[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7C4VM29KCQC98BMJ8BHAD2K":3,"skill-same:01M7C4VM29KCQC98BMJ8BHAD2K":167},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":14,"repositoryHtmlUrl":27,"repositoryHomepageUrl":6,"ownerHtmlUrl":28,"ownerBlogUrl":29,"ownerDescription":6,"updatedAt":30,"install":31,"resources":44},"01M7C4VM29KCQC98BMJ8BHAD2K","agent-routing",null,"Invoke before any git\u002Fgh and before starting another agent: dispatch a worker\u002Freviewer, delegate_task, hand off; who gathers, gemini gatherer, Codex model\u002Feffort, \u002Fmultitask, fan out in parallel. Pick provider\u002Fworktree. NOT for one edit or dependent steps.","EtanHey\u002Fgolems","EtanHey","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F93338119?v=4","01M7C4SKC9HJC3B9TAX0D9MZ1E","Etan Heyman","01M7C4SKCT59Z49TTN4WWDDEYD",0,"Apache-2.0","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: agent-routing\ndescription: \"Invoke before any git\u002Fgh and before starting another agent: dispatch a worker\u002Freviewer, delegate_task, hand off; who gathers, gemini gatherer, Codex model\u002Feffort, \u002Fmultitask, fan out in parallel. Pick provider\u002Fworktree. NOT for one edit or dependent steps.\"\n---\n\n# Agent Routing\n\n> This skill owns fleet routing rules, role selection, delegation checks, review ownership, and model\u002Feffort dispatch policy. `standards\u002Fmodel-roles.json` owns role → model; `\u002Frepogolem` owns launcher mechanics and canon #6 owns launcher law.\n\n> **Auto-dispatch triggers** (canonical in `\u002Forc` C4): batch reads >=3, transcription >=2,\n> web research >=1, or any \"in parallel\" \u002F \"all of these\" phrasing -> fan out sub-agents\n> in the SAME message before asking permission.\n\n## Kickoff contract\n\n- **Claude seats:** invoke `\u002Fagent-routing` with the Skill tool before the first dispatch of a lane: a worker or helper, a `delegate_task`, a hand-off, or a provider\u002Fmodel choice. Make it the first tool call, before any git\u002Fgh\u002Fgrep orientation: live misses open with `git status`\u002F`gh pr` and never load the skill (PR15 trigger eval, 2026-10-19).\n- **Codex seats:** read this SKILL.md body (a file read) before that action; Codex has no Skill tool.\n- **A lead's kickoff brief** to a worker names the skills the worker invokes (`\u002Flarge-plan`, `\u002Fagent-routing`, `\u002Fpr-loop`).\n- Standing ask: \"Once its PRs land, invoke the updated skills at the start of each plan, dispatch or PR action.\" (Etan 2026-10-18, via orc collab)\n- Why: Claude T3 leads made 0 Skill calls on Oct 16–18 while all three lead skills sat in their catalog (skill-edits audit, 2026-10-18).\n\n## Service tier and placement\n\n- Lead launch prompts say: \"Read and follow \u003Crepo>\u002FLEAD.md and the global LEAD.md\" (`~\u002F.codex\u002FLEAD.md`, installed from the private global instruction source). Neither shared instruction body imports the lead layer; delegates' briefs never name it.\n- Every delegate\u002Fsub-agent explicitly passes Claude `fastMode:false` or Codex `serviceTier:\"default\"`. Elevated service requires Etan's explicit request for that task.\n- Harness examples do not authorize elevated service.\n- Lead threads use the repo root checkout (`t3_thread_launch` with `workspaceStrategy:{type:\"root\"}`); only worker threads bind worktrees.\n- Use Theo's official T3 Code nightly; never build or install T3 from our fork.\n\nAttribution: Etan, 2026-10-26.\n\n## Routing rules (SSOT)\n\nThis section is the single source of truth for fleet routing. Fleet canon, global instructions, collab templates and other skills point here; models resolve through `standards\u002Fmodel-roles.json`.\n\n### Gathering\n\n**Antigravity is an underused pool (2026-11-08, Etan, via orc; corrected ruling).**\n\n> \"we gotta be using agy much more, … y'all arent doing what I asked for.\"\n>\n> \"ummm I never said all these need to go there??\"\n>\n> \"claude usage through agy is now natively supported in t3 code, so we can use this too. we now have 2 percent of it left for the next 15 hours so we should deplete it but also from now on we can utilize it (and maybe bench it against normal claude? to see if its on par with normal claude code same models?)\"\n\nDraw on Antigravity much more where it fits: Gemini Flash\u002FArgon AND Claude via agy, now natively supported in T3 Code. There is no blanket Gemini-first routing and no ban on agy Claude\u002FGPT. The 2026-11-01 agy-Claude rule and all existing routing guardrails stand unchanged. The quoted quota\u002Fwindow is the dated snapshot. A same-model bench of agy-Claude versus native Claude Code is pending, owned by orc; no result is claimed.\n\nEasy gathers go to Antigravity Flash-high at any depth (the 2026-11-02 rule below); Codex Luna\u002FTerra or Claude Haiku are the fallback when Antigravity is out (the 2026-10-28 rule below). Other gathers go to Antigravity first when Gemini Argon is out (the 2026-11-01 rule below). Cursor or a `gemini.gather.*` gatherer gathers and verifies; Gemini handles the helper-eligible shapes described in the Role Matrix and decision rules below. A gatherer never implements, reviews or decides. Cursor is Auto-only: never pass a model flag or model field (`default` is Auto); pinned Cursor drains its subscription pool. Route it as T3 `delegate_task` with `target: {providerInstanceId: \"cursor\", model: \"default\"}`. Write one brief file per child; dispatch `Read and follow \u003Cabsolute brief path>`. Each brief says `Report to \u003Ccollab path>, not to the lead thread.` Never use a headless CLI, AutoCursor runner, or background Cursor loop. Never select a named Cursor model; `default` is the T3 Auto sentinel.\n\n### Execution capability\n\nA task that must RUN COMMANDS (shell, ffmpeg\u002Fmedia conversion, builds, tests, installs, or file writes) never goes to a read-only gatherer profile (agy `gatherer`, no shell). The brief names the execution profile and required tool access.\n\nGemini spawns are read-only gathers only (2026-11-02, Etan, via orc): use the `gatherer` profile, or `video-qa` only to describe what frames show, with the QA verdict staying with Claude. Never send Gemini command execution, builds, tests, installs, fixes or file writes (the `shell-worker` profile included); route those to an executor with the required access under § Implementation and review.\n\n### Implementation and review\n\n#### Implementation default (Etan, 2026-10-29)\n\nHistorical quote (superseded by the 2026-10-28 default):\n\n> \"**EXPERIMENT, starting now for a few days: FLIP the roles.** Claude Opus 5.5 IMPLEMENTS and Codex gpt-6.1-sol REVIEWS.\" (Etan 2026-10-18, via orc collab)\n\n- Historical metrics remain: \"Track it in your STATUS: PRs merged per day, hours from open to merge, defects caught after merge. We compare at the end.\" (Etan 2026-10-18, via orc collab)\nEtan (2026-10-28, via orc\u002FPHX): use Sol\u002FTerra\u002FLuna for more of the work so the Claude 5-hour window survives; not all work needs Claude.\n\n- Default implementers (2026-10-29, replaces the 2026-10-28 Sol\u002FTerra default; Etan, via fleet control: Claude quality is generally higher on holistic work, Codex stays heavy on recon): holistic\u002Fproduct work (skill and doc text, UX\u002FUI, product behaviour, multi-file feature work, anything where judgment and coherence matter) → Claude Opus implements. Sonnet\u002FHaiku never implement; they are sub-agents only for trivial edits (changing some words). Narrow technical work (a bounded fix, a parser\u002Fguard\u002FCI script, a mechanical refactor, a dependency bump) → Codex implements (Sol at medium by default; high with one line of why; § Codex seats D, 2026-11-13). Claude also holds the leads. Codex workers keep fresh context and a bounded lifetime (§ Codex seats, 2026-11-13; it supersedes the recon note above), explicit effort and `serviceTier:\"default\"`.\n- Reviewers (2026-10-29) are always the other vendor: Sol reviews a Claude implementation, `claude.judgment` reviews a Codex implementation, with no exception. Gemini\u002FAntigravity gather and never review.\n- UI\u002FUX (2026-10-28): Claude Opus implements and reviews taste; Codex handles only mechanical parts. A fresh taste reviewer and the other-vendor reviewer remain required.\n- Leads are Claude only; Codex never holds a lead seat or long context. \"I think codexes should just be workers and not leads and not hold long context.\" (Etan, 2026-10-28)\n- **Easy work goes to Flash-high, at any depth (2026-11-02, Etan, via orc):** \"really got to put them(agy) in use, at least for the most easy tasks...... 3.8 flash high\" \u002F \"either leads, subagents or subagents of subagents should use them, not lead being agy\". Easy work is a `delegate_task` on provider `antigravity`, model `gemini-3.8-flash`, `options: {thinking_level: \"high\"}`, from a lead, its sub-agents and theirs. Easy work means ONLY gathers, reads, greps, inventories, log\u002Fscreen reads, simple read-only checks and summaries. Flash NEVER implements, builds, fixes, edits files or gives QA verdicts (2026-11-02, Etan, via orc, after a lane had Gemini implement); those keep their implementation and review routes. It excludes judgment recon and hazard recon: anything that decides a deletion or a test edit stays on `claude.judgment` (decision rule 2). The evidence below is why: Flash-High was 20\u002F20 on mixed gathers and 44\u002F44 on visual reads with 0 fabrications, but 0\u002F3 on deletion hazards. Claude verifies a Flash claim before anyone acts on it. Judgment, implementation and review keep their routes in this section. A lead seat is never agy.\n- Luna\u002FTerra\u002FHaiku (2026-10-28): the fallback for easy work when Antigravity is out, and the tier for mechanical edits at any time (never Flash); Sonnet 5.5 is a valid worker tier between Haiku and Opus.\n- **When Gemini Argon is out (2026-11-01, Etan, via orc):** gathers that are not easy work (easy work: the 2026-11-02 rule above) go to `antigravity` with `claude-opus-5-5` or `claude-sonnet-5-5` (Flash is only for easy gathers), never to Codex by default. Antigravity's Claude & GPT pool is separate from the Codex weekly pool and sat unused while Codex drained. Before picking a gatherer, check T3 **Settings > Usage limits**; a Codex gather is the fallback only when Antigravity is out too.\n- Keep the Claude 5h window alive (2026-10-28). Never fast\u002Fpriority: explicitly pass `fastMode:false` \u002F `serviceTier:\"default\"`. This supersedes the dated budget-pacing note; it does not revert at a window reset.\n\n| Work | Implements | Reviews | Plus |\n|---|---|---|---|\n| UX\u002FUI | `claude.judgment` | Codex (`codex.implement`) | — |\n| Security | `codex.security` at effort `high` | `claude.judgment` | a `codex-security` deep scan per security PR |\n| Holistic\u002Fproduct work (skill and doc text, product behaviour, multi-file features, anything where judgment and coherence matter) | `claude.judgment` (Opus; Sonnet\u002FHaiku only as sub-agents for trivial wording edits), 2026-10-29 | Codex Sol (`codex.implement`) | a deletion\u002Ftest-edit decision by `claude.judgment` first (decision rule 2) |\n| Narrow technical work (a bounded fix, a parser\u002Fguard\u002FCI script, a mechanical refactor, a dependency bump, tests, deletions) | Codex Sol at medium by default (high with one line of why), § Codex seats D, 2026-11-13 | `claude.judgment` | a deletion\u002Ftest-edit decision by `claude.judgment` first (decision rule 2) |\n| Ratchets and TLC\u002FTLA+ specs (ratchet rows, budgets, runners and attestation; formal models and their invariants) | `claude.judgment` (Opus) designs and builds, 2026-10-29 (Etan, via orc: highest quality, thought through). Tiny exception: Codex may build a mechanical edit to an existing Opus-designed threshold or row (no new runner, attestation, model or invariant design) | Codex Sol (`codex.implement`) reviews an Opus build; a fresh `claude.judgment` reviews a tiny Codex build, including design conformance | this row wins over \"CI script\" above; any other Codex-built ratchet or model gets an Opus design-conformance review and rework before merge; a lead self-edit also stays inside the Lead Topology cap |\n\nThe reviewer is always the other vendor.