[aepod]

What this adds

ADR-402 makes Ruflo's Agent Teams work as a host-agnostic MCP surface instead of a Claude Code-only feature built on SendMessage/Task. The team bus (create/spawn/send/inbox/broadcast/plan/status/on_stop/shutdown) now lives in Ruflo as team_* MCP tools backed by a filesystem mailbox under .claude-flow/teams/{teamId}/, with team_spawn returning a spawn plan keyed by host (host.grok.spawn, host.claude.taskType, ...) so any host adapter can drive it. Claude Code's teammate-plugin stays as one adapter rather than the only one.

The first non-Claude adapter is Grok Build. npx ruflo init --grok scaffolds .grok/config.toml, .grok/rules/, .grok/agents/*.md, .grok/skills/, and a SubagentStop hook wired to team_on_stop, mirroring the existing init --codex pattern. On Grok Build 1.0.41 the lead calls spawn_subagent directly with the plan Ruflo hands back; it doesn't need Grok to have its own SendMessage-equivalent.

Also included:

• docs/grok/README.md and README-GROK.md: the operator guide and the monorepo/pre-publish guide for using Ruflo from Grok Build, including running it against an external project rather than the ruflo monorepo itself.
• scripts/bench-grok-host-conformance.mjs: a no-LLM conformance bench that checks the MCP/CLI surface Grok is supposed to call (host surface, tool inventory, team lifecycle, swarm, hive-mind, learning loop, neural, CLI parity, init product) — this is what backs the ADR-402 Phase 4 claim, not a live Grok run.
• scripts/probe-host-live.mjs / scripts/probe-grok-host-live.mjs: live, single-machine probes of an actually-installed Grok/Claude/Codex binary and its MCP config. These aren't part of the conformance claim and their sample output isn't included in this PR (see "left out" below).
• scripts/grok-team-bus.mjs, scripts/grok-subagent-stop-hook.mjs: the hook-side implementation the templates install.
• v3/@claude-flow/cli/src/mcp-tools/team-tools.ts and the memory-tools/memory-initializer changes needed to support it (tag-preserving memory export/import round-trips, team_* MCP tool registration).

Two merge conflicts came up bringing this current with upstream/main (v3/@claude-flow/cli/src/mcp-tools/memory-tools.ts and .../src/memory/memory-initializer.ts), both in code this branch didn't originally touch that upstream had since changed independently. Both were resolved to keep upstream's newer behavior (the dbPath-aware storeEntry call with a real success check, and the TTL-aware liveMemoryRowSql() helper) plus this branch's tag-preservation on import — not by picking one side wholesale.

The design record is ADR-402 (v3/docs/adr/ADR-402-host-agnostic-agent-teams.md). It was drafted as ADR-320 on the fork; 320, 400 and 401 are already taken upstream, so it is renumbered here.

Relationship to #3373

#3373 (open, fixes #3372) makes Grok Build a detected host: a row in defaultHostRegistry(), maybeAutoDetectGrok, and the trust instruction. This PR is complementary: it adds the team bus as team_* MCP tools and an explicit init --grok scaffold for Grok's agents, rules, skills and SubagentStop hook. It follows the same line on folder trust. init --grok prints the /hooks-trust / grok --trust step and never writes ~/.grok/trusted_folders.toml. Both PRs edit v3/@claude-flow/cli/src/commands/init.ts, so whichever lands second needs a rebase there; I'm happy to do that one.

How this was tested

All of this was run locally against the merged tree (upstream/main + feat/grok-host, after the fork-local cleanup below), node 22.23.1 / pnpm 11.13.0:

