
|
|
class ProxyFeatures:
|
|
trial_id: str
|
|
task_id: str
|
|
turn_count: int
|
|
final_exit_status: ExitStatus
|
|
has_tool_calls_ever: bool
|
|
assistant_text_length_avg: float
|
|
docker_error_count: int
|
|
|
|
|
|
# Docker / container daemon errors most commonly seen inside exec tool output
|
|
# or in the runner's exception traceback when something blows up at the infra layer.
|
|
_DOCKER_ERROR_PATTERNS = re.compile(
|
|
r"(?:"
|
|
r"docker:\s+error\b"
|
|
r"|cannot\s+connect\s+to\s+the\s+docker\s+daemon"
|
|
r"|error\s+response\s+from\s+daemon"
|
|
r"|no\s+such\s+container"
|
|
r"|container\s+not\s+running"
|
|
r"|docker\.errors\."
|
|
r")",
|
|
re.IGNORECASE,
|
|
)
|
|
|
|
|
|
_EXCEPTION_TO_STATUS = {
|
|
"AgentTimeoutError": ExitStatus.AGENT_TIMEOUT,
|
|
"VerifierTimeoutError": ExitStatus.VERIFIER_TIMEOUT,
|
|
"RewardFileNotFoundError": ExitStatus.REWARD_FILE_NOT_FOUND,
|
|
"RuntimeError": ExitStatus.RUNTIME_ERROR,
|
|
}
|
|
|
|
|
|
def _classify_exit(result_json: dict, reward_passed: bool | None) -> ExitStatus:
|
|
"""Map result.json's exception_info / reward.txt into an ExitStatus."""
|
|
if reward_passed is True:
|
|
return ExitStatus.PASSED
|
|
exc = (result_json.get("exception_info") or {}).get("exception_type")
|
|
if exc in _EXCEPTION_TO_STATUS:
|
|
return _EXCEPTION_TO_STATUS[exc]
|
|
if reward_passed is False:
|
|
return ExitStatus.FAILED_VERIFIER
|
|
# no reward and no recognised exception → bucket as OTHER
|
|
if exc:
|
|
return ExitStatus.OTHER
|
|
return ExitStatus.NO_SESSION
|
|
|
|
|
|
def _read_reward(trial_dir: Path) -> bool | None:
|
|
rt = trial_dir / "verifier" / "reward.txt"
|
|
if not rt.exists():
|
|
return None
|
|
try:
|
|
return float(rt.read_text().strip()) >= 1.0
|
|
except (ValueError, OSError):
|
|
return None
|
|
|
|
|
|
def _read_result_json(trial_dir: Path) -> dict:
|
|
rj = trial_dir / "result.json"
|
|
if not rj.exists():
|
|
return {}
|
|
try:
|
|
return json.loads(rj.read_text())
|
|
except (json.JSONDecodeError, OSError):
|
|
return {}
|
|
|
|
|
|
def _session_path(trial_dir: Path) -> Path | None:
|
|
"""Locate the agent's session.jsonl file under a trial dir."""
|
|
sessions_dir = trial_dir / "agent" / "workspace" / "sessions"
|
|
if not sessions_dir.is_dir():
|
|
return None
|
|
# the legacy runner writes a single tb2-task.jsonl by convention; fall back to
|
|
# any .jsonl if name differs.
|
|
preferred = sessions_dir / "tb2-task.jsonl"
|
|
if preferred.exists():
|
|
return preferred
|
|
candidates = sorted(sessions_dir.glob("*.jsonl"))
|
|
return candidates[0] if candidates else None
|
|
|
|
|
|
def _scan_session(session_path: Path) -> tuple[int, bool, float, int]:
|
|
"""Walk session.jsonl once and emit
|
|
``(turn_count, has_tool_calls_ever, assistant_text_length_avg, docker_error_count)``.
|
|
|
|
Iteration is a single pass over the file; each line is parsed once
|
|
and three counters are updated. Returns 0/False/0.0/0 for an empty
|
|
or unreadable session.
|
|
"""
|
|
turn_count = 0
|
|
has_tool_calls_ever = False
|
|
assistant_lengths: list[int] = []
|
|
docker_errors = 0
|
|
|
|
try:
|
|
with session_path.open() as f:
|
|
for line in f:
|
|
line = line.strip()
|
|
if not line:
|
|
continue
|
|
try:
|
|
r = json.loads(line)
|
|
except json.JSONDecodeError:
|
|
continue
|
|
role = r.get("role")
|
|
if role == "assistant":
|
|
content = r.get("content") or ""
|
|
assistant_lengths.append(len(content))
|
|
if r.get("tool_calls"):
|
|
turn_count += 1
|
|
has_tool_calls_ever = True
|
|
elif role == "tool":
|
|
content = r.get("content") or ""
|
|
docker_errors += len(_DOCKER_ERROR_PATTERNS.findall(content))
|
|
except OSError:
|
|
pass
|
|
|
|
avg_len = (sum(assistant_lengths) / len(assistant_lengths)) if assistant_lengths else 0.0
|
|
return turn_count, has_tool_calls_ever, avg_len, docker_errors
|
|
|
|
|
|
def _trial_task_id(trial_name: str) -> str:
|
|
return trial_name.rsplit("__", 1)[0] if "__" in trial_name else trial_name
|
|
|
|
|
|
def extract_features(trial_dir: str | Path) -> ProxyFeatures:
|
|
"""Extract :class:`ProxyFeatures` from a single trial dir.
