2 Commits

Author SHA1 Message Date
c7d296d2c5
feat: submission-complete — proof capture, CMC hub lineage, live wiring, docker fix (#3)
* fix(docker): deterministic pnpm + CI flag for dashboard image

corepack enable downloaded a nondeterministic 'latest' pnpm (v11+) whose
TTY/CI gates broke the non-interactive build. Pin pnpm@9.15.4, set CI=true,
require the lockfile (drop the silent '|| pnpm install' fallback). Verified
with a full local 'docker build -f infra/Dockerfile.dashboard .' (next build
completes, image exports). Greens the docker CI job.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(cmc): CMC Agent Hub capability/lineage surface (Phase B)

Make the CMC integration verifiable and discoverable, not just asserted —
the gap that kept the CMC Agent Hub lane at PARTIAL. Read-only only; no
execution is ever exposed to the hub.

- configs/cmc/capabilities.json: single source of truth declaring each CMC
  dataset (quotes/OHLCV/Fear&Greed/DEX-liquidity/security/trending/global)
  -> the capability it powers, with the exact cmc-client source and the
  API route / MCP tool that exposes it.
- GET /cmc/capabilities (apps/guardrail-api/src/cmc_capabilities.rs): serves
  the descriptor + a summary (datasets, capabilities, execution_exposed=false).
- MCP resource guardrail://cmc/capabilities mirrors the descriptor.
- agent_card.json: advertise a "CMC Agent Hub" service so the agent is
  discoverable from its ERC-8004 identity document.
- docs/CMC_AGENT_HUB.md + PRIZE_MAP row.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(submission): one-command proof capture + evidence auto-ticker (Phase C)

Turn the unchecked submission checklist into a captured, evidence-backed
artifact in one command — and tick it from real evidence, never by hand.

- python-lab/scripts/tick_checklist.py: reads the SQLite event log, the run
  report, and configs, then flips each proof-artifact checkbox based on whether
  the artifact actually exists, and regenerates an evidence table. Honest by
  construction: items needing a real on-chain tx report "pending (live)" in
  paper mode instead of a false tick. Evidence is scoped to the primary run's
  run_id so the shared-DB kill-switch demo can't shadow it.
- configs/risk_policy.killswitch_demo.json + configs/killswitch_demo.toml: a
  low-threshold (0.05%) policy/config that trips the kill switch on a normal
  paper run, emitting a real KillSwitchTriggered event (shared DB, throwaway
  report so the primary run report is preserved).
- scripts/capture_submission.sh: orchestrates run -> kill-switch demo -> export
  submission.md -> proof-verify -> tick. Offline-safe (paper/mock); flips to
  live when keys + production config are supplied.

Verified end-to-end: a paper capture ticks 10/11 from real evidence; only
Competition registration stays pending (live), exactly as it should.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(live): registration wiring + go-live preflight + go_live.sh (Phase D)

Make the live path turn-key and the on-chain registration proof flow end to
end, so the keyed go-live (the one human-in-the-loop step) is a single safe
command.

- agent-runtime: thread the competition registration tx through RunMeta into
  the run report, the AgentProof, and the AgentReportPublished summary — but
  only for a real LIVE registration. In paper the mock executor's synthetic
  hash is logged as an event for the demo and never presented as an anchored
  on-chain tx, so proof_verify and the checklist stay honest (registration =
  pending until live). The run report now also carries agent_id.
- guardrail-doctor --live: a preflight safety gate — required credentials
  (CMC_API_KEY, BSC_RPC_URL, TWAK transport), production config in live mode,
  and conservative policy caps. Env + file checks only; refuses to green when
  unconfigured. +2 tests.
- tick_checklist.py: registration ticks only on a live tx; a paper mock stays
  pending (live).
- scripts/go_live.sh: preflight -> on-chain reachability -> confirm -> bounded
  live run -> capture proof -> print BscScan links. Requires a typed "GO LIVE"
  confirmation (spends real money) unless --yes.

Workspace builds green, clippy clean, agent-runtime e2e tests pass, paper
capture still ticks 10/11.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* docs: runbook one-command paths, checklist regen note, WHATS_NEW (Phase F)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-16 02:54:00 +03:00
2d2b15c7fe feat: full guardrail-alpha system + CLI modularization + Vercel deploy config
- Rust engine: real bootstrap/wiring preflight (guardrail-agent), execution layer (retry/router/pre-trade/post-trade/reconciliation), guardrail-doctor diagnostics with allowlist-drift guard.

- guardrail-cli: extracted pure helpers into util.rs (split in progress).

- 6 Track-2 skills + ensemble; python-lab analytics (regime/drawdown/montecarlo/dossier/ensemble/journal/seed); clients (ts/py/go/mcp/langchain/web-lite/proof-verifier); alert relay; scenarios; helm/k8s/compose; docs.

- dashboard/vercel.json + docs/VERCEL_DEPLOY.md for Vercel deployment.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-15 11:02:22 +03:00