Summary

NS-16 — Tier-4 plan-readiness audit of Plan-005 (Provider Driver Contract + Capabilities), Plan-006 (Session Event Taxonomy + Audit Log), and Plan-007 remainder (Local IPC + Daemon Control, Phases 4–7). Swaps the audited working copies into the corpus and lands the cross-cutting contract shapes the audit ratified. Plan-005 + Plan-006 stay approved; Plan-007 flips approved → review for the Tier-4 remainder Phases 4–7 design reopen — the Phases 1–3 partial PR sequence (#16/#17/#19) is audit-evidence preserved in §Tier 1 Partial PR Sequence + §Progress Log + §Retroactive Audit Memo.

Audit runbook: docs/operations/plan-implementation-readiness-audit-runbook.md. This is the Tier-4 pass (Tier-3 = NS-15 merged in PR #118; G6 satisfied).

Decisions ratified (D1–D7 + G3 overage)

G5 ephemeral-ref scrub

~100 F-XXX-Y-ZZ ID prefixes + PR #N inline refs stripped from audit-introduced plan-body text per Tier-3 Plan-003 canonical precedent (PR #118). Plan-007 baseline-preserved zones (§Tier 1 Partial PR Sequence, §Progress Log, §Retroactive Audit Memo) retain 13 PR #N refs as legitimate audit-evidence; narrative content preserved verbatim with only ephemeral prefixes scrubbed.

Anchor + path corrections (pre-commit-hook-surfaced)

The audit-introducer subagents emitted citation-shape defects the docs-anchor-check hook (lychee) caught at commit time. Root-cause fixes applied in this PR (not deferred):

• Plan-005 Required-ADRs — phantom ADR-004 citation with non-existent filename slug (004-event-streams-are-private-not-shared.md; real ADR-004 = SQLite/Postgres). Replaced with ADR-017 (Shared Event-Sourcing Scope) — the ADR the body's runtime_node.* driver-event emission actually invokes (line 271). Header + Preconditions + Done-Checklist all updated consistently.
• Plan-005 §CP-005-6 — broken Plan-004 filename slug (004-run-state-machine-and-events.md). Corrected to canonical 004-queue-steer-pause-resume.md.
• Plan-005 §CP-005-5 — two 4-up relative-path traversals (../../../../docs/architecture/contracts/api-payload-contracts.md). Corrected to 1-up (../architecture/contracts/api-payload-contracts.md).
• Plan-006 §Phase 4 narrative — Spec-015 fragment slug (#idempotency-classes not in target). Corrected to canonical #idempotency-classes-and-recovery-behavior.
• Plan-006 §Phase 4 narrative — forbidden .agents/tmp/research/... link target ((.) no-op markdown link to transient research path). Sentence rewritten per AGENTS.md Surface-Forward-Then-Delete anti-pattern (the load-bearing content is already surface-forwarded into this plan + the api-payload-contracts.md §Plan-006 doc-mirror).

File manifest

Tag on merge: plan-readiness-audit-tier-4-complete.

§6 fold (cross-plan-dependencies.md)

• NS-16 :::ready → :::completed; PRs checkbox [x] tier-4.
• NS-17 promoted blocked → ready per the audit-chain edge NS-16 → NS-17 (audit chain remains strictly serialized through NS-21).
• Ready set after this PR: {NS-17}; blocked set: {NS-09, NS-10, NS-18..NS-21}.

Findings ledger — 97 findings, 17 critical-root-causes (5.7:1 collapse ratio)

Per-plan breakdown: Plan-005 = 49 findings (5 critical-roots) · Plan-006 = 15 findings (6 critical-roots) · Plan-007 remainder = 33 findings (7 critical-roots, minus overlap = ~17 distinct closures). Healthy finding-instance / root-cause collapse ratio (~5.7:1) — each consuming phase independently surfaced the same upstream gap from its own vantage, not over-flagging (same pattern as Tier-3 NS-15: 31 findings → 4 roots = 7.75:1).

Root-cause critical → disposition:

Calibration + gates

B1–B6: B1 avg-critical/phase (Plan-005=1.25; Plan-006=1.5; Plan-007=1.75) in band; B2 total 97 (per-plan: 49/15/33 — Plan-005 over upper band, justified by 4-Phase × audit-density) · B3 Tasks:blocking (Plan-005=1.5:1; Plan-006=4.67:1) in band ✓ · B4 user-review ~60–90 min ✓ · B5 7 substantive advisor calls (range expanded vs Tier-3 baseline; Tier-4 contract-density justified) · B6 1 status-flip (Plan-007; 0–1) ✓

G1–G6: G1 skeleton additions-only ✓ · G2 all 17 root criticals dispositioned (full matrix in REVIEW.md) ✓ · G3 non-Tasks diff Plan-005 2.52×, Plan-006 1.42×, Plan-007 1.53× — Plan-005 + Plan-007 FAIL the literal <1.5× rule; ratified as audit-scope-justified per D-G3 overage (100% audit-domain content, zero redesign) · G4 all Tasks Steps → Spec coverage + Verifies-invariant ✓ · G5 0 ephemeral F-XXX-Y-ZZ / PR #N refs in audited plan bodies (Plan-007 §Tier 1 Partial PR Sequence + §Progress Log + §Retroactive Audit Memo preserve 13 PR refs as audit-evidence per baseline carryover; verified against plan-readiness-audit-tier-4-baseline tag) ✓ · G6 Tier-3 committed (NS-15 merged in PR #118) ✓

