Status

status · Project GPS — scan, integrity-check, derive, route

The framework's "where am I / what's next" oracle. It is a cross-cutting meta activity (named by function, not an SDLC seat — prefix-less, like 08-refactor), and per the master plan it is built and run last: it scans everything, so it is correct only once every other skill's shape is final. Give any project to /status and it answers three questions from the files on disk — never from memory:

1. Is the spine sound? — the integrity verdict (PASS, or FAIL naming the offending REQ-ID).
2. What's blocking? — the pending/deferred amendment + surviving [NEEDS CLARIFICATION] counts that 06-release gates on, surfaced early.
3. What do I run next? — the single exact next command, in the framework's vocabulary.

Its graded value is not "a nice summary" — a strong reader summarizes a repo too. It is the deterministic derived state: the integrity verdict + the governance counts + the routed command, computed from the files and emitted machine-readably into CLAUDE.md § Current State so the next seat — and 06's release gate, and the next /status — can act on it, not re-derive it.

The seat's two defining properties

• Sequential — no subagents. shared/subagent-protocol.md: subagents exist for context-isolated verification and a few bounded parallel reads. status is neither — it is a single derive-only read. There is nothing to relax in the harness <SUBAGENT-STOP> (the 04/06/08 precedent, not the 03/05/07 one). Nothing here spawns.
• Non-appender + non-mutator of truth — the honesty gate. status is absent from the amendment-protocol's appender list (00/02/03/08). It never edits the spine or a realization. It writes only generated views — CLAUDE.md § Current State and the AGENTS.md emission — both pure regenerable projections (shared/spine-boundary.md: a generated view is safe to overwrite because its source lives elsewhere). Its correctness proxy is the inverse of every other skill: after it runs, docs/spec/** and every realization are byte-identical. This is status's analog of 05's EXECUTED / 08's behavior-preserved — truth is unchanged.

Operating principle — read the files, derive the state, route the command, touch no truth

• Derive from disk, never from memory. Every claim — the phase, the integrity verdict, the counts, the route — traces to a file that exists (or is absent). If a file is unreadable, say UNKNOWN and name it; never guess, never edit a source to make the picture tidy.
• Report a break; never fix it. An integrity FAIL is surfaced with the offending ID and routed to repair (/00-discovery owns the spine write-path). status editing the spine to "fix" corruption would itself corrupt the declaration boundary — the one thing this seat must never do.
• Route, don't do. The next command names the owning skill; status never plans, designs, builds, reviews, ships, secures, or refactors. It is the GPS, not the driver. (A demonstrably wrong route — the router, not the project — is framework friction: FB entry via the feedback skill; shared/feedback-loop.md § Activation.)

The flow — one mode, one sequential pass, no gate

/status takes no arguments (one repo = one spine). It runs one pass; because it mutates no truth, there is no user-decision gate — it scans and reports. Craft lives in the references; load each as its step begins.

1. SCAN. Enumerate the artifact chain against shared/artifact-map.md — spec → planning (incl. docs/planning/patches/ + the backlog's ## Patches ledger) → design → architecture → src → quality → release → security → refactoring — recording present/absent per sprint. Read-only. Determine the current sprint (the highest sprint-NN with any artifact), the earliest missing phase within it, and any patch in flight (an open planned/in-progress Patches row).
2. INTEGRITY. Run the load-bearing checks → the PASS/FAIL verdict; run the advisory checks → the WARN list. A FAIL names the specific offending entry. (references/integrity-and-governance.md.)
3. GOVERN. Count amendment-log.json rows by disposition (pending · deferred · approved · auto-applied) and the surviving [NEEDS CLARIFICATION] markers in the spine. pending + deferred + markers are the release-blockers — 06-release blocks on any of them; surface the counts tied to that fact. When a ## Verify-live block is declared (WS6), also derive the verify-live coverage (verified / stale / missing) from the declared techs vs their docs/verification/<tech>.md records — 06 G11 blocks on any missing/uncited (this is L7's data, re-read for the human line). No block declared ⇒ omit the line (the on-demand pattern).
4. DERIVE + ROUTE. Apply the priority-ordered router (references/next-command.md): P0 no-spine → P1 integrity-halt → patch-in-flight → the patch's next seat → P2 governance-blocker override → P3 earliest-missing-phase in the current sprint → P4 advisories. First match wins. Produce the single next command.
5. REGENERATE VIEWS. Write the machine-readable derived state into CLAUDE.md § Current State (the emission the grader and 06 read) and re-emit AGENTS.md from the spine Constitution + summary. Both are generated views — never a source of truth. Then print the human report. (references/generated-views.md.)

