Link structured evidence to coverage goals - #7
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 43a72a92ba
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| description = "field input (value omitted)" | ||
| s.coverage_action_trace.append({ | ||
| "tool": tool, | ||
| "description": _short(_redact(description), 120), |
There was a problem hiding this comment.
When a goal evidence window includes drop_file, the existing drop_file trace records the local basename (for example, 2026-layoffs.xlsx onto Upload) as the action description, and this new generic copier writes that description directly into coverage.evidence_refs.actions for both JSON and HTML reports. That exposes local or customer data in the coverage ledger even though these references are meant to contain value-redacted actions; sanitize drop_file the way upload_file only records a count.
AGENTS.md reference: AGENTS.md:L56-L56
Useful? React with 👍 / 👎.
1 participant
What changed
Why
The 0.6 coverage ledger required a narrative explanation but still trusted the agent to summarize what it had actually exercised. Structured references make goal completion auditable against session artifacts without exposing typed or pasted values in action references.
Impact
Coverage reports now answer both whether a goal was marked complete and what session evidence supports that state. Existing goal fields and sessions without explicit goals remain compatible. The core MCP profile remains at 30 tools and about 33,518 estimated context characters.
Validation
python -m pytest -q— 244 passedpython -m argus.bench --target all— 34/34 recall and 3/3 FP resistanceuv build— wheel and sdist built successfully