• pnpm install --frozen-lockfile in v3/ — clean, no lockfile changes needed.
• pnpm build (pnpm -r build, same as the CI build job) — all 25 workspace packages built with no errors.
• pnpm typecheck (pnpm -r typecheck, same as CI) — exit 0, no errors.
• pnpm test (vitest run, same as CI's test job) — full v3 suite. Failures seen: two memory-command tests in @claude-flow/cli/__tests__/commands.test.ts ("should search memory" / "should list memory entries") that need a real memory DB the test file never seeds (memory store is it.skip'd there as "requires live memory service"), plus a handful of graph-analyzer.test.ts and provider-integration.test.ts (Ollama-not-running) failures that vary between the pnpm-hoisted duplicate copies of the same spec file. I reproduced the same two commands.test.ts failures, deterministically, on a clean upstream/main checkout with no grok-host changes at all — they're pre-existing and unrelated to this branch, not something this PR introduces.
• The three extra scripts the CI test job runs from v3/@claude-flow/cli: node scripts/test-mcp-policy-stdio.mjs (10/10 pass), node scripts/smoke-flywheel-proof.mjs (PASSED), node scripts/replay-generation.mjs (PASSED, offline clean-room replay).
• node scripts/bench-grok-host-conformance.mjs (the feature's own bench) — 91/92 checks pass, 0 critical failures, 1 warning. The report ADR-402 points to (docs/benchmarks/grok-host-conformance-latest.md) is this run.
• The CI witness-verify job's check: node plugins/ruflo-core/scripts/witness/verify.mjs --manifest verification/macos/manifest.md.json --json against the built tree returns ok: true, regressed: 0 (89 pass, 28 drift, 0 missing) with upstream's manifest unchanged. This PR doesn't touch verification/.

What was deliberately left out of this branch, and why

The fork branch (weave-logic-ai/ruflo feat/grok-host) carried some things that only make sense there:

• data/clone-data.ledger.json/.proof.json/.rvf: the fork's own clone-tracker bot state, unrelated to this feature. Restored to upstream's version.
• docs/handoff.md: a personal session handoff note. Removed; doesn't exist upstream.
• A handoff Grok skill (.grok/skills/handoff/ and its templates/grok/skills/handoff/ copy). It wrote session notes to docs/handoff.md, a personal workflow rather than default Ruflo behavior, and init --grok would have installed it into every project. Removed along with its references in the bench file list, the init --grok output, and both Grok READMEs.
• docs/benchmarks/{grok-host,host}-live-latest.{json,md}: one-machine output from the live probe scripts (installed binary paths, local MCP config detection). Not referenced by any doc or script, so removed rather than scrubbed.
• docs/benchmarks/grok-host-conformance-latest.{json,md}: this one is referenced (ADR-402, README-GROK.md), so it's kept but regenerated fresh and with the CLI/repo path made repo-relative instead of this machine's absolute worktree path.
• verification/macos/: the fork had a regenerated witness manifest and a local witness run in history.jsonl, both keyed to the fork branch. Upstream's versions are kept; they verify clean against this tree.
• agentdb.rvf / agentdb.rvf.lock (root and v3/@claude-flow/cli/) and the related .gitignore edits: the fork untracked these as a side cleanup. Out of scope here, so they're left exactly as upstream has them.
• Fork issue-tracker ids in the memory-tools / memory-initializer / memory-bridge comments: stripped; the comments themselves stay.
• .grok/config.toml (the ruflo repo's own dev config, not the shipped template under v3/@claude-flow/cli/templates/grok/, which was already portable): switched the local-CLI MCP override to a path relative to the repo root and commented out the ruvnet-brain block instead of hardcoding this machine's home directory, matching the template's {{HOME}}-placeholder convention.

None of this touches the feature's actual surface — the MCP tools, the CLI, the templates, the ADR, or the operator docs are unchanged by the cleanup.

Follow-ups

• The conformance bench and the live-probe scripts overlap in what they check host discovery/config for; worth folding probe-host-live.mjs's installed-binary detection into the conformance bench (or clarifying the split in the scripts' own doc comments) rather than keeping them as three separate mjs entry points.
• verification/macos/manifest.md.json currently carries 28 "drift" entries (content changed since it was last signed, marker still present, not a regression) — worth a regen pass with real release versions once this lands, rather than as part of this PR.
• ADR-402 still needs the "live Claude Task/SendMessage side-by-side" comparison it calls out as unaddressed in the Phase 4 note.

🤖 Generated with RuFlo

[aepod]

Added 0dee8bb81 — feat(grok): show a loaded-status row from the user config.

init --grok appends [ui.status_line] to ~/.grok/config.toml when that block is missing. Grok Build 1.0.41 reads the row from the user file. The script prints whether the Ruflo MCP server, rules, four agents, two skills, the team hook, and folder trust are present.

[proffesor-for-testing]

Thanks Mat — the team-bus direction is right, and the core is in good shape: pnpm install --frozen-lockfile, the cli build and tsc --noEmit are clean (0 errors), the new team-tools / ollama-local-default-model tests pass (9/9), and 250 related existing tests pass on this head and on main (b14c79e) alike. The spawn plan does carry only prompt, description, background, isolation, and path traversal in team names is rejected.

