Tracker Phase 0 — harness first + live-defect fixes

[dean0x]

Initial Request

Devflow's command→agent seam for issue handling is untyped and currently broken in five user-visible ways, and the test harness that Phases 1–3 will lean on does not yet exist in a form that survives an agent rename. /plan #42 parses an issue reference and never fetches it; /debug #42 passes a key (ISSUE:) that the fetch-issue op does not declare; three command-layer sites still invoke or describe gh issue view directly; resolve.mds:378 states a D9 thread-resolution rule that contradicts git.md:96's single authority; and release.md:138 promises a post-release "close milestone" step that does not exist. Phase 0 fixes those defects and lands the harness — the shared dist-preferred agent-source resolver, the two-sided seam test, the D11 guard-class union, the two goldens, and test:integration in CI — against the fixed corpus, so that a red guard in any later phase is unambiguous evidence of a regression rather than of a pre-existing defect.

Phase label (honest, verbatim): "no change to the GitHub rendering of any existing traceability artifact."

This label is deliberately narrower than "no behaviour change". Phase 0 is a fix, not a refactor, precisely because it does change five user-visible behaviours. Claiming otherwise would be the untruthful-traceability-claim defect this phase exists to delete (intent #22).

The five user-visible changes (enumerated with before/after in the issue body and CHANGELOG.md ### Fixed, AC-0.15):

1. /plan #N now actually fetches the issue (it previously consumed a body it never fetched).
2. /debug's issue path works for the first time (ISSUE: → ISSUE_INPUT:).
3. dynamic-plan issue bodies are now wrapped <untrusted-issue-body> at the sink.
4. resolution-summary.md's Tracked = (pending) now names a reason.
5. release.md's untruthful "close milestone" claim is deleted.

Execution: SEQUENTIAL_CODE_AGENTS · context-risk: HIGH · Design artifact: .devflow/docs/design/tracker-phase-0.2026-09-05_1155.md (authoritative; local, gitignored — a 60k-truncated copy is posted as a collapsed comment).

Product Requirements

Depends on

none — first phase (branch from main). Tracking issue: #321.

Acceptance criteria

• AC-0.1 — seam test red on debug.mds before, green after; covers Git and Code fences both directions; excludes Produces/Requires; ≥1 fence scanned per agent type
• AC-0.2 — both goldens exist, captured after A1; npm run test:golden:update documented; CI never auto-regenerates
• AC-0.3 — /plan #42 → OPERATION: fetch-issue; /plan #12 #15 #18 → OPERATION: fetch-issues-batch with ≤50 pinned, TRUNCATED ({n} not processed) beyond it, ## Issues Batch ({n} issues) asserted
• AC-0.4 — no gh issue invocation or descriptive mention in dist/commands/*.md outside a Git spawn fence; the two gh pr view sites recorded as exceptions
• AC-0.5 — resolve.mds D9 matches git.md:96 verbatim, pinned in source and dist/commands/resolve.md
• AC-0.6 — tracker-unavailable manage-debt emits DEGRADED and resolution-summary.md keeps Tracked = (pending) with the reason named at all four sites; every REQUIRED_OP doing remote I/O has a D4 line
• AC-0.7 — every agent-corpus scan asserts registry completeness through the shared resolver; no new guard contains a literal agent path
• AC-0.8 — D11 guards scan the sink class with postingOps.length >= 8 from git.md alone and fail on a known-bad sample
• AC-0.9 — github-status-lines.txt byte-identical before and after Phase 0
• AC-0.10 — every op Output block rendering a remote-sourced field carries <untrusted-issue-body> (non-vacuous, ≥3 ops); no reply/summary template interpolates a remote body field
• AC-0.11 (amended) — fetch-issues-batch declares ISSUE_REFS at git.md:69 and git.md:292; plan.mds:32 renamed; fetch-issue keeps ISSUE_INPUT; debug.mds passes ISSUE_INPUT; both ops in REQUIRED_OPS; fetch-issues-batch out of INTERNAL_OPS with a rationale comment; no ISSUE_NUMBERS literal remains repo-wide, and ISSUE_NUMBER (singular) is untouched at all 14 sites
• AC-0.12 — Guard 2 pins manage-debt: 60000 before the literal moves
• AC-0.13 — test-inventory AC: each new guard named and proven non-vacuous with a red case (process-only — PR checklist) §0.10 / D-B
• AC-0.14 — release.md:138 "close milestone" deleted
• AC-0.15 — the phase label, issue body and CHANGELOG enumerate the five user-visible changes with before/after (process-only — PR checklist) §0.10 / D-B
• AC-0.16 — every dist-scanning guard has a missing-directory case asserting a thrown error with a build hint
• AC-0.17 — all existing suites green; no threshold lowered anywhere ... now mechanised [DR-27a]: tests/guards/numeric-floor-manifest.test.ts compares every pinned numeric floor in tests/** against a committed manifest and fails when any entry decreases (new entries allowed, decreases never). If the manifest guard is not landed in P0-S22, AC-0.17 falls back to (process-only — PR checklist) per D-B
• Prefix-shippability — After phase N: (i) npm run build && npm test && npm run test:integration && npm pack --dry-run are green; (ii) a fresh install from the tarball drives /plan → /implement → /code-review → /resolve → /release on a GitHub project with no new prompt and no new file in git status; (iii) every artifact in the tree has at least one reachable consumer in this phase's end-state — no file, define, list element, frontmatter key or conditional arm exists solely for a later phase (ADR-003); (iv) the frozen github-status-lines.txt fixture is byte-identical to its Phase-0 capture.
• Commit order — Phase 0's commit order A1 → A2 → A3 → A4 is itself an acceptance criterion. A2 (the golden capture) must contain nothing but the two fixtures, and must be captured from the post-A1 tree. A1 edits git.md; capturing first pins a tree Phase 0 immediately moves past (GAP-06 / B-9). Verified by reviewing the commit list on the PR: exactly four subtask commits in that order, A2's diff limited to tests/fixtures/golden/**.

Subtasks

#	Domain	Owned files (exclusive, with anchors)	Steps	Depends on	Gate green before next	Handoff contents
A1	prompt assets (commands, agents, partials) + hooks	src/assets/commands/resolve.mds (:352,:378,:499,:508,:539) · debug.mds (:17,:24,:46,:51) · plan.mds (:27,:32, new Git fence inserted at #### Phase 1: Gate 0 before :63) · dynamic-plan.mds:103 · dynamic-build.mds:73 · _partials/_wave.mds (:4,:9) · implement.mds:73 · release.md:138 · src/assets/agents/git.md (:69,:268-284,:288-300,:292,:294-298,:300-318,:379-411,:723,:922) · src/assets/scripts/hooks/ensure-devflow-init:23	P0-S1 (prep) · P0-S2 … P0-S13	—	G0.1	post-A1 git.md SHA-256 + line count; the list of every line A1 changed in git.md (so A2's capture is spot-checkable); confirmation no test file was touched
A2	tests (fixtures only)	new tests/fixtures/golden/git-agent.md · new tests/fixtures/golden/github-status-lines.txt	P0-S14, P0-S15	A1 committed on the branch	G0.2	the two fixture paths + byte counts + the explicit statement "these fixtures are immutable for the remainder of Phase 0–3; a mismatch means the source is wrong, never the fixture"
A3	tests + CI	tests/helpers.ts · new tests/seams/command-agent-input.test.ts · new tests/goldens/git-agent-golden.test.ts · new tests/goldens/github-status-lines.test.ts · new tests/guards/agent-source-resolver.test.ts · new tests/guards/extended-references.test.ts · new tests/guards/retired-wording.test.ts · new tests/guards/numeric-floor-manifest.test.ts + its committed manifest [DR-27a] · tests/git-agent.test.ts · tests/registry-integrity.test.ts · tests/agent-name-guards.test.ts · tests/agent-frontmatter.test.ts · tests/skill-references.test.ts · tests/installer-new.test.ts · tests/build-mds.test.ts · tests/shell-hooks.test.ts (gitignore fixture only) · tests/init-logic.test.ts (gitignore fixture only) · .github/workflows/ci.yml:23 · package.json (test:golden:update)	P0-S16 … P0-S25	A1, A2	G0.3	every new guard by name, its file, and its proven-red mechanism; the five user-visible behaviour changes with before/after strings; the two recorded exceptions (release.md:85 conventions read; gh pr view at code-review.mds:76-78 / bug-analysis.mds:43-45); the mode ('union' / 'sole') each extractOpSectionFromCorpus consumer names [DR-18]
A4	docs	new docs/reference/platform-assumptions.md · .devflow/features/dynamic-workflow-engine/KNOWLEDGE.md · CHANGELOG.md · the seven drifted citations in .devflow/features/*/KNOWLEDGE.md · the Phase-0 issue body	P0-S26	A1–A3 (sweep the branch's final tree, not the commit — PF-025)	G0.4	— (terminal)

Gates

Gate	Must be green before	Command / assertion
G0.1	A2 starts	npm run build && npm test — full suite green at the post-A1 tree. No new test exists yet; this proves A1 broke nothing.
G0.2	A3 starts	tests/fixtures/golden/git-agent.md byte-equals src/assets/agents/git.md (verify with cmp, not by reading); github-status-lines.txt non-empty and every extracted literal present in the source at the cited line.
G0.3	A4 starts	npm run build && npm test && npm run test:integration green; each new guard demonstrated non-vacuous with a recorded red case (AC-0.13); postingOps.length >= 8 satisfied by git.md alone (AC-0.8); resolveAllAgents() ⊇ getAllAgentNames() = 16 (AC-0.7).
G0.4	PR opens	Prefix-shippability sequence (§14) + github-status-lines.txt byte-identical to its A2 capture (AC-0.9) + git status --porcelain empty.

