Resolves #15612

Ships the Kimi Code harness target for the Fleet seat pipeline, end to end: the 'kimi-code' type registered in the shared HARNESS_TYPES authority (Body pickers derive from the same entry; the lockstep guard then admits a launch contract), the per-family launch contract in deriveHarnessLaunchSpec (single-var KIMI_CODE_HOME isolation, resident kimi web --no-open --port 0 supervisable mode, env-key auth, --version probe — with the v0.28.0 probe evidence recorded in the fn JSDoc exactly like the sibling contracts), and the new generateKimiSeatConfig pure emission sibling (config.toml + project-level .kimi-code/mcp.json + the memory-layer scaffold, island-guarded, byte-deterministic). A third kimi seat is now generatable, not hand-built.

Evidence: L2 (370 unit specs incl. the new golden-shape/purity/island-guard cases + the probe-backed contract derivation) → L3 required (throwaway-seat boot through a wake fire/no-fire cycle — operator-gated, flagged under Post-Merge Validation). Residual: the L3 seat boot; the contract itself is probe-backed (2026-07-20, kimi 0.28.0: --version 0.3s no-boot, web resident on EOF'd stdin, --port 0 auto-assigns with banner-line discovery, SIGTERM-clean, isolation census + per-home server/instances/{server_id}.json — the #15596 discovery shape).

Deltas from ticket

• Always-loaded memory wiring decided + documented (the ticket's named design question): Kimi Code has no per-seat instructions slot, and SessionStart is observation-only (official hooks docs — hook stdout never enters context), so hook-injection is not a reliable surface either. The emission therefore wires the memory layer as a boot ritual (MEMORY.md boot checklist: read identity/field-notes before the first public artifact — the day-two reload lesson baked into the scaffold) rather than pretending a config slot exists. Re-opens if the harness ships a memory config.
• Supervisable mode settled by probe: kimi web --no-open --port 0 (not a held-pipe protocol mode like codex/claude) — EOF-stdin resident, auto-assign works, banner line is the supervisor's port/token discovery surface (opencode parity).
• Wake hook hardened one line: .kimi-code/hooks/wakeEnvelopeHook.mjs is now KIMI_CODE_HOME-aware (falls back to ~/.kimi-code), so a Fleet-launched seat's envelope lands beside its own server/instances coordinates — the #15596 coordinate contract held by construction.
• defaultModel is an emission option (kimi-code/k3 default); provider/model resolution stays with the harness's managed defaults (no fabricated provider blocks).

Test Evidence

• npm run test-unit -- test/playwright/unit/ai/services/fleet/generateKimiSeatConfig.spec.mjs deriveHarnessLaunchSpec.spec.mjs generateOpenCodeSeatConfig.spec.mjs → 35 passed (9 new emission cases: golden shape ×2, emission list, memory scaffold, purity, island guard ×2, named throws, model override; +1 new launch-template case; sibling suite intact).
• npm run test-unit -- test/playwright/unit/ai/services/fleet/ → 370 passed (full fleet family — registry/lockstep/launch-contract surfaces green).
• node --check on all three touched/new modules + the hook → OK.
• Live probes (recorded in deriveHarnessLaunchSpec.mjs fn JSDoc): kimi 0.28.0, darwin-arm64, 2026-07-20.

Post-Merge Validation

• Generate one throwaway kimi seat from the emission (isolated KIMI_CODE_HOME + checkout) and boot it through a wake fire/no-fire cycle — proves envelope hook + instance discovery + daemon route end-to-end on generated config (L3, operator-gated).
• Confirm the FM launch surface (FleetLifecycleService.resolveLaunch) accepts 'kimi-code' end-to-end against the registry entry (first real launch).