\n\nInterim route (Etan 2026-09-30\u002F10-01): Daybreak Blue requires a hardware security key from 2026-10-01. `codex.security` resolves to the interim model until Etan has keys; then one config line switches it back.\n\nSecurity effort is `high` on the interim route (Etan 2026-09-30: \"6.1 Sol high implements\"); a security phase records it like any other phase effort.\nDeep scans run from `\u003Crepo>Codex --scan -E high` (workspace-write, no approvals, computer use off, Locals T3code on by default, Drive read-only and other connectors off, the `codex.security` model, a per-launch Deep Scan cost cap); a bypass seat cannot start one. XS\u002FS security PRs get a diff-scoped Standard scan, M\u002FL or anything touching auth\u002Fguards a full Deep scan; the PR body records the scan's reported token use, and cost is never a reason to skip it.\n\nProven T3 Deep-scan route (try it before the interim below): from the lead, `delegate_task` with `runtimeMode: \"auto\"` on the `codex` instance, gpt-6.1-sol at high effort, serviceTier `default`, scope `.` (Deep scans the whole target). `auto` gives the managed workspace-write profile Deep needs; `full-access` disables it. The child's completion wakes the lead, so no `t3_thread_send` is needed. The start call blocks until the scan aggregates, so don't wait for an early scan id. Also proven end to end: a top-level T3 Codex thread with runtimeMode `auto`. It can't `t3_thread_send` to a `full-access` lead (`runtime_mode_escalation_denied`), so its brief names a report path and the lead reads that report. Proven 2026-10-27 on etanheyman.com #127 and a delegated start probe.\n\n**T3 seats (interim):** If the Deep scan can't start from a T3 seat (\"the parent must provide a managed filesystem permission profile\"), run a Standard scan with its scope widened to every touched auth\u002Fguard path, plus one full-repo Standard pass. Record in the PR body: \"Deep scan unavailable from T3 (interim)\", both scans' scope, and their reported token use. Tell the Claude security reviewer, before its round, that Deep was unavailable and which Standard scopes ran. Cost is never a reason to skip either pass.\n\nWhen Daybreak Blue returns, evaluate its first security PRs: if they show more review rounds or more defects than the prior route, flip the pair (`claude.judgment` implements, Daybreak Blue reviews) and record the evidence in the PR bodies (carried from canon #1, 2026-09-29).\n\n### Inner loop (sequential)\n\n- The implementer goes first. The reviewer is spawned or briefed only after the implementer reports done: its DONE marker or report line, or for a cloud implementer, PR head stable ≥10 min with checks finished.\n- A reviewer never reads a half-finished diff. The implementer may push and open the PR as a **draft** at any point, so the work is visible and linked. They iterate until both are happy; then the implementer marks the draft ready (`gh pr ready \u003CN>`) and runs the rest of `\u002Fpr-loop`. Merge follows the lane's merge authority after the PR loop is happy.\n- The LEAD routes the reviewer; a worker never starts its own. No reviewer pane means ask the lead. The lead is the escalation path, not a gate; its call is final when invoked, and it delegates the review decision to a reviewer worker.\n\nThis inner-loop pair review happens while the PR is a draft, before it is marked ready. `\u002Fpr-loop` bot and PR reviewers are separate and start at the ready flip. Pane mechanics live in `\u002Fcollab-monitor` section \"Completion -> Reviewer Handoff\".\n\n### Review sizing\n\nReview is sized to risk; the PR's `size:*` label sets the tier.\n\n- \"XS\u002FS: 1 reviewer, 1 round.\" (Etan 2026-10-18, via orc collab) That reviewer is the other vendor, per the table above; UI\u002Fstyle PRs follow the UI flow below.\n- M: up to 2 rounds, and round 2 only for BLOCKING findings (canon #9).\n- \"Full panel only for L, data or security paths, and that panel = Opus + Cursor working as a team. **NO Gemini Argon as a reviewer** (it fabricates).\" (Etan 2026-10-18, via orc collab) The panel is a fresh `claude.judgment` reviewer plus a read-only Cursor diff pass, giving one shared verdict; the Cursor pass is a read-only input to the Opus verdict, not a reviewer. When the implementer is `claude.judgment` (holistic\u002Fproduct or UX\u002FUI work), the table's Codex reviewer stays on, so cross-vendor review still holds. No Gemini model ever reviews, at any size.\n- UI\u002Fstyle PRs: \"a Codex worker uses computer use to record the changed states (default, hover, clicked, …), then hands the recording to an Opus reviewer.\" \"That reviewer loads the frontend-design skill (`frontend-design:frontend-design`, the latest Anthropic one) and judges the UX\u002FUI.\" (Etan 2026-10-18, via orc collab) The recorder resolves `codex.implement` and reports functional defects; proof requirements now follow [pr-loop Visual proof](..\u002Fpr-loop\u002FSKILL.md#visual-proof).\n- UI flow (Etan FRONTEND RULE, 2026-10-28): A Claude-implemented visual slice gets the red-team Claude (taste, BOTH skills) plus the table's Codex reviewer (cross-vendor, canon #1); Codex plumbing gets the red-team Claude, which is already the other vendor. Both reviews cite the Visual proof.\n  The red-team Claude is a fresh `claude.judgment`, never the implementer, loading `frontend-design:frontend-design` and `emil-design-eng`; normal size-based round caps and the full panel remain.\n\nClaude keeps the official auto-updating `frontend-design:frontend-design` plugin;\nCodex\u002FCursor use the pinned `frontend-design` skill. These names each denote one\nprovider route (Etan, 2026-10-27).\n\n### Records\n\nEach PR body records its implementer, review rounds, and bot\u002Freviewer defects. Claude leads orchestrate and route work through visible panes.\n\n### Model pins and dispatch\n\n- Fresh boots run `claude.judgment` at 1M via the bare launcher pin. The pin follows the role's current model and the pin is never removed; it prevents a prior session's model persisting.\n- Fable only via explicit per-invocation selection; a prior session's model never persists into the next.\n- Every non-Cursor Agent\u002FWorkflow\u002FTask spawn pins its model explicitly, resolved from roles. Cursor is the Auto-only exception.\n- Codex model list (Etan 2026-10-28): gpt-6.1-sol, gpt-6-luna and gpt-6.1-terra are allowed; the older two-model restriction is superseded. \"**Never gpt-5.6-\\*** (luna, sol, terra)\" \"gpt-6-astra only when Etan asks for it.\"\n  - Resolve role defaults through § Model roles; the 2026-10-28 Luna\u002FTerra fan-out permission is policy. Dispatch still respects configured candidate gates until the separate RT-2 model-role evaluation changes them.\n  - A request to use a `gpt-5.6-*` model (\"it's cheaper\") is refused, and any such pin you own is removed (2026-10-18 ruling).\n  - Astra needs Etan's own explicit ask; a lead's or worker's suggestion does not count (2026-10-18 ruling).\n  - Daybreak Blue's return to `codex.security` is not on this list. It needs Etan's ruling against the list before the swap (open question, 2026-10-18).\n- Effort is per `\u002Flarge-plan` phase; declare effort + why and pass it explicitly at dispatch. Codex effort caps: § Codex seats, rule D.\n- Claude concurrency (lead\u002Fworker caps): fleet canon #5 owns the numbers.\n- Usage is managed by default-pinning and dispatch-counting, not by usage-blocking buckets.\n\n### Codex seats (2026-11-13)\n\nEvery Codex dispatch follows this section. It is the only home of these rules; other skills point here.\n\nEtan (2026-11-12, via orc): \"I want y'all to start being a little more conservative with codex or finding the way to optimize it because I barely use codex or the chat GPT sub outside of T3 code and I want it to last longer.\"\n\nEvidence (orc's Codex usage audit, 6–12 Nov 2026, both Macs, 1,321 sessions): 13.3 billion tokens in 6.5 days, 98% of it cached input, the same big context re-read on every model call. So the burn is turn count times context size, not effort level alone. Sessions of 2 hours or more were 52% of everything; reading, searching and running scripts were 31%; review was 7%. One xhigh Codex lead seat (470 turns) was 6.4% by itself. Source: orc's 2026-11-13 03:05 post in `orchestrator\u002Fdocs.local\u002Fcollab\u002FCONVERGENCE.SYNCED.md`; report `orchestrator\u002Fdocs.local\u002Freports\u002F2026-11-13\u002FCODEX-USAGE-AUDIT.md`.\n\nThe rules, verbatim (orc, 2026-11-13 03:05; they replace rule 3 of orc's 02:46 post):\n\n- **A.** Codex seats are short. One task, one fresh child, done. No Codex lead seats. No Codex thread is resumed for a new task: start a fresh child with a short brief file instead (a resumed thread re-reads everything it ever saw on every call). A child that passes about 2 hours or 40 turns writes a five-line handoff and stops; a fresh child continues from the handoff.\n- **B.** A Codex child never waits inside a turn: no sitting through a test suite, a hook run, a build, an install or CI. It starts the long command in the background, writes where the result will land, and ends its turn; a script or a Sonnet or Haiku sub-agent watches; a fresh child reads the result. (B as written by orc, 2026-11-13.)\n- **C.** Nothing that is only reading, searching, tabulating, running scripts or tests goes to Codex (that was 31% of the week). Cursor, Gemini gatherers, Sonnet and Haiku do it.\n- **D.** Effort: Sol at medium is the default builder; Sol at high needs one line of why in the brief; xhigh needs orc's word in the inbox first; test, QA and script roles are never above low on any Codex model. Reviews: Terra at medium for small diffs.\n- **E.** Review stays on Codex untouched: it was 7% and it is what the subscription is for.\n- **F.** Keep the one-line daily count in your plan (turns dispatched to Codex, by purpose). orc reruns this audit next Thursday and names the lanes that did not change.\n\nHow they bind the rest of this skill:\n\n- A tightens § Lead Topology #5: a same-PR fix round may reuse a Codex child only while it is under both caps (one PR, about 2 h or 40 turns); past either cap, the five-line handoff goes to a fresh child.\n- C bans tasks, not activities: a Codex builder or reviewer still reads, searches and runs the checks its own task needs (a long run follows B). It withdraws the read-heavy Codex `recon` child in [references\u002Fmodel-and-effort.md](references\u002Fmodel-and-effort.md) and the Codex gather fallbacks in § Gathering and the 2026-10-28 tiers (Luna\u002FTerra for easy work, a Codex gather when Antigravity is out): those fall back to Cursor, Sonnet or Haiku instead.\n- D wins over `\u002Flarge-plan` § Choosing effort per phase for every Codex brief, and replaces the 02:46 \"Terra at low or medium for bounded and mechanical work\" line: the Codex builder is Sol at medium, bounded work included. A small diff is size XS or S; its Codex review runs on Terra at medium (allowed in § Model pins and dispatch). Larger reviews keep the table's reviewer, and a high effort carries its one line of why. Security implementation keeps its `high` (§ Implementation and review); its why line names the security route.\n- F's line goes in the plan's STATUS.md, one per day, for example `Codex 2026-11-13: 6 turns (review 4, build 2)`.