Requesting changes for the items below. Each (R) was reproduced on head 0dee8bb81; (C) is from reading the code, cross-checked by a second reviewer (Codex).

Blocking

1. init --grok can corrupt a valid ~/.grok/config.toml (R). grok-generator.ts:136-149 detects an existing status line only by the literal header line [ui.status_line], then appends a new block. Existing valid configs using an inline table ([ui] + status_line = {…}), dotted keys (status_line.type = …), [ ui.status_line ] or [ui."status_line"] all become invalid TOML (tomllib: Cannot declare ('ui','status_line') twice). Grok would then presumably fail to read the user config.
2. Silent write to the user's home (R). grok-generator.ts:245-254 writes (or creates) ~/.grok/config.toml on every init --grok, with no opt-out, no dry-run, and the path is not listed in the init summary. Suggest: make it opt-in (e.g. --grok-statusline) or print the snippet for the user to paste; if writing, parse and merge with a TOML library.
3. Mailboxes are per agent name, not per team (R). team-tools.ts:148,253,450-458 — team_inbox validates team and then ignores it. A message sent to coder on team A was returned by team_inbox for coder on team B. team_status also counts other teams' mail.
4. Lost updates on team.json (R). team-tools.ts:169-180,341-366 and scripts/grok-team-bus.mjs:94-96 do read-modify-write with no lock and no atomic rename. 30 parallel team_spawn calls left 13/30 members registered. Parallel SubagentStop hooks can drop plan advances the same way.
5. The SubagentStop hook never advances the plan (R). The spawn plan sets description: "role:agent" (team-tools.ts:183-188); the hook falls back to description as the agent name (grok-subagent-stop-hook.mjs:51-72), and on-stop rejects coder:alice (: is not allowed). Stdio is ignored, so it fails silently — the plan stayed at ['ready','pending'].

Major

6. The advertised status row can't be reached (R). host-statusline.mjs:120 expects two skills (agent-teams-grok, handoff) but only agent-teams-grok ships, so after init the best output is RuFlo │ missing skills 1/2.
7. User-level status command runs a repo-relative script in every repo (C). grok-generator.ts:123 installs sh -c 'test -f scripts/host-statusline.mjs && node scripts/host-statusline.mjs' into the user config, so any repo opened in Grok that ships scripts/host-statusline.mjs gets it executed — unless Grok gates status-line commands on folder trust. Please pin it to an absolute path or a ruflo subcommand.
8. team_shutdown doesn't close the team, and maxAgents isn't enforced (C). team_shutdown only sets status (:650-661), but spawn and send check only that the team exists (:341, :390). maxAgents is stored (:305) and never checked at spawn.
9. Fresh install can't use what init advertises (C). The generated MCP config launches ruflo@latest (templates/grok/config.toml:12-28, which itself notes a local build is needed until a release ships team_*), while init.ts:449 prints team_create / team_spawn as next steps.
10. Windows (C). The status command requires sh, and a Windows home path inserted into a TOML basic string keeps backslashes (\U…), which is an invalid escape (grok-generator.ts:115-135).

Minor

• Broadcasts sent before any member exists go to _broadcast/ (:413-426), which team_inbox can never read (names must start alphanumeric) (R).
• priority is unvalidated: "abc" produced NaN_msg_….json; filenames sort lexically, so 10_ sorts before 2_ (R).
• A mailbox directory that is a symlink is followed on write (:417-431) (C).
• Two concurrent inbox drains throw ENOENT on renameSync; activeTeam() picks the first active team when several are active (C).

Scope

• The memory tag-filter / export v1.1 changes (memory-bridge.ts, memory-initializer.ts, memory-tools.ts) and the provider-precedence / Ollama default-model changes (agent-execute-core.ts:190,201,508-523) aren't Grok-related and change behaviour for existing users: with an explicit provider set, a stray OPENROUTER_API_KEY no longer wins, and claude-* model IDs now map to the Ollama default. Could those go in their own PR? (The tag filter also loads up to 100k rows into memory.)
• scripts/* are byte-identical copies of templates/grok/scripts/*, and the repo-root .grok/ and docs/grok/README.md have already drifted from the templates. The 1,785-line benchmark JSON, 1,101-line bench script and 464-line README-GROK.md would be easier to review separately.

