π― Target Workflow
Agentic Workflow AI Credit Spend Optimizer (.github/workflows/agentic-token-optimizer.lock.yml)
Selected as the only workflow in the repository with no prior optimization history and the longest average agent runtime (~8.4 min/run). All other candidates were optimized between 2026-05-28 and 2026-06-05 (within the 14-day exclusion window). Pre-aggregated token data (top-workflows.json, all-runs.json) returned empty results for the current 7-day window; workflow selection and evidence derive directly from the GitHub Actions API.
π
Analysis Period & Runs Audited
- Period: 2026-05-20 β 2026-06-08 (19 days)
- Runs analyzed: 14 (all weekday runs since workflow creation)
- Conclusions: 14 Γ
success (100% reliability β no reliability waste)
- **Actual AI-credit (redacted) unavailable (log pipeline returned empty); metrics derived from job-wall-clock timing as proxy
π° Spend Profile
| Metric |
Value |
| Runs analyzed |
14 |
| Avg run duration |
8.4 min |
| Min / Max duration |
6.2 min / 12.7 min |
| Duration std-dev |
Β±1.9 min (2Γ spread between best and worst) |
| Estimated Actions minutes (14 runs) |
~117 min |
| Cadence |
5 runs/week (weekdays) |
| Avg AI credits / run |
N/A (log pipeline empty β see caveats) |
| Long-run outliers (>avg+1Ο) |
2 runs: 12.7 min, 12.3 min |
The 2Γ spread between short (6.2 min) and long (12.7 min) runs indicates the optimizer performs significantly more work when pre-aggregated data is absent (common scenario) and falls back to sequential GitHub API exploration.
Per-run timing table (14 runs)
π§ Tool Configuration (from source)
| Tool |
Config |
Assessment |
bash |
all patterns |
Keep β primary tool for file reads and gh api calls |
github |
mode: gh-proxy, toolsets: [issues] |
Keep β used to list existing issues for context |
repo-memory |
branch memory/token-audit, max 102KB/file |
Keep β reads daily snapshots and writes optimization log |
No tools are candidates for removal: all three are used in successful runs.
π Ranked Recommendations
1. Condense the "Data Access Guidelines" section (~8β12% savings)
Evidence: The ## Data Access Guidelines section spans ~40 lines and contains three separate code blocks demonstrating the same gh api --jq pattern (extract-only, list runs, pipe-with-filter, plus a negative example). The AI agent reads this meta-instruction on every turn. A tighter 12-line version conveying the same rules would reduce per-turn prompt token consumption.
Action: Replace the multi-block example section with a single consolidated block:
# Always filter with --jq or | jq; never load unfiltered responses
gh api "repos/$REPO/contents/.github/workflows/my.md" --jq '.content' | base64 -d
gh api "repos/$REPO/actions/workflows/ID/runs?per_page=10" \
--jq '.workflow_runs[] | {id, name, conclusion, run_started_at}'
# Chain multi-step reads in one bash block; prefer --jq for simple filtersRemove the # β block (negative example) as it adds tokens without changing behavior.
Estimated savings: ~25 lines removed from 319-line prompt β 8% reduction β ~5β8% credit savings per run.
2. Add inline sub-agent for Phase 3+4 β workflow source analysis (~15β20% savings)
Evidence: Phases 3 (Read Workflow Source) and 4 (Structural Optimization Checks) are fully independent from Phase 2 (runtime behavior analysis) after Phase 1 selects the target workflow name. Both phases perform extractive work β reading a markdown file, listing sections, checking for ## agent: blocks, and scoring sub-agent opportunities β tasks well-suited to a smaller model.
Sub-agent score:
| Dimension |
Score / Max |
Rationale |
| Independence |
2 / 3 |
Needs only workflow name from Phase 1; otherwise standalone |
| Small-model adequacy |
3 / 3 |
Extractive: read markdown, count lines per section, check patterns |
| Parallelism |
2 / 2 |
Can run concurrently with Phase 2 behavioral analysis |
| Size |
2 / 2 |
Reading a 200β320 line file + structured scoring |
| Total |
9 / 10 |
Strong candidate |
Action: Add the following block between Phase 1 and Phase 2 in agentic-token-optimizer.md:
## agent: source-analyzer
You receive `workflow_path` (e.g. `.github/workflows/foo.md`) and `repo` variables.
