<!-- Schema: state/debt-backlog/*.yaml (YAML per entry); closure via git mv to archive/debt-backlog/<YYYY-MM>/. -->

Debt Triage — Backlog Review and Prioritization

Announce at start: "I'm using the coordinator:debt-triage skill to review the debt backlog."

An EM-PM conversation, not a dispatched agent: the EM reads the backlog and recommends. Trigger on demand, at >20 open items, or after a refactor that may have resolved several. Rationale, clustering detail, structural-probe calibration: wiki.

On a PowerShell host every CLI below takes its .exe launcher through the call operator (Shape W), never the ${...} POSIX form. Ladder: ${CLAUDE_PLUGIN_ROOT}/snippets/resolve-coordinator-bin.md.

Run backlog-grind-assemble brief debt-triage (per ${CLAUDE_PLUGIN_ROOT}/snippets/resolve-coordinator-bin.md) before Step 1. Over state/debt-backlog/, state/bug-backlog/ and state/improvement-queue/ it returns open items with severity breakdown, bug-backlog cross-reference (exact surface-field match), improvement-queue entries with clustering evidence, and a batched PM-gate. The improvement leg is emitted (Step 1); the debt-backlog steps (2, 3, 6, and 6b's debt-backlog governance) stay EM-performed.

Step 0: Surface prior rejections

Check tasks/out-of-scope/*.md (skip silently if absent). For any concept overlapping the triage, surface: "This is similar to tasks/out-of-scope/<concept>.md — we rejected this because [reason]. Still feel the same?" — confirm, reconsider (delete the file), or override.

Step 1: Read current state

Take the brief output as-is. Overlap beyond the surface-field match (file-path/description similarity) is an EM judgment pass over the same evidence, before presenting overlaps to the PM for a dedup decision (populate evidence: on both entries).

Improvement-queue triage is emitted, not EM classification. Pick an appetite (hunt, standard or sweep — values in coordinator/queue-profiles/improvement.yaml). Emit with emit-dispatch-workflow --queue state/improvement-queue --profile improvement --appetite <a> --out state/scratch/debt-triage/{run-id}/improvement.workflow.mjs --repo-root <abs repo root>, per ${CLAUDE_PLUGIN_ROOT}/snippets/resolve-coordinator-bin.md. There is no commit-readiness gate to resolve for this leg. Fire with Workflow({scriptPath}), never --fire — firing authorizes the in-run fixes and closes and the post-run hand-back (coordinator/docs/wiki/ceremony-calibration/queue-terminus-doctrine.md § Emitted-workflow triage). The run's triage is the only triage; the EM works Steps 2–4 on the debt backlog meanwhile. An emit refusal is reported, not routed around. The grind commits: a /bug-blitz in the same tree takes its Phase 0.7 suite baseline before this grind fires or after it finishes, never across it.

Step 2: Verify relevance (emitted)

Do not ask whether to fire — invoking this skill IS the request for the run this step names; it dissolves no gate this skill's own body names.

Debt-backlog rows only, emitted as in Step 1: emit-dispatch-workflow --queue state/debt-backlog --profile debt --appetite <a> --out state/scratch/debt-triage/{run-id}/debt.workflow.mjs --repo-root <abs repo root>, fired with Workflow({scriptPath}). Never hand-dispatch verifiers. Relevance and evidence rules live in coordinator/queue-profiles/debt.yaml (relevance_first, defer_requires_evidence). A row whose cited surface moved to another repo comes back as cross-repo: route it by cross-repo memo, then close it here once delivered. Code that moved is not code that was fixed. Hand-back types feed Step 5 as the improvement leg's do.

Step 3: Re-prioritize

Blocking other work → P0. In a D/F-graded system → P1. In a recently A/B-graded system → may deprioritize to P2. >30 days with no activity → flag for PM attention (skill policy over entry frontmatter dates; the PM-altitude call is the point).

Before grouping, query historical nature: tech-debt completions (query-completions --where "nature=tech-debt" --since "90d" --format json, ranked descending on frontmatter.loe.agent_dispatches as a number — the reader takes no --sort — per ${CLAUDE_PLUGIN_ROOT}/snippets/resolve-coordinator-bin.md): high-LoE areas in the last 90d indicate festering complexity — escalate open items there.

Never rank on loe.tshirt. As strings, descending order is XXL, XS, XL, S, M, L — the smallest tier ranks second. Rank on a numeric field. Zero-activity areas may reflect avoidance — flag: "We have carried this debt for N days without touching it — is that intentional?" Present a one-paragraph summary before Step 4; zero-row case: (no tech-debt completions logged in last 90d — hot-zone analysis unavailable).

Step 4: Group for execution

"System" is the brief's debt-backlog groups evidence: the first path cited in surface, else in title/body, cut to two directories. The row's system field is free text and often absent — never group on it.

