[agentic-token-optimizer] Daily Agentic Workflow AIC Usage Audit — Optimization Recommendations

[github-actions]

🎯 Target Workflow

Daily Agentic Workflow AIC Usage Audit (agentic-token-audit) — selected as highest-AIC eligible workflow. All workflows were optimized within 14 days; this one has the longest elapsed time since review (2026-06-23, 10 days ago) and the second-highest 7-day AIC spend.

📊 Spend Profile (2026-06-27 → 2026-07-03, 5 runs)

Metric	Value
Total AIC	1,052.33
Avg AIC / run	210.47
Total tokens (1 instrumented run)	653,717
Avg turns / run (instrumented)	~16
Total action minutes	43 min
Error rate	0 / 5 runs

Token coverage is limited: 4 of 5 runs reported null token usage. AIC figures are complete; per-token analysis is based on run §28662299837.

Observability insight (medium severity — execution drift): the workflow varied from 0 to 17 turns across runs, flagged as "changing task shape or unstable prompts."

---

🏆 Ranked Recommendations

1. Trim RunData Schema Table · 5–8 AIC/run

Phase 1 documents 14 RunData fields; the processing script uses only 6 (workflow_name, aic, token_usage, turns, error_count, warning_count, url). The 8 unused fields (effective_tokens, duration, branch, display_title, head_sha, logs_path, classification, event) add ~180 prompt tokens per run.

Action: Replace the 14-field table with a 6-field version. Remove the deprecation note on effective_tokens.

---

2. Merge Phases 1 + 2 Into One Execution Block · 10–15 AIC/run

Phase 2 (persist snapshot + update rolling-summary) contains only 2 file operations that are trivially appended to the Phase 1 Python script. The current structure forces the agent to:

1. Write + run process_audit.py
2. Read Phase 2 instructions
3. Copy audit_snapshot.json and update rolling-summary.json (separate turns)

Action: Extend the Phase 1 script spec to also write YYYY-MM-DD.json to repo-memory and update rolling-summary.json. Remove Phase 2 as a separate section.

Evidence: ~2 turns saved per run; at 16 turns/run that is ~12% reduction.

---

3. Batch Both Charts Into One Python Script · 8–12 AIC/run

Phase 3 describes two charts sequentially, causing the agent to write and run two separate Python scripts and make two upload_asset calls across ~4 turns before it can build the issue.

Action: Replace with a single instruction to write gen_charts.py that creates both PNGs in one execution, then make both upload_asset calls before Phase 4.

Evidence: Current sequential description and repeated PYTHONPATH reminder suggest 2 extra scripting turns per run.

---

4. Consolidate PYTHONPATH Instructions · 3–5 AIC/run

The PYTHONPATH prefix is mentioned 3 times: once as a bullet in "Chart requirements," once as a verbatim example command, and implicitly required in Phase 1. This causes agents to re-read and re-derive the same path across turns.

Action: Add one setup note at the top of Phase 1: export PYTHONPATH=/tmp/gh-aw/token-audit/site-packages${PYTHONPATH:+:$PYTHONPATH}. Remove the repeated mentions from Phase 3.

---

5. Compress Phase 4 Report Template · 5–8 AIC/run

The Phase 4 template is ~50 lines of verbatim markdown with explicit placeholder strings loaded into context on every Phase 4 turn.

Action: Replace with a compact requirements list (5 bullet points) naming each required section and its key fields. Remove explicit placeholder names — instruct the agent to embed chart URLs inline as they are returned.

Evidence: Current template is ~600 prompt tokens. A requirements list reduces this to ~150 tokens.

---

🔧 Structural Optimization — Sub-Agent Candidates

The workflow has no existing ## agent: blocks and 4 major prompt sections.

Phase 1 — Python Data Processing · Score 9/10 (Strong)

Dimension	Score	Rationale
Independence	3	Only needs pre-downloaded workflow-logs.json
Small-model adequacy	2	Extractive Python aggregation with defined output schema
Parallelism	2	First phase, starts immediately
Size	2	30+ line script spec

Why a smaller model fits: group-by, aggregate, sort, write JSON — fully specified in/out. A Haiku-class model handles this class of task reliably.

Proposed invocation (add to prompt, replacing Phase 1 prose):

## agent: process-logs
model: haiku
task: |
  Read /tmp/gh-aw/token-audit/workflow-logs.json. Filter status=="completed".
  Group by workflow_name; compute run_count, total_ai_credits, avg_ai_credits,
  total_tokens, avg_tokens, total_turns, avg_turns, total_action_minutes,
  error_count, warning_count. Sort desc by total_ai_credits.
  Write audit_snapshot.json (schema: {date, period_days, overall, workflows}).
  Copy to /tmp/gh-aw/repo-memory/default/YYYY-MM-DD.json.
  Update rolling-summary.json (last 90 entries).
  Treat null aic/token_usage as 0.

Estimated additional savings: 15–20 AIC/run (smaller model for merged Phases 1+2).

Phase 3 — Chart Generation · Score 7/10 (Moderate)

Dimension	Score
Independence	1 (needs Phase 1 output)
Small-model adequacy	3 (mechanical matplotlib)
Parallelism	1
Size	2

Recommend implementing Recommendation 3 (batched script) first; revisit sub-agent extraction if turns remain high.

---

💡 Summary of Expected Savings

Recommendation	Est. AIC/run saved
1. Trim schema table	5–8
2. Merge Phases 1+2	10–15
3. Batch charts	8–12
4. Consolidate PYTHONPATH	3–5
5. Compress template	5–8
Total (without sub-agent)	31–48
+ Phase 1 sub-agent	+15–20
Total (with sub-agent)	46–68

Conservative estimate: ~40 AIC/run saved (19% reduction from 210 avg AIC/run).

---

⚠️ Caveats

• Turn data covers only 1 of 5 runs; savings estimates assume it is representative.
• "Execution drift" insight (0–17 turns) likely reflects metadata gaps in 4 runs rather than genuine zero-turn executions.
• Sub-agent invocation requires ## agent: block support in the runtime.

Run-level evidence

Run	Date	AIC	Turns	Tokens	API Calls
§28376941750	2026-06-29	188.90	—	—	10
§28446558922	2026-06-30	229.14	—	—	10
§28520469613	2026-07-01	198.38	—	—	10
§28592118147	2026-07-02	212.66	—	—	10
§28662299837	2026-07-03	223.25	17	653,717	10

References: §28662299837 · §28592118147 · §28376941750

Generated by Agentic Workflow AIC Usage Optimizer · 282.8 AIC · ⊞ 21.6K · ◷

• expires on Jul 10, 2026, 3:16 PM UTC