The integrity verdict — load-bearing vs advisory (the spine write-path's rule #4)

The verdict is PASS / FAIL. Only load-bearing checks flip it — the spine's ID→file→block map is internally consistent and the amendment log is machine-readable (what every skill and the release gate depend on). Advisory findings WARN but never fail the verdict (they are realization drift, owned elsewhere). Full assertions in the reference; the split:

Load-bearing (FAIL the verdict, naming the offending entry): L1 every registry File resolves · L2 each registry REQ's leaf contains its delimited block (### REQ-NNN: … <!-- /REQ-NNN -->) · L3 no orphan blocks (every block has a registry row) · L4 no duplicate REQ-IDs · L5 amendment-log.json is valid JSON with an amendments array · L6 every REQ eval-block dataset: path resolves (agent-system; vacuously PASS with no eval block) · L7 verify-live records resolve (WS6: every ## Verify-live tech has a resolving, cited docs/verification/<tech>.md; no orphan record; no uncited claim; vacuously PASS when nothing is declared). The load-bearing set mirrors the emitted verify-spine.py FAIL checks — the two must never diverge on the integrity verdict.

Advisory (WARN only — realization drift, not spine corruption): A1 ID zero-padding/well-formedness · A2 ADR index (adr/README.md) ↔ ADR-NNN.md files in sync (03's) · A3 backlog ledger ↔ registry exactly-once (01's) · A4 registry Status (stated/derived) ↔ each block's <!-- source --> line (00's) · A5 an approved Tier-2's resolved_by resolves to an existing ADR · A6 patch-lane pressure — ≥3 consecutive done patches since the last planned sprint, or any escalated row → "this cadence is a sprint — run /01-planner plan-sprint N / consider /08-refactor assess" (advisory, never a block).

The next command — the priority-ordered router (first match wins)

The full table is references/next-command.md; the shape:

• P0 no docs/spec/specification.md → /00-discovery (no spine — start the chain).
• P1 integrity FAIL → halt normal routing; lead with the broken entry; repair routes to /00-discovery.
• Patch-in-flight (after P1, before P2) — an open ## Patches row routes to the patch's next seat: planned → /04-builder (the funnel) · in-progress + no qa-report-patch-NNN.md → a fresh /05-reviewer · qa FIX/BLOCK → /04-builder (fix pass) · qa SHIP + no release-report-patch-NNN.md → /06-release · escalated → /01-planner plan-sprint N / /00-discovery reflect (+ the A6 advisory). One patch in flight finishes before sprint routing resumes.
• P2 otherwise ship-ready and a governance blocker exists → route to resolve it (/00-discovery reflect for a pending/deferred amendment; /00-discovery for a surviving marker) instead of /06-release.
• P3 the normal chain — route to the earliest missing phase in the current sprint (this one rule subsumes forward progress and gap/backfill): no backlog → /01-planner · no design → /02-designer sprint N · no system.md → /03-architect init · no specs → /03-architect sprint N · no src → /04-builder sprint N · no qa → /05-reviewer sprint N · qa FIX/BLOCK → /04-builder sprint N · qa SHIP, no release → /06-release sprint N · no security → /07-security sprint N · sprint complete → /01-planner plan-sprint N+1 (or project complete → /00-discovery reflect).
• P4 advisories (surfaced, primary only when a sprint is otherwise complete) — refactor signals → /08-refactor assess; reflection overdue → /00-discovery reflect.