CI / merge order

• No checks appear in the PR UI because this is a fork PR: 28 workflow runs on 0dee8bb8 are action_required, waiting for maintainer approval.
• #3513 is stacked on a0f8cd71e without 0dee8bb81; the only rebase conflict is scripts/probe-host-live.mjs (see the review there).

_{Review method: three agent reviewers built each PR in isolated worktrees and ran the new and related suites plus live CLI/MCP probes, compared against a main baseline; Codex CLI then cross-checked the findings and added the (C) items. — Dragan (QE)}

[aepod]

Thanks, Dragan. That was a thorough review, and the reproductions made the fixes easy to test. I force-updated the branch with a single commit on main (fa61df2). Point by point:

Blocking

1. ~/.grok/config.toml corruption. init now parses the user file with @iarna/toml, which the cli already depends on. If ui.status_line exists in any form (header, [ ui.status_line ], [ui."status_line"], dotted keys, inline table), init leaves it alone. Otherwise it appends the block, re-parses the result, and writes only if the new file parses, holds exactly our row, and leaves every other value unchanged. It writes through a temp file and a rename. A file it can't parse, or can't merge safely (for example ui = {…} as an inline table), is not touched, and init prints the snippet instead. grok-init.test.ts covers all of your cases. It also checks the merged output with Python tomllib when that's available.
2. Silent home write. Off by default now. Plain init --grok writes only inside the project and prints the snippet to paste. --grok-statusline opts in, --dry-run writes nothing, and the summary lists every path written, including anything under ~/.grok.
3. Mailboxes per team. Mail now lives in .claude-flow/teams/<team>/mailbox/<agent>/. team_inbox requires team, and team_status counts only that team. The A/B case you reproduced is a test.
4. Lost updates. Every team.json change takes a cross-process lock (mkdir team.lock, with stale-lock recovery) and is written through a temp file and a rename. There is now one implementation, used by the MCP tools, the CLI and the hook (see Scope below). New test: 15 CLI processes plus 15 MCP calls spawn at the same time, and all 30 register. I also removed the lock temporarily to check the test catches the bug: it failed with 16/30.
5. SubagentStop never advanced. The spawn description is now <role>:<agent>@<team>, and the hook parses it. The hook calls the bus directly instead of a child process whose output was discarded. If it can't record a team subagent (unknown team, shut-down team, agent on several teams), it prints the reason to stderr and exits 1. Subagents outside any team still exit 0 with no output. Tests: the plan goes to ['done','ready'], the ambiguous and missing-team cases fail with a message, and 12 hooks running in parallel lose no updates.

Major

6. Status row needed handoff. Removed. That was a personal skill that doesn't ship. After init, the row reports everything as loaded except folder trust, which is set in Grok (there's a test for this).
7. Repo-relative command. --grok-statusline installs a copy at ~/.grok/ruflo/host-statusline.mjs, and the command is node "<that absolute path>". It needs no sh, and no repo can supply its own script. If the user already has their own status row, init doesn't install anything.
8. Shutdown and maxAgents. team_shutdown closes the team: spawn, send, broadcast, plan and on-stop refuse afterwards. Inbox and status still work, so leftover mail can be drained. team_spawn enforces maxAgents; re-spawning an existing member doesn't take a new seat.
9. ruflo@latest. The generated project config pins the MCP server to the version that ran init (ruflo@<version>). That release is the one with team_*, since init and the tools ship together. Run from a source checkout, init points at that checkout's bin/cli.js and prints a note saying so.
10. Windows. The status command no longer needs sh. Every value inserted into TOML (home directory, CLI path) is escaped as a TOML string, and there's a test with C:\Users\… paths. On Windows the npx launch uses cmd /c npx, the same as .mcp.json.

Minor

• A broadcast with no members is now refused with a clear error. Mail is never parked in an unreadable _broadcast/.
• priority must be an integer from 0 to 999, and the filename prefix is zero-padded (002_…), so priority 10 sorts after 2. "abc" is rejected.
• Mailbox paths are created one directory at a time and a symlink anywhere in the path is refused. Message files are opened with wx.
• A drain claims each message by renaming it into archive/ before reading it. Two drains running at once each get a separate set of messages and neither throws (test: 40 messages, two processes).
• Several teams can be active at once. Every call names its team, and the hook resolves the team from the description as described in item 5. I removed activeTeam().