\n\n## Model roles\n\nRoles live in `standards\u002Fmodel-roles.json`. From the golems checkout, resolve with\n`node scripts\u002Fmodel-roles.mjs \u003Crole> --field model|alias|launcher_tier`\n(select one field). Never hardcode a role-owned model name. Generated launcher\ncommands must keep the resolver substitution, not today's resolved literal. Read the role's status\nand gate before dispatch: `codex.subagent.mechanical` remains a candidate, with Etan’s\n2026-10-04 exception for Codex-internal mechanical children and named `packet` only. Watchers are\nthe second exception — \"anything that must sit and watch (a PR, CI, a file)\" (Etan 2026-10-18, via orc collab) —\nresolved with `node scripts\u002Fmodel-roles.mjs codex.subagent.mechanical --use watcher`. Workers\nremain on `codex.implement`; see the model-and-effort reference for the scoped exception.\nThe Routing rules (SSOT) section owns dispatch policy and the role config owns current model defaults; older model-selection recipes in the\nreferences are pending PR 2b migration and cannot override the config.\nEffort is not in the model-roles config; each `\u002Flarge-plan` phase declares effort + why, and every dispatch passes it explicitly (Codex `-E` \u002F `effort:`).\nThe launcher refuses a worker spawn without an explicit effort. Codex prompted\u002Fworker launches require `-E \u003Clevel>` or `GOLEM_EFFORT`; bare interactive launches use Codex config. For Claude, `default` means omit the flag: Claude sub-agents inherit the session's effort unless their agent frontmatter sets `effort:` ([Claude Code sub-agents: Supported frontmatter fields](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fsub-agents#supported-frontmatter-fields)). Gemini uses `launcher_tier`.\n\n## Read Map\n\n- Choosing a Codex model or effort, comparing model cost\u002Fcontext, dispatching a Codex child, or verifying its runtime? Read [references\u002Fmodel-and-effort.md](references\u002Fmodel-and-effort.md). Use the role config for current defaults and candidate gates; that reference supplies runtime verification and detailed procedures pending PR 2b.\n- Launching, reusing, monitoring, recovering, or closing a cmux worker lane? Read [references\u002Fdelegation-operations.md](references\u002Fdelegation-operations.md). Its monitor-loop and isolation mechanics are the cmux adapter.\n- Creating or auditing a collab, diagnosing a routing failure, or copying a routing template? Read [references\u002Fverification-and-incidents.md](references\u002Fverification-and-incidents.md).\n- Fanning out independent units, or asked about Cursor `\u002Fmultitask`? Read [references\u002Ffan-out-engines.md](references\u002Ffan-out-engines.md) (recipes, gotchas, GUI prompt contract, dispatch hygiene).\n- Dispatching child work from a T3 thread (`delegate_task`, review rounds, pending questions, PR watch)? Read [references\u002Ft3-native.md](references\u002Ft3-native.md). It owns the T3 child lifecycle (dispatch, wait, isolation, verify, settle), not delegation-operations.md: for a T3 child skip that file's monitor loop and no-worktree route.\n\nRead every reference triggered by the mission before dispatch. The live reference owns its detailed procedure; Routing rules (SSOT) and the role config take precedence over conflicting routing or model-selection recipes.\n\n## Role Matrix\n\n| Tool | Role | Does | Never does |\n|---|---|---|---|\n| **Cursor** | Gather | SQL, file\u002Fcode scans, grep, read-only lookups and audits | Changes files, implements, opens PRs, decides |\n| **`codex.implement`** | Implement \u002F review | Narrow technical implementation (Sol at medium by default, § Codex seats D) and cross-vendor review per SSOT (2026-10-29) | Research, data gathering, orchestration; a task that is only reading, searching, tabulating or running scripts or tests (§ Codex seats C). The reading and checks inside its own build or review task are part of that task |\n| **`codex.security`** | Implement (security) | Security implementation per SSOT | Gathering, orchestration, reviewing its own work |\n| **`gemini.gather.text` gatherer** (`{repo}Gemini -m $(node scripts\u002Fmodel-roles.mjs gemini.gather.text --field launcher_tier)`; this matches the launcher default) | Gather (text) | Doc\u002Flink\u002Fcopy audits, inventories\u002Fcounts, doc fetch+quote, local digests, BrainLayer recall | Running commands, implementing, reviewing, deciding (including a deletion or a test edit), UX\u002FUI judgment |\n| **`gemini.gather.visual` gatherer** (`{repo}Gemini -m $(node scripts\u002Fmodel-roles.mjs gemini.gather.visual --field launcher_tier)`; Pro-High via `-m pro` only when explicitly requested) | Gather (visual) | Frame\u002Fscreenshot reads, OCR, video state changes; the default image reader for `\u002Fqa-video` (its one `claude.judgment` judge judges this text and spot-checks ≤2 frames per finding) | Running commands, implementing, reviewing, deciding, UX\u002FUI judgment |\n| **`claude.judgment`** | Orchestrate \u002F judgment | Coordinates, talks to users, decides, synthesizes, monitors, queries BrainLayer; holistic\u002Fproduct and UX\u002FUI implementation and other-vendor pair review per SSOT | Bulk reads\u002FSQL; narrow technical implementation (Codex's lane) |\n\nDecision rules:\n\n1. Read-only query, scan, search, audit, or lookup -> Antigravity Flash-high per the 2026-11-02 SSOT rule when it is easy work (fallback: the 2026-10-28 tiers or Cursor; hazard recon: rule 2); for docs\u002Flink\u002Fcopy\u002Finventory gathers only, a `gemini.gather.text` gatherer; Pro-High only when explicitly requested; never a Low tier; single-fact sub-agent lookups follow the bounded exception below.\n2. Anything that decides a code deletion or a test edit (callers=0, dispatch sites, tool→test maps, \"safe to delete\") -> `claude.judgment`. A gatherer's grep may feed it; the gatherer never concludes it. This applies to Cursor gathers too.\n3. Code or file changes follow the Routing rules (SSOT) table: UX\u002FUI -> `claude.judgment`; security -> `codex.security`; holistic\u002Fproduct work (skill and doc text, product behaviour, multi-file features) -> Claude Opus (Sonnet\u002FHaiku only as sub-agents for trivial wording edits); narrow technical work (bounded fix, parser\u002Fguard\u002FCI script, mechanical refactor, dependency bump) -> Codex Sol at medium by default (§ Codex seats D); ratchets and TLC\u002FTLA+ specs -> Opus designs and builds (the table's ratchet row owns the tiny exception); the reviewer is always the other vendor (2026-10-29).\n4. Coordination, synthesis, monitoring, or decisions -> `claude.judgment`; pane mechanics and bounded verifiers follow the cheap sub-agent rule below.\n5. Mixed gather + implement work -> the gatherer (Cursor or Gemini) returns read-only findings; coordinating Claude records them under `docs.local\u002F`; the implementer selected by Routing rules (SSOT) implements from that handoff. A deletion or test edit gets its `claude.judgment` decision (rule 2) before implementation.\n6. Independent parallel units -> § Fan-out engine chooses the engine; Routing rules (SSOT) owns Cursor model selection.\n7. For video extraction\u002Fprocessing that runs commands, select the execution profile under § Execution capability first. Video QA, debrief or review -> `\u002Fqa-video` (2026-11-12, Etan, via orc): the requesting seat dispatches ONE `claude.judgment` judge as a delegate child (never a top-level thread) on the machine that holds the recording; the judge runs the extraction scripts, sends the frame windows to gatherers (Antigravity Flash-high, else `visual-batch.py`, else TOOLING FAILURE; never a Claude reader), judges their text over time and spot-checks ≤2 frames per reported finding; missing measurements become a script row, never eyeballing. Standalone frame OCR or bulk screenshot reads outside that pipeline -> a `gemini.gather.visual` gatherer; from an Agent-tool context, dispatch `visual-gatherer`; it runs `gemini.gather.visual` headless; use Pro-High only when explicitly requested.\n8. UX\u002FUI and design judgment stays on `claude.judgment`. Open-ended research: a Gemini gatherer may draft; the lead verifies before it reaches Etan. A gatherer never implements, reviews, merges, or decides.\n\n**Evidence (skill-creator eval, 2026-09-25; visible cmux workers, mechanical answer keys, lead-scored):** 40 bounded text-gather tasks: Flash-Low 40\u002F40 and Opus 5.5 40\u002F40, 0 fabricated claims each; higher Gemini effort on text gave the same accuracy 2.8–4.7× slower. 20 mixed tasks (8 image reads, 4 video): Opus 20\u002F20, Pro-High 20\u002F20 (2:35), Flash-High 20\u002F20 (6:17), Flash-Low 19\u002F20 (missed counting distinct screens across a video). An open-ended Pro research draft had a dead citation, a stale \"recent\" item, and missed the key release. Limits: screening sample (n=60) on one Mac; not evidence for judgment, design, review, or code. 2026-09-25 ruling: Low tiers are not used for gathering. On a real narrated QA review Flash-Low found 15–16\u002F22 vs Flash-High 21\u002F22, with an invented quote.\n\nVisual re-bench 2026-09-30 (22 visual items incl. 10 harder: near-dup screens, one-frame flash, 10 px OCR, contact-sheet diffs; n=2, headers verified): Flash-High 44\u002F44, 0 fabrications, ~11 min per batch; Pro-High 43\u002F44, ~3 min, fabricated its self-reported timing in 2\u002F2 runs → visual tier = Flash-High.\n\n### Evidence: deletion reachability\n\n**Evidence for rule 2 (cmuxlayer CX-3 real recon slices; one repo, 28 tools, scored against a truth table from fresh `rg`):** callers=0 for 28 tools about to be deleted: Opus surfaced 3\u002F3 deletion hazards; Flash-High and Haiku 0\u002F3 (Flash-Low, a lower tier, was not run on this slice), and Flash-High called all 28 safe, which would have broken a by-name engine accessor used by 23 test files; Pro made 59 false positives. Tool→test-file map errors: Opus 1, Flash-High 16, Pro 21, Haiku 23. Docs\u002Flink audit: Opus, Flash-High and Flash-Low 0 errors (82 s, 87 s, ≈270 s); Pro 47 false positives from a gitignored build directory; Haiku 20 errors (16 misses, 4 substring-match false positives). The eval above passed callers=0 because its tasks were literal greps against a known answer key; deciding a deletion needs reachability reasoning (by-name lookups, policy tables, agent-facing text).\n\n## Cheap sub-agents\n\nEasy work goes to Antigravity Flash-high first (the 2026-11-02 rule above), including one-shot\nread-only verifiers. This section covers two cases: the fallback when Antigravity is out, and the packaged\nagents below whose own tools (BrainLayer MCP, cmux MCP, deploy steps) a `delegate_task` child may not\ncarry. In those cases every `claude.judgment` seat (leads and workers) delegates parity tasks to\n`claude.subagent.cheap`; pass its resolved alias as the Agent tool's `model`:\n`node scripts\u002Fmodel-roles.mjs claude.subagent.cheap --field alias`.\n\n- **brain-worker:** single-fact recall (\"when did X merge\", \"what did Etan rule on Y\"); use one cheap child instead of the judgment seat's own context.\n- **orc-helper:** spawn\u002Fresume\u002Fsend\u002Fread\u002Fwait\u002Fclose pane mechanics; the packaged agent's `role:` pins its alias through the drift lint.\n- **Verifiers (one-shot), when Antigravity is out:** rerun a claimed command once, check now whether an artifact exists, or read a report's DONE marker.\n- **Deploy hops:** release → install → restart → live probe → seat reconnect go to the\n  `deploy-local-prod` sub-agent (Agent tool `subagent_type: \"deploy-local-prod\"`, model pinned to the\n  resolved `claude.subagent.cheap` alias). The lead routes the hop and verifies the artifact it reports.\n  Deploy hops stay on the MBP\u002FM4; the M1 is status-only for deploys. Test placement follows the M1 test-rig adapter (§ Dispatch Boundaries).