[agentic-token-optimizer] Design Decision Gate: switch off Opus, skip agent on checkout failure (~20 AIC/run) #70904

Target workflow

Design Decision Gate 🏗️ (.github/workflows/design-decision-gate.md). It is the #2 workflow by AIC over the last 7 days. #1 (Matt Pocock Skills Reviewer) and #3 (Avenger) were optimized within the last 14 days, so they were skipped.

Analysis period

Last 7 days, 5 runs analyzed (2 success, 3 failure). All five were pull_request events.

Cost profile

Ranked recommendations

1. Use a cheaper model (est. −15 to −20 AIC/run)

• Action: replace model: copilot/claude-opus-5.5 with a mid-tier model (for example a Sonnet-class model).

• Evidence:

  • The task is mostly classification (does the PR need an ADR?), and the final output is a short comment or a templated ADR draft.
  • A successful run produces only ~4.2k output tokens but costs ~28.5 AIC.
  • Check ADR draft quality on the next few runs and revert if it drops.

2. Skip the agent when PR branch checkout fails (est. −32 AIC per failed run)

• Action: make the agent job fail fast, or conditionally skip, when Checkout PR branch fails. Alternatively, have the pre-fetch step detect a missing branch and set skip_reason.

• Evidence:

  • Runs §39618696347 and §39614442359 failed at Checkout PR branch. The error was git fetch failed with exit code 128 for branch chore/remove-dead-functions-... on PR 70898.
  • Despite the failure, they were billed about 32 AIC each, the highest per-run AIC in the sample. About 64 of the 121.7 AIC (53%) went to failed runs.

• Caveat: I did not confirm why the fetch failed (a deleted or renamed branch is plausible). Treat the cause as unconfirmed.

3. Align the turn budget and prompt (est. −2 to −4 AIC/run)

• Action: the frontmatter has max-turns: 30, but the prompt says "Hard Turn Budget: 20" and "Total ≤ 7". Set max-turns to about 12 and remove the duplicated stop rules. The Mandatory Stopping Criteria and Efficiency Rules overlap with the turn table.
• Evidence: successful runs have 225 session events, far more than the 7-turn plan implies.

4. Narrow bash: ["*"] (reliability and token hygiene, low savings)

• Action: the prompt only needs find, cat and mkdir. Use a restricted allowlist (for example ["find","cat","mkdir","ls"]).
• Caveat: I did not audit per-tool usage in the run logs. Confirm no other command is used in a successful run before narrowing.

Structural optimizations

None recommended. The prompt has only about 4 sections, and the steps depend on each other (summary → PR details → ADR check → output). Moving them to inline sub-agents would add overhead without savings. The prefetch step already runs in setup steps, so there is no repeated setup prefix to extract.

Caveats

• Only 5 runs are in the window, and 3 of them failed (one in 7s, with 0 AIC).
• The AIC savings estimates are not measured. They assume a roughly 3–5× cheaper model class.
• Turn counts are missing from the run data.
• Recommendation 2 assumes failed-checkout runs are billed as observed. The Pi engine run appears to continue past the checkout error.

References: §39618696347, §39616146839, §39615568992

Generated by Agentic Workflow AIC Usage Optimizer · copilot · auto · 23.9 AIC · ⊞ 11K · ◷

• expires on Oct 28, 2026, 7:00 AM UTC-08:00