[agentic-token-optimizer] Optimize PR Sous Chef: reduce AIC via duplicate-line cleanup, redundant skip-table removal, and working-set growth mitigation

[github-actions]

Target workflow

PR Sous Chef (.github/workflows/pr-sous-chef.md) — highest total AIC among workflows not optimized in the last 14 days (324.4 total AIC over 11 runs in the last 7 days; avg 29.5/run). Runner-up candidates (Claude Code User Docs Review, Repository Quality Improvement Agent, Daily File Diet) were single-run outliers or already optimized recently.

Analysis period

Last 7 days, 11 runs analyzed (10 success, 1 failure — agent_logic failure kind, run §36149578895).

Cost profile

Metric	Value
Total AIC (11 runs)	324.41
Avg AIC/run	29.49
Total raw tokens	5,404,980
Avg raw tokens/run	~491,362
Avg turns/run	0 (not populated)
Avg sub-agent invocations/run	11.9
Avg working-set rebuild factor	9.0x
Avg cumulative input tokens/run	488,827
Total GitHub API calls	~220 (≈20/run)
Safe items produced (11 runs)	17

Ranked recommendations

1. Remove duplicated instruction line (~2–3% AIC savings/run, ~0.6–0.9 AIC/run)

Evidence: Line 466–467 of the prompt body contains an exact duplicate sentence:

- Every `add_comment` must include `pr_number`. Never emit `add_comment` without a numeric target field.
- Every `add_comment` must include `pr_number`. Never emit `add_comment` without a numeric target field.

This line is re-sent as static prompt context on every one of the 11 runs analyzed with no functional benefit — pure duplicated tokens. Action: delete one of the two identical lines.

2. Finish the remove_redundant_context_v1 experiment rollout (~3–5% AIC savings/run once graduated)

Evidence: The workflow already has a live experiment (remove_redundant_context_v1) testing removal of the "Required skip rules per PR" backup table, on the stated rationale that the same 3 conditions are already deterministically enforced by the fetch-prs prefilter step and re-verified by the pr-processor sub-agent — i.e., the table is redundant on every run regardless of variant. The experiment requires min_samples: 20; only 11 runs have been observed in the analysis window. Action: no prompt change recommended until the experiment reaches its min_samples threshold and a decision is made per its existing mann_whitney analysis — flagging this here only to confirm the audit is aware of the in-flight test and does not duplicate or contradict it.

3. Investigate working-set rebuild-factor variance (~5–10% AIC savings/run on high-rebuild runs)

Evidence: Rebuild factor ranges from 4.3x to 14.6x across the 11 runs, with the two highest-AIC runs (48.9 AIC / 21 invocations / 14.6x rebuild; 32.6 AIC / 16 invocations / 13.4x rebuild — the latter being the failed run) also showing the highest sub-agent invocation counts and cumulative input tokens (867K–927K). Runs with fewer eligible/nudged PRs (safe_items_count=1) cluster around 4–7x rebuild factor and 6–9 invocations. This suggests each additional PR processed via the pr-processor sub-agent re-sends a large amount of shared context rather than reusing it. Action: confirm the pr-processor sub-agent invocation only passes "the PR number and compact context" as instructed (step 7 already mandates this) — audit whether the compact JSON payload size grows with PR count, and consider trimming failed_checks/reviewThreads fields passed per-PR to only what's needed for the current PR being processed.

4. Investigate the single agent_logic failure (~9% of runs at 100% partial-completion risk)

Evidence: Run §36149578895 completed 5 safe outputs (2× update_pull_request, 2× add_comment, 1× create_issue) but was still marked conclusion: failure with failure_kind: agent_logic and error_count: 1, at the highest token cost of the sample (678,188 tokens, 32.6 AIC). Wasted spend on a failed run compounds cost without full nudge value delivered. Action: pull the run's error/warning log for this specific run to identify the root cause (likely a missing precondition check or unhandled sub-agent response) and add a targeted guard.

Caveats

• Sample size is limited to 11 runs in a 7-day window; turns field is not populated (0 for all runs), so turn-based iteration analysis was not possible — recommendations rely on AIC, token, and working-set metrics instead.
• The workflow already has one active experiment in flight (remove_redundant_context_v1); this audit intentionally avoids proposing an overlapping change to the same section.
• No new inline sub-agents are recommended: the workflow already has one (pr-processor) handling the clear per-PR-independent task, and no other section met the "no existing agent: blocks" precondition for additional inline sub-agent recommendations.
• The single failed run's root cause was not fully diagnosed here (would require deeper log inspection); flagged as recommendation Add workflow: githubnext/agentics/weekly-research #4 for follow-up rather than a confirmed fix.

Raw per-run metrics (11 runs)

AIC	Invocations	Rebuild factor	Cumulative input tokens	API calls	Safe items	Conclusion
32.61	16	13.35x	675,668	20	5	failure
36.72	15	12.78x	545,288	20	1	success
18.94	10	7.53x	315,659	22	1	success
30.88	13	10.75x	521,920	18	1	success
21.15	6	4.29x	241,474	20	1	success
48.89	16	11.80x	926,833	20	2	success
23.91	7	4.48x	261,839	20	1	success
10.17	6	4.68x	115,124	22	1	success
43.96	21	14.59x	867,679	18	1	success
31.59	12	9.15x	548,628	20	1	success
25.58	9	5.64x	356,987	20	2	success

References: §36149578895 (failed run), §36147251830, §36140522853 (highest AIC success run)

Generated by Agentic Workflow AIC Usage Optimizer · copilot · auto · 72.7 AIC · ⊞ 10.8K · ◷

• expires on Oct 2, 2026, 7:02 AM UTC-08:00