\n\nAnything that sits, polls or waits (CI, a PR, a file appearing, a sleep-then-check) is the mechanical\ntier, never the lead; a one-shot check stays a cheap verifier above. From a Claude seat, pass\n`node scripts\u002Fmodel-roles.mjs claude.subagent.mechanical --field alias` as the Agent tool's `model`; for\na Codex watcher, use `node scripts\u002Fmodel-roles.mjs codex.subagent.mechanical --use watcher --field model`.\n`claude.subagent.cheap` keeps brain-worker recall and orc-helper, plus verifiers when Antigravity is out, per its 2026-09-29 evidence.\n\nKeep `claude.judgment` for multi-source or decision-grade history, deletion decisions,\nPR-gating reviews, and UX\u002FUI judgment. Reserve judgment → 3× cheap fan-out for\nhigh-stakes history only: it costs 3.4× and runs 2.6× slower. Pass this rule into\nevery judgment worker's brief. This is the bounded exception to general lookup routing.\n\n### Evidence: sub-agent and cleanup routing\n\n`skill-creator\u002Fdocs.local\u002Fevals\u002F2026-09-29-subagent-routing\u002FRESULTS.md`: plain brain-worker recall parity at 0.46×; orc-helper 4\u002F4 at 0.47×; verifiers 0 wrong at 0.55× (bounded cases; multi-source losses remain).\n`skill-creator\u002Fdocs.local\u002Fevals\u002F2026-09-29-cleanup-routing\u002FROUTING-R2.md`: deletion hazards Opus 6\u002F6 vs every Sol arm 4\u002F6; `codex.implement` effort medium, never default xhigh (n=2 per arm\u002Fcase).\n\n## Fan-out engine\n\nOnce the work is known to be N independent units, pick the parallelism engine. Full recipes and\nevidence: [references\u002Ffan-out-engines.md](references\u002Ffan-out-engines.md).\n\n**Cursor `\u002Fmultitask` is an in-editor GUI slash command, not a `cursor-agent` CLI feature.** Sent to\na headless agent, it degrades to one plain prompt answered by ONE agent that may claim it ran in\nparallel. A terminal\u002Fcmux orchestrator never uses it.\n\n| Task shape | Engine |\n|---|---|\n| Human in the Cursor editor, independent read-heavy prompts | Cursor `\u002Fmultitask` (GUI; `\\|\\|\\|`-separated; no write locking) |\n| Cursor child work | T3 `delegate_task`, `target: {providerInstanceId: \"cursor\", model: \"default\"}`; one pointer brief per child, report to collab |\n| Inside a Claude session; let the parent decide the split | Claude Workflow \u002F Agent tools |\n| Lead in a T3 thread; child work (implement, review, recon) owned by that thread | T3 `delegate_task` (§ T3-native dispatch) |\n| One coherent edit, or step B needs step A's output | None: run serially in one agent |\n\nFan out the read-heavy discovery; make the changes in one agent that holds the whole picture.\nVisible cmux fleets: § Dispatch Boundaries › cmux adapter.\n\n## T3-native dispatch\n\nA lead in a T3 thread dispatches child work through the T3 MCP, not panes. The lifecycle, as rules: [references\u002Ft3-native.md](references\u002Ft3-native.md).\n\n1. **Discover:** call `orchestrator_capabilities` first; pin provider, model and effort in `target` from its live catalog (canon #5).\n   For ordinary Codex implementers, recon and reviewers, explicitly name `target.providerInstanceId: \"codex-worker\"` every time.\n   `codex` stays for: security (implementers + scans), UI \u002F computer-use recorders, browser \u002F web-QA workers and Etan (Codex has no lead seats).\n   A Codex child inherits its parent's Codex instance. A Claude lead naming only the codex driver gets whichever codex instance T3 lists first; always name the instance.\n   Remember: the role picks the model, the work type picks the instance. Codex reviewers on security or full-panel PRs stay on `codex`; ordinary Codex reviewers use `codex-worker`.\n   Otherwise each Codex worker thread starts codex-security, computer-history, node_repl and cua_repl (~24.5 MiB, 4 processes), which T3 keeps until settle\u002Fidle release.\n   Recreate, upgrade verification and missing-instance fallback: [Codex instances](references\u002Ft3-native.md#codex-instances).\n2. **Dispatch:** follow the [worktree launch guard](references\u002Fworktree-launch.md) for exact-base preflight and preparation checks; `delegate_task` with `mode: \"async\"`; keep the `taskId`; one stable `clientRequestId` per round. Edits get `isolation: \"worktree\"`, which starts from the parent's committed HEAD, so the brief names absolute paths.\n3. **Wait:** completion wakes the parent. End the turn; never poll or spawn a watcher for it.\n4. **Settle:** at the PR's terminal state (merged or handed off; never at the first DONE — a PR in review rounds keeps its implementer) (a child with no PR, such as a reviewer or recon: as soon as its report is verified), read the report, verify the artifact (PR URL and head, files) before relaying, then conclude the child, `task_cancel` its `taskId`, `unwatch_pull_request` its PR, then `t3_thread_organize {action: \"settle\", threadId: \u003CchildThreadId>}`. Never omit threadId: an omitted threadId settles the calling lead thread. See § Goal Contract for provider release timing and the worktree caveat.\n5. **Review rounds:** every round is a new `delegate_task` carrying prior findings; never `t3_thread_send` to the `childThreadId`.\n6. **Questions:** a child's pending question is answered by the lead, never forwarded to Etan as a routing question.\n7. **Workers:** `delegate_task` children of their lead, `isolation:\"worktree\"` for code, never top-level threads. `t3_thread_launch`\u002F`create_threads` are only for a separate conversation explicitly requested by Etan. Cross-project work goes through that repo’s lead: `delegate_task` runs only in the caller’s project. \"I thought codex workers were just like implementer threads ... we usually just delegate workers and not put them in threads, right?\" (Etan, 2026-10-28)\n   Messages to a RUNNING thread use `t3_thread_send` mode `auto`, never `queue`. Never settle, re-snooze or unsnooze a thread Etan snoozed himself.\n8. **PRs:** `link_pull_request` every PR; the parent owns `watch_pull_request`.\n9. **Who and how many:** Routing rules (SSOT) picks the vendors; fresh implementer per PR and the Codex cap live in § Lead Topology #5.\n10. **Scheduled tasks:** a thread that exists for a `schedule_task` run ends every run by snoozing itself to its next run time (`t3_thread_organize {action: \"snooze\", threadId: \u003Cits own threadId>, snoozedUntil: \u003Cnext run, ISO>}`), so it never sits idle in Etan's sidebar between runs. Write that as the last step of the scheduled prompt. If Etan snoozed the thread himself, his snooze stands (#7). (Etan, 2026-11-13, via orc.)\n\n## Dispatch Boundaries\n\n- Cursor model selection lives in Routing rules (SSOT).\n- Claude Workflow\u002FAgent-tool fan-out is read-only recon, verification, or synthesis except audio-dashboard builds.\n- T3 child work follows § T3-native dispatch.\n- Easy standalone read-only lanes go to Antigravity Flash-high (the 2026-11-02 rule in Routing rules (SSOT)); the 2026-10-28 tiers or Cursor are the fallback.\n- Skills, hooks, agent definitions, and global agent settings are skillCreator-domain work. If skillCreator is not already in the loop, stop and request its audit before patching.\n- **A lane blocked by a guard never changes that guard.** No edit to the gate, its allowlist, policy or tests to get its own action through. It reports the block (the exact command and the refusal) to its lead, and the gate's owner decides. (Etan, 2026-11-13, via orc, after a lane opened a PR widening the human-confirm gate to post its own approved comment.)\n\n- Superseded placement: \"**The M1 Pro becomes a TEST RIG only.** Development stays on the MBP. … Nothing changes on the M1 for this yet.\" (Etan 2026-10-18, via orc collab) The 2026-10-28 placement below now governs.\n- M1 placement (Etan, 2026-10-28): \"I want the m1 pro to be used for tests.\" Full suites, E2E and XCUITest default to M1 Pro: ≥10 GiB guard, max 3 heavy lanes, never during a sync window, PHX admission required. Quick focused unit tests stay local.\n- Dispatching to the other Mac (linked T3 environment, remote code writes, project ids, admission ledger)? Read [references\u002Fcross-machine.md](references\u002Fcross-machine.md). It replaces the m1-router launch transport.\n\n### cmux adapter\n\nApplies to visible cmux fleets only; T3 children follow § T3-native dispatch.\n\n| Task shape | Engine |\n|---|---|\n| Visible, multi-vendor, long-running, human-watchable workers | cmux fleet (`\u002Fcmux-agents`) |\n\n- Visible-worker launch law lives in fleet canon #6; `\u002Frepogolem` owns invocation mechanics and [delegation operations](references\u002Fdelegation-operations.md#launcher-boundary) owns the `--fast` prohibition.\n- Codex workers edit only their delegated worktree; their Claude lead owns acceptance.\n- Pane monitoring and reviewer handoff: `\u002Fcollab-monitor`.\n\n### M1 test-rig adapter\n\nActive placement follows the 2026-10-28 M1 rule in § Dispatch Boundaries.\n\n- Codex workers only, until a recorded Claude-in-T3 proof passes (real `brain_search`, one skill invoked, no hook errors). Full suites, E2E and XCUITest use M1 Pro by default under the guard, lane cap, sync-window exclusion and PHX admission above.\n- Each dispatch requires a same-turn BrainBar probe: the socket accepts and `brain_search` answers. A failed probe is reported to the lead; the work stays on the MBP.\n- Leads never move. Vendor routing, inner-loop order and PR-body records are unchanged.\n\n## Lead Topology\n\nDomain leads are orchestrators one tier below orc:\n\n1. Leads delegate implementation according to Routing rules (SSOT); fleet canon #7 owns their worker-monitor guard.\n2. Lead goals preserve orchestration duties: delegate, maintain health gates, synthesize, and verify.\n3. A lead is a managed `agent_id` with `role:\"orchestrator\"` and left-column placement.\n4. Tiny lead self-edits are capped at \u003C=20 changed lines, one single-purpose change, and zero new files. They require an isolated worktree plus same-post collab disclosure of what changed, why urgent, and line count. Urgency alone never qualifies; everything else follows Routing rules (SSOT).\n5. **One PR = one fresh implementer.** A new PR gets a new implementer (a new `delegate_task` or a new pane, fresh context), even when a healthy idle worker exists in the same repo\u002Fworkspace\u002Frole. Only that PR's own fixes (its review rounds) may reuse it. Dependent stacks are no exception: each stack layer is its own PR, so it gets its own implementer, handed a compact checkpoint; the 2026-10-18 rules supersede R6's dependent-stack exception. Why: \"it re-uses workers instead of settling and opening new ones, like that ship 2246 chat that had 25 prs, 24 merged\" (Etan 2026-10-18, via orc collab).\n   - **Lifetime cap:** \"No Codex session lives past one PR, or ~100 turns. After that, start a fresh delegate with a short brief and a compact checkpoint. Never re-inject full plans or logs.\" (Etan 2026-10-18, via orc collab) Tightened 2026-11-13: about 2 h or 40 turns (§ Codex seats, rule A).\n   - **Reviewers:** \"reviewers fresh per implementer (same PR may reuse them)\" (Etan 2026-10-18, via orc collab).\n   - **Concurrency:** \"at most 4 Codex workers running at once per lead\" (Etan 2026-10-18, via orc collab). A fifth queues until one settles; the Claude dispatch cap in § Model pins and dispatch is separate.\n   - **Settle, then open new:** keep the implementer through its own PR's review rounds, not past them. At the PR's terminal state (merged or handed off, artifacts harvested) settle it, never at its first DONE; the next PR opens a new one.\n   - **cmux adapter:** canon #7's \"minimize pane churn\" is cmux fd-leak hygiene. It means no needless pane open\u002Fclose, never keeping one worker across PRs.\n\n6. A lead's model is written in its lane brief, signature line and claim post; a mid-sprint tier change names its reason (budget | quality | stall).