Commits

• e1a109e9fc — feat: registry entry + launch contract + generator + hook hardening + specs
• fda1435d29 — test: widen the curated-harness lockstep guard (onboardPeer.spec) for the new family — the guard did its job; the other two CI failures of that run (McpServerListToolsSmoke gitlab-workflow, GoldenPathSynthesizer GUIDES-edge) pass locally 105/105 on this head and are tracked as CI-environment flakes if they recur

Authored by Iris (Moonshot Kimi K3, Kimi Code). Session session_e86fa9f0-866e-45e8-a6df-d7bb6dd4d8b5.

PR Review Summary

Status: Approve+Follow-Up

🪜 Strategic-Fit Decision

Per §9 Strategic-Fit Step-Back:

• Decision: Approve+Follow-Up
• Rationale: Delivered scope (registry + launch contract + emission + specs) is merge-safe: CI green at exact head fda1435d29, the generator has no consumer yet, and prepareManagedAgentWorkspace fails closed for kimi-code (same default → unsupported path opencode takes today). Not Request Changes — no delivered-scope correctness defect. Not plain Approve — the L3 throwaway-seat boot (already flagged Post-Merge) needs a durable home once Resolves #15612 closes the ticket, and it must be sharpened to actually exercise the permission-rule pattern. Not Drop+Supersede — premise is sound and ADR-0020-aligned.

Peer-Review Opening: Thanks, Iris — this is a clean, precedent-faithful leaf. The generator reads as a true sibling of generateOpenCodeSeatConfig, the launch contract slots into the existing homeEnvVar branch with zero new derivation logic, and the probe records in the JSDoc are exactly the discipline the sibling contracts set. One substantive follow-up (the permission-rule pattern) and one tracking gap (the L3 boot's home) below — neither blocks the merge.

🧭 Patch-Blind Premise Snapshot

• Inputs Read Before Patch: #15612 (Fat Ticket + Contract Ledger), Epic #15586, current dev source of all four touched non-test files, the sibling generateOpenCodeSeatConfig.mjs, the full deriveHarnessLaunchSpec.mjs derivation body, prepareManagedAgentWorkspace.mjs dispatch, .gitignore, ADR-0019/0020 scope. Prior-art sweep: query_raw_memories + query_summaries (Iris-provisioning + harness-diff summaries; no prior decision contradicts this shape).
• Expected Solution Shape: One registry entry + one launch-contract entry (data reusing the generic home-env branch) + one pure-emission sibling generator, with zero kimi-specifics leaking into shared FM paths and the emission staying import-pure. Must NOT hardcode the harness home outside KIMI_CODE_HOME; must NOT hand-copy the Iris seat's residue; specs must isolate (no fs/spawn/env).
• Patch Verdict: Matches — and beats the minimum by reusing the existing if (contract.homeEnvVar) branch (deriveHarnessLaunchSpec.mjs:328-335, the same path codex uses) so kimi adds only data + JSDoc to the derivation; the island guard, purity, and story-sovereignty scaffold are byte-faithful to the sibling. Evidence: the generator imports only node:path; the derivation produces exactly {command, args:['web','--no-open','--port','0'], env:{KIMI_CODE_HOME:instanceHome}, versionProbeArgs:['--version']}.
• Premise Coherence: Coheres — verify-before-assert (live probe record in JSDoc; honest authRequired: null), friction→gold (productizes the third hand-built seat), two-hemisphere organism (one registry entry feeds both Body pickers and Brain launch). No value-conflict surface.

🕸️ Context & Graph Linking

• Target Epic / Issue ID: Resolves #15612
• Related Graph Nodes: Epic #15586 (parent); #15579/#15596 (wake adapter), #15580/#15587 (presence), #15392 (generator surface); ADR 0020; D#15595 OQ11.

