Give CautionDrafter the frame tallies RunDebriefer already sees

[xmap]

Summary

CautionDrafter proposes standing "Caution" advisories on Assets/Procedures from terminal Run events, but CautionDrafterPayload never carried frame tallies, unlike its sibling RunDebriefer, which already receives them via extract_capture_progress and correctly flags DataSuspect. CautionDrafter was structurally blind to frame-count shortfalls: the data existed on the terminal event, but never reached the LLM call.

• Add capture_progress: dict[str, int] | None = None to CautionDrafterPayload, mirroring RunDebriefPayload's field of the same name.
• Wire extract_capture_progress(event) into the subscriber, same helper RunDebriefer already uses.
• System prompt: a shortfall is a fact about ONE Run; RunDebriefer already records it as DataSuspect; a Caution is a claim about an ASSET, which a single Run cannot establish a pattern for. A shortfall alone must never originate a proposal, and NoAction is correct for an isolated shortfall. Refusing discards nothing: the tallies live permanently on the terminal event and RunDebriefer's verdict is a durable Decision.
• The existing lookback-then-propose ProposeSupersede branch (shortfall + a matching existing Active Caution) is kept, but demoted to an explicitly secondary path, not the primary mechanism, because the pilot currently has zero Caution streams and it would otherwise be dead on arrival.
• capture_progress absence stays ordinary (not every Run is a witnessed capture); never compare a saved count with a collected count (verbatim rule from run_debrief.py, different upstream instruments).
• Module docstring names cross-Run recurrence as a deliberate v1 scope boundary, not a TODO: this agent reads one terminal event at a time and cannot yet distinguish "this Run was short" from "this Asset is short on a third of its Runs by a consistent margin." Grounded in a pilot measurement (2026-08-21): of 974 completions, all declare the same 1541-frame expectation; 616 met it and 364 fell short by ~11 frames on average, a systematic offset, not noise. Sourcing that recurrence signal without a new port/lookup is the next slice.
• CAUTION_DRAFTER_PROMPT_TEMPLATE_ID is unchanged (per design, pre-PromptTemplate-aggregate); its one-line description in KNOWN_PROMPT_TEMPLATES was updated to note the payload now carries capture_progress. Note: the ID is stable across this change even though the underlying prompt bytes differ before/after, which is the known provenance limitation of the stopgap registry.

Deliberately does not enable origination from a shortfall pattern. This slice is awareness-only by design: the agent sees the tallies and can reason about them in a corroboration role, but nothing in this change lets a shortfall, by itself or by recurrence, produce a new Caution proposal on its own.

Out of scope (unchanged): no new port/lookup/query/projection, no change to RunDebriefer or extract_capture_progress itself, no change to the 5-choice vocabulary/severity anchors/tier quotas, and the unrelated agent_name repr bug in regenerate_run_debrief/handler.py is left alone.

Test plan

• uv run pytest apps/api/tests/unit apps/api/tests/architecture -q — 45320 passed, 631 skipped
• uv run pytest apps/api/tests/integration -q — 1273 passed, 1 skipped
• make lint — clean
• make typecheck — 0 errors
• New tests pin payload/request-building behavior only (frame tallies reach the built request when present, None when absent, both the shortfall-with-matching-Caution and shortfall-with-no-existing-Caution payload shapes), never asserting what the model chooses.

🤖 Generated with Claude Code

[github-actions]

Coverage report

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
apps/api/src/cora/agent/prompts
__init__.py
caution_drafter.py
apps/api/src/cora/agent/subscribers
caution_drafter.py
Project Total

This report was generated by python-coverage-comment-action