The machine-readable emission (what the grader and 06 read)

CLAUDE.md § Current State carries the derived state — derived status only (methodology is framework-level, never restated here). Exact shape in templates/current-state.md; the graded fields:

## Current State
<!-- GENERATED by /status — derived status only. Do not hand-edit; re-run /status. -->

- **Spine integrity:** PASS            <!-- or: FAIL — REQ-014: registry File `capabilities/x.md` does not resolve -->
- **Sprint:** 2 · **Phase:** quality (qa verdict: SHIP)
- **Amendments:** 0 pending · 0 deferred (2 approved · 1 auto-applied)   <!-- 06-release blocks on any pending/deferred -->
- **Open [NEEDS CLARIFICATION]:** 0
- **Verify-live:** 2 verified · 0 stale · 0 missing   <!-- WS6: ONLY when a `## Verify-live` block is declared; omit the whole line otherwise (the on-demand pattern) -->
- **Next command:** `/06-release sprint 2`

*Derived by /status from docs/spec/ + the artifact chain. Not a source of truth.*

Write-path (the read-only guarantee — follow exactly)

• Writes ONLY two generated views: CLAUDE.md § Current State (derived status) and AGENTS.md (re-emitted from the spine Constitution + summary — full re-emit each run: idempotent in-sync, self-healing on drift, per references/generated-views.md).
• Never writes docs/spec/** (the spine) or any realization (docs/planning/**, docs/design/**, docs/architecture/**, src/**, docs/quality|release|security|refactoring/**). An integrity break is reported, not repaired.
• Reference, never copy. The report and the emission cite REQ-NNN/ADR-NNN; they never paste requirement prose (shared/spine-boundary.md).
• The honesty invariant: after /status, git diff shows only CLAUDE.md and AGENTS.md changed — the spine and every realization are byte-identical. If you cannot honor that, you are doing something status must not do.

Progress checklist (copy this and track as you go)

• SCAN — artifact chain enumerated vs shared/artifact-map.md; current sprint + earliest-missing-phase determined (read-only)
• INTEGRITY — load-bearing checks → PASS/FAIL (FAIL names the offending REQ-ID); advisory checks → WARN list
• GOVERN — amendment dispositions counted (pending/deferred/approved/auto-applied) + surviving markers; pending+deferred+markers surfaced as the 06-release blockers; verify-live coverage (verified/stale/missing) derived iff a ## Verify-live block is declared (else the line is omitted)
• DERIVE + ROUTE — the priority router applied (P0 → P1 halt → P2 blocker-override → P3 earliest-missing → P4 advisory); one exact next command produced
• REGENERATE VIEWS — CLAUDE.md § Current State written (the machine emission: integrity · sprint/phase · counts · next command); AGENTS.md re-emitted from the spine (Constitution verbatim + recomputed summary)
• Human report printed (Spine · Governance · Sprint History matrix · Integrity · Advisories · Next command)
• Honesty: git diff shows only CLAUDE.md + AGENTS.md changed — docs/spec/** + every realization byte-identical (the read-only proxy); no source edited to "fix" a finding

Reads / Writes

Reads: docs/spec/specification.md + capabilities/** + design-intent.md + architecture-constraints.md + amendment-log.json · docs/planning/backlog.md (incl. the ## Patches ledger) + sprints/** + patches/** · docs/design/** · docs/architecture/system.md + adr/** + specs/** · src/** · docs/quality/** · docs/release/** · docs/security/** · docs/refactoring/** · .claude/rules/*-guardrails.md · git state. Writes (generated views ONLY): CLAUDE.md § Current State · AGENTS.md. Never docs/spec/** or any realization.

References (load when the step needs them)

• references/next-command.md — the full P0–P4 decision table + the routing craft + the Sprint History render.
• references/integrity-and-governance.md — the load-bearing (L) + advisory (A) assertion tables + the governance counts (dispositions + surviving markers) and how they tie to 06's block.
• references/generated-views.md — the CLAUDE.md § Current State + AGENTS.md re-emission contracts (read-only to truth; the generated-view-vs-maintained-artifact rule).
• shared/artifact-map.md — the cross-skill index status verifies reality against; repo-root-relative.
• shared/spec-amendment-protocol.md — the dispositions counted + the write-path rule #4 (/status integrity-checks the spine); repo-root-relative.
• shared/spine-boundary.md — generated view vs maintained artifact (why writing the two views is not mutating truth); repo-root-relative.
• shared/subagent-protocol.md — why status stays sequential (it is neither verification nor a parallel read); repo-root-relative.
• shared/live-source-verification.md — L7 (verify-live records) joins the load-bearing set (the parity contract); the conditional verify-live coverage line; repo-root-relative.

Next skill

status is informational — the report's Next command is what you run next. There is no downstream artifact and no gate; re-run /status anytime to re-orient.