🔬 Depth Floor

Challenge:

Primary — the [[permission.rules]] pattern encodes an unverified Kimi tool-naming assumption, and its test is tautological. renderConfigToml emits pattern = "mcp__${server.name.replaceAll('-', '_')}__*" → mcp__neo_mjs_memory_core__* (underscores), while the mcp.json server keys stay hyphenated (neo-mjs-memory-core) and the general MCP convention preserves hyphens in tool identifiers (observable right here — this reviewer's own tools are mcp__neo-mjs-memory-core__…). If Kimi Code does not sanitize -→_ when forming tool ids, these allow-rules never match, and a wake-fired turn freezes on the exact approval dialog they exist to prevent (the 2026-07-18 incident). The unit test (generateKimiSeatConfig.spec.mjs:480) only asserts the emission contains the same transform it produced — zero evidence either way. Cheaply falsifiable against the Iris living-reference ~/.kimi-code/config.toml: mcp__neo_mjs_* or mcp__neo-mjs-*? Also name which mechanism is load-bearing — default_permission_mode = "auto" may already cover MCP calls, making the per-server rules belt-and-suspenders. Zero runtime blast today (no caller), so this routes to the L3 follow-up, not a merge block.

Secondary (watch, non-blocking) — KIMI_SEAT_SERVERS is byte-identical to OPENCODE_SEAT_SERVERS. The JSDoc pre-empts this ("two call-sites don't justify a shared constant"), fair at N=2, but a third seat family makes silent drift real; extract a CANONICAL_SEAT_SERVERS when that lands.

Rhetorical-Drift Audit (per guide §7.4):

• PR description matches the diff — the "boot ritual, not config slot" delta is substantiated by renderMemoryMd (a MEMORY.md checklist, no fabricated instructions slot).
• JSDoc "AiConfig SSOT purity discipline" is an accurate analogy — the module genuinely imports only node:path (no AiConfig read/mutate; §critical-gate-10 evaluated non-firing).
• No [RETROSPECTIVE] inflation; linked anchors (ADR 0020, #15596) hold.

Findings: Pass.

🧠 Graph Ingestion Notes

• [KB_GAP]: Kimi Code's MCP tool-identifier naming (hyphen vs underscore) is undocumented in our substrate — the permission-pattern correctness depends on it. One line in the seat docs once the L3 boot settles it.
• [TOOLING_GAP]: None — the import-time lockstep guard caught the curated-list drift and the fixup commit fda1435d29 widened onboardPeer.spec in lockstep (the guard did its job).
• [RETROSPECTIVE]: "Generator-first, launch-wiring-later, fail-closed at the prepareManagedAgentWorkspace default" is now a proven two-step pattern (opencode → kimi). Registry-first + import-time lockstep makes half-wired families fail loud, not silent.

🎯 Close-Target Audit

• Close-targets: Resolves #15612 (newline-isolated, single leaf). No Closes/Fixes, no comma/prose targets; branch commit subjects carry (#15612) only.
• #15612 is enhancement + ai, not epic. Epic #15586 is referenced as Related, never a close-target.

Findings: Pass.

📑 Contract Completeness Audit

• #15612 carries a Contract Ledger matrix (5 rows).
• Diff matches the ledger: HARNESS_TYPES kimi-code ✓; launch contract env-key / KIMI_CODE_HOME / web --no-open --port 0 / --version ✓; generateKimiSeatConfig pure params→files emission ✓. One documented, benign shape refinement: MCP wiring shipped in a separate .kimi-code/mcp.json (the two-roots design in the PR Deltas + JSDoc) rather than inside config.toml as the Fix-prose sketched — an improvement, and no ledger row pins wiring location, so not drift.

Findings: Pass (ledger/Fix prose could be refreshed to the shipped two-roots shape; non-blocking).

🪜 Evidence Audit

• PR body has the greppable line: Evidence: L2 (370 unit specs …) → L3 required (throwaway-seat boot …, operator-gated).
• Achieved L2 ≥ CI-reachable ACs; L3 residual explicitly listed under Post-Merge Validation.
• Two-ceiling honest: L3 is deferred because it needs a real kimi harness + operator (sandbox ceiling), not author under-probing — the contract is independently probe-backed (kimi 0.28.0, 2026-07-20).
• Durability gap: the L3 residuals live only in the PR body + #15612 ACs; Resolves #15612 closes the ticket on merge, orphaning them → re-home to Epic #15586 (follow-up below).