\n7. **Leads never do ops themselves:** \"Leads never do ops themselves: deploy hops go to the `deploy-local-prod` sub-agent; mechanical checks and waits go to Haiku or gpt-6-luna.\" (Etan 2026-10-19, via orc) A deploy, install, restart or reconnect hop is a `deploy-local-prod` dispatch; a sit\u002Fpoll\u002Fwait task is a `claude.subagent.mechanical` or `codex.subagent.mechanical --use watcher` dispatch (§ Cheap sub-agents). Resolve both through `scripts\u002Fmodel-roles.mjs`; never write the model names. The lead's part is routing the hop and verifying the artifact it reports; the lead never runs the hop, the probe or the wait in its own session.\n8. **The lead shape (default, 2026-11-13).** Fable orchestrates a few Opus per domain; Opus hands reading and mechanical work to Sonnet or Haiku. A domain lead is a Fable seat that dispatches a few Opus (`claude.judgment`) workers for its domain's judgment and holistic work. Each Opus hands its reading to Sonnet (`claude.subagent.cheap`) and its mechanical work (sit, poll, wait, bounded checks) to Haiku (`claude.subagent.mechanical`), per § Cheap sub-agents. Fable is selected explicitly at each lead launch (§ Model pins and dispatch); the bare boot pin stays on `claude.judgment`, and #6 still records the lead's model. orc launches Fable leads: `model-pin-gate` allows a Fable target only from an apex seat, so a lead does not spawn Fable workers. The Routing rules (SSOT) table is unchanged: Codex keeps narrow technical work and other-vendor review, and gathers keep their routes.\n   - **Concurrency:** fleet canon #5 owns the numbers; this shape changes none. A Fable lead takes one of canon #5's Claude lead tracks and its place in the shared ceiling with leads. \"A few Opus\" means at most canon #5's `claude.judgment` worker cap running at once; the rest queue. Sonnet and Haiku draw on their own canon #5 caps.\n   - **Caveat:** it depends on the lead, and Saturday's weave should say whether Fables are better as advisers to Opus leads (orc's routing of card 5). Etan's words: \"Yeah, I think it depends on the lead, but I think some leads could be opposes [Opuses]. Maybe most of them should be fable but I think the weave should help us figure out if maybe fables should be more of advisors that come in to correct or give extra insight to opus leads before they make opus workers or soul [Sol] workers as well as maybe sometimes they should utilize fable for workers so we’ll need to see, but for now I think we should adopt that default\" (Etan, 2026-11-13, dashboard card 5, via orc). Until that weave (2026-11-14) rules, a lead may stay on Opus when its domain calls for it; the brief names why.\n\n## Goal Contract\n\nEvery complex or multi-hour dispatch uses one absolute file-backed goal containing the full user\nmission, constraints, green\u002Fno-green criteria, report path, and exact DONE marker. Name the CLI,\nexecution profile and required tool access, matched to the task under § Execution capability. Supersede an\nimplementer's goal only within its own PR (its review rounds); a new PR gets a fresh implementer (§ Lead Topology #5). A lane may\nclose only as verified `DONE`, file-backed `BLOCKED`\u002F`NOT_GREEN`, or `TRANSFERRED` with successor\nevidence. A DONE marker is only a prompt to verify the contracted artifact.\n\n**Stay on task:** a worker reads only its brief, the files it names, and repo instruction files. Never read other agents' reports, inboxes, briefs or collab threads not addressed to it unless the brief explicitly allows it.\n\n**Capability mismatch:** if the worker's profile\u002Ftools cannot do the task, stop and report one line to its lead: `BLOCKED: \u003Ctask> needs \u003Ccapability>; profile \u003Cx> lacks it`. Never substitute other work or explore.\n\n**Questions go to the lead:** \"workers never ask Etan routing questions\" (Etan 2026-10-18, via orc collab). A worker\nunsure about routing, scope, model or reviewer asks its lead. T3: the child's pending question, which the lead reads with\n`t3_pending_request_read` and answers with `t3_pending_request_respond` ([t3-native § 6](references\u002Ft3-native.md#6-pending-questions)).\ncmux adapter: one addressed collab line to the lead. The lead answers from this skill and escalates only a decision Etan\nalone can make, as an `@etan` ask in the collab. Etan alone decides only true one-way doors (deleting\ndata\u002Fcreds\u002Faccounts, money, messaging people); a tag, release, install, deploy or upgrade ships without asking (Etan\n2026-10-19, via orc; `\u002Fpr-loop` § Hard Gates).\n\n**Settle at the terminal state:** \"settle idle workers\" (Etan 2026-10-18, via orc collab). When a lane closes (verified\n`DONE`, filed `BLOCKED`\u002F`NOT_GREEN`, or `TRANSFERRED`), the lead settles its worker. T3: conclude the task (no new\nsends), `task_cancel` its `taskId` so queued runs stop, and `unwatch_pull_request` the lane's PR, then `t3_thread_organize {action: \"settle\", threadId: \u003CchildThreadId>}`. Never omit threadId: an omitted threadId settles the calling lead thread.\nThis releases its provider session (Claude: MCP children exit at once; shared Codex: the thread unloads and its MCPs close after Codex's unload delay); `task_cancel` alone does not detach.\nSettle once nothing still needs the child's worktree (a pending reviewer or fix round): T3's hourly cleanup removes settled threads' worktree checkouts (it keeps the branch).\nSource: T3 `storageCleanup.ts:85-90,1017-1029,1236-1261`. cmux adapter: harvest,\nthen close the pane through the cmuxlayer MCP (`close_surface`). Idle with no artifact pending = settle now, not later.\nA PR still in its review rounds is not idle; its implementer stays for those fixes (§ Lead Topology #5). The implementer\nremoves its own worktree at DONE (`\u002Fpr-loop` § Cleanup Receipt): `git worktree remove` (no `--force`),\nthen `git worktree prune`. Before lane close or settle, the lead verifies removal proof: the path is absent\nfrom `git worktree list --porcelain`. Dirty files or local-only commits stay; name the path and reason\nin the receipt and escalate to the lead. No proof or escalated exception = invalid receipt.\n\nIf the user questions why work is happening or corrects the route, pause spawning and patching,\nexplain the evidence-backed state, and store the correction separately.\n\n## Cross-Skill Ownership\n\n- `\u002Frepogolem`: launcher flags, defaults, resume behavior, and raw escape hatches.\n- `\u002Fpr-loop`: branch through draft PR to ready-for-review; Routing rules (SSOT) above owns the inner-loop pair, which runs on the draft.\n- `\u002Fcollab-monitor`: cmux\u002Fcollab-mail only: pane worker monitoring and reviewer handoff mechanics. T3 children use completion notifications instead (§ T3-native dispatch).\n- `\u002Fqa-video`: video QA\u002Fdebrief pipeline; one `claude.judgment` judge child runs where the recording lives, `gemini.gather.visual` gatherers (delegate_task \u002F visual-batch.py) read the frames, the judge judges their text and spot-checks.\n- `\u002Fwhats-new`: tool-surface changes; they do not revise Routing rules (SSOT) by implication.\n\n## Installed design skill constraints\n\nThese fleet constraints apply independently of the usage proposal below:\n\n- `improve-animations` is audit mode only (`quick` or `standard`). Its `execute`\n  mode is **never** used in the fleet: it self-dispatches and self-reviews.\n  Write plans only under `docs.local\u002Fanimation-plans\u002F`, never a tracked `plans\u002F`\n  directory. Send the plans to the lead, who routes implementation and review.\n  Any authorized read-only helpers are capped at two and follow normal model\u002Ftier\n  rules (default service, no fast; Cursor stays Auto-only).\n- `animate` must use the repo's existing dependencies. `pick-ui-library` is not\n  installed: never fetch it or a replacement library to satisfy that reference.\n\n## Design skills usage — FRONTEND RULE, 2026-10-28\n\nEtan: \"Split: Claude for visual slices\" (2026-10-28, via orc\u002FPHX).\nClaude (`claude-opus-5-5`, `fastMode:false`), loaded with BOTH\n`frontend-design:frontend-design` and `emil-design-eng`, implements taste-heavy\nvisual slices; Codex does the plumbing. This is the taste-critical exception to\nthe Claude budget hold; keep visual slices small. A fresh red-team Claude\n(`claude.judgment`, never the implementer) with BOTH skills reviews every visual\nPR and cites [pr-loop's required proof](..\u002Fpr-loop\u002FSKILL.md#visual-proof).\nCodex computer-use recorders may produce the artifacts and functional checks.\nClaude uses the official plugin; Codex\u002FCursor use vendored `frontend-design`.\nPlugin namespacing denotes one route, not a duplicate.\n\nSequence (lead guidance): `frontend-design` sets the holistic direction before planning\nand reviews the assembled full page at the end. `emil-design-eng` handles each\ninteractive component\u002Fsection inside that direction: feedback, keyboard\u002Ffocus,\nloading\u002Fempty\u002Ferror states, gestures, interruption and reduced motion. Holistic\nlayout\u002Ftokens win on visual cohesion; Emil governs local interaction craft.\n\n- VoiceBar and app UIs: holistic → every page\u002Fsection with Emil → holistic review\n  of the complete navigation\u002Fsettings flow. Adapt web recipes to native conventions.\n- BrainLayer, cmuxlayer, VoiceLayer, Golems and etanheyman.com websites: holistic\n  frontend-design pass → `\u002Flarge-plan` → each interactive row's implementer gets\n  both skills, the shared tokens and required states → holistic assembled-page pass.\n- Use `animate` to implement justified motion, `review-animations` explicitly for\n  a motion diff, `improve-animations` to audit\u002Fwrite plans, and\n  `find-animation-opportunities` for restrained proposals. Do not load the full suite\n  into every worker; no animation is a valid decision. Give Emil a concrete task\n  because a bare invocation only greets. `animate` references the uninstalled\n  `pick-ui-library`; use existing repo dependencies and never fetch it. The\n  audit-only, docs.local output and never-execute constraints above always apply.\n\nClaude keeps the official auto-updating\nfrontend-design plugin and the five vendored Emil skills. Codex uses all six\nvendored skills plus the upstream drift\u002Fupdater path. Cursor has the six skills\nfor routed read-only review only, never implementation. Gemini\u002Fagy receive none.\nAdditional Emil skills await Etan's decision.\n","# Agent Routing\n\n> This skill owns fleet routing rules, role selection, delegation checks, review ownership, and model\u002Feffort dispatch policy. `standards\u002Fmodel-roles.json` owns role → model; `\u002Frepogolem` owns launcher mechanics and canon #6 owns launcher law.\n\n> **Auto-dispatch triggers** (canonical in `\u002Forc` C4): batch reads >=3, transcription >=2,\n> web research >=1, or any \"in parallel\" \u002F \"all of these\" phrasing -> fan out sub-agents\n> in the SAME message before asking permission.\n\n## Kickoff contract\n\n- **Claude seats:** invoke `\u002Fagent-routing` with the Skill tool before the first dispatch of a lane: a worker or helper, a `delegate_task`, a hand-off, or a provider\u002Fmodel choice. Make it the first tool call, before any git\u002Fgh\u002Fgrep orientation: live misses open with `git status`\u002F`gh pr` and never load the skill (PR15 trigger eval, 2026-10-19).\n- **Codex seats:** read this SKILL.md body (a file read) before that action; Codex has no Skill tool.\n- **A lead's kickoff brief** to a worker names the skills the worker invokes (`\u002Flarge-plan`, `\u002Fagent-routing`, `\u002Fpr-loop`).\n- Standing ask: \"Once its PRs land, invoke the updated skills at the start of each plan, dispatch or PR action.