/implement spawn notes

• Spawn one Code agent per subtask, in order: A1 → A2 → A3 (optionally A3a → A3b) → A4. Never two concurrently.
• CREATE_PR: false on A1, A2, and A3 (and A3a when the split is taken); CREATE_PR: true on A4, the terminal agent, which opens the PR with the complete test-inventory table and the RED proofs from all subtasks (implement.mds:409). A4 must therefore receive every prior handoff.
• HANDOFF_REQUIRED: true on every subtask but the last (A1, A2, A3 — and A3a when the split is taken). A4 is terminal and produces no handoff.
• Every Code agent reads .devflow/docs/handoff-{branch_slug}.md via HANDOFF_FILE; each writes its handoff contents (column 7 above) after its own gate passes, never before.
• Commit with an explicit pathspec of the subtask's owned files only: git commit -- <owned paths> (PF-020's validated cheap mitigation; per-commit provenance survives the squash-merge).
• Standing instruction in every Code prompt: "a golden mismatch means the source is wrong, never the fixture." (H2.)

Execution strategy

SEQUENTIAL_CODE_AGENTS — 4 subtasks (A1 → A2 → A3 → A4). Context risk: HIGH.

Rationale — three independent forcings (any one sufficient)

1. Commit order is itself an acceptance criterion (§F.5-I). A1 edits git.md; capturing the golden first pins a tree Phase 0 immediately moves past (GAP-06). A single Code agent has no mechanism that guarantees commit ordering inside one run; sequential agents make each ordinal a spawn boundary.
2. A2 must touch nothing else. The golden capture is a byte-exact 59,376-character copy plus a hand-extracted frozen fixture. Interleaving it with any behaviour change reproduces the exact defect GAP-06 exists to prevent, and the "never regenerate alongside a behaviour change" rule (B-05) is unenforceable inside one agent's context.
3. Working set. ~30 files / ~1,400–1,900 changed lines spanning four unrelated domains (prompt text, fixtures, test harness, docs). A3 alone requires simultaneous knowledge of 12 test files' internal conventions.

Why PARALLEL_CODE_AGENTS is rejected (feature-wide, PF-020 closure rule)

PF-020's 2026-08-30 extension binds — "disjointness has to hold over each agent's READ/EXECUTE closure, not merely its WRITE set." Three facts collapse every candidate partition:

1. npm run build is rm -rf dist && … (package.json:20). Every Code agent's verification step deletes and rebuilds the one shared dist/. A sibling agent mid-suite observes a missing dist tree — and Phase 0 deliberately makes every dist-scanning guard fail loud on a missing directory (AC-0.16, R3). Parallel agents therefore manufacture nondeterministic red suites that look exactly like real regressions.
2. /implement has no worktree isolation — grep for worktree in implement.mds returns zero hits. All Code agents share one worktree, one git index, one dist/, one node_modules, one vitest run.
3. The test corpus spans every partition. The seam test, the D11 sink-class guard, the containment test and the frozen fixture all read git.md ∪ dist/commands/*.md ∪ dist/skills/git/references/** — the union of every proposed subtask's outputs. Disjoint write sets, 100% overlapping read closure.

This disqualifier is independent of write-set disjointness, so it holds for all four phases. It is a tooling limitation, not a plan limitation [DR-23, ALREADY-ADDRESSED].

PR title

fix(traceability): repair issue-seam defects and land the tracker test harness

Implementation Plan

Design artifact (authoritative): .devflow/docs/design/tracker-phase-0.2026-09-05_1155.md — local and gitignored. A copy truncated at the 60k comment cap is posted as a collapsed comment: #322 (comment). Run with /implement .devflow/docs/design/tracker-phase-0.2026-09-05_1155.md. Tracking issue: #321.

Test Inventory (AC-0.13)

Each new guard landed in Phase 0, named with its file and proven non-vacuous via a known-bad inline sample (mechanic 2, per H10).

Guard	File	What it guards	Non-vacuity red case (mechanic 2)
Command→agent seam: forward check	tests/seams/command-agent-input.test.ts	Every KEY: passed in a caller fence is declared in the op's **Input:** — fixes debug.mds ISSUE: / ISSUE_INPUT: mismatch (AC-0.1)	Inline corpus with OPERATION: fetch-issue\nISSUE: 42\n → exactly 1 violation naming ISSUE
Command→agent seam: reverse check	tests/seams/command-agent-input.test.ts	Every required **Input:** value declared by an op is passed by at least one caller fence	Same inline corpus
Command→agent seam: third direction (producer)	tests/seams/command-agent-input.test.ts	Every issue_capture_contract() value has a caller-side producer	Same inline corpus
resolveAllAgents() ⊇ getAllAgentNames()	tests/guards/agent-source-resolver.test.ts	Agent roster completeness — 16 agents (AC-0.7)	Would fail if an agent is registered in plugins.ts but has no source file
resolveAgentSource dist-preferred	tests/guards/agent-source-resolver.test.ts	When dist/agents/ exists, resolver prefers it over src/assets/agents/	Synthetic dist tree with a different content → dist version returned
resolveAgentSource build hint	tests/guards/agent-source-resolver.test.ts	Throws with `/Run \`npm run build\ `` when neither src nor dist file exists	resolveAgentSource('_nonexistent_agent_') → throws containing the build hint
extractOpSectionFromCorpus 'sole' mode	tests/guards/agent-source-resolver.test.ts	Throws on duplicate anchor across corpus files (DR-18)	Two-file corpus with duplicate ## Operation: X heading → throws naming both file paths
extractOpSectionFromCorpus 'union' mode	tests/guards/agent-source-resolver.test.ts	Returns all matching sections with a match count (DR-18)	Same two-file corpus → matchCount = 2, not 1
git-agent golden	tests/goldens/git-agent-golden.test.ts	tests/fixtures/golden/git-agent.md is byte-identical to src/assets/agents/git.md	Fixture absent → throws with regeneration command hint
github-status-lines frozen fixture	tests/goldens/github-status-lines.test.ts	Frozen fixture byte-identical; test:golden:update refuses to overwrite without --unfreeze	node scripts/update-golden.js github-status-lines → non-zero exit + §0.2 rule text printed
Guard 1: REQUIRED_OPS coverage (17 ops)	tests/git-agent.test.ts	All 17 operations exist in git.md including fetch-issue and fetch-issues-batch (AC-0.11)	Removing either new op from REQUIRED_OPS → test fails naming the missing op
Guard 2: manage-debt: 60000 bound	tests/git-agent.test.ts	manage-debt section contains the 60000-char archive threshold before Phase 2 moves it (AC-0.12)	Removing the literal → test fails with "missing 60000-char archive threshold"
INTERNAL_OPS: fetch-issues-batch removed	tests/registry-integrity.test.ts	INTERNAL_OPS allowlist is ['learn-conventions'] only — fetch-issues-batch removed with a rationale comment (AC-0.11, SG-11)	Re-adding fetch-issues-batch without updating the test → reverse-check finds it unlisted in compiled commands
DIST_FILES non-vacuity	tests/build-mds.test.ts §20	DIST_FILES.length === 14 (13 compiled hosts + release.md, SG-13)	Would fail if dist/commands/ is absent (throw) or if release.md is removed from the constant
compliance_gate() adoption guard	tests/build-mds.test.ts §20	≥6 compiled hosts import the {compliance_gate()} partial (P0-S22)	hostsScanned >= 6 — fails if all imports are removed
gh-issue scope guard	tests/build-mds.test.ts §21	No gh issue invocation or descriptive mention in any DIST_FILE outside a Git spawn fence (AC-0.4)	Inline corpus OPERATION: fetch-issue\ngh issue view 42\n → violation detected (not inside spawn fence)
Retired wording: ISSUE_NUMBERS	tests/guards/retired-wording.test.ts	ISSUE_NUMBERS (renamed to ISSUE_REFS in Phase 0) is absent from all Phase-0 output files (GAP-32, AC-0.17)	Synthetic file seeded with ISSUE_NUMBERS → wouldFlag === true
Numeric floor manifest	tests/guards/numeric-floor-manifest.test.ts	No pinned numeric floor in tests/** decreases vs the committed manifest tests/fixtures/numeric-floors.json (AC-0.17, DR-27a)	Copy of manifest with one entry decremented → allPatternsExist assertion fails
Extended references resolver	tests/guards/extended-references.test.ts	Every ## Extended References row in any skill resolves to a real file on disk	Synthetic skill with a nonexistent reference path → violations array non-empty

[dean0x]

Design artifact: tracker-phase-0.2026-09-05_1155.md (excerpt truncated at the 60k comment cap — the local file is authoritative)

---

type: design-artifact
version: 1
status: APPROVED
issue: 322
title: "Tracker Phase 0 — harness first + live-defect fixes"
slug: tracker-phase-0
created: 2026-09-05T11:55:00Z
execution-strategy: SEQUENTIAL_CODE_AGENTS
context-risk: HIGH

Tracking issue: #321 · Depends on: none (first phase) · Next phase: Phase 1 (#323)

This artifact is self-contained by construction. .devflow/docs/ is gitignored, so a pointer to another artifact resolves to "not committed; ask the author" (PF-037 / H9). Every cross-phase rule this phase depends on is inlined verbatim in §14 Normative Appendix. Do not follow links for normative content; there are none.