Scope

• The memory tag-filter / export v1.1 changes and the provider-precedence / Ollama default-model changes are out of this PR. They'll come as their own PR. So will the conformance bench, its JSON report, and README-GROK.md.
• The drift: the repo-root scripts/grok-*, scripts/host-statusline.mjs, .grok/ and docs/grok/README.md copies are deleted. templates/grok/ is the only source. The MCP tools now import templates/grok/scripts/grok-team-bus.mjs instead of carrying a TS copy, so the CLI and MCP paths can't drift either. scripts/probe-host-live.mjs stays, and it tells you to run init --grok in the checkout first.
• I also classified team_* in the capability brain. Without it, guidance-brain.test.ts failed on this branch.

Verification (fork head vs. main fa61df2): pnpm install --frozen-lockfile, the cli build, and tsc --noEmit all pass with 0 errors. The new suites pass: team-tools 12 tests and grok-init 21. On the full cli vitest run the branch has 7 failing tests and main has 7. Six fail on both, all in memory or dual-init tests (#2226, commands memory search/list, #2545 ×2, #2636). The seventh on main is #3286 auto-memory, which passes here. The seventh on the branch is guidance-brain, fixed by the capability-brain change above; it now passes alone and in the related suites.

One thing I couldn't verify: whether Grok runs hook commands through a shell on Windows. The SubagentStop hook still uses node "$CLAUDE_PROJECT_DIR/…", like the Claude hooks do.

[aepod]

One follow-up to the reply above: I force-updated to 19c1afe17, still a single commit, now on current main (3dce4664d).

• Stale-lock race. Your feat(teams): Codex and generic command hosts for the team bus (ADR-402) #3513 review flagged a race in breaking a stale lock, and the new team-store lock here had the same bug: a waiter that found an old lock deleted it outright, so it could remove a lock another process had just taken. The lock is now a team.lock file holding an owner token. A stale lock is renamed aside and its token re-checked, and it's put back if a live holder re-took it in between. Release deletes the lock only if it still holds the releaser's own token. New test: "breaks a stale lock left by a crashed holder". The team suite now has 13 tests and passed five runs in a row; team, init and guidance-brain together pass 42/42, and tsc shows 0 errors.
• Rebase. The only new commit on main (fix(gateway): align ChatGPT tool annotations with behavior #3525) touches plugins/ruflo-x-gateway only; the patch applies byte-identically.

[proffesor-for-testing]

Round 2 on 19c1afe17. Thanks Mat, this is a big improvement: all ten round-1 items are fixed, and we reproduced each fix. One new packaging issue blocks, and it's a one-line fix; once that's in, I'm happy to approve with nits.

Round-1 item	Result
1. TOML corruption	Fixed. Run from a CLI installed out of an npm pack tarball, with a temp HOME: inline table, dotted keys, [ ui.status_line ] and [ui."status_line"] stay byte-identical. Inline ui = {…}, [[ui]] and unparsable files fall back to the snippet. Merged files parse with python tomllib.
2. Home write	Fixed. Plain init --grok and --dry-run write nothing under ~/.grok.
3. Per-team mailbox	Fixed. Mail sent to coder on A doesn't show in B's team_inbox.
4. Lost updates	Fixed. 15 CLI + 15 MCP parallel spawns register 30/30, and a stale lock is broken. See N2/N3 for what's left.
5. SubagentStop	Fixed. coder:alice@H1 advances the plan; ambiguous, missing and shut-down teams print to stderr and exit 1; non-team descriptions exit 0.
6–9	Fixed. The status row is reachable once the folder is trusted. The status script lives at an absolute path under ~/.grok/ruflo/ and is refreshed on re-run. Shutdown closes spawn/send/broadcast/plan/on-stop. maxAgents is enforced. The MCP pin is ruflo@<version>.
10. Windows	Read only, not run: values are escaped as TOML strings and sh is gone.
Minors	Fixed. Empty broadcasts are refused. Priority abc/-1/1000/1.5 is refused and filenames are zero-padded. Whitespace-only messages are refused.

Blocking

N1: the claude-flow package doesn't ship the templates the team tools now load at runtime (R). The root package.json files list includes v3/@claude-flow/cli/dist/** but not v3/@claude-flow/cli/templates/**, and main has no templates directory at all today, so this PR is the first to depend on it. From an npm pack of the root package:

• mcp exec -t team_create fails with Cannot find module …/templates/grok/scripts/grok-team-bus.mjs;
• init --grok fails with "Grok templates not found".