\" (Etan 2026-10-18, via orc collab)\n- Why: Claude T3 leads made 0 Skill calls on Oct 16–18 while all three lead skills sat in their catalog (skill-edits audit, 2026-10-18).\n\n## Service tier and placement\n\n- Lead launch prompts say: \"Read and follow \u003Crepo>\u002FLEAD.md and the global LEAD.md\" (`~\u002F.codex\u002FLEAD.md`, installed from the private global instruction source). Neither shared instruction body imports the lead layer; delegates' briefs never name it.\n- Every delegate\u002Fsub-agent explicitly passes Claude `fastMode:false` or Codex `serviceTier:\"default\"`. Elevated service requires Etan's explicit request for that task.\n- Harness examples do not authorize elevated service.\n- Lead threads use the repo root checkout (`t3_thread_launch` with `workspaceStrategy:{type:\"root\"}`); only worker threads bind worktrees.\n- Use Theo's official T3 Code nightly; never build or install T3 from our fork.\n\nAttribution: Etan, 2026-10-26.\n\n## Routing rules (SSOT)\n\nThis section is the single source of truth for fleet routing. Fleet canon, global instructions, collab templates and other skills point here; models resolve through `standards\u002Fmodel-roles.json`.\n\n### Gathering\n\n**Antigravity is an underused pool (2026-11-08, Etan, via orc; corrected ruling).**\n\n> \"we gotta be using agy much more, … y'all arent doing what I asked for.\"\n>\n> \"ummm I never said all these need to go there??\"\n>\n> \"claude usage through agy is now natively supported in t3 code, so we can use this too. we now have 2 percent of it left for the next 15 hours so we should deplete it but also from now on we can utilize it (and maybe bench it against normal claude? to see if its on par with normal claude code same models?)\"\n\nDraw on Antigravity much more where it fits: Gemini Flash\u002FArgon AND Claude via agy, now natively supported in T3 Code. There is no blanket Gemini-first routing and no ban on agy Claude\u002FGPT. The 2026-11-01 agy-Claude rule and all existing routing guardrails stand unchanged. The quoted quota\u002Fwindow is the dated snapshot. A same-model bench of agy-Claude versus native Claude Code is pending, owned by orc; no result is claimed.\n\nEasy gathers go to Antigravity Flash-high at any depth (the 2026-11-02 rule below); Codex Luna\u002FTerra or Claude Haiku are the fallback when Antigravity is out (the 2026-10-28 rule below). Other gathers go to Antigravity first when Gemini Argon is out (the 2026-11-01 rule below). Cursor or a `gemini.gather.*` gatherer gathers and verifies; Gemini handles the helper-eligible shapes described in the Role Matrix and decision rules below. A gatherer never implements, reviews or decides. Cursor is Auto-only: never pass a model flag or model field (`default` is Auto); pinned Cursor drains its subscription pool. Route it as T3 `delegate_task` with `target: {providerInstanceId: \"cursor\", model: \"default\"}`. Write one brief file per child; dispatch `Read and follow \u003Cabsolute brief path>`. Each brief says `Report to \u003Ccollab path>, not to the lead thread.` Never use a headless CLI, AutoCursor runner, or background Cursor loop. Never select a named Cursor model; `default` is the T3 Auto sentinel.\n\n### Execution capability\n\nA task that must RUN COMMANDS (shell, ffmpeg\u002Fmedia conversion, builds, tests, installs, or file writes) never goes to a read-only gatherer profile (agy `gatherer`, no shell). The brief names the execution profile and required tool access.\n\nGemini spawns are read-only gathers only (2026-11-02, Etan, via orc): use the `gatherer` profile, or `video-qa` only to describe what frames show, with the QA verdict staying with Claude. Never send Gemini command execution, builds, tests, installs, fixes or file writes (the `shell-worker` profile included); route those to an executor with the required access under § Implementation and review.\n\n### Implementation and review\n\n#### Implementation default (Etan, 2026-10-29)\n\nHistorical quote (superseded by the 2026-10-28 default):\n\n> \"**EXPERIMENT, starting now for a few days: FLIP the roles.** Claude Opus 5.5 IMPLEMENTS and Codex gpt-6.1-sol REVIEWS.\" (Etan 2026-10-18, via orc collab)\n\n- Historical metrics remain: \"Track it in your STATUS: PRs merged per day, hours from open to merge, defects caught after merge. We compare at the end.\" (Etan 2026-10-18, via orc collab)\nEtan (2026-10-28, via orc\u002FPHX): use Sol\u002FTerra\u002FLuna for more of the work so the Claude 5-hour window survives; not all work needs Claude.\n\n- Default implementers (2026-10-29, replaces the 2026-10-28 Sol\u002FTerra default; Etan, via fleet control: Claude quality is generally higher on holistic work, Codex stays heavy on recon): holistic\u002Fproduct work (skill and doc text, UX\u002FUI, product behaviour, multi-file feature work, anything where judgment and coherence matter) → Claude Opus implements. Sonnet\u002FHaiku never implement; they are sub-agents only for trivial edits (changing some words). Narrow technical work (a bounded fix, a parser\u002Fguard\u002FCI script, a mechanical refactor, a dependency bump) → Codex implements (Sol at medium by default; high with one line of why; § Codex seats D, 2026-11-13). Claude also holds the leads. Codex workers keep fresh context and a bounded lifetime (§ Codex seats, 2026-11-13; it supersedes the recon note above), explicit effort and `serviceTier:\"default\"`.\n- Reviewers (2026-10-29) are always the other vendor: Sol reviews a Claude implementation, `claude.judgment` reviews a Codex implementation, with no exception. Gemini\u002FAntigravity gather and never review.\n- UI\u002FUX (2026-10-28): Claude Opus implements and reviews taste; Codex handles only mechanical parts. A fresh taste reviewer and the other-vendor reviewer remain required.\n- Leads are Claude only; Codex never holds a lead seat or long context. \"I think codexes should just be workers and not leads and not hold long context.\" (Etan, 2026-10-28)\n- **Easy work goes to Flash-high, at any depth (2026-11-02, Etan, via orc):** \"really got to put them(agy) in use, at least for the most easy tasks...... 3.8 flash high\" \u002F \"either leads, subagents or subagents of subagents should use them, not lead being agy\". Easy work is a `delegate_task` on provider `antigravity`, model `gemini-3.8-flash`, `options: {thinking_level: \"high\"}`, from a lead, its sub-agents and theirs. Easy work means ONLY gathers, reads, greps, inventories, log\u002Fscreen reads, simple read-only checks and summaries. Flash NEVER implements, builds, fixes, edits files or gives QA verdicts (2026-11-02, Etan, via orc, after a lane had Gemini implement); those keep their implementation and review routes. It excludes judgment recon and hazard recon: anything that decides a deletion or a test edit stays on `claude.judgment` (decision rule 2). The evidence below is why: Flash-High was 20\u002F20 on mixed gathers and 44\u002F44 on visual reads with 0 fabrications, but 0\u002F3 on deletion hazards. Claude verifies a Flash claim before anyone acts on it. Judgment, implementation and review keep their routes in this section. A lead seat is never agy.\n- Luna\u002FTerra\u002FHaiku (2026-10-28): the fallback for easy work when Antigravity is out, and the tier for mechanical edits at any time (never Flash); Sonnet 5.5 is a valid worker tier between Haiku and Opus.\n- **When Gemini Argon is out (2026-11-01, Etan, via orc):** gathers that are not easy work (easy work: the 2026-11-02 rule above) go to `antigravity` with `claude-opus-5-5` or `claude-sonnet-5-5` (Flash is only for easy gathers), never to Codex by default. Antigravity's Claude & GPT pool is separate from the Codex weekly pool and sat unused while Codex drained. Before picking a gatherer, check T3 **Settings > Usage limits**; a Codex gather is the fallback only when Antigravity is out too.\n- Keep the Claude 5h window alive (2026-10-28). Never fast\u002Fpriority: explicitly pass `fastMode:false` \u002F `serviceTier:\"default\"`. This supersedes the dated budget-pacing note; it does not revert at a window reset.\n\n| Work | Implements | Reviews | Plus |\n|---|---|---|---|\n| UX\u002FUI | `claude.judgment` | Codex (`codex.implement`) | — |\n| Security | `codex.security` at effort `high` | `claude.judgment` | a `codex-security` deep scan per security PR |\n| Holistic\u002Fproduct work (skill and doc text, product behaviour, multi-file features, anything where judgment and coherence matter) | `claude.judgment` (Opus; Sonnet\u002FHaiku only as sub-agents for trivial wording edits), 2026-10-29 | Codex Sol (`codex.implement`) | a deletion\u002Ftest-edit decision by `claude.judgment` first (decision rule 2) |\n| Narrow technical work (a bounded fix, a parser\u002Fguard\u002FCI script, a mechanical refactor, a dependency bump, tests, deletions) | Codex Sol at medium by default (high with one line of why), § Codex seats D, 2026-11-13 | `claude.judgment` | a deletion\u002Ftest-edit decision by `claude.judgment` first (decision rule 2) |\n| Ratchets and TLC\u002FTLA+ specs (ratchet rows, budgets, runners and attestation; formal models and their invariants) | `claude.judgment` (Opus) designs and builds, 2026-10-29 (Etan, via orc: highest quality, thought through). Tiny exception: Codex may build a mechanical edit to an existing Opus-designed threshold or row (no new runner, attestation, model or invariant design) | Codex Sol (`codex.implement`) reviews an Opus build; a fresh `claude.judgment` reviews a tiny Codex build, including design conformance | this row wins over \"CI script\" above; any other Codex-built ratchet or model gets an Opus design-conformance review and rework before merge; a lead self-edit also stays inside the Lead Topology cap |\n\nThe reviewer is always the other vendor.\n\nInterim route (Etan 2026-09-30\u002F10-01): Daybreak Blue requires a hardware security key from 2026-10-01. `codex.security` resolves to the interim model until Etan has keys; then one config line switches it back.\n\nSecurity effort is `high` on the interim route (Etan 2026-09-30: \"6.1 Sol high implements\"); a security phase records it like any other phase effort.\nDeep scans run from `\u003Crepo>Codex --scan -E high` (workspace-write, no approvals, computer use off, Locals T3code on by default, Drive read-only and other connectors off, the `codex.security` model, a per-launch Deep Scan cost cap); a bypass seat cannot start one. XS\u002FS security PRs get a diff-scoped Standard scan, M\u002FL or anything touching auth\u002Fguards a full Deep scan; the PR body records the scan's reported token use, and cost is never a reason to skip it.\n\nProven T3 Deep-scan route (try it before the interim below): from the lead, `delegate_task` with `runtimeMode: \"auto\"` on the `codex` instance, gpt-6.1-sol at high effort, serviceTier `default`, scope `.` (Deep scans the whole target). `auto` gives the managed workspace-write profile Deep needs; `full-access` disables it. The child's completion wakes the lead, so no `t3_thread_send` is needed. The start call blocks until the scan aggregates, so don't wait for an early scan id. Also proven end to end: a top-level T3 Codex thread with runtimeMode `auto`. It can't `t3_thread_send` to a `full-access` lead (`runtime_mode_escalation_denied`), so its brief names a report path and the lead reads that report. Proven 2026-10-27 on etanheyman.com #127 and a delegated start probe.\n\n**T3 seats (interim):** If the Deep scan can't start from a T3 seat (\"the parent must provide a managed filesystem permission profile\"), run a Standard scan with its scope widened to every touched auth\u002Fguard path, plus one full-repo Standard pass. Record in the PR body: \"Deep scan unavailable from T3 (interim)\", both scans' scope, and their reported token use. Tell the Claude security reviewer, before its round, that Deep was unavailable and which Standard scopes ran. Cost is never a reason to skip either pass.