---

1. Problem Statement

Devflow's command→agent seam for issue handling is untyped and currently broken in five user-visible ways, and the test harness that Phases 1–3 will lean on does not yet exist in a form that survives an agent rename. /plan #42 parses an issue reference and never fetches it; /debug #42 passes a key (ISSUE:) that the fetch-issue op does not declare; three command-layer sites still invoke or describe gh issue view directly; resolve.mds:378 states a D9 thread-resolution rule that contradicts git.md:96's single authority; and release.md:138 promises a post-release "close milestone" step that does not exist. Phase 0 fixes those defects and lands the harness — the shared dist-preferred agent-source resolver, the two-sided seam test, the D11 guard-class union, the two goldens, and test:integration in CI — against the fixed corpus, so that a red guard in any later phase is unambiguous evidence of a regression rather than of a pre-existing defect.

Phase label (honest, verbatim): "no change to the GitHub rendering of any existing traceability artifact."

This label is deliberately narrower than "no behaviour change". Phase 0 is a fix, not a refactor, precisely because it does change five user-visible behaviours. Claiming otherwise would be the untruthful-traceability-claim defect this phase exists to delete (intent #22).

The five user-visible changes (enumerated with before/after in the issue body and CHANGELOG.md ### Fixed, AC-0.15):

1. /plan #N now actually fetches the issue (it previously consumed a body it never fetched).
2. /debug's issue path works for the first time (ISSUE: → ISSUE_INPUT:).
3. dynamic-plan issue bodies are now wrapped <untrusted-issue-body> at the sink.
4. resolution-summary.md's Tracked = (pending) now names a reason.
5. release.md's untruthful "close milestone" claim is deleted.

---

2. Acceptance Criteria

Verbatim from §F.3, with the pinning test for each. (process-only — PR checklist) marks an AC that has no mechanical test and must not be given a fake one (§0.10 / Gate-2 decision D-B); a test that asserts a document's shape rather than the system's behaviour is the vacuous guard PF-018 names.

AC	Text (verbatim §F)	Pinning test	Status
AC-0.1	seam test red on debug.mds before, green after; covers Git and Code fences both directions; excludes Produces/Requires; ≥1 fence scanned per agent type	tests/seams/command-agent-input.test.ts (3 directions + per-type scanned > 0); RED proof vs debug.mds:51 recorded in the PR	covered
AC-0.2	both goldens exist, captured after A1; npm run test:golden:update documented; CI never auto-regenerates	tests/goldens/*.test.ts + loadGolden throw-on-absent	covered
AC-0.3	/plan #42 → OPERATION: fetch-issue; /plan #12 #15 #18 → OPERATION: fetch-issues-batch with ≤50 pinned, TRUNCATED ({n} not processed) beyond it, ## Issues Batch ({n} issues) asserted	build-mds.test.ts (plan.md contains both ops); git-agent.test.ts Guard 2 ≤50 + TRUNCATED + ## Issues Batch ({n} issues)	covered
AC-0.4	no gh issue invocation or descriptive mention in dist/commands/*.md outside a Git spawn fence; the two gh pr view sites recorded as exceptions	new gh issue scope guard over DIST_FILES with the exception list	covered
AC-0.5	resolve.mds D9 matches git.md:96 verbatim, pinned in source and dist/commands/resolve.md	git-agent.test.ts:167-168 (existing) + new caller guard on source and dist	covered
AC-0.6	tracker-unavailable manage-debt emits DEGRADED and resolution-summary.md keeps Tracked = (pending) with the reason named at all four sites; every REQUIRED_OP doing remote I/O has a D4 line	git-agent.test.ts new manage-debt D4 it; four-site (pending) guard; github-status-lines.txt	covered
AC-0.7	every agent-corpus scan asserts registry completeness through the shared resolver; no new guard contains a literal agent path	tests/guards/agent-source-resolver.test.ts + "no literal agent path in any new guard" grep	covered
AC-0.8	D11 guards scan the sink class with postingOps.length >= 8 from git.md alone and fail on a known-bad sample	git-agent.test.ts:359-425 repointed to the union, >= 8 unchanged, known-bad sample	covered
AC-0.9	github-status-lines.txt byte-identical before and after Phase 0	frozen-fixture equality	covered
AC-0.10	every op Output block rendering a remote-sourced field carries <untrusted-issue-body> (non-vacuous, ≥3 ops); no reply/summary template interpolates a remote body field	containment forward guard (≥3 ops) + reply/summary negative	covered
AC-0.11 (amended)	fetch-issues-batch declares ISSUE_REFS at git.md:69 and git.md:292; plan.mds:32 renamed; fetch-issue keeps ISSUE_INPUT; debug.mds passes ISSUE_INPUT; both ops in REQUIRED_OPS; fetch-issues-batch out of INTERNAL_OPS with a rationale comment; no ISSUE_NUMBERS literal remains repo-wide, and ISSUE_NUMBER (singular) is untouched at all 14 sites	REQUIRED_OPS +2; INTERNAL_OPS −1 with rationale; repo-wide ISSUE_NUMBERS = 0; ISSUE_NUMBER count unchanged	covered
AC-0.12	Guard 2 pins manage-debt: 60000 before the literal moves	git-agent.test.ts Guard 2 manage-debt: 60000	covered
AC-0.13	test-inventory AC: each new guard named and proven non-vacuous with a red case	the test-inventory table in the issue body	(process-only — PR checklist) §0.10 / D-B
AC-0.14	release.md:138 "close milestone" deleted	release.md in DIST_FILES; 'close milestone' absent	covered
AC-0.15	the phase label, issue body and CHANGELOG enumerate the five user-visible changes with before/after	CHANGELOG ### Fixed ×5 with before/after; phase label	(process-only — PR checklist) §0.10 / D-B
AC-0.16	every dist-scanning guard has a missing-directory case asserting a thrown error with a build hint	missing-directory case per dist-scanning guard	covered
AC-0.17	all existing suites green; no threshold lowered anywhere	full suite green; threshold-diff review: no numeric literal in tests/ decreased in the PR diff — now mechanised [DR-27a]: tests/guards/numeric-floor-manifest.test.ts compares every pinned numeric floor in tests/** against a committed manifest and fails when any entry decreases (new entries allowed, decreases never). If the manifest guard is not landed in P0-S22, AC-0.17 falls back to (process-only — PR checklist) per D-B	covered (mechanical) [DR-27a]

Standing AC — prefix-shippability (§F.5-G, verbatim, applies at the Phase-0 boundary)

After phase N: (i) npm run build && npm test && npm run test:integration && npm pack --dry-run are green; (ii) a fresh install from the tarball drives /plan → /implement → /code-review → /resolve → /release on a GitHub project with no new prompt and no new file in git status; (iii) every artifact in the tree has at least one reachable consumer in this phase's end-state — no file, define, list element, frontmatter key or conditional arm exists solely for a later phase (ADR-003); (iv) the frozen github-status-lines.txt fixture is byte-identical to its Phase-0 capture.

Clause (iii)'s Phase-0 consequence: - **PR link line**: / - **Branch token**: emission is deliberately excluded — producer and consumer contract both land in P2; shipping the producer alone violates (iii). Proven by a negative grep: neither literal appears in git.md after P0.

Clause (ii) additionally requires, once before merge, a manual install of the packed tarball into a scratch $HOME and the five-command walk-through.

Standing AC — commit order (§F.5-I)

Phase 0's commit order A1 → A2 → A3 → A4 is itself an acceptance criterion. A2 (the golden capture) must contain nothing but the two fixtures, and must be captured from the post-A1 tree. A1 edits git.md; capturing first pins a tree Phase 0 immediately moves past (GAP-06 / B-9). Verified by reviewing the commit list on the PR: exactly four subtask commits in that order, A2's diff limited to tests/fixtures/golden/**.

Phase-0 exit gate: AC-0.17 with the floors named explicitly (>= 8, >= 50, 60000, toHaveLength(11) / toBe(13) / toBe(14), the 3,072-byte charter cap) · AC-0.9 · F.5-G (i)–(iv).

---

3. Scope

Included in Phase 0

• Five live prompt-text defect fixes (/plan fetch wiring, /debug spawn key, three gh issue command-layer sites, the D9 divergence, release.md's untruthful milestone claim).
• Untrusted-content containment at the sink for fetch-issue / fetch-issues-batch, plus Principle 8 and the no-echo rule widened to every provider and every comment op.
• ISSUE_NUMBERS → ISSUE_REFS rename at three sites (ISSUE_NUMBER singular untouched).
• fetch-issues-batch bounded ≤50 with TRUNCATED, and rewritten as a single GraphQL query in the same A1 commit [DR-07].
• manage-debt's missing D4 clause plus the four (pending) restatements, moved as a set.
• The two goldens, captured in their own commit from the post-A1 tree.
• The harness: shared dist-preferred agent-source resolver, loadGolden, extractOpSectionFromCorpus with explicit union/sole modes [DR-18], gitAgentSinkCorpus(), parseFences/isAgentBlock; the two-sided seam test; the D11 guard-class union; DIST_FILES; four guard-gap closures; the non-contiguous v3 .gitignore fixture.
• npm run test:integration added to CI — a Phase-1 blocker landed in Phase 0.
• Docs: docs/reference/platform-assumptions.md (created), the dynamic-workflow-engine KB amendment, CHANGELOG ### Fixed ×5, the two recorded exceptions, the seven corrected citations.

