[agentic-token-optimizer] Daily Agentic Workflow AIC Usage Audit — optimization opportunities

[github-actions]

Target workflow: Daily Agentic Workflow AIC Usage Audit (agentic-token-audit.md)
Reason selected: Active daily workflow; current display name has no "Token" suffix; not logged under its current name in the last 14 days. Token footers on recent issues show 5.7M–7.1M tokens per run despite processing zero-value data every time.

Analysis period: 2026-05-28 → 2026-06-09 · 10 runs audited

---

📊 Spend Profile

Metric	Value
Runs audited	10
Run conclusions	10/10 success
Retries	1 (2026-06-08, producing 3 duplicate issues in one day)
Avg run duration	~5.6 min (range 4.5–7.5 min)
Token footprint (from issue footer)	5.7M–7.1M tokens/run
Useful data produced	0 — every run reports zero AIC / zero tokens
upload_assets job	Skipped in 8 of 10 runs; ran in 2 with empty-data charts

AIC billing figures from audit snapshots are all 0 because gh aw logs --start-date -1d returns an empty .runs array on every run — a persistent data-pipeline gap that means each run burns 5–7 M context tokens for zero useful output.

---

🏆 Ranked Recommendations

#1 — Add explicit early-exit after Phase 1 when runs is empty

Estimated savings: 20–30% AIC/run

Evidence: 10 of 10 audited runs detected len(runs) == 0. Despite having nothing to report, every run executes the full four-phase prompt:

• Phase 1: writes and runs a Python script (correct)
• Phase 2: copies an empty snapshot file (wasted)
• Phase 3: attempts chart generation (skipped 8×, or generates empty chart 2×)
• Phase 4: fills the entire report template with N/A placeholders

Action: After Phase 1, add a guard:

If `len(runs) == 0`, immediately publish a compact issue:
  title: "Daily AIC Usage Audit — {YYYY-MM-DD}"
  body: "No agentic workflow runs were captured in the last 24 hours (`gh aw logs` returned an empty `.runs` array). No snapshot or charts generated."
Then stop — skip Phases 2, 3, and 4.

Skipping Phase 2–4 on the empty-data path eliminates the full template-filling exercise, chart-generation attempts, and snapshot copy on every run.

---

#2 — Consolidate Phases 1 and 2 into a single Python script

Estimated savings: 5–8% AIC/run

Evidence: Phase 2 is described as a standalone prompt section, but its entire content is:

Read the snapshot from audit_snapshot.json. Copy it to /tmp/gh-aw/repo-memory/default/YYYY-MM-DD.json. Also maintain rolling-summary.json.

This is three lines of Python that can be the final block of the Phase 1 script, removing the need for the agent to process a separate phase instruction, write another bash call, and report completion.

Action: Merge Phase 2 into the Phase 1 Python script. Replace the Phase 2 section header and body with a single sentence:

At the end of the Phase 1 script, also write `audit_snapshot.json` to
`/tmp/gh-aw/repo-memory/default/{today}.json` and update `rolling-summary.json`.

---

#3 — Fix the inconsistent issue title format

Estimated savings: minimal AIC, high reliability impact

Evidence: 5 most recent issues use 4 different title patterns:

Issue	Title
#112	Daily AIC Usage Audit — 2026-06-09
#108	Daily AI Credit Spend Audit — 2026-06-08
#107	Daily AI Credit Spend Audit — 2026-06-08
#106	Daily AI Credit Spend Audit — 2026-06-08
#96	Daily Token Usage Audit — 2026-06-05

The agent re-invents the title every run. This makes issue search and de-duplication unreliable.

Action: Add an explicit instruction in Phase 4:

Use this exact title: "Daily AIC Usage Audit — {YYYY-MM-DD}"

---

#4 — Fix incorrect period reporting

Estimated savings: reliability

Evidence: Issue #107 (run §27136881236) reported Period: last 30 days (2026-05-09 to 2026-06-08) instead of last 24 hours. The period calculation is ambiguous and the agent uses whatever window it derives from the log file.

Action: Harden the period instruction in Phase 4:

Always write: **Period**: last 24 hours (YYYY-MM-DD to YYYY-MM-DD)
Use today's UTC date for the end and yesterday's UTC date for the start.
Do not derive the period from the log data range.

---

#5 — Simplify the PYTHONPATH setup instruction in Phase 3

Estimated savings: 2–4% AIC/run

Evidence: Phase 3 currently requires the PYTHONPATH variable to be set on each individual Python command:

Set PYTHONPATH=/tmp/gh-aw/token-audit/site-packages${PYTHONPATH:+:$PYTHONPATH} for every Python command that imports pandas, matplotlib, or seaborn

This pattern appears as an inline reminder repeated in the chart-generation instructions, adding token overhead and occasionally being misapplied (one run set the wrong PYTHONPATH for the first command).

Action: Replace the repeated reminder with a single export at the top of Phase 3:

Before any chart Python command, run once:
export PYTHONPATH=/tmp/gh-aw/token-audit/site-packages${PYTHONPATH:+:$PYTHONPATH}

---

⚠️ Root-Cause Note (out of scope for prompt optimization)

Every run reports zero data because gh aw logs --start-date -1d consistently returns an empty .runs array. Until this data-pipeline gap is fixed, all prompt optimizations above reduce wasted work on an empty payload — they do not restore the audit's core function. Investigating why the log download always returns no runs (rate limiting? collection window mismatch? pipeline misconfiguration?) is a prerequisite for the workflow to be useful.

References:

• §27209129613 — 2026-06-09, 7.5 min, 7.1M tokens, 0 runs captured
• §27136881236 — 2026-06-08, attempt 2, wrong period reported
• §26954882996 — 2026-06-04, 4.7 min, 0 runs captured

---

🔍 Caveats

• AIC billing figures are unavailable (pre-aggregated all-runs.json and daily snapshots all show 0); savings estimates are based on observed run token footprints (5.7M–7.1M/run from issue footers) and prompt section size.
• All 10 audited runs produce zero-value reports; behavior with actual run data is unobserved.
• The data-pipeline fix is outside prompt scope and may change the effective savings once data flows correctly.

Generated by Agentic Workflow AIC Usage Optimizer · ● 9.5M · ◷

• expires on Jun 16, 2026, 3:39 PM UTC