[agentic-token-optimizer] Reduce Avenger AIC: skip duplicate CI fixes and cap guidance-test loops #76889

Target workflow

Avenger (.github/workflows/avenger.md, hourly CI fixer). Selected as the highest-AIC workflow not optimized in the last 14 days: 96.2 AIC across 3 runs. Dead Code Removal Agent (107.9 AIC) was excluded because it was optimized on 2026-11-09.

Analysis period

• Runs: 3 Avenger runs, 2026-11-11 12:30–14:26 UTC, all success.
• The pre-aggregated sample covers only 11:53–14:38 UTC today. This is a short-window read, not a 7-day trend.

Cost profile

What drives the cost

Main-branch CI failed on 7 consecutive completed runs (08:58–14:09 UTC). The Avenger agent job ran in 22 of its last 24 hourly runs, and each run repeated the same work:

• Byte-budget loop: test_debug_playbook_agent_consumption_budget fails because .github/aw/debug-playbook.md is 26,387 bytes against a 26,000-byte limit. Two runs made 14 and 15 bash calls re-checking size and re-running the guidance test. The third run made none.
• Duplicate fixes: Two open issues propose the same trim, [avenger] Trim debug playbook to fit guidance size budget #76859 (12:39 UTC) and [avenger] Trim debug playbook under byte budget #76873 (13:36 UTC). Both PRs fell back to issues because they touched protected files.
• Repeated filtered logs: get_job_logs was integrity-filtered twice per run (6 total), and the agent retried it.
• Nonexistent tool names: The agent called github and safeoutputs as tool names, failing 1–2 times per run.

The cheapest run (43038561248, 18.0 AIC, 9 tool calls) classified the failure from job metadata as "release alignment, not code-fixable" and called noop. Early classification roughly halved its cost.

Ranked recommendations

1. Skip when an open fix already exists (~20 AIC/run avg; up to ~30 on affected runs)

   • Action: Before any edits, search open issues and PRs with the [avenger] prefix for the same failing check name. If one is open, call noop with a link to it.
   • Evidence: [avenger] Trim debug playbook to fit guidance size budget #76859 and [avenger] Trim debug playbook under byte budget #76873 are both open for the same failure. 2 of 3 runs duplicated that work.

2. Cap guidance-test iteration (~10 AIC/run avg; ~15 on loop runs)

   • Action: The source caps only make test-unit (Step 7). Add the same cap for make test-guidance and single-test reruns: at most 2 runs after a fix, then stop and report. Do not loop on wc -c size checks.
   • Evidence: 14 and 15 loop calls in 2 of 3 runs.
   • Overlaps with rejig docs #1, so do not add the two savings together.

3. Stop retrying integrity-filtered logs (~3 AIC/run)

   • Action: On the first integrity denial from get_job_logs, classify the failure from job and step names, as run 43038561248 did. Do not retry.
   • Evidence: Integrity-filter friction cost was 2.6–3.6 AIC/run. All 6 filtered calls were get_job_logs.

4. Use correct tool names (~1–2 AIC/run)

   • Action: State that GitHub MCP tools are called by their function name (for example actions_list), and that safe-output functions are called by their own names (for example create_pull_request, noop). Never invoke github or safeoutputs as tool names.
   • Evidence: "Tool 'github' does not exist" and "Tool 'safeoutputs' does not exist", 1–2 times per run.

Caveats

• The sample is 3 runs from one 2.5-hour window, all with the same failure. That is below the 5-run minimum, so no tool-removal recommendation is made.
• Savings are estimates from request-proportional AIC, not per-call measurements.
• These changes lower cost per run but do not fix CI. The root cause (debug-playbook over budget) needs a human-reviewed change.
• No sub-agent or setup-prefix change is recommended. The cost is in validation loops that need main-agent context.

References: §43029233076 · §43020983525 · §43038561248

Generated by Agentic Workflow AIC Usage Optimizer · copilot · routed: haiku55 xhigh · ~100 AIC · ⊞ 5.8K · ◷

• expires on Nov 18, 2026, 7:25 AM UTC-08:00