Deferred to later phases (explicitly not in Phase 0)

Deferred	Phase	Why not now
git.md → git.mds generator host, dist/agents/git.md	1	Phase 0 lands the resolver so the rename needs zero test edits
Count literals → name manifests	1	H4 mitigation; Phase 0 only introduces DIST_FILES as a name array
Contract/mechanics split, references/tracker/github/{op}.md, _tracker.mds, byte budget, installer overlay	2	Needs the MDS plumbing from Phase 1
- **PR link line**: / - **Branch token**: producers	2	Clause (iii): producer and consumer contract land together
learn-conventions / publication-gate / decision-markers reference cuts	2	Byte-budget work, needs the generated subtree
Init-selected provider, features.tracker, Tracker agent, tracker.md, Jira, Linear, _mcp.md, --emit	3	Nothing tracker-shaped ships in Phase 0
SG-7's second git.md cut (8 PR/review op bodies)	filed follow-up (D-H)	SG-8: post-review-summary / post-resolution-summary mechanics move only in a PR that moves their guards, never as a size optimisation

Excluded from the feature entirely

• The carve-out-v4 parallel branch and the "3a-first" choreography — DELETED (§F.6).
• learn-tracker, any interactive question step, and the DEGRADED reason interactive setup required — run /plan in an interactive session.
• .devflow/tracker.md (project-local) and any gitignore carve-out v4; there is no commit step.
• Any test requiring a live MCP server; the manual tarball walk-through stays manual (clause ii).
• npm run test:golden:update in CI — a CI job that regenerates a golden is a golden that asserts nothing.

---

4. Gap Analysis Results

The GAP/EC items this phase mitigates. Each row: id · one-line · the step that mitigates it.

Id	One-line	Mitigating step
GAP-01	The PF-026 saving is ~3× overstated; every later byte constant must derive from a measured baseline	P0-S1 (baseline 59,376 / 71,562 recorded), P0-S23 (baselines pinned as named constants)
GAP-03	The always-loaded cross-cutting D4/D11 text is GitHub-specific while §F.1 declares it provider-independent — two authorities on the secret-redaction path	P0-S8 (its cited exemplar: the live D9 divergence collapsed into git.md's single authority)
GAP-05	fetch-issue / fetch-issues-batch have no untrusted-content containment, and Phase 0 wires them into /plan and dynamic-plan (CRITICAL)	P0-S10, P0-S11
GAP-06	Phase 0 invalidates the golden Phase 1 is measured against, and no artifact protects byte-stable GitHub status lines across all four phases	P0-S14 (a), P0-S15 (b), P0-S23; commit-order AC (A1 before A2)
GAP-07	scanned > 0 is the wrong non-vacuity assertion for the Phase-1 rename: 15 of 16 agents survive it while coverage of git silently disappears (CRITICAL)	P0-S17
GAP-08	ISSUE_REF carries two incompatible definitions, collides with ISSUE_REFS, and the ISSUE_NUMBER-is-kept decision was dropped in synthesis	P0-S2, P0-S3, P0-S16
GAP-13	~13 new DEGRADED literals with three synonyms for one condition and no consolidated registry	P0-S15 (the GitHub-path literals frozen); §14's canonical table + retired synonyms
GAP-14	AC-0.4 fails on two gh issue sites no Phase-0 deliverable owns; and Phase 0 is mislabelled "no behaviour change" while shipping five	P0-S5, P0-S21, P0-S26 (CHANGELOG ×5 + honest label)
GAP-15	ISSUE_PR_LINK / ISSUE_BRANCH_TOKEN are consumed but no op produces them, no sink validates them	P0-S16 direction 3 (producer direction, run against the plan-side capture list in P0)
GAP-21	Guard classes must move with the text, not be lowered or rewritten in the same PR as the corpus; three bound guards break silently in Phase 2	P0-S18, P0-S19
GAP-22	Three phases edit the same count literals	P0-S21 introduces DIST_FILES as a name array, never a count (H4)
GAP-23	The only automated check behind AC-1.3 is excluded from CI	P0-S25
GAP-32	The PF-025 sweep list omits files, its phase is a range, and its counts disagree	P0-S22 (one shared retired-wording grep, allowlist narrowed once per phase), P0-S26
GAP-39	dynamic-plan.mds:103 is mischaracterised as a delegation; AC-0.4's deliberate gh pr exclusion is unrecorded	P0-S5 (free-text-op conversion), P0-S21 (exception list), P0-S26 (recorded)
GAP-43	Batch mode has no output AC, and the ISSUE_REFS → ISSUE_REF/ISSUE_ID relationship in batch mode is unstated	P0-S4, P0-S12
GAP-46	Six platform and MCP assumptions cannot be asserted by any test in this repo and are undocumented	P0-S26 item 2 (creates docs/reference/platform-assumptions.md with the four confirmed Claude Code facts)
GAP-49	AC-2.6 pins a literal guard count; Phase 0 alone moves it	§0.1-R2: replace count literals with "green and the guard count has not decreased"; P0-S20 raises REQUIRED_OPS by name, not by number
GAP-51	Citation drift across the source reports; a wrong line number is a missing artifact at the call site	P0-S26 item 6 (seven corrected citations, incl. git.md:292 / E9)
EC-17	platform-assumptions.md row-present assertion is referenced by a later Section-3 output-length test	P0-S26 item 2 (the file must exist with the rows before that test can be non-vacuous)
EC-53	Hook output must be independent of the developer's real $HOME	§0.1-R4 applied to P0-S24's fixtures: temp HOME, seeded, never empty (PF-018)
EC-54	macOS /private realpath divergence	P0-S24: fixtures write paths using the same string passed as cwd; assert both raw and realpath forms
Missing-test (ii)	No two-sided command→agent spawn-key test exists	P0-S16
Missing-test (iii)	No guard over gh issue in the deployed command corpus	P0-S21
Missing-test (iv)	compliance_gate adoption is unguarded; SKILL_CHECK_HOSTS covers only 2 of 6 importers	P0-S22
Missing-test (v)	The 21 ## Extended References tables are unguarded	P0-S22

---

5. Execution Strategy

SEQUENTIAL_CODE_AGENTS — 4 subtasks (A1 → A2 → A3 → A4). Context risk: HIGH.

Rationale — three independent forcings (any one sufficient)

1. Commit order is itself an acceptance criterion (§F.5-I). A1 edits git.md; capturing the golden first pins a tree Phase 0 immediately moves past (GAP-06). A single Code agent has no mechanism that guarantees commit ordering inside one run; sequential agents make each ordinal a spawn boundary.
2. A2 must touch nothing else. The golden capture is a byte-exact 59,376-character copy plus a hand-extracted frozen fixture. Interleaving it with any behaviour change reproduces the exact defect GAP-06 exists to prevent, and the "never regenerate alongside a behaviour change" rule (B-05) is unenforceable inside one agent's context.
3. Working set. ~30 files / ~1,400–1,900 changed lines spanning four unrelated domains (prompt text, fixtures, test harness, docs). A3 alone requires simultaneous knowledge of 12 test files' internal conventions.

Why PARALLEL_CODE_AGENTS is rejected (feature-wide, PF-020 closure rule)

PF-020's 2026-08-30 extension binds — "disjointness has to hold over each agent's READ/EXECUTE closure, not merely its WRITE set." Three facts collapse every candidate partition:

1. npm run build is rm -rf dist && … (package.json:20). Every Code agent's verification step deletes and rebuilds the one shared dist/. A sibling agent mid-suite observes a missing dist tree — and Phase 0 deliberately makes every dist-scanning guard fail loud on a missing directory (AC-0.16, R3). Parallel agents therefore manufacture nondeterministic red suites that look exactly like real regressions.
2. /implement has no worktree isolation — grep for worktree in implement.mds returns zero hits. All Code agents share one worktree, one git index, one dist/, one node_modules, one vitest run.
3. The test corpus spans every partition. The seam test, the D11 sink-class guard, the containment test and the frozen fixture all read git.md ∪ dist/commands/*.md ∪ dist/skills/git/references/** — the union of every proposed subtask's outputs. Disjoint write sets, 100% overlapping read closure.

This disqualifier is independent of write-set disjointness, so it holds for all four phases. It is a tooling limitation, not a plan limitation [DR-23, ALREADY-ADDRESSED].

Context-risk driver

A3. tests/build-mds.test.ts is 1,386 lines and tests/shell-hooks.test.ts is 2,677; an agent that reads both in full plus git-agent.test.ts, registry-integrity.test.ts and agent-name-guards.test.ts has consumed its budget before writing the seam test. The A3a/A3b split below is the sanctioned response.

Serialization points, ranked (all four phases mutate rank 1)

Rank	File	P0	P1	P2	P3
1	src/assets/agents/git.md → git.mds	A1 (9 edit anchors)	S2 (whole file)	T1/T2 (10 ops + cross-cutting)	3a-4 (preamble)
2	tests/build-mds.test.ts	A3 (DIST_FILES, :990, SKILL_CHECK_HOSTS)	S3 (count → name manifests)	T5 (partial count)	3a/3b/3c (manifest rows)
3	src/targets/claude-code/installer.ts	—	S1 (agent resolver)	T4 (overlay)	—
4	tests/git-agent.test.ts	A3 (resolver, D11 union, bounds)	S3 (repoint — no-op if A3 done right)	T2 (corpus follows the text)	—
5	src/core/assets.ts	—	S1 (compiledAgentsDir)	T4 (compiledSkillRefsDir)	—

Rank 2 is the only one Phase 1 defuses (count literals → name manifests, GAP-22); ranks 1 and 3–5 stay serialization points for the life of the feature.

