Tracking: 2026-08-09 execution plan — token burn P0, merge-queue drain, Ruby cleanup (#309)

Start with the dependency plan and Step 0/Step 1 blockers, especially PR #280 and the Ruby-touching changes that precede #309. For the refactor scope, read bin/agent-workflow-seam-doctor, bin/push-downstream, and the listed test files, then run the relevant validation and RuboCop checks. Done means the staged Ruby cleanup and its verification are complete without the documented merge conflicts.

Execution-order tracking issue from the 2026-08-09 whole-surface triage (105 open issues, 33 open PRs). Priorities: (1) stop token burn ASAP, (2) Ruby cleanup ASAP, (3) merge near-done PRs first so the Ruby refactor doesn't create mass conflicts, (4) everything else after.

Dependency plan id: aw-triage-2026-08-09-tokens-ruby-v1 (coordinator-local artifact, 51 typed edges). Ordering below is the dependency-ordered worklist from that graph.

References are intentionally plain (#NNN) — nothing here should auto-close.

Step 0 — Stop the token bleeding (P0)

• Merge PR #396 (approved, green — queued ahead of #395 in the same file domain)
• Merge PR #362 (approved, green — same domain)
• Review + merge PR #395 (fixes #393). Highest-leverage token fix: one idle 15-min Goal monitor burned ~309M tokens (~5,256 credits) in 8 days, 97% cached-context reloads on "no change" wakes.
• Start #398 (replay-safe usage telemetry receipts) — critical path: #399 budgets, #400 measurement, and the #402 pilot all consume its receipts.
• #399 hierarchical token budgets (after #398 lands).
• One combined policy-guard PR after Domain A drains: #397 (no ad-hoc lanes without an issue/PR), #401 (coordination opt-in for single-target work), #403 (block cross-task lane takeover — 246M-token coordinator-wake trace), plus the accepted #402 decision (single canonical target as default batch shape). One PR, not four — they all edit the same contract files.
• #400 policy edit now (default coordinator route below Sol/xhigh; 67% of a 6.1B-token window was Sol/xhigh coordination); measured validation waits for #398. PR #348 conflicts and touches the same routing doc — rebase or fold in.

Step 1 — Drain the merge queue (before any Ruby churn)

"Green" = the 3 hosted checks only; the seam's independent-review gate still applies, and #387/#378/#384 are +3–5k lines.

Standalone, parallel-safe — land whenever

• PR #328 (+46/−12)
• PR #342
• PR #347 (new skill; unblocks #325/#326/#332)
• PR #346 (new skill; unblocks #325/#326/#354)
• PR #341
• PR #343 (hooks, self-contained)
• PR #373
• PR #383 (docs-only — merge early: it is the Ruby packaging plan #309 should follow)
• PR #301 (dependabot; .github/workflows is a human-review path — maintainer merge)

Domain A — batch-contract files (skills/pr-batch/SKILL.md, workflows/pr-processing.md, goal-completion-contract-test.rb; 15 PRs collide)

Serial order:

• PR #396 (also in Step 0)
• PR #362 (also in Step 0)
• PR #395 (also in Step 0)
• PR #388
• PR #339
• PR #378
• PR #387
• PR #322 (rebase first — conflicting)

Domain B — installer/bin/

• PR #360 (flake-cluster fix; also stabilizes validate for the Ruby wave)
• PR #336
• PR #321 (contributor branch — verify fork CI actually ran; green-looking rollups can hide never-run fork CI)
• PR #384 (+5k, real review; touches seam-doctor + push-downstream = Ruby hot files)

Domain C — closeout-evidence-replay

• PR #371
• PR #269 (rebase — conflicting)
• PR #291 (changes requested — address or close)

Domain D — pr-ci-readiness / merge-assurance / pr-merge-submit (Ruby hot files)

• PR #329
• PR #330
• PR #312
• PR #314 (rebase — conflicting)
• PR #288 (rebase — conflicting)

Decide, don't drift

• PR #289 (changes requested) — address or close
• PR #259 (oldest, conflicting) — address or close
• PR #280 — the elephant. 98 files, conflicting, restructures the entire bin/ Ruby surface into bin/agent_workflows_operation/*, overlapping ~90% of the #309 refactor territory. Merge-after-rework, split, or close before #309 starts.
• #355 — seam declares merge_queue_only but main has no merge queue, so agent merges are unreachable. Fix before any agent-driven merge wave (or defer if merging by hand).

Step 2 — Ruby cleanup (#309), staged

Preconditions, in order:

1. Ruby-touching PRs above merged or closed (#396, #360, #330, #329, #312, #384, #371/#269, #314, #288, #291, and the #280 decision).
2. Dev-loop fixes so refactor verification is trustworthy:
   • #363 (validate dirty-tree gate skips RuboCop — breaks the refactor loop)
   • Locale/encoding cluster: #331, #327, #352 (verify — 6239fd0 may have fixed part), #349, #351, #350
   • Flaky-test remnants: #338, #364, #361, #345, #260
3. Staged refactor, production helpers before tests, biggest offenders first (offense counts from the #309 baseline):
   • bin/agent-workflow-seam-doctor (172) then its test (214)
   • bin/push-downstream (114) then its test (160)
   • skills/post-merge-audit/bin/completed-batch-audit-receipt-test.rb (121)
   • skills/pr-batch/bin/pr-ci-readiness-test.rb (114)
   • Re-enable default RuboCop metrics cops incrementally; commit a machine-readable baseline first.
   • Align split boundaries with the #383 gem-extraction plan to avoid refactoring the same code twice.

Step 3 — Deferred backlog (after Ruby wave), batch by theme

• Merge-gate hardening: #249, #179, #257, #251, #266, #293 (PR #314), #367, #276/#268 (PR #343)
• Contract-test hardening: #340, #284, #357, #227, #225, #226, #244 leftovers
• #151 chain: #333 → #334 → #335, with PR #348 rebased in
• Skills/docs tail: #375, #389, #369, #394 (reserved: AbanoubGhadban), #358, #359; follow-ups #325/#326/#332/#354 unblock when #346/#347 merge
• Bigger designs: #392 (task-scoped review packets — also token-saving), #296, #319, #258, #189, #275, #248/#336, #368, #376, #353 → #186 (Batch id: line blocked by prompt-size slack)

Maintainer decisions needed

1. #280 disposition (merge-after-rework / split / close) — biggest conflict bomb; blocks #309 start.
2. #355 — enable a real merge queue vs. change merge_submission.mode.
3. #402 / #401 — single-canonical-target default and coordination-as-opt-in.
4. #400 — approve coordinator default-route drop now.
5. #289, #291, #259 — address or close.

Risks

• Mergeable states decay as the queue drains; re-check per merge. Expect #322/#348/#314/#288/#269/#259 to need rebases regardless.
• Contributor branches (#321, #280) can show green rollups when fork CI never ran — verify before merge-readiness.
• Hosted checks ≠ review: every non-trivial PR still needs the seam's independent review gate.

Triage provenance: single-operator mode (coordination_backend: n/a); phase-2 capacity split intentionally not generated — run $plan-pr-batch on a chosen slice for launch prompts.