Test plan

• pnpm install --frozen-lockfile provisioned the worktree; all pre-commit + commit-msg hooks green (gitleaks: no leaks; lychee: 603 links / 532 OK / 0 errors / 71 excluded; docs-corpus-checks: governance cite-walk; lint-staged: prettier; commitlint: passed).
• 0 ephemeral F-XXX-Y-ZZ / PR #N tags in Plan-005 + Plan-006 + Plan-007 plan-bodies (Plan-007 baseline preserved zones retain 13 PR refs as legitimate audit-evidence per Tier-3 Plan-003 precedent).
• Mermaid coherence: §6 ready/blocked sets consistent with edge set; NS-16 → completed, NS-17 promoted.
• Anchor + path corrections verified — git show :path | grep confirms 0 remnants of ADR-004, 004-run-state-machine, 4-up traversal, #idempotency-classes standalone, or transient .agents/tmp/ link targets in the audited Plan-005 + Plan-006 surfaces.

Codex review resolution

Round 4 — 182ac84

Eleven findings (5×P1) resolved in 182ac84; all eleven review threads replied + resolved. Headline P1: F9 — daemon.start was modeled as an in-daemon IPC handler, which is impossible (a stopped daemon hosts no IPC server to receive the call). Removed DaemonStart* from the daemon-side surface (lifecycle schemas, IPC handlers, Tier-4 method-name table, client SDK list, renderer mutating-op gate, and the action:"start" union + shared DaemonLifecycleParams in api-payload-contracts.md); cold-boot is now the ai-sidekicks daemon start CLI process-spawn path (runtime-resolved daemon path + detached spawn + DaemonHelloAck wait, T-007r-3-4). This re-aligns Plan-007 to Spec-023:76's "spawned child process of the shell" framing — the plan had drifted from the spec, not the reverse. Other P1s: F11 (DaemonKeyStore real impl + composition-root injection at Plan-022 Tier 5 per CP-007-8, not R2 — no tier inversion), F1 (R1 registers only daemon.*+settings.*; the other five namespaces attach from their owning plans), F10 (banner read is daemon-availability-tolerant), F2 (evidence_pr: 124).

Round 5 — 34ee7f7

Seven findings (1×P1) resolved in 34ee7f7; all seven threads replied + resolved — 45/45 threads resolved, mergeStateStatus: CLEAN. Headline P1 R5-6: the round-2 stub_signature signed only the compacted-row payload projection, leaving the denormalized scalar columns (id, session_id, sequence, occurred_at, category, type, actor) — read by SQL filters and envelope reconstruction — unbound; an at-rest scalar edit could forge an actor/type while both stub_signature and the anchor over the frozen row_hash still verified. The verifier now adds a scalar-binding check (each scalar column byte-equals its payload-projection counterpart) → a new failureMode: 'stub_scalar_mismatch', an additive-MINOR enum extension 10→11 per ADR-018 §Decision #8. Other findings: R5-1 (DaemonStatusReadResult API mirror reconciled to the canonical processState shape), R5-2 (Spec-007 amended to disambiguate daemon start as a process-spawn capability, not a daemon.start IPC handler), R5-3 (separate DaemonStopRequest/DaemonRestartRequest mirror schemas carrying idleDrainDeadlineMs), R5-4 (Plan-005 DriverClient interface enumerates exactly the 7 client-facing methods, not the stale 11), R5-5 (Plan-005 storage summary lists all four Phase-2 SQLite tables), R5-7 (R2-T6 re-scoped to the emission-side securityCritical boundary R2 owns).

Round-4 deferrals, now resolved. The two round-4 "named but not chased" divergences are closed: (a) the api-payload-contracts.md DaemonStatusReadResult mirror is reconciled to the canonical processState shape (R5-1); (b) the Plan-007 CLI §Files inventory (one-vs-three daemon-lifecycle.ts) is fixed in 8631a1d, which also reconciled the broader R3 §Files block to its authoritative Tasks — added exit-codes.ts (T-3-3) + settings.ts (T-3-6), fixed the daemon-status.ts citation (T-3-3→T-3-5), and removed the session-*/meta.ts orphans (the CLI session surface is owned by Plan-001 Phase 5; --version/--help are clipanion Cli-builder built-ins). docs/specs/007-local-ipc-and-daemon-control.md:74 is now amended (R5-2) — a real defect in the approved spec's IPC-method listing, not an intentional carry-forward.

Remaining deferrals (each owned by its own pending audit). DaemonStart survives in the capability sense (the shell boots the daemon, realized by spawn) in three Tier-8 peer artifacts revisited by their own audits: docs/specs/023-desktop-shell-and-renderer.md:76,221, docs/plans/023-desktop-shell-and-renderer.md:62, docs/operations/local-daemon-runbook.md. Amending a peer plan/spec/runbook to restate something already correct in the capability sense is out of this PR's scope.

Refs: Plan-005, Plan-006, Plan-007, Spec-006, Spec-007, ADR-017, ADR-018, BL-134

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com