The @claude-flow/cli tarball (which ruflo depends on) does include templates and works end to end, so this only hits npx claude-flow users, a published entry point (3.48.0). Fix: add v3/@claude-flow/cli/templates/** to the root files list.

Strongly suggested

• N2: stale-lock break can still admit two holders (R, narrow). grok-team-store.mjs:163-185: waiter W1 reads the stale token; meanwhile W2 breaks the lock and C takes a fresh one; W1 then renames C's live lock aside, D takes the free slot with wx, and restoring C's lock fails. A scripted replay of that ordering left C and D both holding the lock. It needs a crashed holder plus three waiters. Suggest: if the token changed, back off and retry instead of renaming.
• N3: lock wait (10 s) is shorter than the stale window (30 s) (R). After a kill -9 of a holder, spawns in the next ~20 s time out at 10 s, and the SubagentStop hook's own timeout is 5 s (subagent-stop-team.json:9). The wait also uses a synchronous Atomics.wait inside the MCP server, which blocks it (store:16-17,115).

Minor / nits

• --grok-statusline replaces a symlinked ~/.grok/config.toml (dotfiles setups) with a regular file at mode 0600 (grok-generator.ts:216-221) (R).
• An integer above 2^53 in the user config crashes init with "Do not know how to serialize a BigInt" after the project files are written. The user config is untouched (grok-generator.ts:162,250) (R).
• The status row matches text instead of parsing: trusted = true in a comment reads as trusted, and an MCP server with enabled = false still reads as ruflo (host-statusline.mjs:54-76) (R).
• @iarna/toml is TOML 0.5, so some valid TOML 1.0 (mixed-type arrays) falls back to the snippet. That's safe.
• Re-creating a team with force keeps its old mailbox (grok-team-bus.mjs:152-158) (C).
• Priority accepts 1e2 / 0x10.
• Stray changes: .claude/helpers/.helpers-version bump and router.js mode 644→755.

Build: pnpm install --frozen-lockfile passes, the cli build passes, and tsc --noEmit shows 0 errors. The swarm/shared/memory/neural/cli-core deps need to be built first. team-tools, grok-init and guidance-brain pass 42/42. Merges cleanly onto current main.

_{Round-2 method: an agent reviewer rebuilt the PR, re-ran the round-1 reproductions (including from an npm packed install), and Codex CLI cross-checked; new Codex claims were verified before inclusion. (R) = reproduced, (C) = read from code. — Dragan (QE)}

[aepod]

This addresses the blocker plus everything under "strongly suggested" and "minor."

Blocking

1. N1 (templates not shipped by the root package). Added v3/@claude-flow/cli/templates/** to the root package.json files array. Verified by npm pack at repo root, extracting the tarball, and confirming v3/@claude-flow/cli/templates/grok/scripts/grok-team-bus.mjs is actually in it this time.

Strongly suggested

2. N2 (stale-lock break can admit two holders). The bug was W1 acting on a token it read before re-checking it was still stale, then renaming a live lock aside instead of failing safe. It now re-reads the token immediately before breaking/renaming; if the token changed since the stale read, it backs off and retries the wait loop instead of touching the lock. Covered by a monkeypatched unit test that forces the race window, plus a real concurrent-child-process test (crashed holder, three waiters) asserting exactly one holder at a time.

3. N3 (wait window shorter than stale window, plus a sync Atomics.wait blocking the MCP server). Reconciled the timings to 4s wait / 4.5s stale threshold, comfortably inside the SubagentStop hook's 5s budget. Replaced the synchronous Atomics.wait poll with an async backoff loop threaded through the whole call chain (store, team-bus, the hook script, team-tools.ts), so a caller waiting on a lock no longer blocks the MCP server from serving anyone else.

Minor

4. --grok-statusline now writes through a symlinked ~/.grok/config.toml instead of replacing it with a plain file.
5. Config values above 2^53 are validated before any project files are written, so a BigInt-serialization failure can't leave a half-written init anymore. It either fails with a clear error naming the offending key, or it doesn't touch disk at all.
6. host-statusline.mjs parses the TOML with @iarna/toml (already a dependency here) instead of text-matching, so a trusted = true in a comment or a enabled = false server no longer confuses it.
7. Force-recreating a team now clears the old mailbox instead of carrying it over.
8. Priority parsing rejects 1e2 and 0x10 now too; plain base-10 integers only, same as it already rejected abc/-1/1000/1.5.
9. Reverted the stray .helpers-version bump and the router.js mode change (644→755). Neither belonged to this PR; both look like helper auto-refresh noise from a concurrent session.