1. Read the file: `gh api "repos/$repo/contents/$workflow_path" --jq '.content' | base64 -d`
2. Return a JSON object with:
- `sections`: list of all `##`-level headings in order
- `has_subagents`: true if any `## agent:` block exists
- `configured_tools`: list of tools in YAML frontmatter
- `lines_per_section`: map of heading β line count
- `setup_prefix_repeated`: any shell snippet appearing verbatim in 2+ section openings
3. Output only the JSON object β no prose.
In the main prompt, replace Phases 3+4 with: "Use the source-analyzer sub-agent output for structural analysis; do not re-read the workflow source."
Estimated savings: 2β3 main-agent turns delegated to a smaller model β ~15β20% credit reduction for source-analysis phases.
3. Add a no-data fast-path at Phase 1 (~5β8% savings on empty-data runs)
Evidence: All 14 analyzed runs show the log pipeline returning empty data. Issue bodies confirm the agent probes top-workflows.json, all-runs.json, and daily snapshot files sequentially before pivoting to GitHub API β adding 3β5 turns of empty reads per run.
Action: Add a single conditional gate at the top of Phase 1:
### Data Availability Check
Run: `jq '.runs | length' /tmp/gh-aw/token-audit/all-runs.json`
- If > 0: proceed with pre-aggregated data as described below.
- If == 0: skip reading `top-workflows.json` and daily snapshots entirely;
go directly to GitHub API target selection.
This replaces 3+ sequential empty file reads with one check.
Estimated savings: ~3 turns eliminated on empty-data days β ~5β8% per run; proportionally larger on 12+ min outlier runs.
4. Collapse Phase 4 scoring guidance into the table (~5% savings)
Evidence: Phase 4 contains a 4-dimension scoring table plus two prose subsections ("Scoring guidance" with threshold rules + "Smaller models are a good fit for" bulleted list) totaling ~35 lines. These are consulted every run even when no sub-agent analysis is triggered.
Action: Embed threshold column into the scoring table; remove standalone prose blocks:
| Dimension |
Meaning |
Max |
Good fit when β₯ |
| Independence |
Runs without other sections' output |
3 |
2 |
| Small-model adequacy |
Extractive / classificatory / formatting |
3 |
2 |
| Parallelism |
Can run concurrently |
2 |
1 |
| Size |
Substantial enough to justify agent call |
2 |
1 |
Total β₯ 6 β recommend; 4β5 β moderate; < 4 β keep in main agent.
Estimated savings: ~25 lines removed β ~5% credit savings per run.
ποΈ Structural Optimization
Inline Sub-Agent: source-analyzer (Phases 3+4)
As detailed in Recommendation #2. This is the highest-confidence structural change:
- Task: Read target workflow
.md source; return structured JSON (sections, tool config, verbosity, setup-prefix patterns).
- Why smaller model fits: Fully extractive β read one markdown file, identify headings, check for string patterns, count lines. No strategic synthesis or cross-referencing required.
- Score: 9/10 (Independence 2, Small-model 3, Parallelism 2, Size 2).
- Main prompt change: Replace
## Phase 3 β Read Workflow Source + ## Phase 4 β Structural Optimization Checks with a sub-agent call block; consume returned JSON artifact in Phase 5 synthesis.
- No existing sub-agents β this would be the first
## agent: block added to this workflow.
β οΈ Caveats
- No AI-credit data available. The
gh aw logs pipeline returned empty results for all 14 analyzed runs. Savings estimates are proportional reductions inferred from wall-clock durations, not measured AIC values.
- Duration proxy limitations. Wall-clock time includes ~1β2 min of container startup and artifact upload overhead. Agent computation time is shorter; savings figures are conservative.
- 14-run sample. Two long-run outliers (12+ min) drove the Β±1.9 min std-dev. Recommendations 1, 3, and 4 disproportionately benefit those outlier runs.
- Reliability is strong (14/14 success). No reliability-driven optimizations are warranted.
