[agentic-token-optimizer] Optimize PR Description Updater: remove unused tools, tune diff chunk size

Target Workflow

PR Description Updater (.github/workflows/pr-description-caveman.md) — highest AIC among workflows not optimized in the last 14 days (last touched 2026-06-18). Excludes "Token"-named workflows per policy.

Analysis Period

Last 7 days, 12 runs, all success (0 failures, 0 retries).

Cost Profile

AIC scales almost linearly with sub-agent invocation count: the two highest-AIC runs (§32852854725, §32853314722) hit 31–32 invocations and 37.2–37.9 AIC, while the cheapest run (§32853462356) used only 4 invocations for 8.2 AIC — a ~4.5x AIC spread driven mostly by diff size/chunk count.

Ranked Recommendations

1. Remove unused github MCP toolset — est. 3 AIC/run savings
The workflow declares tools.github: {mode: gh-proxy, toolsets: [default]}, but the prompt body never calls a github(...) tool — all repository context comes from bash (git diff, git log) captured in the setup step, and the only write is update_pull_request (a safe-output, not the github MCP tool). The default toolset injects a large function-schema payload into every turn's context, rebuilt each of the ~13 avg invocations. Evidence: no github( tool-call references anywhere in pr-description-caveman.md; all 12 analyzed runs relied solely on the pre-fetched diff/commit files and inline sub-agents.
Action: remove the github: tool block entirely (keep permissions: pull-requests: read etc. as-is; they gate the safe-output, not the removed MCP tool).

2. Remove unused cli-proxy: true — est. 1 AIC/run savings
cli-proxy: true is declared but the prompt never issues a gh CLI command (bash allowlist only covers git diff/log/cat/ls/wc/split/head/tail, and none of those go through cli-proxy). This tool adds its own schema/gateway overhead on every turn.
Action: drop cli-proxy: true from tools:.

3. Reduce chunk-call overhead for large diffs — est. 4–5 AIC/run on large-PR runs
Diff chunking splits every PR diff into fixed 400-line chunks, each requiring a separate chunk-analyzer sub-agent call. The 4 highest-AIC runs all had 17+ invocations vs. 4–7 for the cheapest runs — invocation count (proportional to ceil(diff_lines/400)) is the dominant cost driver, not diff content complexity.
Action: raise the chunk size from 400 to 600–700 lines. This proportionally cuts sub-agent call count (and their fixed per-call schema/prompt overhead) by roughly 30–40% for large diffs while staying well under chunk-analyzer's per-call context budget (Haiku handles 600-line unified diff slices comfortably).

Optional Structural Optimization — Setup Prefix

Not warranted: the setup work (diff fetch, chunking, manifest generation) already lives entirely in the single steps: shell block before the agent starts, so there is no duplicated setup logic across prompt sections to extract.

Inline Sub-Agent Note

The workflow already uses two inline sub-agents (chunk-analyzer, pr-description-synthesizer) as its core architecture — per policy, no additional sub-agents are recommended.

Caveats

• 12 runs is a modest sample; conclusions on chunk-size tuning should be re-validated after the change lands across a wider mix of PR sizes.
• Removing the github tool and cli-proxy assumes no other silent dependency exists in imported shared components (shared/reporting.md was checked and contains no github(/gh calls either).
• AIC/invocation correlation is observational, not causal; a portion of the increase at high invocation counts may also reflect genuinely larger, more complex diffs needing more synthesis tokens.

References: §32852854725, §32853314722, §32853278481

Generated by Agentic Workflow AIC Usage Optimizer · copilot · auto · 61.3 AIC · ⊞ 11K · ◷

• expires on Sep 1, 2026, 7:21 AM UTC-08:00