Target Workflow
Release (workflows/release.md) — the only non-Token, agent-based workflow not yet ineligible after applying the 14-day exclusion rule to all pure-CI and maintenance workflows. Selected because it has a prompt-heavy structure with clear consolidation opportunities.
Note: All workflows in the optimization log were audited within the past 14 days. No pre-aggregated token data was available in the 7-day window (all snapshots contain zero completed agentic runs). The analysis below is based on prompt structure and GitHub Actions run-level timing.
Analysis Period & Runs
| Attribute |
Value |
| Period |
2026-05-14 → 2026-05-21 (7 days) |
| Runs audited |
2 |
| Conclusions |
2 × success (100%) |
| Errors / Warnings |
0 / 0 |
| Avg agent job duration |
~2 min 10 s |
| Estimated turns/run |
2–4 |
References: §26163733067 · §25873066389
Token Profile
Token-level telemetry was not available in the pre-downloaded log data. Estimates below are derived from prompt line counts and observed agent durations.
| Metric |
Estimate |
| Prompt body size |
~129 lines / ~2,100 tokens |
| "What to Write" section |
105 lines (81% of prompt) |
| Avg tokens/run (estimated) |
~3,500–5,000 |
| Cache efficiency |
Unknown (short runs suggest warm cache likely) |
| Runs/week |
~2 |
Ranked Recommendations
1 · Merge redundant writing-rule subsections — Est. –500 tokens/run
Problem: The ## What to Write section (105 lines) contains three subsections that heavily overlap:
Release-note anti-patterns to avoid (6 bullets)Writing quality constraints (5 bullets)Exclude internal-only changes unless they materially affect users (6 bullets)
All three convey the same core directive: focus on user-visible impact; omit internal housekeeping. The redundancy trains the model to re-process the same constraint in three passes, inflating input tokens with no added precision.
Action: Collapse all three into a single ### Writing rules block of 5–6 bullets, removing ~35 lines (~560 tokens).
Example collapsed block:
### Writing rules
- Focus on user-visible impact (behavior, capability, compatibility, cost). Omit CI, build, test, and internal housekeeping.
- Prefer one synthesized bullet over several low-value commit-level bullets.
- Start each bullet with the outcome; add condition/scope after.
- Avoid vague filler ("various fixes"), hype, and contributor mentions.
- Do not repeat the same fact across Highlights and Upgrade Notes.
Evidence: Lines 91–129 of release.md (anti-patterns + quality + exclude sections) are functionally equivalent to the above 5 bullets.
2 · Condense semver guidance — Est. –150 tokens/run
Problem: The semver block (~15 lines) gives separate prose guidance for major, minor, patch, and 0.y.z releases. For a changelog-generation task driven by a concrete diff, the distinction between minor and patch rarely changes the agent's behavior beyond tone — the diff is the primary signal.
Action: Reduce to a 4-line rule:
Infer scope from the tag (major/minor/patch) and the diff. Call out breaking changes for major; capabilities for minor; fixes for patch. For 0.y.z, still name any breakage plainly.
Saves ~11 lines (~175 tokens).
3 · Compress the selection rubric — Est. –100 tokens/run
Problem: The include/exclude selection rubric (6 bullets, ~10 lines) is wordy and partially duplicates the writing rules.
Action: Reduce to 2–3 tight bullets:
Include: anything a user notices by using, configuring, upgrading, or debugging the workflow.
Omit: CI, build, test, refactoring, internal plumbing, and generated-file updates.
Collapse related low-level commits into one user-facing outcome.
Saves ~7 lines (~110 tokens).
Structural analysis: inline sub-agent candidates
The Release prompt has 3 major sections (Data Available, What to Write, Output Requirements) with no existing ## agent: blocks. Sub-agent scoring:
| Section |
Independence |
Small-model |
Parallelism |
Size |
Total |
| Read & summarize diff files |
3 |
2 |
2 |
1 |
8 |
| Generate release notes |
1 |
1 |
0 |
2 |
4 |
| Update CHANGELOG.md |
2 |
2 |
1 |
1 |
6 |
Diff summarization (score 8): Reading commit_subjects.tsv and changed_files.txt and producing a structured impact summary is extractive and classification-based — strong fit for a smaller model. However, the workflow runs only ~2×/week, making the overhead of an agent call less justifiable. Not recommended at current frequency.
CHANGELOG update (score 6): Sequential on the notes output; cannot run concurrently. Keep in main agent.
Verdict: No sub-agent additions recommended for the Release workflow at this time. Revisit if run frequency increases.
Caveats
- Only 2 runs were available for analysis; behavior on large diffs (many commits, many changed files) may differ.
- Token telemetry was unavailable — savings estimates are derived from prompt line-count proxies (~16 tokens/line average).
- The Release workflow is manually triggered and low-frequency; absolute weekly savings are modest. The primary benefit of these changes is prompt clarity and faster future iterations.
Generated by Agentic Workflow Token Usage Optimizer · ● 11.2M · ◷
Target Workflow
Release (
workflows/release.md) — the only non-Token, agent-based workflow not yet ineligible after applying the 14-day exclusion rule to all pure-CI and maintenance workflows. Selected because it has a prompt-heavy structure with clear consolidation opportunities.Analysis Period & Runs
References: §26163733067 · §25873066389
Token Profile
Token-level telemetry was not available in the pre-downloaded log data. Estimates below are derived from prompt line counts and observed agent durations.
Ranked Recommendations
1 · Merge redundant writing-rule subsections — Est. –500 tokens/run
Problem: The
## What to Writesection (105 lines) contains three subsections that heavily overlap:Release-note anti-patterns to avoid(6 bullets)Writing quality constraints(5 bullets)Exclude internal-only changes unless they materially affect users(6 bullets)All three convey the same core directive: focus on user-visible impact; omit internal housekeeping. The redundancy trains the model to re-process the same constraint in three passes, inflating input tokens with no added precision.
Action: Collapse all three into a single
### Writing rulesblock of 5–6 bullets, removing ~35 lines (~560 tokens).Example collapsed block:
Evidence: Lines 91–129 of
release.md(anti-patterns + quality + exclude sections) are functionally equivalent to the above 5 bullets.2 · Condense semver guidance — Est. –150 tokens/run
Problem: The semver block (~15 lines) gives separate prose guidance for major, minor, patch, and
0.y.zreleases. For a changelog-generation task driven by a concrete diff, the distinction between minor and patch rarely changes the agent's behavior beyond tone — the diff is the primary signal.Action: Reduce to a 4-line rule:
Saves ~11 lines (~175 tokens).
3 · Compress the selection rubric — Est. –100 tokens/run
Problem: The include/exclude selection rubric (6 bullets, ~10 lines) is wordy and partially duplicates the writing rules.
Action: Reduce to 2–3 tight bullets:
Saves ~7 lines (~110 tokens).
Structural analysis: inline sub-agent candidates
The Release prompt has 3 major sections (Data Available, What to Write, Output Requirements) with no existing
## agent:blocks. Sub-agent scoring:Diff summarization (score 8): Reading
commit_subjects.tsvandchanged_files.txtand producing a structured impact summary is extractive and classification-based — strong fit for a smaller model. However, the workflow runs only ~2×/week, making the overhead of an agent call less justifiable. Not recommended at current frequency.CHANGELOG update (score 6): Sequential on the notes output; cannot run concurrently. Keep in main agent.
Verdict: No sub-agent additions recommended for the Release workflow at this time. Revisit if run frequency increases.
Caveats