References: Β§27023961142 Β· Β§26646567699 Β· Β§26407979066
Generated by Agentic Workflow AI Credit Spend Optimizer Β· β 14.4M Β· β·
π― Target Workflow
Agentic Workflow AI Credit Spend Optimizer (
.github/workflows/agentic-token-optimizer.lock.yml)Selected as the only workflow in the repository with no prior optimization history and the longest average agent runtime (~8.4 min/run). All other candidates were optimized between 2026-05-28 and 2026-06-05 (within the 14-day exclusion window). Pre-aggregated token data (
top-workflows.json,all-runs.json) returned empty results for the current 7-day window; workflow selection and evidence derive directly from the GitHub Actions API.π Analysis Period & Runs Audited
success(100% reliability β no reliability waste)π° Spend Profile
The 2Γ spread between short (6.2 min) and long (12.7 min) runs indicates the optimizer performs significantly more work when pre-aggregated data is absent (common scenario) and falls back to sequential GitHub API exploration.
Per-run timing table (14 runs)
π§ Tool Configuration (from source)
bashgh apicallsgithubmode: gh-proxy, toolsets:[issues]repo-memorymemory/token-audit, max 102KB/fileNo tools are candidates for removal: all three are used in successful runs.
π Ranked Recommendations
1. Condense the "Data Access Guidelines" section (~8β12% savings)
Evidence: The
## Data Access Guidelinessection spans ~40 lines and contains three separate code blocks demonstrating the samegh api --jqpattern (extract-only, list runs, pipe-with-filter, plus a negative example). The AI agent reads this meta-instruction on every turn. A tighter 12-line version conveying the same rules would reduce per-turn prompt token consumption.Action: Replace the multi-block example section with a single consolidated block:
Remove the
# βblock (negative example) as it adds tokens without changing behavior.Estimated savings: ~25 lines removed from 319-line prompt β 8% reduction β ~5β8% credit savings per run.
2. Add inline sub-agent for Phase 3+4 β workflow source analysis (~15β20% savings)
Evidence: Phases 3 (Read Workflow Source) and 4 (Structural Optimization Checks) are fully independent from Phase 2 (runtime behavior analysis) after Phase 1 selects the target workflow name. Both phases perform extractive work β reading a markdown file, listing sections, checking for
## agent:blocks, and scoring sub-agent opportunities β tasks well-suited to a smaller model.Sub-agent score:
Action: Add the following block between Phase 1 and Phase 2 in
agentic-token-optimizer.md:In the main prompt, replace Phases 3+4 with: "Use the
source-analyzersub-agent output for structural analysis; do not re-read the workflow source."Estimated savings: 2β3 main-agent turns delegated to a smaller model β ~15β20% credit reduction for source-analysis phases.
3. Add a no-data fast-path at Phase 1 (~5β8% savings on empty-data runs)
Evidence: All 14 analyzed runs show the log pipeline returning empty data. Issue bodies confirm the agent probes
top-workflows.json,all-runs.json, and daily snapshot files sequentially before pivoting to GitHub API β adding 3β5 turns of empty reads per run.Action: Add a single conditional gate at the top of Phase 1:
This replaces 3+ sequential empty file reads with one check.
Estimated savings: ~3 turns eliminated on empty-data days β ~5β8% per run; proportionally larger on 12+ min outlier runs.
4. Collapse Phase 4 scoring guidance into the table (~5% savings)
Evidence: Phase 4 contains a 4-dimension scoring table plus two prose subsections ("Scoring guidance" with threshold rules + "Smaller models are a good fit for" bulleted list) totaling ~35 lines. These are consulted every run even when no sub-agent analysis is triggered.
Action: Embed threshold column into the scoring table; remove standalone prose blocks:
Total β₯ 6 β recommend; 4β5 β moderate; < 4 β keep in main agent.
Estimated savings: ~25 lines removed β ~5% credit savings per run.
ποΈ Structural Optimization
Inline Sub-Agent:
source-analyzer(Phases 3+4)As detailed in Recommendation #2. This is the highest-confidence structural change:
.mdsource; return structured JSON (sections, tool config, verbosity, setup-prefix patterns).## Phase 3 β Read Workflow Source+## Phase 4 β Structural Optimization Checkswith a sub-agent call block; consume returned JSON artifact in Phase 5 synthesis.## agent:block added to this workflow.gh aw logspipeline returned empty results for all 14 analyzed runs. Savings estimates are proportional reductions inferred from wall-clock durations, not measured AIC values.References: Β§27023961142 Β· Β§26646567699 Β· Β§26407979066