[agentic-token-optimizer] Optimize Daily Go Test Parallelizer: cut context rebuild and cadence (~25 AIC/run) #71459

Target workflow

Daily Go Test Parallelizer (.github/workflows/daily-go-test-parallelizer.md). Selected as the highest-AIC workflow not optimized in the last 14 days. Avenger is excluded because it was optimized on 2026-10-15. The next entry, Matt Pocock Skills Reviewer, was optimized 2026-10-14, which is also inside the 14-day window.

Analysis period

7-day window ending 2026-10-23. Only 2 runs are in all-runs.json, so this audit is below the 5-run minimum. The workflow-run list shows 7 of the last 8 runs succeeded and 1 failed after about 46s. Run-level data comes from the 2 audited runs.

Cost profile

Ranked recommendations

1. Reduce context rebuild by batching reads (~15–20 AIC/run)

About 95% of input tokens are re-sent context across 23–37 invocations. The AIC range tracks the invocation count (23 vs 37).

• Action: Add an explicit step: read all selected files in one parallel block, or with a single bash call such as grep -nE 'func Test|t\.(Setenv|Chdir|Parallel)|os\.(Setenv|Chdir)|\bvar\b' <files>. Use that pre-filter to classify files, and only open the full file for candidates.
• Action: Say "do not re-read edited files; verify through git diff" to prevent repeat reads.

2. Lower the batch size or cadence (~8–12 AIC/run, or more per day)

The schedule is every 2h (about 12 runs/day) but the title, description and prompt say "daily". At 25 files per run, a single pass over the repo needs many runs, and each run costs about 54 AIC.

• Action: Either change the schedule to daily, or keep the cadence and add early exit. Once a full wrap-around yields no safe candidates, store {"last_file":..., "no_change_streak":N} and noop immediately when the streak is long.
• Action: Skip the run when the cache shows all files have been analyzed and no *_test.go has changed since the last pass.

3. Run go test ./... once, scoped (~3–6 AIC/run)

The prompt requires go test -race per package and then go test ./..., which overlaps. A repo-wide test run adds tool-output tokens and wall-clock time.

• Action: Run go test -race on modified packages only. Run go test ./... only if more than one package changed. Pipe output through tail -n 40.

4. Fix prompt ambiguity that invites extra iteration (~1–2 AIC/run)

• The "Batched analysis agent" section references a sub-agent but defines none. Replace it with a direct instruction: "Analyze the batch yourself and keep a compact JSON note".
• Move the exclusion list above "Analyze safety" so the rules are read before analysis.

Structural optimizations

None recommended. The prompt has no repeated setup prefix, and an inline sub-agent would not pay off for a single-batch analysis that depends on main-agent edits.

Caveats

• Only 2 runs are in the window, so the savings estimates are approximate. They assume invocation counts fall toward the lower run (about 23 calls).
• Recommendation 2 depends on a maintainer decision about intended cadence.
• The failure run 39939907542 (46s) was not analyzed.

References: §39971096234, §39955435424, §39939907542

Generated by Agentic Workflow AIC Usage Optimizer · copilot · auto · ~20 AIC · ⊞ 11K · ◷

• expires on Oct 30, 2026, 6:59 AM UTC-08:00