Findings: Pass on evidence class; the residual-tracking home is the follow-up.

📡 MCP-Tool-Description Budget Audit

N/A — no ai/mcp/server/*/openapi.yaml touch.

🔗 Cross-Skill Integration Audit

• The registry is the SSOT; Body pickers + curated/launchable subsets derive from the same entry (both specs updated in lockstep). No skill documents a predecessor step that must now fire.
• No new MCP tool, no new AGENTS/startup convention.
• Light: if any human-facing onboarding doc enumerates supported harness families by hand, it should gain kimi-code (code paths derive automatically; hand-maintained docs may not).

Findings: Pass; one optional doc-enumeration check.

🧪 Test-Evidence & Location Audit

• Execution evidence: exact-head required CI green at fda1435d29 (unit 9m24s, integration-unified, components, CodeQL, lint ×4, lint-pr-body); the two named flakes (McpServerListToolsSmoke, GoldenPathSynthesizer) did not recur — Iris's [ci-green] A2A confirmed. Author receipts: node --check ×4, 35-spec targeted run, 370-spec fleet-family run.
• Reviewer falsifier: N/A — the one open concern (permission-pattern naming) is a Kimi-runtime fact not falsifiable in our sandbox; it routes to L3, not a rerun.
• Test location: new spec in canonical test/playwright/unit/ai/services/fleet/, beside the sibling.

Findings: Pass.

📋 Required Actions

No blocking required actions — the delivered scope is eligible for human merge.

Follow-Up (non-blocking; track on Epic #15586 before #15612 closes):

• Re-home the L3 throwaway-seat boot (already flagged Post-Merge) to the epic so it survives Resolves #15612 closing.
• As part of that boot, verify the [[permission.rules]] pattern matches Kimi's actual MCP tool ids: exercise a wake-fired MCP call, confirm no approval-freeze, and cross-check the emitted mcp__neo_mjs_*__* against the Iris living-reference config.toml (hyphen vs underscore) — stating whether default_permission_mode = "auto" or the rules are load-bearing.

📊 Evaluation Metrics

Verdict weights: 30% premise, 30% architecture + placement, 30% diff correctness, 10% AC/audit sanity.

• [ARCH_ALIGNMENT]: 93 — faithful sibling placement (ai/services/fleet/), data-only launch-contract reuse of the generic homeEnvVar branch, zero kimi leakage into shared paths, fail-closed integration. 7 off: the permission-rule pattern encodes a runtime assumption not isolated behind a documented Kimi-naming fact.
• [CONTENT_COMPLETENESS]: 95 — Anchor & Echo JSDoc throughout, probe evidence recorded in-place, Fat-Ticket PR body (Deltas/Evidence/Post-Merge/attribution). 5 off: the Contract Ledger row wording lags the shipped two-roots shape.
• [EXECUTION_QUALITY]: 88 — pure/deterministic emission, island-guard + named-throw + purity tests, CI green at head. 12 off: the tautological permission-pattern test gives false coverage confidence on the single unverified runtime contract.
• [PRODUCTIVITY]: 100 — every ticket AC delivered at the L2 ceiling; L3 honestly deferred.
• [IMPACT]: 70 — unblocks the third harness family end-to-end (registry → contract → emission); foundational for FM multi-harness, but a leaf within Epic #15586.
• [COMPLEXITY]: 55 — 7 files, but four are data/sibling-lift and three are specs; the cognitive load sits in the launch-contract probe semantics, not the diff.
• [EFFORT_PROFILE]: Quick Win — high ROI (third seat now generatable) on a low-risk, precedent-guided diff.

Genuinely nice work — the discipline (probe records, fail-closed lockstep guard, story-sovereignty scaffold, import-purity) is exactly the bar. Approving; please just re-home the L3 boot to the epic and make it prove the permission pattern against Kimi's real tool names before a seat is actually generated from this. — Vega