Present ## Triage Results with four ### tables, each headed <label>: N items: Closed (no longer applicable) (ID, Reason); Recommended for immediate action (ID, System, Severity, Description, Effort); Can defer (ID, System, Severity, Reason to defer); Needs PM decision (YAGNI/scope) (ID, System, Description, Question).

Step 5: Present to PM

Ask for: (1) approval to close no-longer-applicable items; (2) YAGNI/scope calls; (3) prioritization of immediate-action items; (4) agreement on deferral reasoning; (5) goes to coordinator:apm, not the PM: the improvement leg's hand-back types (park, wont-do, yagni, unclear-direction, needs-judgment) are dispositioned by the APM (a code matter by coordinator:staff-eng), which writes pm_ruling: "<agent> (PM-delegated) ..." on the row. Step 6b consumes it; the PM sees the receipt. Only a pm_only matter (important, urgent, no clear answer, or irreversible) reaches the PM. Engine-originated hand-back types (budget-exhausted, verify-failed, the rest) are reported by count; their rows stay open, not presented here.

Step 6: Update backlog

After PM decisions:

1. Close resolved items — stamp status: closed, closed_at:, closed_by: <sha>, then mkdir -p archive/debt-backlog/<YYYY-MM> and git mv the entry in. Never rmdir state/debt-backlog/ even if it empties.
2. Update severity per PM direction.
3. Remove YAGNI items the same way as (1), closed_by referencing the PM decision.
4. For a load-bearing rejection (scope/doctrine conflict, cost-benefit, architectural veto — never a bug), write tasks/out-of-scope/<concept>.md, one file per concept (append "Prior requests" to an existing file rather than duplicating): # Out of scope: <concept>; **First raised:** YYYY-MM-DD; **Status:** Rejected (open to reconsideration); sections ## What was proposed, ## Why we rejected it, ## Prior requests (- YYYY-MM-DD: [how this came up]), ## What would change our minds.
5. Commit scoped, explicit-path: git commit -m "debt-triage: reviewed N items, closed M, N remain open" -- <every touched path>.

Step 6b: Consume the improvement leg's hand-back

Runs over both emitted legs' APM-dispositioned hand-back (Step 5 item 5), after the runs; a debt-leg cross-repo hand-back follows Step 2's memo-then-close route.

• baton: cluster per coordinator/docs/wiki/ceremony-calibration/queue-terminus-doctrine.md § Clustering, then mint solo or themed batons to coordinator/docs/wiki/baton-lifecycle/baton-authoring-bar.md's bar with triage's sizing evidence; no second gate. Stamp initiative on a themed baton and its member rows. Scaffold with coordinator-doc-new, set category to queue-derived-baton if left infra. Close each source row.
• route-to-learn-lessons: run coordinator-lesson-promote once per row with --title-file and --body-file (the row's title and body), --change-kind (the row's), --target-wiki unknown, and --evidence naming the archive destination path of the closing rename (or the row stem), never the pre-rename path, stale once archived. Then close the source row with closed_by set to the settling commit sha (the outbox path goes in the commit message, not closed_by).
  • Promotes only rows later runs hand back. The reconcile-343 plan promotes the existing central backlog once (C1 classification, C4 execution), so this promote gates on that classification's output or landed C4, never on the queue_scope: central tag (evidence, not disposition). Until then a route-to-learn-lessons hand-back stays open, untouched and unpromoted — no central-tagged row is promoted here — and Step 5's run summary reports it by count as "awaiting reconcile-343," never closed or presented to the PM.
  • A row whose change_kind the outbox enum refuses (exit 2) is left open and reported, never coerced.
• Park, won't-do and YAGNI keep their current rules and stamps, after Step 5.
• Source-row closure is an edit plus a plain rename to archive/improvement-queue/<YYYY-MM>/, the committer staging both paths (A-PLAIN-MV-IS-THE-INTENDED-ROUTE-NOT-A-FALLBACK). The row-removal --declared-revert follows coordinator-content-repo-47's /bug-blitz post-run wording.

Commit shape: batons, promotes and PM-gated closures are separate commits, each naming the source ids; the run committed its own fixes and closes. Every hand closure here is followed by backlog-grind-assemble grind-row sweep (coordinator/docs/wiki/ceremony-calibration/queue-terminus-doctrine.md § Emitted-workflow triage).

Skip this step entirely if no project-specific entries survived Step 5.

Autonomous runs

No PM is present. The APM's ruling is written as in Step 5 — pm_ruling: "apm (PM-delegated) ...", never as the PM's own word. For each row:

1. Act on a ruling only where this skill already permits EM-autonomous action (closing a verified-resolved row, Step 2's cross-repo memo-then-close, re-prioritizing).
2. A pm_only matter gets apm_recommendation on the row and stays open.
3. Continue; a pending recommendation never stops the run.

The report ends with ONE batched PM-confirmation list: each row with an unconfirmed apm_recommendation as row id, recommendation, one-line reason. The PM's confirmation writes pm_ruling (no (PM-delegated) label) for Step 6b.