fix(log-watcher): remove branch filter; fold high-cost alert into primary report

[adamhenson]

Addresses two issues flagged in #18:

High - branch filter removed

The branches: [main] filter on the workflow_run trigger caused the watcher to silently skip all PR-branch agent runs, which is the common case for gh-aw. Removing the filter means the workflow fires for any branch, matching the behavior of the Cost Tracker and the rest of the agentics sample pack.

Medium - double-issue removed

Step 8 previously created a second issue for high-cost failures even when Step 7 had already opened a primary diagnosis issue (the no-PR path). This violated the "One report per run" guideline and would spam notifications. The high-cost callout is now folded into the body of the report that Step 7 already posts, keeping exactly one comment or issue per triggering run.

Closes the two bot findings on #18.

[github-actions]

Finding

• workflows/log-watcher.md:10 removes the branches: main filter from the workflow_run trigger. That means the watcher will now fire for every completed agent-implement / agent-pr-fix run on any branch. If those workflows run on feature or release branches, this will generate extra diagnosis comments/issues for runs you likely meant to ignore, and can create duplicate noise for the same logical agent run. Keep the branch filter unless cross-branch monitoring is intentional.

No other correctness issues stood out in this diff. The Step 8 change is directionally good because it avoids creating a second issue for high-cost failures.

[adamhenson]

Fair point - both concerns are valid and the real answer is that this should be user-configurable. Updated in 03ccd95: the branches filter is now a commented-out example with a note explaining the trade-off (omit to watch all branches including PRs, uncomment to restrict). The default behavior is correct for the common gh-aw case without silently surprising users who want branch filtering.

[github-actions]

Review Summary

This documentation change correctly addresses both issues from #18:

✅ Branch filter removal - Clear comment with trade-off explanation
✅ High-cost alert consolidation - Properly folded into Step 7's report to maintain "one report per run"

Found one minor clarity issue with variable naming consistency between Step 4 and Step 8. Otherwise the changes are clear, accurate, and achieve the stated goals.

Generated by Agent: Code Review for issue #19

[github-actions]

Findings:

• src/run.ts:71 and src/inputs.ts:19: workflow_run_id mode still treats the status input as the source of truth for rawConclusion, but that input defaults to success. If a caller follows the companion-workflow docs and omits status, failed/cancelled runs will be ingested as success. In this mode, read github.event.workflow_run.conclusion directly or make status required.

• src/token-extractor.ts:148 and src/comment.ts:160: the Codex JSONL path uses total_token_usage.input_tokens and also cached_input_tokens. Your own type/comment says cached_input_tokens is a subset of input_tokens, so cached prompt tokens get counted twice in cost calculations and also distort the cache-hit-rate math. Subtract cached tokens from the billable input bucket before pricing, or store uncached input separately.

Overall the rest of the flow looks solid, but I couldn’t run the test suite in this checkout because dependencies are not installed.