Split policy (Gate-2 decision D-G)

Split	Kind	Status
A3 → A3a / A3b	in-PR subtask split, no merge consequence	PRE-AUTHORISED. The implementing agent may take it whenever A3 shows context strain, without asking
P0a / P0b	PR split (file-set seam: P0a = /plan, /debug, dynamic-plan, implement wiring not touching git.md; P0b = harness + manage-debt D4 + goldens)	REQUIRES APPROVAL — changes the issue count. Do not take it unilaterally
A1 / A2 apart	—	FORBIDDEN. Splitting by commit kind across a PR boundary recreates the golden-provenance defect across a PR boundary

Authoring order ≠ commit order (load-bearing)

Author P0-S16 (seam test), P0-S17 (resolver) and P0-S18 (D11 union) first in the working tree to obtain RED; then git add -p / reorder so A1 lands as commit 1. Record each RED run in the PR body.

AC-0.1 demands the seam test be red on debug.mds before the fix, but the commit order puts A1 before A3. Reconciliation (H10): A3 proves non-vacuity against a synthetic known-bad corpus fixture (a fake dist/commands/*.md fence passing ISSUE: instead of ISSUE_INPUT:), matching the "fail on a known-bad sample" idiom AC-0.8 already uses. An agent must never revert A1's fix to demonstrate red. Recorded as a test-inventory item under AC-0.13.

---

6. Subtask Breakdown

#	Domain	Owned files (exclusive, with anchors)	Steps	Depends on	Gate green before next	Handoff contents
A1	prompt assets (commands, agents, partials) + hooks	src/assets/commands/resolve.mds (:352,:378,:499,:508,:539) · debug.mds (:17,:24,:46,:51) · plan.mds (:27,:32, new Git fence inserted at #### Phase 1: Gate 0 before :63) · dynamic-plan.mds:103 · dynamic-build.mds:73 · _partials/_wave.mds (:4,:9) · implement.mds:73 · release.md:138 · src/assets/agents/git.md (:69,:268-284,:288-300,:292,:294-298,:300-318,:379-411,:723,:922) · src/assets/scripts/hooks/ensure-devflow-init:23	P0-S1 (prep) · P0-S2 … P0-S13	—	G0.1	post-A1 git.md SHA-256 + line count; the list of every line A1 changed in git.md (so A2's capture is spot-checkable); confirmation no test file was touched
A2	tests (fixtures only)	new tests/fixtures/golden/git-agent.md · new tests/fixtures/golden/github-status-lines.txt	P0-S14, P0-S15	A1 committed on the branch	G0.2	the two fixture paths + byte counts + the explicit statement "these fixtures are immutable for the remainder of Phase 0–3; a mismatch means the source is wrong, never the fixture"
A3	tests + CI	tests/helpers.ts · new tests/seams/command-agent-input.test.ts · new tests/goldens/git-agent-golden.test.ts · new tests/goldens/github-status-lines.test.ts · new tests/guards/agent-source-resolver.test.ts · new tests/guards/extended-references.test.ts · new tests/guards/retired-wording.test.ts · new tests/guards/numeric-floor-manifest.test.ts + its committed manifest [DR-27a] · tests/git-agent.test.ts · tests/registry-integrity.test.ts · tests/agent-name-guards.test.ts · tests/agent-frontmatter.test.ts · tests/skill-references.test.ts · tests/installer-new.test.ts · tests/build-mds.test.ts · tests/shell-hooks.test.ts (gitignore fixture only) · tests/init-logic.test.ts (gitignore fixture only) · .github/workflows/ci.yml:23 · package.json (test:golden:update)	P0-S16 … P0-S25	A1, A2	G0.3	every new guard by name, its file, and its proven-red mechanism; the five user-visible behaviour changes with before/after strings; the two recorded exceptions (release.md:85 conventions read; gh pr view at code-review.mds:76-78 / bug-analysis.mds:43-45); the mode ('union' / 'sole') each extractOpSectionFromCorpus consumer names [DR-18]
A4	docs	new docs/reference/platform-assumptions.md · .devflow/features/dynamic-workflow-engine/KNOWLEDGE.md · CHANGELOG.md · the seven drifted citations in .devflow/features/*/KNOWLEDGE.md · the Phase-0 issue body	P0-S26	A1–A3 (sweep the branch's final tree, not the commit — PF-025)	G0.4	— (terminal)

File-set disjointness holds across all four (A1 = src/assets/** + hooks, A2 = tests/fixtures/golden/**, A3 = tests/** + CI, A4 = docs/KB). Not required for SEQUENTIAL, but it makes each commit's pathspec trivially derivable — use git commit -- <owned paths>.

Optional A3 split (pre-authorised, in-PR only, no merge consequence):

• A3a = resolver + loadGolden + seam test + git-agent.test.ts + registry-integrity.test.ts → P0-S16, S17, S18, S19, S20, S23
• A3b = build-mds.test.ts guards + gitignore fixture + Extended-References guard + retired-wording grep + numeric-floor manifest + CI → P0-S21, S22, S24, S25

Do not split A1/A2 apart.

Gates

Gate	Must be green before	Command / assertion
G0.1	A2 starts	npm run build && npm test — full suite green at the post-A1 tree. No new test exists yet; this proves A1 broke nothing.
G0.2	A3 starts	tests/fixtures/golden/git-agent.md byte-equals src/assets/agents/git.md (verify with cmp, not by reading); github-status-lines.txt non-empty and every extracted literal present in the source at the cited line.
G0.3	A4 starts	npm run build && npm test && npm run test:integration green; each new guard demonstrated non-vacuous with a recorded red case (AC-0.13); postingOps.length >= 8 satisfied by git.md alone (AC-0.8); resolveAllAgents() ⊇ getAllAgentNames() = 16 (AC-0.7).
G0.4	PR opens	Prefix-shippability sequence (§14) + github-status-lines.txt byte-identical to its A2 capture (AC-0.9) + git status --porcelain empty.

/implement spawn notes

• Spawn one Code agent per subtask, in order: A1 → A2 → A3 (optionally A3a → A3b) → A4. Never two concurrently.
• CREATE_PR: false on A1, A2, and A3 (and A3a when the split is taken); CREATE_PR: true on A4, the terminal agent, which opens the PR with the complete test-inventory table and the RED proofs from all subtasks (implement.mds:409). A4 must therefore receive every prior handoff.
• HANDOFF_REQUIRED: true on every subtask but the last (A1, A2, A3 — and A3a when the split is taken). A4 is terminal and produces no handoff.
• Every Code agent reads .devflow/docs/handoff-{branch_slug}.md via HANDOFF_FILE; each writes its handoff contents (column 7 above) after its own gate passes, never before.
• Commit with an explicit pathspec of the subtask's owned files only: git commit -- <owned paths> (PF-020's validated cheap mitigation; per-commit provenance survives the squash-merge).
• Standing instruction in every Code prompt: "a golden mismatch means the source is wrong, never the fixture." (H2.)

---

7. Implementation Plan

P0-S1 — Branch, baseline capture, green-suite proof (prep, no commit) → A1

• Goal: Establish the phase branch and record the measured byte baseline every later budget constant derives from.
• Files: none modified.
• Change: git switch -c feat/tracker-phase-0 from main@e726874. Record in the PR body: git.md 938 L / 59,376 ch, skills/git/SKILL.md 283 L / 9,236 ch, skills/worktree-support/SKILL.md 92 L / 2,950 ch, total 1,313 L / 71,562 ch [verified].
• Ids: GAP-01, §F.4 — §C.4's 71,090 / 58,904 are wrong by 472 ch; every Phase-2 constant derives from 71,562 / 59,376.
• Verify: all four gates green at HEAD before any edit.

P0-S2 — Correct the debug.mds spawn key → A1

• Goal: Make /debug #42 pass the key the op actually declares.
• Files: debug.mds:51 (ISSUE: {issue number} → ISSUE_INPUT: {issue reference}), :46 ("fetch the GitHub issue" → "fetch the issue"), :17, :24 (- GitHub issue: "#42" → - Issue reference: "#42").
• Ids: AC-0.1, GAP-08, PF-024 class, intent Evaluate: Should /plan-next-steps command have dedicated sub-agent? #13.
• RED-first: seam test (P0-S16) RED before — ISSUE ∉ fetch-issue's **Input:** list (git.md:261 [verified]) — GREEN after. Non-vacuity: seam test asserts ≥1 fence scanned per agent type.

P0-S3 — ISSUE_NUMBERS → ISSUE_REFS at all three repo-wide sites → A1

• Goal: Rename the batch input key while the op is still dead, so it is free rather than a later seam break.
• Files: git.md:69 (Operations table row), git.md:292 [drift: impl-2 cites :291; verified :292], plan.mds:32.
• Change: three literal replacements. ISSUE_NUMBER (singular) untouched at all 14 Code-spawn sites, code.md:35, github-api.md:142.
• Ids: AC-0.11 (amended), GAP-08, E9, GAP-51.
• Verify: grep -rn ISSUE_NUMBERS src/ dist/ → 0 hits (pre-state: 3 source hits + 1 in dist/commands/plan.md). Seam-test key comparison is exact-match, never startsWith — the two identifiers differ by one character.

