[agentic-token-optimizer] Package Specification Enforcer: cut ~60-80 AIC/run (context rebuild, model cost) #70539

Target workflow

Package Specification Enforcer (.github/workflows/spec-enforcer.md). It is the highest-AIC workflow in the 7-day window that has no optimization in the last 14 days. Higher-ranked entries were excluded: Claude Code User Documentation Review was optimized on 2026-10-09, and Avenger on 2026-10-15.

Analysis period and runs

• Window: the last 7 days (all-runs.json).
• Runs analyzed: 1, §39415378641. It succeeded.
• Prior daily snapshots from 2026-09-14 to 2026-10-02 show about 2–5 AIC per run, and 0 AIC on some days.
• This run cost roughly 20–50x that baseline.

Cost profile

The engine and model are Codex with openai/gpt-5.6-sol. Cache efficiency could not be measured from the data. About 94% of the tokens are re-read context, so per-step context size is the main cost driver.

Ranked recommendations

1. Cut context growth by capping per-package output (about 40–60 AIC/run).

   • Phase 2 Step 3 runs cat pkg/<package>/spec_test.go. It also reads the full README and runs grep ... | head -50. All of these stay in context for every later call.
   • Instead, have the main agent read only grep -n "^func Test" spec_test.go.
   • Pipe go test -v output through tail -n 60 before passing it to test-output-classifier.
   • Evidence: 33 invocations rebuilt a context that grew 59x.

2. Reduce steps per package (about 15–25 AIC/run).

   • The prompt has many sequential bash calls per package: find, git log, grep, grep package, find spec_test, cat, build, test, and a cache write.
   • Combine them into one && chain per package.
   • Run go build only when go test fails to compile, since go test already compiles.
   • Evidence: 33 steps with a trajectory efficiency of 0.22.

3. Reconsider the model choice (savings uncertain).

   • Earlier runs cost about 4 AIC. The current model: openai/gpt-5.6-sol pin is a plausible cause of the jump, but this is not verified. The audit data does not show an earlier model.
   • Compare against the previous model or a cheaper tier on 3+ runs before keeping the pin.

4. Lower max-turns: 100 to about 40 as a runaway guard.

   • The run needed 33 steps, so there is headroom.
   • This is a safety net and not an expected saving.

Structural optimizations

• Inline sub-agents: the workflow already has two (rotation-state-recoverer, readme-spec-extractor/test-output-classifier references). I recommend no additional ones.
• Shared setup prefix: none warranted. The sections do not repeat opening calls.

Tool usage

• The configured tools are bash: "*", edit, cache-memory, cli-proxy and github (default toolsets).
• The agenticworkflows and github tools did not clearly appear in this run's usage data.
• I would not remove any tool on one run, because the rule requires at least 5 runs. github toolsets: re-check once more runs are available.

Caveats

• Only one run is in the window, so the recommendations are low-confidence.
• A larger package batch (the 2–3 round-robin packages, or enforce_all) probably explains part of the cost.
• Run §39415378641 is the only evidence cited.

References: §39415378641

Generated by Agentic Workflow AIC Usage Optimizer · copilot · auto · 22.7 AIC · ⊞ 11K · ◷

• expires on Oct 27, 2026, 7:00 AM UTC-08:00