\n\nWhen Daybreak Blue returns, evaluate its first security PRs: if they show more review rounds or more defects than the prior route, flip the pair (`claude.judgment` implements, Daybreak Blue reviews) and record the evidence in the PR bodies (carried from canon #1, 2026-09-29).\n\n### Inner loop (sequential)\n\n- The implementer goes first. The reviewer is spawned or briefed only after the implementer reports done: its DONE marker or report line, or for a cloud implementer, PR head stable ≥10 min with checks finished.\n- A reviewer never reads a half-finished diff. The implementer may push and open the PR as a **draft** at any point, so the work is visible and linked. They iterate until both are happy; then the implementer marks the draft ready (`gh pr ready \u003CN>`) and runs the rest of `\u002Fpr-loop`. Merge follows the lane's merge authority after the PR loop is happy.\n- The LEAD routes the reviewer; a worker never starts its own. No reviewer pane means ask the lead. The lead is the escalation path, not a gate; its call is final when invoked, and it delegates the review decision to a reviewer worker.\n\nThis inner-loop pair review happens while the PR is a draft, before it is marked ready. `\u002Fpr-loop` bot and PR reviewers are separate and start at the ready flip. Pane mechanics live in `\u002Fcollab-monitor` section \"Completion -> Reviewer Handoff\".\n\n### Review sizing\n\nReview is sized to risk; the PR's `size:*` label sets the tier.\n\n- \"XS\u002FS: 1 reviewer, 1 round.\" (Etan 2026-10-18, via orc collab) That reviewer is the other vendor, per the table above; UI\u002Fstyle PRs follow the UI flow below.\n- M: up to 2 rounds, and round 2 only for BLOCKING findings (canon #9).\n- \"Full panel only for L, data or security paths, and that panel = Opus + Cursor working as a team. **NO Gemini Argon as a reviewer** (it fabricates).\" (Etan 2026-10-18, via orc collab) The panel is a fresh `claude.judgment` reviewer plus a read-only Cursor diff pass, giving one shared verdict; the Cursor pass is a read-only input to the Opus verdict, not a reviewer. When the implementer is `claude.judgment` (holistic\u002Fproduct or UX\u002FUI work), the table's Codex reviewer stays on, so cross-vendor review still holds. No Gemini model ever reviews, at any size.\n- UI\u002Fstyle PRs: \"a Codex worker uses computer use to record the changed states (default, hover, clicked, …), then hands the recording to an Opus reviewer.\" \"That reviewer loads the frontend-design skill (`frontend-design:frontend-design`, the latest Anthropic one) and judges the UX\u002FUI.\" (Etan 2026-10-18, via orc collab) The recorder resolves `codex.implement` and reports functional defects; proof requirements now follow [pr-loop Visual proof](..\u002Fpr-loop\u002FSKILL.md#visual-proof).\n- UI flow (Etan FRONTEND RULE, 2026-10-28): A Claude-implemented visual slice gets the red-team Claude (taste, BOTH skills) plus the table's Codex reviewer (cross-vendor, canon #1); Codex plumbing gets the red-team Claude, which is already the other vendor. Both reviews cite the Visual proof.\n  The red-team Claude is a fresh `claude.judgment`, never the implementer, loading `frontend-design:frontend-design` and `emil-design-eng`; normal size-based round caps and the full panel remain.\n\nClaude keeps the official auto-updating `frontend-design:frontend-design` plugin;\nCodex\u002FCursor use the pinned `frontend-design` skill. These names each denote one\nprovider route (Etan, 2026-10-27).\n\n### Records\n\nEach PR body records its implementer, review rounds, and bot\u002Freviewer defects. Claude leads orchestrate and route work through visible panes.\n\n### Model pins and dispatch\n\n- Fresh boots run `claude.judgment` at 1M via the bare launcher pin. The pin follows the role's current model and the pin is never removed; it prevents a prior session's model persisting.\n- Fable only via explicit per-invocation selection; a prior session's model never persists into the next.\n- Every non-Cursor Agent\u002FWorkflow\u002FTask spawn pins its model explicitly, resolved from roles. Cursor is the Auto-only exception.\n- Codex model list (Etan 2026-10-28): gpt-6.1-sol, gpt-6-luna and gpt-6.1-terra are allowed; the older two-model restriction is superseded. \"**Never gpt-5.6-\\*** (luna, sol, terra)\" \"gpt-6-astra only when Etan asks for it.\"\n  - Resolve role defaults through § Model roles; the 2026-10-28 Luna\u002FTerra fan-out permission is policy. Dispatch still respects configured candidate gates until the separate RT-2 model-role evaluation changes them.\n  - A request to use a `gpt-5.6-*` model (\"it's cheaper\") is refused, and any such pin you own is removed (2026-10-18 ruling).\n  - Astra needs Etan's own explicit ask; a lead's or worker's suggestion does not count (2026-10-18 ruling).\n  - Daybreak Blue's return to `codex.security` is not on this list. It needs Etan's ruling against the list before the swap (open question, 2026-10-18).\n- Effort is per `\u002Flarge-plan` phase; declare effort + why and pass it explicitly at dispatch. Codex effort caps: § Codex seats, rule D.\n- Claude concurrency (lead\u002Fworker caps): fleet canon #5 owns the numbers.\n- Usage is managed by default-pinning and dispatch-counting, not by usage-blocking buckets.\n\n### Codex seats (2026-11-13)\n\nEvery Codex dispatch follows this section. It is the only home of these rules; other skills point here.\n\nEtan (2026-11-12, via orc): \"I want y'all to start being a little more conservative with codex or finding the way to optimize it because I barely use codex or the chat GPT sub outside of T3 code and I want it to last longer.\"\n\nEvidence (orc's Codex usage audit, 6–12 Nov 2026, both Macs, 1,321 sessions): 13.3 billion tokens in 6.5 days, 98% of it cached input, the same big context re-read on every model call. So the burn is turn count times context size, not effort level alone. Sessions of 2 hours or more were 52% of everything; reading, searching and running scripts were 31%; review was 7%. One xhigh Codex lead seat (470 turns) was 6.4% by itself. Source: orc's 2026-11-13 03:05 post in `orchestrator\u002Fdocs.local\u002Fcollab\u002FCONVERGENCE.SYNCED.md`; report `orchestrator\u002Fdocs.local\u002Freports\u002F2026-11-13\u002FCODEX-USAGE-AUDIT.md`.\n\nThe rules, verbatim (orc, 2026-11-13 03:05; they replace rule 3 of orc's 02:46 post):\n\n- **A.** Codex seats are short. One task, one fresh child, done. No Codex lead seats. No Codex thread is resumed for a new task: start a fresh child with a short brief file instead (a resumed thread re-reads everything it ever saw on every call). A child that passes about 2 hours or 40 turns writes a five-line handoff and stops; a fresh child continues from the handoff.\n- **B.** A Codex child never waits inside a turn: no sitting through a test suite, a hook run, a build, an install or CI. It starts the long command in the background, writes where the result will land, and ends its turn; a script or a Sonnet or Haiku sub-agent watches; a fresh child reads the result. (B as written by orc, 2026-11-13.)\n- **C.** Nothing that is only reading, searching, tabulating, running scripts or tests goes to Codex (that was 31% of the week). Cursor, Gemini gatherers, Sonnet and Haiku do it.\n- **D.** Effort: Sol at medium is the default builder; Sol at high needs one line of why in the brief; xhigh needs orc's word in the inbox first; test, QA and script roles are never above low on any Codex model. Reviews: Terra at medium for small diffs.\n- **E.** Review stays on Codex untouched: it was 7% and it is what the subscription is for.\n- **F.** Keep the one-line daily count in your plan (turns dispatched to Codex, by purpose). orc reruns this audit next Thursday and names the lanes that did not change.\n\nHow they bind the rest of this skill:\n\n- A tightens § Lead Topology #5: a same-PR fix round may reuse a Codex child only while it is under both caps (one PR, about 2 h or 40 turns); past either cap, the five-line handoff goes to a fresh child.\n- C bans tasks, not activities: a Codex builder or reviewer still reads, searches and runs the checks its own task needs (a long run follows B). It withdraws the read-heavy Codex `recon` child in [references\u002Fmodel-and-effort.md](references\u002Fmodel-and-effort.md) and the Codex gather fallbacks in § Gathering and the 2026-10-28 tiers (Luna\u002FTerra for easy work, a Codex gather when Antigravity is out): those fall back to Cursor, Sonnet or Haiku instead.\n- D wins over `\u002Flarge-plan` § Choosing effort per phase for every Codex brief, and replaces the 02:46 \"Terra at low or medium for bounded and mechanical work\" line: the Codex builder is Sol at medium, bounded work included. A small diff is size XS or S; its Codex review runs on Terra at medium (allowed in § Model pins and dispatch). Larger reviews keep the table's reviewer, and a high effort carries its one line of why. Security implementation keeps its `high` (§ Implementation and review); its why line names the security route.\n- F's line goes in the plan's STATUS.md, one per day, for example `Codex 2026-11-13: 6 turns (review 4, build 2)`.\n\n## Model roles\n\nRoles live in `standards\u002Fmodel-roles.json`. From the golems checkout, resolve with\n`node scripts\u002Fmodel-roles.mjs \u003Crole> --field model|alias|launcher_tier`\n(select one field). Never hardcode a role-owned model name. Generated launcher\ncommands must keep the resolver substitution, not today's resolved literal. Read the role's status\nand gate before dispatch: `codex.subagent.mechanical` remains a candidate, with Etan’s\n2026-10-04 exception for Codex-internal mechanical children and named `packet` only. Watchers are\nthe second exception — \"anything that must sit and watch (a PR, CI, a file)\" (Etan 2026-10-18, via orc collab) —\nresolved with `node scripts\u002Fmodel-roles.mjs codex.subagent.mechanical --use watcher`. Workers\nremain on `codex.implement`; see the model-and-effort reference for the scoped exception.\nThe Routing rules (SSOT) section owns dispatch policy and the role config owns current model defaults; older model-selection recipes in the\nreferences are pending PR 2b migration and cannot override the config.\nEffort is not in the model-roles config; each `\u002Flarge-plan` phase declares effort + why, and every dispatch passes it explicitly (Codex `-E` \u002F `effort:`).\nThe launcher refuses a worker spawn without an explicit effort. Codex prompted\u002Fworker launches require `-E \u003Clevel>` or `GOLEM_EFFORT`; bare interactive launches use Codex config. For Claude, `default` means omit the flag: Claude sub-agents inherit the session's effort unless their agent frontmatter sets `effort:` ([Claude Code sub-agents: Supported frontmatter fields](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fsub-agents#supported-frontmatter-fields)). Gemini uses `launcher_tier`.\n\n## Read Map\n\n- Choosing a Codex model or effort, comparing model cost\u002Fcontext, dispatching a Codex child, or verifying its runtime? Read [references\u002Fmodel-and-effort.md](references\u002Fmodel-and-effort.md). Use the role config for current defaults and candidate gates; that reference supplies runtime verification and detailed procedures pending PR 2b.\n- Launching, reusing, monitoring, recovering, or closing a cmux worker lane? Read [references\u002Fdelegation-operations.md](references\u002Fdelegation-operations.md). Its monitor-loop and isolation mechanics are the cmux adapter.\n- Creating or auditing a collab, diagnosing a routing failure, or copying a routing template? Read [references\u002Fverification-and-incidents.md](references\u002Fverification-and-incidents.md).