P0-S4 — Wire /plan to actually fetch issues → A1

• Goal: Make plan.mds's issue-mode parse reach a Git op instead of consuming a body it never fetched.
• Files: plan.mds:27/:32 (grammar + ISSUE_REFS), and a new Git spawn fence inserted at #### Phase 1: Gate 0 before :63 — the only placement that makes :60 ("references … a detailed issue") and :70 ("multi-issue: present unified scope … after individual discovery") non-vacuous. Candidate (a) — end of ## Input at :33 — is rejected for that reason.
• Change: single-ref → OPERATION: fetch-issue / ISSUE_INPUT:; multi-ref → OPERATION: fetch-issues-batch / ISSUE_REFS:. Capture per §F.5-A: ISSUE_REF, ISSUE_ID, ISSUE_URL, ISSUE_CONTENT, ACCEPTANCE_CRITERIA.
• Ids: AC-0.3, AC-0.11, GAP-43.
• Verify: registry Guard 6 forward check (registry-integrity.test.ts:391-487) green with the new OPERATION: values; seam test asserts every passed key ∈ the op's declared inputs and every declared input is passed.
• Note [DR-07]: this step makes fetch-issues-batch live. Its N+1 must therefore be fixed in the same commit group — see P0-S12. Shipping the caller without the fix is the choice that would need justifying, not the reverse.

P0-S5 — Route the three remaining gh issue command-layer sites through the Git agent → A1

• Goal: Remove every raw gh issue invocation and descriptive mention from the command layer so AC-0.4 is satisfiable.
• Files: dynamic-plan.mds:103; dynamic-build.mds:73; _partials/_wave.mds:9 and :4.
• Change: dynamic-plan.mds:103 — the call at :100-105 is already inside { agentType: "Git" }, so this is a free-text-op conversion, not a delegation (GAP-39 corrects §2.6). dynamic-build.mds:73 — "A GitHub issue body (fetch via gh issue view <number>)" → "fetched via the Git agent". _wave.mds:9 — "a Git agent may pre-fetch…" → a mandatory once-per-wave Git pre-fetch; :4 "reading the GitHub issues" → "reading the issues".
• Ids: AC-0.4, GAP-14, GAP-39.
• Verify: the P0-S21 guard RED before, GREEN after; 'may pre-fetch' absent from _wave.mds. The two gh pr view sites (code-review.mds:76-78, bug-analysis.mds:43-45) are an explicit allowlisted PR-hosting exception.

P0-S6 — Name the owner of implement.mds's dangling cross-reference → A1

• Files: implement.mds:73 — ISSUE_NUMBER: GitHub issue number (if provided or created by the issue-first gate in step 1c) → names the owner explicitly (setup-task's issue-first step in the Git agent).
• Ids: retired-wording grep ("step 1c").
• Verify: the P0-S22 retired-wording allowlist for Phase 0 excludes "step 1c"; seam test unaffected (ISSUE_NUMBER is a kept Code-spawn key).

P0-S7 — Delete release.md's untruthful post-release claim → A1

• Files: release.md:138 — delete 6. **Post-release steps** — version bump to next dev, close milestone, etc. (or the milestone clause only).
• Ids: AC-0.14, intent feat(cli): add --verbose flag for clean default init output #22 (untruthful traceability claims are deleted, not implemented).
• Verify: build-mds.test.ts:1278-1292 is the only test reading release.md today (pins COMMIT_LIST / SHIPPED_ISSUES / backlink-shipped-issues) — unaffected; retired-wording guard adds close milestone.

P0-S8 — Collapse the live D9 divergence into git.md's single authority → A1

• Files: resolve.mds:378 — last line of the resolve-review-threads fence (:370-378). Current text verified verbatim: D9: resolve threads only when VERIFICATION_STATUS == PASS and verdict is FIXED/FALSE_POSITIVE/BY_DESIGN with cited evidence." → the git.md:96 literal: ONLY when VERIFICATION_STATUS == PASS AND verdict == FIXED AND commit_sha non-empty.
• Ids: AC-0.5, intent 🔒 SECURITY (Long-term): Implement cryptographic script signing #9, GAP-03's cited exemplar.
• Verify: new caller-side guard extended to dist/commands/resolve.md (compiled ≈ :426) asserting byte-equality against the git.md:96 extract, already pinned at git-agent.test.ts:167-168 [drift: plan-1/E10 cite :163; the it( is :162, the assertion :167-168 — verified]. Nothing pins the caller side today — the guard is RED against the pre-fix tree.

P0-S9 — Add manage-debt's missing D4 clause and the four (pending) restatements → A1

• Files: git.md:379-411 — insert **Degradation (D4):** between **Process:** (:384-394) and **Output:** (:397). resolve.mds:352 (add the DEGRADED / Tracked = (pending) branch with the reason named) plus :244, :499, :508, :539 — all verified live; all must move together or the artifact contract splits.
• Ids: AC-0.6, GAP-06 (this is the edit that invalidates a pre-A1 golden — hence A1 before A2).
• Verify: guard "every REQUIRED_OP doing remote I/O has a **Degradation (D4):** line" RED before (manage-debt has zero DEGRADED occurrences today), GREEN after; the four (pending) sites asserted as a set, not individually.

P0-S10 — Move untrusted containment into the ops' Output blocks → A1

• Goal: Contain attacker-authored issue text at the sink, in the same change that opens the caller (PF-027).
• Files: git.md:268-284 (fetch-issue Output — wrap {title} :270, ### Description :273-274, ### Acceptance Criteria :276-277, ### Dependencies :279-280); git.md:300-318 (fetch-issues-batch Output — same containment for :302-310) [drift: P0-S15 lists :302-319, P0-S10 said :300-319; verified :300-318].
• Change: <untrusted-issue-body>…</untrusted-issue-body> + the one-line data-only note, sentence source verbatim from _wave.mds:14.
• Ids: AC-0.10, GAP-05 (CRITICAL — P0-S4/P0-S5 are exactly what make attacker text flow into /plan's Design agents).
• Deps: P0-S4, P0-S5 (the callers that open the path).
• Verify: forward guard over git.md ∪ references — every op whose Output renders a remote-sourced field carries the token; non-vacuous with ≥3 ops; negative arm that no reply/summary template interpolates a remote body field.

P0-S11 — Widen Principle 8 and the no-echo rule to every provider and every comment op → A1

• Files: git.md:922 (Principle 8 — today covers <external-thread> only) → widen to remote issue/comment bodies for every provider. git.md:723 → extend the no-echo rule beyond post-resolution-summary to every tracker comment op.
• Ids: AC-0.10, SG-2, GAP-05, intent Evaluate: Should /debug command have dedicated sub-agent? #11.
• Verify: same forward/negative guard pair as P0-S10; the widened sentence is an assertion literal.

P0-S12 — Bound fetch-issues-batch and rewrite it as a single GraphQL query → A1 [DR-07]

• Files: git.md:288-300 (**Input:** :292, **Process:** :294-298).
• Change (bound): add ≤50 bound + TRUNCATED ({n} not processed).
• Change (rewrite) [DR-07]: rewrite git.md:294-298 step 2 — today "2. Fetch each issue via gh issue view {number} --json number,title,body,labels,assignees,milestone,comments" — as one GraphQL query with the ≤50 bound and TRUNCATED ({n} not processed), in the same A1 commit that wires the caller (P0-S4). The original plan deferred the rewrite to P2-S6 on the grounds that "a dead op is the cheapest place to fix an N+1"; P0-S4 wires the op live in the same commit group, so the cheap window is consumed rather than used. Left deferred, /plan #a #b … issues up to 50 sequential gh issue view calls at B10(10)'s 1 s spacing (3 s at X-RateLimit-Remaining < 50) — up to ~150 s of user-visible latency and 50 API calls where one query suffices — shipping through Phases 0 and 1. ADR-003 forbids relocating known-defective text unchanged; it does not require deferring the fix. Fixing in place here and moving the fixed text verbatim in P2-S6 satisfies ADR-003 strictly better, and A2 captures the golden from the post-A1 tree so the rewrite is inside the golden and P2's move stays a pure verbatim move. P2-S6's fetch-issues-batch row therefore becomes a plain verbatim move — the rewrite clause is deleted from it (recorded here so the Phase-2 author does not re-apply it).
• Ids: AC-0.3, intent v0.3.2: UX improvements - simplified output, unified commands, streamlined statusline #4, intent Evaluate: Should /plan-next-steps command have dedicated sub-agent? #13, GAP-43; [DR-07].
• Verify: Guard 2 (git-agent.test.ts:79-159) gains a ≤50 + TRUNCATED pin for this op; RED before. New edge-case row moved into Phase 0 [DR-07]: "gh issue view absent from the batch op", scoped to git.md, proven RED against the pre-fix tree (the literal is present at :294-298 today). This row was previously Phase 2's GAP-26 assertion; it is now Phase 0's and must not be duplicated in Phase 2.

P0-S13 — Remove the dead .root-gitignore-configured-v2 fast-path → A1

• Files: src/assets/scripts/hooks/ensure-devflow-init:23 — [ -f "$_DEVFLOW_DIR/.root-gitignore-configured-v2" ] while both writers stamp v3 and ensure-root-gitignore:113 deletes the v2 marker on success ⇒ the branch is dead.
• Ids: ADR-003 (delete a guard for a now-impossible state); §B.4 "retained despite the cancellation". Orphaned by the carve-out-v4 cancellation; it has no other owning phase.
• Verify: shell-hooks.test.ts HOOK_SCRIPTS (:20-44 [verified: 24 entries; plan-1 cites :20-46]) bash -n syntax check stays green; either the branch is deleted or an assertion records why it survives.

P0-S14 — Capture tests/fixtures/golden/git-agent.md → A2

• Files: new tests/fixtures/golden/git-agent.md — full src/assets/agents/git.md at the post-A1 tree.
• Ids: AC-0.2, GAP-06(a). Resolves G-2 in favour of a committed fixture: option (b) — recompiling a pinned SHA — silently passes if the SHA moves, exactly the failure it must prevent.
• Deps: P0-S2..P0-S13 all committed.
• Verify: its Phase-0 consumer is a byte-equality assertion src/assets/agents/git.md === loadGolden('git-agent.md') — trivially true today, but a reachable consumer (clause iii) that catches unreviewed git.md drift, and becomes the compile-equality subject in P1 with no test edit.