|
|
|
|
Raises :class:`FileNotFoundError` when ``trial_dir`` does not exist
|
|
or is not a directory.
|
|
"""
|
|
p = Path(trial_dir)
|
|
if not p.is_dir():
|
|
raise FileNotFoundError(p)
|
|
|
|
result = _read_result_json(p)
|
|
reward_passed = _read_reward(p)
|
|
|
|
session = _session_path(p)
|
|
if session is None:
|
|
turn_count = 0
|
|
has_tool_calls_ever = False
|
|
avg_len = 0.0
|
|
docker_errors_session = 0
|
|
else:
|
|
turn_count, has_tool_calls_ever, avg_len, docker_errors_session = _scan_session(session)
|
|
|
|
# also scan exception traceback for docker errors
|
|
exc_tb = (result.get("exception_info") or {}).get("exception_traceback") or ""
|
|
docker_errors = docker_errors_session + len(_DOCKER_ERROR_PATTERNS.findall(exc_tb))
|
|
|
|
final = _classify_exit(result, reward_passed)
|
|
|
|
return ProxyFeatures(
|
|
trial_id=p.name,
|
|
task_id=_trial_task_id(p.name),
|
|
turn_count=turn_count,
|
|
final_exit_status=final,
|
|
has_tool_calls_ever=has_tool_calls_ever,
|
|
assistant_text_length_avg=avg_len,
|
|
docker_error_count=docker_errors,
|
|
)
|
|
|
|
|
|
def _find_attempt_root(trial_dir: Path) -> Path:
|
|
"""Same logic as stability_bucket: discriminate by ``verifier/`` subdir."""
|
|
if not trial_dir.is_dir():
|
|
raise NotADirectoryError(trial_dir)
|
|
has_trial_children = any(p.is_dir() and "__" in p.name and (p / "verifier").is_dir() for p in trial_dir.iterdir())
|
|
if has_trial_children:
|
|
return trial_dir
|
|
nested = [p for p in trial_dir.iterdir() if p.is_dir()]
|
|
if len(nested) == 1:
|
|
return nested[0]
|
|
return trial_dir
|
|
|
|
|
|
def extract_trial_dir(trial_dir: str | Path) -> dict[str, ProxyFeatures]:
|
|
"""Extract :class:`ProxyFeatures` for every trial under ``trial_dir``.
|
|
|
|
Accepts either the legacy jobs_dir or the dated subdir; returns
|
|
``{trial_id: ProxyFeatures}``.
|
|
"""
|
|
root = _find_attempt_root(Path(trial_dir))
|
|
out: dict[str, ProxyFeatures] = {}
|
|
for d in sorted(root.iterdir()):
|
|
if not d.is_dir() or "__" not in d.name:
|
|
continue
|
|
if not (d / "result.json").exists():
|
|
continue
|
|
if not (d / "verifier").is_dir():
|
|
continue
|
|
out[d.name] = extract_features(d)
|
|
return out
|
|
|
|
|
|
def main(argv: list[str] | None = None) -> int:
|
|
import argparse
|
|
from collections import Counter
|
|
|
|
ap = argparse.ArgumentParser(description=__doc__)
|
|
ap.add_argument("--trial-dir", required=True, help="legacy jobs_dir or dated subdir")
|
|
ap.add_argument("--json", default=None, help="optional JSON dump path")
|
|
args = ap.parse_args(argv)
|
|
|
|
feats = extract_trial_dir(args.trial_dir)
|
|
by_status = Counter(f.final_exit_status.value for f in feats.values())
|
|
|
|
print(f"trial_dir: {args.trial_dir}")
|
|
print(f"trials observed: {len(feats)}")
|
|
print("\nExit status breakdown:")
|
|
for status, n in sorted(by_status.items(), key=lambda x: -x[1]):
|
|
print(f" {status:24s} {n}")
|
|
|
|
print("\nTurn count quantiles (incl. wall-cap-zero):")
|
|
turns = sorted(f.turn_count for f in feats.values())
|
|
if turns:
|
|
for q, p in (("min", 0.0), ("p25", 0.25), ("p50", 0.5), ("p75", 0.75), ("max", 1.0)):
|
|
idx = min(len(turns) - 1, int(p * len(turns)))
|
|
print(f" {q}: {turns[idx]}")
|
|
|
|
if args.json:
|
|
with open(args.json, "w") as f:
|
|
json.dump(
|
|
{
|
|
tid: {
|
|
"task_id": x.task_id,
|
|
"turn_count": x.turn_count,
|
|
"final_exit_status": x.final_exit_status.value,
|
|
"has_tool_calls_ever": x.has_tool_calls_ever,
|
|
"assistant_text_length_avg": x.assistant_text_length_avg,
|
|
"docker_error_count": x.docker_error_count,
|
|
}
|
|
for tid, x in sorted(feats.items())
|
|
},
|
|
f,
|
|
indent=2,
|
|
)
|
|
return 0
|
|
|
|
|
|
if __name__ == "__main__":
|
|
raise SystemExit(main())
|