[agentic-token-optimizer] Issue Monster: cut per-run AIC by batching safe-output calls and trimming prompt #70278

Target workflow

Issue Monster (.github/workflows/issue-monster.md). It is the highest-AIC workflow in the 7-day window (118.7 AIC, ahead of Daily Go Test Parallelizer at 96.2) and was last optimized 2026-09-14, outside the 14-day exclusion.

Analysis period

Last 7 days, 7 runs, all success. Engine is Pi, model copilot/gpt-5.4. turns is not reported for these runs, so I use agent invocations from working_set instead. Only per-run summary data was audited. Per-run MCP/tool call traces are not in the pre-downloaded data. The aggregate mcp_tool_usage covers all workflows, so I could not attribute tool counts to Issue Monster.

Cost profile

Cost tracks invocation count. The 10-invocation run cost 23.98 AIC and 333K tokens. The 5-invocation runs cost 13.8–14.5 AIC and 113–123K tokens. Rebuild excess (re-sent context) is ~80% of input tokens, for example 168K of 207K in §39271700732. Every run makes 3 add_comment calls and 3 assign_to_agent calls. These are independent but appear to be issued across several turns.

Ranked recommendations

1. Batch the independent safe-output calls into a single response (~3–5 AIC/run)

Add this to the prompt (the Important note in step 5, or the issue-monster-token-budget skill):

After selecting issues, emit all assign_to_agent and add_comment calls together in one response. Do not call them one at a time or interleave them with reads.

Evidence: each run makes 6 safe-output calls for 3 issues, and invocations vary from 5 to 10 for identical work. The run-to-run variance (13.8 to 24.0 AIC) is mostly extra turns, each of which re-sends the context.

2. Make body-first validation stricter, so issue_read is the exception (~2–4 AIC/run)

Pre-activation already supplies 600-character body snippets for the top 8 candidates. Step 3 still permits issue_read for any "ambiguous" excerpt. Tighten it to: "Call issue_read only if a selected issue is not in the pre-fetched context. Never call it for the retry-blocked list. If an excerpt is ambiguous, skip the issue and pick the next candidate."

Evidence: the high-invocation runs (10 invocations, 333K tokens) look like extra read round-trips. I could not confirm this per run, because tool traces were not available.

3. Trim the prompt body (~1–2 AIC/run)

The prompt is ~280 lines and is re-sent on every invocation, so each cut is paid about 6 times per run.

• Delete the "Filtering Applied" and "Scoring System" bullet lists in step 1. This is the pre-activation job's logic. The agent only needs the sorted list. Replace with one line: "The list is already filtered and sorted by priority; use it as is."
• Collapse the "Topic Separation Examples" and the two integrity-blocked examples in step 3 into one short rule each.
• Remove the issue-monster-report-formatting skill. This workflow produces no report, only comments and noop, so the skill is unused context.
• Remove the Success Criteria block, which duplicates steps 2–5.

Estimated reduction: ~1.5K tokens per invocation.

4. Skip step 1a when no candidate has a task or plan label (~0.5–1 AIC/run)

Add a guard: "Only run step 1a if a selected issue has the task or plan label. Otherwise skip it." Without this, the agent may fetch parent and sibling issues unnecessarily.

5. Consider a smaller model (needs a trial; do not apply blindly)

The task is mostly extractive and classificatory: pick up to 3 topically distinct issues from a sorted list and call two tools per issue. Trial a smaller model (for example a mini-tier model) for 1–2 days and compare assignment quality using the existing issue_assigned eval. Potential saving is large but unmeasured.

Housekeeping

The single_issue_scoped eval asks whether "at most one issue was assigned per run", but the workflow assigns up to 3 (max: 3). Every run assigns 3. Fix the eval wording so it does not report false failures.

Structural optimizations

• Setup prefix extraction: not warranted. The sections do not repeat opening tool calls.
• Inline sub-agents: not recommended. The workflow is a short, sequential select-then-act flow and each section depends on the previous one. The existing skill: blocks are not sub-agents, and nothing is independent enough to justify an agent call.

Estimated savings

Recommendations 1–4 combined: ~6–10 AIC/run (~35–55% of the 16.96 average). The conservative target is ~8 AIC/run, bringing the average to roughly the 13.8–14.5 AIC floor of the best observed runs. This is based on the best-run floor, not a measurement.

Caveats

• 7 runs sampled. Per-run tool call traces were unavailable, so the causes in Add workflow: githubnext/agentics/weekly-research #2 and Add workflow: githubnext/agentics/weekly-research #4 are inferred from invocation counts and token variance.
• Some runs may legitimately need more reads when candidates are ambiguous.
• The recommendation to remove the report-formatting skill assumes no report is ever produced. Verify before deleting.

References: §39252836306, §39271700732, §39246657132

Note: issue-monster.md has no source: frontmatter entry in the portion I read, so it can be edited directly. If it is provenance-managed upstream, apply these changes upstream first.

Generated by Agentic Workflow AIC Usage Optimizer · copilot · auto · 24.8 AIC · ⊞ 11K · ◷

• expires on Oct 26, 2026, 7:01 AM UTC-08:00