[agent-deep-dive] Agent deep dive: gh-aw-duplicate-issue-detector - 2026-02-21

[github-actions]

Agent Deep Dive: gh-aw-duplicate-issue-detector

Date: 2026-02-21
Runs analyzed: 20 total (19 success, 0 failure, 1 other)

Run Outcomes

Run	Date	Conclusion
run 22262186059	2026-02-21 18:40:28 UTC	in_progress
run 22262123419	2026-02-21 18:36:11 UTC	success
run 22261695145	2026-02-21 18:07:59 UTC	success
run 22261332574	2026-02-21 17:44:12 UTC	success
run 22261212227	2026-02-21 17:36:07 UTC	success
run 22260988485	2026-02-21 17:21:22 UTC	success
run 22260149052	2026-02-21 16:23:46 UTC	success
run 22260097115	2026-02-21 16:20:14 UTC	success
run 22260090112	2026-02-21 16:19:39 UTC	success
run 22259979430	2026-02-21 16:11:54 UTC	success
run 22259925085	2026-02-21 16:08:26 UTC	success
run 22259356081	2026-02-21 15:28:33 UTC	success
run 22259287573	2026-02-21 15:23:50 UTC	success
run 22250428437	2026-02-21 04:36:08 UTC	success
run 22250139566	2026-02-21 04:15:59 UTC	success
run 22249657517	2026-02-21 03:42:01 UTC	success
run 22249593200	2026-02-21 03:37:04 UTC	success
run 22249574853	2026-02-21 03:35:35 UTC	success
run 22249253216	2026-02-21 03:11:59 UTC	success
run 22248898492	2026-02-21 02:47:23 UTC	success

Failure Analysis

No failed runs were present in the analyzed 20-run window.

Additional observed artifacts:

• /tmp/gh-aw/deep-dive/errors.json reported total_matches: 0.
• /tmp/gh-aw/deep-dive/manifest.json was empty ([]).
• /tmp/gh-aw/deep-dive/logs/ directory was not present.

Tool Call Patterns

Sampled agent-job logs from successful runs:

• Run 22262123419 (job 64402003271)

• Run 22261695145 (job 64400937447)

• Run 22261332574 (job 64400051749)

• Average turns per run: 18 (using successful tool-call sequence length in sampled runs: 17, 24, 13)

• Most-called tools: github-search_issues (26 total), github-issue_read (16 total), safe output tool (safeoutputs-add_comment/safeoutputs-noop) once per run

• Anomalies: no tool failures observed; each sampled run ends with exactly one safe output emission (output_types: add_comment or output_types: noop)

Behavioral Observations

• Prompt sequence adherence is consistent with the workflow definition in .github/workflows/gh-aw-duplicate-issue-detector.md: context fetch (generate_agents_md), issue read, multiple issue searches, and one terminal safe output.
• Search behavior is broad and query-variant heavy in all sampled runs (many semantically similar search_issues calls per issue).
• The workflow reliably emits a single formatted output type and does not exceed safe-output limits.

Recurring Patterns

Patterns appearing in 2+ sampled runs:

1. Initial generate_agents_md call followed by issue_read for the triggering issue.
2. Burst of search_issues queries with parameter variation (is:issue, is:closed, phrase variants).
3. Follow-up issue_read calls on candidate issue numbers from search results.
4. Exactly one final safe output call (add_comment when a match is found, noop otherwise).

Generated by Agent Deep Dive

• expires on Mar 7, 2026, 6:45 PM UTC

[strawgate]

let's update the deep-dive prompt, we want it to determine if the tool calls are excessive or inappropriate or if changes to the prompt might improve the runtime, reduce the number of tool calls, move things along faster, etc