P0-S15 — Capture tests/fixtures/golden/github-status-lines.txt → A2

• Files: new tests/fixtures/golden/github-status-lines.txt, extracted from these exact sources (all verified live):
  • git.md cross-cutting: :23-28 (STOP / THROTTLED ({n} not processed) / < 10 / < 50), :33, :36, :54 TRACEABILITY: DEGRADED (redaction unavailable), :55 SCRUB: N [type:count,…], :56 SECRET-EXPOSED (rotate {type} credential — the source file still holds it), :57 $DEVFLOW_BODY / $DEVFLOW_BODY_RAW
  • ensure-pr-ready :140-149 (:144, :145, :147 ### Status: READY | BLOCKED) · validate-branch :174-191 (:190) · setup-task :238-252 (:244 ### Traceability, :245, :246)
  • fetch-issue :270-283 (:282-283 ### Suggested Branch / {type}/{number}-{slug}) · fetch-issues-batch :302-318 (:302 ## Issues Batch ({n} issues))
  • post-review-summary :369-374 · manage-debt :399-408 · check-ci-status :429-439 · create-release :467-473 · gather-release-evidence :495-506 · learn-conventions :570-582 · fetch-review-threads :613-632 · resolve-review-threads :682-692 · post-resolution-summary :742-745 · check-merge-readiness :773-775 · backlink-shipped-issues :822-830 · ensure-traceable-issue :865-869 · post-wave-report :905-908
  • Marker formats (Guard-5 literals): <!-- devflow:review-summary cycle:{N} ts:, <!-- devflow:resolution-summary ts:, <!-- devflow:shipped v, <!-- devflow:wave-report wave:
  • code.md:93, :95, :99 · dynamic-build.mds:522, :524 · resolve.mds:244, :352, :499, :508, :539, :619
• Ids: AC-0.9, GAP-06(b), GAP-13, §F.5-F. Frozen at Phase 0, never regenerated through Phase 3.
• Deps: P0-S9 (the (pending) restatements must be final before freeze), P0-S14.
• Verify: byte-equality asserted at the end of every phase (AC-0.9 / AC-1.11 / AC-2.1 / AC-3.1). CI must never auto-regenerate. The extractor is a pure function over this source list, so the test and the fixture derive from one enumeration (risk P0-a).

P0-S16 — Two-sided seam test, three directions → A3 (A3a)

• Files: new tests/seams/command-agent-input.test.ts. Header framing copied verbatim from tests/resolve/duplicate-verdict.test.ts:4-15 (the repo's only two-sided producer/consumer test).
• Change: take fencePattern = /```[^\n]*\n([\s\S]*?)```/g and the isGitBlock predicate (/Agent\(subagent_type="Git"/ || /agentType:\s*"Git"/) verbatim from registry-integrity.test.ts:449-456; generalise per agent type (Git and Code at minimum); harvest all passed keys with /^([A-Z_][A-Z0-9_]*): /gm. Corpus = DIST_FILES (14).
  • Index once [DR-24]: build the op → section map once per corpus in beforeAll and index into it per fence — never re-scan the corpus per fence. With ~40+ Git fences today and a corpus reaching git.md + ~30 generated references by 3c, a per-fence scan multiplies extractOpSectionFromCorpus's throw-on-missing-anchor into dozens of identical failures. Non-vacuity: assert the map's size >= 15, matching the existing opsCovered floor.
  • Direction 1 (forward): keys ∖ {OPERATION, COMPLIANCE, WORKTREE_PATH} ⊆ the identifiers named in that op's **Input:** line, via extractOpSectionFromCorpus(corpus, op, { mode: 'sole' }) [DR-18] — the **Input:** lookup must resolve to the single contract authority in git.md; unioning three providers' sections would let a key declared by only one provider satisfy the forward check. Exact-match, never startsWith. This direction is RED on debug.mds:51's ISSUE: before A1.
  • Direction 2 (reverse): every non-optional identifier in an op's **Input:** is passed by at least one caller fence.
  • Direction 3 (producer): every value named in issue_capture_contract() — ISSUE_REF, ISSUE_ID, ISSUE_URL, ISSUE_CONTENT, ACCEPTANCE_CRITERIA — has a greppable producer in an op **Output:** block. In P0 this runs against the plan-side capture list; from P2 against the compiled _tracker.mds define.
  • Exclusions, asserted as a literal set with a rationale comment: **Produces:** and **Requires:** are not a contract (PF-039, B10(13)) — they are a phase-ordering DAG naming principal upstream state, not a spawn-block field contract.
• Ids: AC-0.1, GAP-08, GAP-15 (direction 3), missing-test (ii); [DR-18], [DR-24].
• Verify / non-vacuity: expect(fencesScanned.get('Git')).toBeGreaterThan(0) and the same for 'Code'; expect(opsCovered.size).toBeGreaterThanOrEqual(REQUIRED_OPS.length - INTERNAL_OPS.size) (≥15). Known-bad sample: inline fixture 'OPERATION: fetch-issue\nISSUE: 42\n' must produce exactly one violation naming ISSUE.

P0-S17 — Shared dist-preferred agent-source resolver + registry completeness → A3 (A3a)

• Files: tests/helpers.ts — add resolveAgentSource(name) and resolveAllAgents() beside requireDistFile (:29), mirroring its throw-with-a-build-hint, never skip contract (:25-38). Repoint: tests/git-agent.test.ts:18 (GIT_AGENT_PATH), tests/registry-integrity.test.ts:393 [drift: plan says :392, which is distCommandsDir], tests/registry-integrity.test.ts:49-56 ("every agent in DEVFLOW_PLUGINS exists as src/assets/agents/{name}.md" — hard-fails the moment git.md becomes git.mds; not in the Phase-1 repoint list), tests/agent-name-guards.test.ts:746-750 (resolves each dist roster entry to path.join(AGENTS_DIR, '{slug}.md') and calls readFrontmatterModel, which throws on a missing file at :403-409).
• Change: dist-preferred (dist/agents), src-fallback (src/assets/agents), ENOENT-tolerant on the dist side only; do not relax the .md filter on the src side. Surface is resolveAgentSource(name): { path, content, origin: 'dist' | 'src' } and resolveAllAgents(): Map<name, {path, content, origin}> (C15: resolvedAgentNames() is [...resolveAllAgents().keys()]; the origin field is what the Phase-1 dist-preferred/src-fallback tests assert on). New assertion, used by every consumer: expect([...resolveAllAgents().keys()]).toEqual(expect.arrayContaining(getAllAgentNames())) — 16 today, 16 after the P1 rename, 17 after 3a.
• Also lands here [DR-18]: extractOpSectionFromCorpus(corpus: Array<{path,content}>, op, { mode: 'union' | 'sole' }) in tests/helpers.ts, with no default mode. 'sole' throws when the anchor matches in more than one corpus file, naming both paths; 'union' returns the concatenation with a match count. Rationale to record in the header: a first-match implementation makes the D11 guard under-count — silently lowering the effective floor without touching the literal 8, the precise evasion R2/H3 exist to prevent — while a concatenating implementation makes the seam test's exact-match key check permissive. Either choice is silently wrong for one caller, so neither is the default.
• Ids: AC-0.7, AC-0.16, GAP-07 (CRITICAL); [DR-18].
• Verify: new tests/guards/agent-source-resolver.test.ts — RED against a fixture tree where git is removed from the resolvable set; RED with a missing-dist case asserting a thrown error with a build hint. Guard: no new test file contains a literal src/assets/agents/git.md path outside the documented src-fallback sites. Two new unit cases [DR-18]: 'sole' throws on a two-file corpus with duplicate anchors (message names both paths); 'union' returns both sections and a match count of 2.
• Anti-pattern named explicitly: scanned > 0 over the agent corpus. Two scanned > 0-shaped assertions already exist and would stay green at 15 (B-11 / GAP-07). Rationale recorded verbatim in the helper header: 15 of 16 agents survive scanned > 0 while coverage of git silently disappears.

P0-S18 — D11 guard-class union, floor unchanged at >= 8 → A3 (A3a)

• Files: tests/git-agent.test.ts:359-383 (forward, >= 8), :384-406 (reverse), :407-425 (bypass INLINE_BODY_RE) [drift: plan says 359-424; bypass ends at :425].
• Change: refactor the sink class to gitAgentSinkCorpus() = git.md ∪ compiledSkillRefsDir() with the dist sibling ENOENT-tolerant. Floor stays >= 8 and must remain satisfied by git.md alone in Phase 0. Keep the non-vacuity probe at :419-424 verbatim. Replace git-agent.test.ts:24-30 extractOpSection (returns '' on a missing marker — a vacuity hole) with the throwing corpus-spanning extractOpSectionFromCorpus (shape from helpers.ts:49), keeping the local name so diffs stay reviewable.
• Mode [DR-18]: the D11 forward guard names { mode: 'union' } — a posting op's D11 reference may live in the moved mechanics file from Phase 2 onward, so the corpus must union. The mode is passed explicitly at the call site with a one-line comment naming the reason; the union's match count is asserted so a silent single-file collapse is red.
• Ids: AC-0.8, GAP-21, B-2, §C.5's flagged coupling ("cut feat: Add project-agnostic release automation workflow #1 removes 4 of the 8 posting ops"); [DR-18].
• Deps: P0-S17.
• Verify: existing >= 8 still green from git.md alone; the bypass guard's known-bad sample retained; add a second known-bad sample for the union corpus (a git.md copy with Comment-sink scrub (D11) stripped from one posting op).

