[agentic-token-optimizer] Optimization: Release workflow — prompt condensation, tool scoping, and CHANGELOG sub-agent

[github-actions]

Target Workflow

Release — selected as the highest-value agentic workflow with the oldest meaningful optimization entry (2026-05-19, 3,079 tokens analyzed). Subsequent optimization log entries from 2026-05-20–2026-05-26 all show 0 tokens analyzed, suggesting no effective re-analysis was performed. The prompt body is 1,201 words / 130 lines — the longest of any readable agentic workflow source in this repo.

Data note: All pre-aggregated token snapshots (top-workflows.json, daily .json files) report zero tokens across all workflows for the current 7-day window. Recommendations below are derived from static prompt analysis and run history.

---

Analysis Period

Metric	Value
Period	2026-05-14 → 2026-05-27 (13 days)
Runs available	2 (2026-05-14, 2026-05-20)
All conclusions	✅ success
Avg run duration	~4 min
Token data	Not available (all snapshots at 0)

Token Profile

Metric	Value
Total tokens (last meaningful audit)	~3,079
Avg tokens/run (estimated)	~1,540
Avg turns/run	Not available
Cache efficiency	Not available
Prompt body size	1,201 words, 130 lines
Configured tools	bash: ["*"] (all bash tools)

---

Ranked Recommendations

1. Condense Overlapping Prompt Sections — Est. ~350 tokens/run

The ## What to Write section (106 lines) contains three partially overlapping subsections that cover the same exclusion logic from different angles:

• "Selection rubric for whether a change belongs in the release notes" (8 bullets) — include/exclude criteria
• "Exclude internal-only changes unless they materially affect users" (10 bullets) — overlaps heavily with the rubric
• "Release-note anti-patterns to avoid" (8 bullets) — several overlap with writing quality constraints

Action: Merge the selection rubric and exclusion list into one 6-bullet decision table. Move the anti-patterns into a two-line addendum at the end of the quality constraints.

Evidence: Both the rubric and the exclusion section explicitly list: CI/build/test changes, commit SHAs, contributor mentions, and dependency bumps as exclusions. This duplication adds ~30 lines that fire on every run.

Estimated savings: ~300–400 tokens/run from reduced prompt size.

---

2. Scope bash Tool List — Est. ~150–250 tokens/run

The workflow declares tools: bash: ["*"], which exposes documentation for every available bash sub-tool to the agent on each run. Based on the prompt and setup step, the agent only needs bash (for git, gh, and file operations).

Action: Replace bash: ["*"] with an explicit allowlist:

tools:
  bash:
    - bash

Evidence: The prompt contains no instructions referencing view, edit, create, grep, or glob tools — only bash commands. Both observed runs completed successfully with standard git/gh operations.

Estimated savings: ~150–250 tokens/run (fewer tool descriptions loaded into context).

---

3. Compress Semver-Awareness Section — Est. ~80 tokens/run

The semver guidance block spans 20+ lines to describe 4 tag patterns. These can be replaced by a 5-line table:

Semver guidance by tag type:
- major (vX.0.0): lead with breaking changes and migration steps
- minor (vX.Y.0): lead with new capabilities and behavior changes
- patch (vX.Y.Z): lead with fixes and reliability improvements
- 0.y.z: treat as pre-1.0; still name breaking changes explicitly

The verbose form adds no information beyond what the agent already knows from semver semantics.

---

Structural Optimization

Inline Sub-Agent: CHANGELOG Update

The ## What to Write section bundles two distinct tasks: (1) writing the release notes body, and (2) updating CHANGELOG.md and pushing the commit. The second task is purely mechanical — read the file, insert a formatted entry, commit, push.

Scoring:

Dimension	Score	Rationale
Independence	2/3	Needs final release body; can run once main agent confirms
Small-model adequacy	3/3	Pure formatting + git operations — no strategic synthesis
Parallelism	1/2	Sequential (needs notes first)
Size	2/2	Multi-step: read, write, commit, push
Total	8/10	Strong candidate

Why a smaller model fits: Updating CHANGELOG.md is template insertion — the format is fixed (Keep a Changelog), the content is already derived, and the only operation is a structured file edit followed by git commit && git push. No cross-referencing or synthesis required.

Proposed change: Replace the CHANGELOG update instructions in ## What to Write with an inline sub-agent block after the release body is finalized:

## agent: Update CHANGELOG

Insert a Keep a Changelog entry for `${RELEASE_TAG}` into `CHANGELOG.md`.
- Format: `## [x.y.z] - YYYY-MM-DD` above the previous version entry
- Sections: only `Added`, `Changed`, `Fixed`, `Removed`, `Security` — omit empty ones
- Content: the user-facing bullets from the finalized release body
- Commit: `docs: update changelog for ${RELEASE_TAG}` as `github-actions[bot]`
- Push: `git push origin HEAD:${RELEASE_TARGET}` (skip if RELEASE_TARGET is a commit SHA)

Estimated savings on main agent: ~200–300 tokens/run (removes 30+ lines of git/CHANGELOG instructions from main context); remainder charged to a smaller model.

---

Summary

Recommendation	Est. savings/run	Effort
Condense overlapping prompt sections	~350 tokens	Medium
Scope bash tool list	~200 tokens	Low
Compress semver section	~80 tokens	Low
CHANGELOG sub-agent	~250 tokens (main model)	Medium
Total	~880 tokens/run

---

Caveats

• Token data is unavailable across all snapshots; savings estimates are based on prompt line/word counts and typical tool-registration overhead.
• Only 2 runs were available for analysis; estimates cannot be validated statistically.
• Prompt condensation must preserve all behavioral requirements — review proposed merges manually before applying.
• The sub-agent recommendation assumes the gh-aw framework supports inline agent blocks; verify before implementing.
• Tool scope narrowing should be validated in a test run to confirm no required tool is removed.

References: §26163733067 · §25873066389

Generated by Agentic Workflow Token Usage Optimizer · ● 10.2M · ◷

• expires on Jun 3, 2026, 3:44 PM UTC