[agentic-token-optimizer] Release workflow: prompt consolidation and CHANGELOG sub-agent opportunity

[github-actions]

Target Workflow

Release (release.md) — selected as the highest-impact AI workflow eligible for optimization (all data sources returned zero-token snapshots this cycle; selection based on available GitHub run history and prompt analysis).

Analysis Period

Metric	Value
Period analyzed	2026-05-14 → 2026-06-02
Runs audited	3
All conclusions	✅ success
Avg agent job duration	~113 s
Token data available	❌ (all-runs.json empty this cycle)

Token Profile

Token telemetry was unavailable this cycle (pre-aggregated files empty). Prompt-based estimation:

Metric	Estimate
Prompt word count	~1,200 words
Approx prompt tokens	~1,800 tokens/run
Configured tools	bash only
Inline sub-agents	None
Avg turns	Not measured (agent ran 74–156 s across 3 runs)

---

Ranked Recommendations

1. Consolidate redundant writing-guidance sections (~300–500 tokens saved/run)

Problem: The ## What to Write section contains seven overlapping sub-sections that each partially cover the same ground — "semver-aware", "user-facing", "Selection rubric", "Exclude internal-only changes", "Release-note anti-patterns", and "Writing quality constraints". The anti-patterns list (11 bullets) and writing-quality list (8 bullets) both rephrase the same core principle: translate implementation changes into user impact.

Action: Merge these into two concise blocks:

1. A 4–6 bullet principles block replacing "semver-aware" + "user-facing" + "Exclude internal-only" sections.
2. A single avoid sentence consolidating the anti-patterns and quality constraints.

Estimated savings: 200–300 words removed → ~300–450 tokens per run.

Evidence: The current prompt dedicates >650 words to writing-quality guidance alone. Three successful runs with agent durations of 74–156 s suggest the model does not need this level of hand-holding.

---

2. Extract CHANGELOG update as an inline sub-agent (~150–250 tokens saved/run)

Problem: The CHANGELOG.md update task is a distinct, extractive operation appended at the end of a long creative synthesis task. The main agent must carry the full Keep a Changelog format rules, section semantics, and git commit/push instructions throughout its context while also writing release notes.

Action: Add an ## agent: block for the CHANGELOG step with a small-model (haiku-class) sub-agent:

## agent: update-changelog
model: small
input:
  - /tmp/gh-aw/release-data/semver_context.json
  - /tmp/gh-aw/release-data/commit_subjects.tsv
  - the release body written by the main agent (passed as env var)

Update CHANGELOG.md: add/replace a Keep a Changelog entry for ${RELEASE_TAG} with sections Added/Changed/Fixed/Removed based on the release body. Commit with message "docs: update changelog for ${RELEASE_TAG}" and push to ${RELEASE_TARGET} unless it is a raw SHA.

Why a small model fits: The task is purely formatting-and-appending: read existing changelog, insert one structured block at the right position, commit. No cross-source synthesis required.

Score breakdown:

Dimension	Score
Independence (can run after release notes are final)	3/3
Small-model adequacy (template formatting + git)	3/3
Parallelism (sequential but off critical path)	1/2
Size (substantial enough to justify)	2/2
Total	9/10

Estimated savings: Removes ~100–150 words of CHANGELOG instructions + git commit details from the main agent context → ~150–250 tokens/run. Main agent also completes faster, reducing turn count.

---

3. Tighten the "Data Available" preamble (~50–100 tokens saved/run)

Problem: The ## Data Available section lists 6 files with full path descriptions. The agent already receives these paths from the setup step environment; the descriptions mostly repeat the filenames.

Action: Reduce to a single-line path list without descriptions:

## Context Files
/tmp/gh-aw/release-data/{current_release,previous_release,semver_context}.json  
/tmp/gh-aw/release-data/{commit_subjects.tsv,changed_files.txt,workflow_sources.txt}

Estimated savings: ~60–80 tokens/run.

---

Caveats

• Token telemetry was unavailable this cycle (all-runs.json and daily snapshots reported zero). Estimates are derived from prompt word count and average agent duration only.
• Only 3 runs were available for analysis; broader sampling could surface additional patterns.
• The Release workflow runs infrequently (3 runs across ~3 weeks), so per-run savings compound slowly. Improvements benefit correctness and prompt maintainability as much as cost.

References: §26780840710 · §26163733067 · §25873066389

Generated by Agentic Workflow Token Usage Optimizer · ● 7.9M · ◷

• expires on Jun 9, 2026, 3:51 PM UTC