P0-S19 — Pin manage-debt: 60000 before the literal moves → A3 (A3a)

• Files: tests/git-agent.test.ts:97-103 — insert the pin after the post-wave-report cap it (Guard 2 block = :79-159).
• Ids: AC-0.12, GAP-21, intent Evaluate: Should /plan-next-steps command have dedicated sub-agent? #13. Must land before the literal moves in Phase 2.
• Verify: RED against a git.md sample with 60000 stripped from manage-debt (git.md:387: "archive if > 60000 chars").

P0-S20 — REQUIRED_OPS / INTERNAL_OPS corrections → A3 (A3a)

• Files: tests/git-agent.test.ts:52-68 — REQUIRED_OPS += 'fetch-issue', 'fetch-issues-batch' (15 entries today → 17). tests/registry-integrity.test.ts:397-405 — remove 'fetch-issues-batch' from INTERNAL_OPS with a one-line rationale comment so the deletion is not read as accidental (SG-11); 'learn-conventions' stays.
• Ids: AC-0.11, SG-11, GAP-49 (the guard count rises by name, never by a pinned literal).
• Deps: P0-S4.
• Verify: registry Guard 6 reverse check green; RED if plan.mds's spawn is reverted.

P0-S21 — DIST_FILES + bring release.md into the four ALL_HOSTS-scoped guards → A3 (A3b)

• Files: tests/build-mds.test.ts:47 (ALL_HOSTS) — add DIST_FILES = [...ALL_HOSTS.map(h => h + '.md'), 'release.md']; extend :283-300 (escaped-brace), :990-1022, the spawn-scoped COMPLIANCE guard and the un-expanded-call-site guard.
• Change: also the new AC-0.4 guard — "no gh issue invocation or descriptive mention in any DIST_FILES entry outside a Git spawn fence" — with the two gh pr view PR-hosting exceptions allowlisted by name in a commented exception list.
• Ids: AC-0.4, GAP-14, GAP-22, SG-13, missing-test (iii). G-6/OD-8: the DIST_FILES vs ALL_HOSTS divergence is permanent and goes in the KB (P0-S26), not only here.
• Deps: P0-S5, P0-S7.
• Verify: RED against the pre-S5 tree (dynamic-plan.mds:103's free-text gh issue view); non-vacuity DIST_FILES.length === 14.

P0-S22 — Guard-gap closures: compliance_gate adoption, build-mds.test.ts:990, Extended References, retired-wording grep, numeric-floor manifest → A3 (A3b)

• Files:
  • tests/build-mds.test.ts:930-933 (SKILL_CHECK_HOSTS [verified: only code-review + plan]) — add the missing bug-analysis assertion + an all-hosts compliance_gate adoption guard over the 6 importers (bug-analysis.mds:27, code-review.mds:43, dynamic-build.mds:49, implement.mds:54, plan.mds:141, resolve.mds:104) with the release.md inlined-gate exception recorded by name. Non-vacuity hostsScanned === 6.
  • tests/build-mds.test.ts:990 — the title advertises a COMPLIANCE: ${ guard the body (:991-1022) does not contain (it asserts COMPLIANCE: {enabled). Rename the title or assert the real literal (dist/commands/dynamic-build.md:210 is literally COMPLIANCE: ${COMPLIANCE}). Record in the defect list.
  • new tests/guards/extended-references.test.ts — existence guard over all 21 ## Extended References tables (skills/git/SKILL.md:265-273). Ships a generated-path exception from the outset — references/tracker/ never exists in src, so a naive guard breaks in Phase 2. Assert the exception list is non-empty and each entry justified.
  • new tests/guards/retired-wording.test.ts — one shared grep guard with a per-phase allowlist narrowed once per phase, never a new grep per phase (GAP-32). Phase-0 entries: ISSUE_NUMBERS, ISSUE: {issue, close milestone, may pre-fetch, step 1c. Assert allowlist size and corpus size.
  • new [DR-27a] tests/guards/numeric-floor-manifest.test.ts + a committed manifest (tests/fixtures/numeric-floors.json) — every pinned numeric floor in tests/** is enumerated with its file, its it title and its value; the guard fails when any manifest entry's live value decreases. New entries are allowed; decreases never are. This mechanises AC-0.17, whose only defence was previously a reviewer's attention, and it is the sole defence against the H3 "just lower 8" pressure Phase 2 deliberately creates. Seed the manifest with the named floors: >= 8 (D11), >= 50 (subagent_type), 60000 (manage-debt), toHaveLength(11), toBe(13), toBe(14), the 3,072-byte charter cap, the >= 2 marker-occurrence floors. Non-vacuity: a seeded decrement of any one manifest entry goes red. If this guard is not landed, AC-0.17 moves into §0.10's process-AC list and is marked (process-only — PR checklist) — it must not be given a fake test.
• Ids: AC-0.8, AC-0.13, AC-0.17, missing-tests (iv)+(v), GAP-32; [DR-27a].
• Deps: P0-S21.
• Verify: each guard proven RED with a named known-bad sample (AC-0.13 is a test-inventory AC).

P0-S23 — Golden fixture infrastructure → A3 (A3a)

• Files: tests/helpers.ts — loadGolden(name) beside requireDistFile (:29), mirroring its throw-with-a-command-hint, never skip contract. package.json:29 area — add test:golden:update. new tests/goldens/git-agent-golden.test.ts and new tests/goldens/github-status-lines.test.ts (the latter also carries the Phase-0 byte baselines as named constants with derivations in comments — C6: Phase 2's tests/tracker/byte-budget.test.ts does not exist yet, and plan-1's "before any text moves" requirement can be met nowhere else).
• Change: the <<</>>> byte-boundary marker device from mds-proto/drive.mjs:23-26 in the mismatch message. Do not self-heal — §C.3-F forbids CI auto-regeneration. No snapshot API (toMatchSnapshot = 0 hits repo-wide; keep it that way; no __snapshots__ exists).
• Change [DR-03] — the update script takes a required named target. npm run test:golden:update must not regenerate both fixtures. As originally written, the single sanctioned command for the single sanctioned regeneration (P2-S16) would overwrite github-status-lines.txt as a side effect — and because that commit is by design "reviewed as a text diff" with the reviewer's attention on the git-agent.md diff, a silently-refreshed frozen fixture reads as noise, leaving AC-0.9 / AC-1.11 / AC-2.1 / AC-3.1 asserting byte-equality against a fixture quietly re-derived from the tree under test. Therefore:
  1. The script takes a required named target: npm run test:golden:update -- git-agent. With no target it exits non-zero and prints usage.
  2. Target github-status-lines refuses unless an explicit --unfreeze argument is also passed, and on refusal prints the §0.2 lifecycle rule verbatim ("frozen at Phase 0, never regenerated through Phase 3").
  3. A subprocess guard in tests/goldens/github-status-lines.test.ts asserts the refusal, mirroring the build-mds.test.ts:495-531 spawnSync shape, non-vacuous against a seeded invocation that omits --unfreeze.
  4. §0.2's Update-ritual sentence reads "writes the named fixture; github-status-lines.txt is refused through Phase 3" — not "writes both fixtures". Recorded in §14's lifecycle table.
• Ids: AC-0.2, GAP-06, §F.5-F; [DR-03].
• Deps: P0-S14, P0-S15.
• Verify: RED with a one-byte mutation of git.md; RED with a missing fixture (message names npm run test:golden:update -- git-agent); RED when the frozen target is invoked without --unfreeze.

P0-S24 — Non-contiguous v3 .gitignore fixture → A3 (A3b)

• Files: tests/shell-hooks.test.ts:1203-1414 and tests/init-logic.test.ts:320-438 (v3 describe begins :320). Existing v2→v3 cases: shell :1336, :1352, :1364; TS :349, :413-438.
• Change: add a fixture where the v3 line is non-contiguous — this repo's own v3 line sits at .gitignore:57, after two unrelated blocks [verified].
• Ids: §B.4 "retained despite the cancellation" — carve-out v4 is cancelled; the fixture still hardens the live v3 writers every user runs. EC-53 (temp HOME, seeded never empty), EC-54 (raw and realpath forms).
• Verify: RED against the current shell/TS writers if either assumes adjacency.

P0-S25 — test:integration in CI → A3 (A3b)

• Files: .github/workflows/ci.yml:23 — append npm run test:integration in the same job (the tarball build reuses npm run build's output).
• Context [verified]: vitest.integration.config.ts exists (testTimeout: 300000) and package.json:29 defines the script, but CI runs only npm run build + npm test, and vitest.config.ts explicitly excludes tests/integration/** — so tests/integration/pack-install.test.ts, the only automated check behind AC-1.3, never runs.
• Ids: AC-1.9, GAP-23. This is a Phase-1 blocker landed in Phase 0.
• Verify: the CI run for this PR shows the integration step executing and passing.

Stays local-only, deliberately, with the reason recorded: npm run test:golden:update (a CI regeneration makes the golden assert nothing) · any test requiring a live MCP server (GAP-46's six platform/MCP assumptions cannot be asserted by any test in this repo — they live in docs/reference/platform-assumptions.md as assumption · date verified · observable symptom if it drifts, with a capability → observable-symptom table that als

…truncated — full report in the local design artifact .devflow/docs/design/tracker-phase-0.2026-09-05_1155.md (not committed; ask the author)