[agentic-token-optimizer] Optimize [aw] Failure Investigator (6h): unused MCP toolset + sub-agent consolidation

[github-actions]

Target Workflow

[aw] Failure Investigator (6h) (.github/workflows/aw-failure-investigator.lock.yml, source aw-failure-investigator.md)

Selected because it is the single highest-AIC workflow in the last 7 days (top-workflows.json), has not been optimized in the last 14 days (last touched: none found in optimization-log.json), and its name does not contain "Token".

Analysis Period + Runs Analyzed

• 7-day window snapshot: 1 run, AIC 254.60, 89,361 raw tokens (top-workflows.json).
• Historical daily snapshots: 2026-07-07 (AIC 105.21), 2026-07-08 (AIC 145.90) — AIC has grown ~2.4x since early July for a functionally unchanged prompt.
• API-confirmed run history (last 15 scheduled runs): 14 success, 1 failure (§32641065351, driver-exit in the agent job, unrelated to prompt/tool config — CLI execution crash).
• Deep-dived the latest full run §32730678157 using local logs under .github/aw/logs/run-32730678157/usage/.

Cost Profile

Metric	Value
Total AIC (7d window)	254.60
Avg AIC/run (7d)	254.60
Raw token usage (latest run)	89,361 (21,904 input / 67,457 output)
Cache read tokens (latest run)	6,562,270
Cache write tokens (latest run)	223,595
Turns / invocations (latest run)	66 invocations, rebuild_factor 2.56x
Peak vs cumulative input tokens	peak 8,541 / cumulative 21,904
Action minutes (latest run)	26

The rebuild_factor of 2.56x means the agent's context was rebuilt roughly 2.5x the size of its peak working set across the run's 66 tool invocations — a direct AIC cost driver from repeated context reconstruction rather than raw prompt size.

Ranked Recommendations

1. Remove the unused pull_requests GitHub MCP toolset — est. 5–8% AIC savings/run

• Action: change tools.github.toolsets from [actions, issues, pull_requests] to [actions, issues].
• Evidence: the prompt body, all three inline sub-agents (failure-classifier, issue-matcher, cluster-evidence-extractor), and the deterministic pre-fetch script never reference pull-request data — investigation is scoped entirely to workflow runs, jobs, logs, and issues. The only pull_requests-adjacent references in the source are the static pull-requests: read permission and exclude_pull_requests: true in the pre-fetch API filter (which explicitly excludes PR-triggered runs, the opposite of needing PR tools). Loading an unused MCP toolset still adds tool-definition tokens to every context rebuild, compounding with the observed 2.56x rebuild factor.
• Caveat: only 1 full run was available in the 7-day window for direct AIC comparison; recommend validating the saving over the next 3–5 runs.

2. Consolidate failure-classifier and issue-matcher into a single sub-agent call — est. 8–12% AIC savings/run

• Action: merge the two sequential small-model sub-agent invocations in Step 1 (failure-classifier then issue-matcher) into one sub-agent that both clusters failures by signature/severity and matches clusters to existing_tracking_issues in a single pass, returning one combined JSON object (clusters + matched + gaps).
• Evidence: both agents consume the same failures payload from the single pre-fetch file, run back-to-back with no intervening dependency on external state, and both use model: small. Each sub-agent call re-serializes/re-sends its instructions and schema into a fresh context, contributing directly to the measured 66 invocations and 2.56x rebuild factor for this run. Combining them keeps the same small-model division of labor while cutting one full invocation round-trip per run.
• Caveat: verify combined output schema doesn't reduce match-confidence quality; keep the two-step logic internally in the merged agent's prompt to preserve behavior.

3. Trim redundant "read the pre-fetch payload" reminders in the prompt body — est. 2–4% AIC savings/run

• Action: Step 0 already states "Read failed_run_ids, failures, and existing_tracking_issues once... Do not re-read this file; keep the parsed data in context." The same payload fields (failures, existing_tracking_issues) are then re-described in Steps 1 and 4. Consolidate the field descriptions into Step 0 only, and have Steps 1/4 reference "the parsed payload from Step 0" instead of re-listing field names.
• Evidence: sections ### 0, ### 1, and ### 4 each independently restate which payload fields to use, adding duplicate instructional tokens to every context rebuild without changing agent behavior.

Caveats

• Sample size is limited: only 1 full run fell inside the 7-day analysis window; historical snapshots (2026-07-07/08) were used to confirm cost trend but predate current experiment/tone-variant tooling, so exact percentage savings should be treated as directional, not guaranteed.
• The workflow already runs an active tone_variant A/B experiment (issue [ab-advisor] Experiment campaign for aw-failure-investigator: A/B test tone_variant #36105); tone changes should not be conflated with the structural/tool recommendations above — no changes to ## Tone Variant Instructions are proposed here.
• The one observed failure (§32641065351) was a driver-level Claude Code CLI crash, not caused by prompt/tool configuration — no reliability fix is recommended for it.
• No inline sub-agents are proposed as new additions since the workflow already contains three (failure-classifier, issue-matcher, cluster-evidence-extractor); recommendation Add workflow: githubnext/agentics/weekly-research #2 above is a consolidation of two existing sub-agents, not a new addition.

References:

• §32730678157
• §32641065351
• §28925815752

Generated by Agentic Workflow AIC Usage Optimizer · copilot · auto · 83.1 AIC · ⊞ 11K · ◷

• expires on Aug 31, 2026, 7:17 AM UTC-08:00