Not touched, both intentionally: the @iarna/toml 0.5-vs-1.0 mixed-array fallback you flagged as safe, and round-1 item 10 (Windows), which was already verified fixed.

Build: npm run build and tsc --noEmit both clean. grok-init, team-tools, and guidance-brain suites: 41/41. Full @claude-flow/cli suite: 3816 passed, 20 failed. All 20 are pre-existing, in files that need @claude-flow/memory/@claude-flow/mcp/@claude-flow/codex built first in a fresh checkout; none touch grok/team/statusline code (checked by name and by grep).

[proffesor-for-testing]

Round 3 on 37147fa99: approving, with nits. Thanks Mat.

Verified by reproduction

• N1 (packaging): fixed. We ran npm pack --ignore-scripts at the repo root (the "Exit handler" error didn't occur here), and the tarball includes all four templates/grok/scripts/*.mjs. We installed it into a temp project with a sandboxed HOME: init --grok --grok-statusline, mcp exec -t team_create and team_spawn all succeed.
• Minor items: a symlinked ~/.grok/config.toml is written through and the link is kept. An integer above 2^53 is reported (naming the key) before any project file is written. The status line now parses TOML: # trusted = true doesn't count and enabled = false shows as missing. create --force clears the mailbox (3 files → 0). 1e2 and 0x10 are rejected through both the CLI and MCP.
• N3 wiring: the lock wait is async, and every caller awaits it (team-tools.ts, the stop hook, the CLI).
• Build and tests: tsc --noEmit passes, grok-init / team-tools / guidance-brain pass 49/49, and the branch merges cleanly onto current main.
• Stray files: nothing under .claude/ differs from main.

Follow-up suggestions (not blocking)

1. The lock is never refreshed while held, and it counts as stale after 4 s (grok-team-store.mjs:30-31). A holder that stalls for more than 4 s (SIGSTOP, laptop sleep, or the hook's 5 s timeout killing it mid-update) gets its lock broken by the next writer, and one of the two updates is lost. We reproduced this with a 6 s holder inside updateTeam: both writers read counter=0, and the final value was 1. The N2 race is also still reachable, but only with pauses injected between the token re-check and the rename (:219-229). 100 natural runs with 3 and 6 waiters showed no overlap. Checking that the pid in the lock token is still alive (process.kill(pid, 0)) before breaking the lock would close both on a single host.
2. The >2^53 check at grok-generator.ts:435 also rejects the case where the user already has their own [ui.status_line], where nothing would be serialized. Falling back to printing the snippet would keep that case working as before.
3. Nit: the header regex at host-statusline.mjs:111 rejects ] inside a quoted key, so a trusted folder like /x/a]b reads as untrusted.

_{Round-3 method: an agent reviewer rebuilt the head, re-ran the round-2 reproductions (including a root npm pack install), and ran timing-injected lock races; Codex CLI cross-checked, and we verified its claims before including them. — Dragan (QE)}

[aepod]

Rebased onto 3.56.9 (55e2c90). The three reviewed commits applied cleanly. One new commit, c6472d0: analyze-json-output.test.ts (new on main) failed because loading the MCP tool table pulled in @iarna/toml via grokTemplatesRoot; that now lives in init/grok-templates.ts. Build, typecheck, full suite (no test fails that passes on main) and team suites 72/72 pass.

[aepod]

Rebased onto main b93144803. New head 6aaaf82f8.

• Conflict: README.md only. Main added the "An origin you can inspect" section in the same place as the Grok Build line. I kept both, with the Grok line after main's section.
• Code: no changes. Range-diff shows commits 2-4 identical.
• ADR-402: still free on main (main has 400, 401, 404-409).
• Test V3 Packages: 9165 passed / 4 failed / 228 skipped of 9397. Every failing file also fails on main in the same environment (main: 9118 / 7 / 228 of 9353).
• Root Test Suite: no failing file that main doesn't also fail.
• Also: pnpm -r build clean, grok-init.test.ts 26/26, team-tools.test.ts 18/18.

Gates ran in clean Linux containers (node 22 for V3 Packages, node 20 for the root suite), compared against the same job on main in the same environment. Main shows some failing files there that pass in real CI, so the claim is only that this branch adds none.

#3513 is stacked on this one, so this can merge first.