Reduce scheduled workflow AI-credit burn and stagger the midnight batch

[Copilot]

The shared 1000-AIC pool was exhausted account-wide, causing scheduled agentic runs to hard-fail with 403 Maximum AI credits exceeded before doing useful work. The main burn driver was the Semantic Function Refactoring workflow’s expensive nightly sweep, compounded by several scheduled workflows firing in the same time window.

• Constrain the highest-burn workflow

  • Added explicit per-run and daily budgets to semantic-function-refactor.
  • Lowered runtime cost by switching from claude-opus-4-8 to claude-sonnet-4.6.
  • Reduced turn budget to keep the run bounded.

• Replace whole-repo analysis with a bounded deterministic slice

  • Added a pre-step that selects a small rotating package slice and writes the scoped file list to /tmp/gh-aw/agent/semantic-function-refactor/.
  • Updated the workflow prompt to treat that precomputed slice as the source of truth.
  • Explicitly forbids widening the scan or spawning subagent fan-out for duplicate sweeps.

• Stagger the affected schedule cluster

  • Moved the implicated scheduled workflows off the same burst window so they no longer contend for the shared AIC pool at once:
    • semantic-function-refactor
    • auto-triage-issues
    • detection-analysis-report
    • duplicate-code-detector
    • unbloat-docs

• Lock the behavior with a focused compile test

  • Added a workflow compilation test that verifies the semantic refactor workflow emits the intended schedule, budget guardrails, model selection, and bounded precompute step.

Example of the new bounded precompute pattern:

max-daily-ai-credits: 300
max-ai-credits: 300
max-turns: 30

steps:
  - name: Precompute semantic refactor slice
    run: |
      mkdir -p /tmp/gh-aw/agent/semantic-function-refactor
      # select a small rotating package slice
      # write targets.txt and go-files.txt for the agent

# Prompt contract:
# - only analyze /tmp/gh-aw/agent/semantic-function-refactor/go-files.txt
# - do not rescan all of pkg/
# - do not spawn subagents

• Fixes [aw-failures] [aw-fix] P0: Shared AI-credit budget exhausted (1006/1000 AIC) — 403 hard-fails agentic runs account-wide #46500

[Copilot]

Pull request overview

Reduces scheduled AI-credit usage by bounding the semantic-refactoring workflow and staggering related schedules.

Changes:

• Adds semantic-refactoring budgets, a cheaper model, and rotating scope precomputation.
• Staggers five scheduled workflows.
• Adds compilation regression coverage.

Show a summary per file

File	Description
.github/workflows/semantic-function-refactor.md	Adds cost and scope controls.
.github/workflows/semantic-function-refactor.lock.yml	Compiles semantic workflow changes.
.github/workflows/auto-triage-issues.md	Staggers triage schedule.
.github/workflows/auto-triage-issues.lock.yml	Compiles triage schedule.
.github/workflows/detection-analysis-report.md	Staggers analysis schedule.
.github/workflows/detection-analysis-report.lock.yml	Compiles analysis schedule.
.github/workflows/duplicate-code-detector.md	Staggers duplicate scan.
.github/workflows/duplicate-code-detector.lock.yml	Compiles duplicate-scan schedule.
.github/workflows/unbloat-docs.md	Staggers documentation workflow.
.github/workflows/unbloat-docs.lock.yml	Compiles documentation schedule.
pkg/workflow/semantic_function_refactor_workflow_test.go	Tests compiled cost guardrails.

Review details

Tip

Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

• Files reviewed: 11/11 changed files
• Comments generated: 4
• Review effort level: Medium