\n- Fanning out independent units, or asked about Cursor `\u002Fmultitask`? Read [references\u002Ffan-out-engines.md](references\u002Ffan-out-engines.md) (recipes, gotchas, GUI prompt contract, dispatch hygiene).\n- Dispatching child work from a T3 thread (`delegate_task`, review rounds, pending questions, PR watch)? Read [references\u002Ft3-native.md](references\u002Ft3-native.md). It owns the T3 child lifecycle (dispatch, wait, isolation, verify, settle), not delegation-operations.md: for a T3 child skip that file's monitor loop and no-worktree route.\n\nRead every reference triggered by the mission before dispatch. The live reference owns its detailed procedure; Routing rules (SSOT) and the role config take precedence over conflicting routing or model-selection recipes.\n\n## Role Matrix\n\n| Tool | Role | Does | Never does |\n|---|---|---|---|\n| **Cursor** | Gather | SQL, file\u002Fcode scans, grep, read-only lookups and audits | Changes files, implements, opens PRs, decides |\n| **`codex.implement`** | Implement \u002F review | Narrow technical implementation (Sol at medium by default, § Codex seats D) and cross-vendor review per SSOT (2026-10-29) | Research, data gathering, orchestration; a task that is only reading, searching, tabulating or running scripts or tests (§ Codex seats C). The reading and checks inside its own build or review task are part of that task |\n| **`codex.security`** | Implement (security) | Security implementation per SSOT | Gathering, orchestration, reviewing its own work |\n| **`gemini.gather.text` gatherer** (`{repo}Gemini -m $(node scripts\u002Fmodel-roles.mjs gemini.gather.text --field launcher_tier)`; this matches the launcher default) | Gather (text) | Doc\u002Flink\u002Fcopy audits, inventories\u002Fcounts, doc fetch+quote, local digests, BrainLayer recall | Running commands, implementing, reviewing, deciding (including a deletion or a test edit), UX\u002FUI judgment |\n| **`gemini.gather.visual` gatherer** (`{repo}Gemini -m $(node scripts\u002Fmodel-roles.mjs gemini.gather.visual --field launcher_tier)`; Pro-High via `-m pro` only when explicitly requested) | Gather (visual) | Frame\u002Fscreenshot reads, OCR, video state changes; the default image reader for `\u002Fqa-video` (its one `claude.judgment` judge judges this text and spot-checks ≤2 frames per finding) | Running commands, implementing, reviewing, deciding, UX\u002FUI judgment |\n| **`claude.judgment`** | Orchestrate \u002F judgment | Coordinates, talks to users, decides, synthesizes, monitors, queries BrainLayer; holistic\u002Fproduct and UX\u002FUI implementation and other-vendor pair review per SSOT | Bulk reads\u002FSQL; narrow technical implementation (Codex's lane) |\n\nDecision rules:\n\n1. Read-only query, scan, search, audit, or lookup -> Antigravity Flash-high per the 2026-11-02 SSOT rule when it is easy work (fallback: the 2026-10-28 tiers or Cursor; hazard recon: rule 2); for docs\u002Flink\u002Fcopy\u002Finventory gathers only, a `gemini.gather.text` gatherer; Pro-High only when explicitly requested; never a Low tier; single-fact sub-agent lookups follow the bounded exception below.\n2. Anything that decides a code deletion or a test edit (callers=0, dispatch sites, tool→test maps, \"safe to delete\") -> `claude.judgment`. A gatherer's grep may feed it; the gatherer never concludes it. This applies to Cursor gathers too.\n3. Code or file changes follow the Routing rules (SSOT) table: UX\u002FUI -> `claude.judgment`; security -> `codex.security`; holistic\u002Fproduct work (skill and doc text, product behaviour, multi-file features) -> Claude Opus (Sonnet\u002FHaiku only as sub-agents for trivial wording edits); narrow technical work (bounded fix, parser\u002Fguard\u002FCI script, mechanical refactor, dependency bump) -> Codex Sol at medium by default (§ Codex seats D); ratchets and TLC\u002FTLA+ specs -> Opus designs and builds (the table's ratchet row owns the tiny exception); the reviewer is always the other vendor (2026-10-29).\n4. Coordination, synthesis, monitoring, or decisions -> `claude.judgment`; pane mechanics and bounded verifiers follow the cheap sub-agent rule below.\n5. Mixed gather + implement work -> the gatherer (Cursor or Gemini) returns read-only findings; coordinating Claude records them under `docs.local\u002F`; the implementer selected by Routing rules (SSOT) implements from that handoff. A deletion or test edit gets its `claude.judgment` decision (rule 2) before implementation.\n6. Independent parallel units -> § Fan-out engine chooses the engine; Routing rules (SSOT) owns Cursor model selection.\n7. For video extraction\u002Fprocessing that runs commands, select the execution profile under § Execution capability first. Video QA, debrief or review -> `\u002Fqa-video` (2026-11-12, Etan, via orc): the requesting seat dispatches ONE `claude.judgment` judge as a delegate child (never a top-level thread) on the machine that holds the recording; the judge runs the extraction scripts, sends the frame windows to gatherers (Antigravity Flash-high, else `visual-batch.py`, else TOOLING FAILURE; never a Claude reader), judges their text over time and spot-checks ≤2 frames per reported finding; missing measurements become a script row, never eyeballing. Standalone frame OCR or bulk screenshot reads outside that pipeline -> a `gemini.gather.visual` gatherer; from an Agent-tool context, dispatch `visual-gatherer`; it runs `gemini.gather.visual` headless; use Pro-High only when explicitly requested.\n8. UX\u002FUI and design judgment stays on `claude.judgment`. Open-ended research: a Gemini gatherer may draft; the lead verifies before it reaches Etan. A gatherer never implements, reviews, merges, or decides.\n\n**Evidence (skill-creator eval, 2026-09-25; visible cmux workers, mechanical answer keys, lead-scored):** 40 bounded text-gather tasks: Flash-Low 40\u002F40 and Opus 5.5 40\u002F40, 0 fabricated claims each; higher Gemini effort on text gave the same accuracy 2.8–4.7× slower. 20 mixed tasks (8 image reads, 4 video): Opus 20\u002F20, Pro-High 20\u002F20 (2:35), Flash-High 20\u002F20 (6:17), Flash-Low 19\u002F20 (missed counting distinct screens across a video). An open-ended Pro research draft had a dead citation, a stale \"recent\" item, and missed the key release. Limits: screening sample (n=60) on one Mac; not evidence for judgment, design, review, or code. 2026-09-25 ruling: Low tiers are not used for gathering. On a real narrated QA review Flash-Low found 15–16\u002F22 vs Flash-High 21\u002F22, with an invented quote.\n\nVisual re-bench 2026-09-30 (22 visual items incl. 10 harder: near-dup screens, one-frame flash, 10 px OCR, contact-sheet diffs; n=2, headers verified): Flash-High 44\u002F44, 0 fabrications, ~11 min per batch; Pro-High 43\u002F44, ~3 min, fabricated its self-reported timing in 2\u002F2 runs → visual tier = Flash-High.\n\n### Evidence: deletion reachability\n\n**Evidence for rule 2 (cmuxlayer CX-3 real recon slices; one repo, 28 tools, scored against a truth table from fresh `rg`):** callers=0 for 28 tools about to be deleted: Opus surfaced 3\u002F3 deletion hazards; Flash-High and Haiku 0\u002F3 (Flash-Low, a lower tier, was not run on this slice), and Flash-High called all 28 safe, which would have broken a by-name engine accessor used by 23 test files; Pro made 59 false positives. Tool→test-file map errors: Opus 1, Flash-High 16, Pro 21, Haiku 23. Docs\u002Flink audit: Opus, Flash-High and Flash-Low 0 errors (82 s, 87 s, ≈270 s); Pro 47 false positives from a gitignored build directory; Haiku 20 errors (16 misses, 4 substring-match false positives). The eval above passed callers=0 because its tasks were literal greps against a known answer key; deciding a deletion needs reachability reasoning (by-name lookups, policy tables, agent-facing text).\n\n## Cheap sub-agents\n\nEasy work goes to Antigravity Flash-high first (the 2026-11-02 rule above), including one-shot\nread-only verifiers. This section covers two cases: the fallback when Antigravity is out, and the packaged\nagents below whose own tools (BrainLayer MCP, cmux MCP, deploy steps) a `delegate_task` child may not\ncarry. In those cases every `claude.judgment` seat (leads and workers) delegates parity tasks to\n`claude.subagent.cheap`; pass its resolved alias as the Agent tool's `model`:\n`node scripts\u002Fmodel-roles.mjs claude.subagent.cheap --field alias`.\n\n- **brain-worker:** single-fact recall (\"when did X merge\", \"what did Etan rule on Y\"); use one cheap child instead of the judgment seat's own context.\n- **orc-helper:** spawn\u002Fresume\u002Fsend\u002Fread\u002Fwait\u002Fclose pane mechanics; the packaged agent's `role:` pins its alias through the drift lint.\n- **Verifiers (one-shot), when Antigravity is out:** rerun a claimed command once, check now whether an artifact exists, or read a report's DONE marker.\n- **Deploy hops:** release → install → restart → live probe → seat reconnect go to the\n  `deploy-local-prod` sub-agent (Agent tool `subagent_type: \"deploy-local-prod\"`, model pinned to the\n  resolved `claude.subagent.cheap` alias). The lead routes the hop and verifies the artifact it reports.\n  Deploy hops stay on the MBP\u002FM4; the M1 is status-only for deploys. Test placement follows the M1 test-rig adapter (§ Dispatch Boundaries).\n\nAnything that sits, polls or waits (CI, a PR, a file appearing, a sleep-then-check) is the mechanical\ntier, never the lead; a one-shot check stays a cheap verifier above. From a Claude seat, pass\n`node scripts\u002Fmodel-roles.mjs claude.subagent.mechanical --field alias` as the Agent tool's `model`; for\na Codex watcher, use `node scripts\u002Fmodel-roles.mjs codex.subagent.mechanical --use watcher --field model`.\n`claude.subagent.cheap` keeps brain-worker recall and orc-helper, plus verifiers when Antigravity is out, per its 2026-09-29 evidence.\n\nKeep `claude.judgment` for multi-source or decision-grade history, deletion decisions,\nPR-gating reviews, and UX\u002FUI judgment. Reserve judgment → 3× cheap fan-out for\nhigh-stakes history only: it costs 3.4× and runs 2.6× slower. Pass this rule into\nevery judgment worker's brief. This is the bounded exception to general lookup routing.\n\n### Evidence: sub-agent and cleanup routing\n\n`skill-creator\u002Fdocs.local\u002Fevals\u002F2026-09-29-subagent-routing\u002FRESULTS.md`: plain brain-worker recall parity at 0.46×; orc-helper 4\u002F4 at 0.47×; verifiers 0 wrong at 0.55× (bounded cases; multi-source losses remain).\n`skill-creator\u002Fdocs.local\u002Fevals\u002F2026-09-29-cleanup-routing\u002FROUTING-R2.md`: deletion hazards Opus 6\u002F6 vs every Sol arm 4\u002F6; `codex.implement` effort medium, never default xhigh (n=2 per arm\u002Fcase).\n\n## Fan-out engine\n\nOnce the work is known to be N independent units, pick the parallelism engine. Full recipes and\nevidence: [references\u002Ffan-out-engines.md](references\u002Ffan-out-engines.md).\n\n**Cursor `\u002Fmultitask` is an in-editor GUI slash command, not a `cursor-agent` CLI feature.** Sent to\na headless agent, it degrades to one plain prompt answered by ONE agent that may claim it ran in\nparallel. A terminal\u002Fcmux orchestrator never uses it.\n\n| Task shape | Engine |\n|---|---|\n| Human in the Cursor editor, independent read-heavy prompts | Cursor `\u002Fmultitask` (GUI; `\\|\\|\\|`-separated; no write locking) |\n| Cursor child work | T3 `delegate_task`, `target: {providerInstanceId: \"cursor\", model: \"default\"}`; one pointer brief per child, report to collab |\n| Inside a Claude session; let the parent decide the split | Claude Workflow \u002F Agent tools |\n| Lead in a T3 thread; child work (implement, review, recon) owned by that thread | T3 `delegate_task` (§ T3-native dispatch) |\n| One coherent edit, or step B needs step A's output | None: run serially in one agent |\n\nFan out the read-heavy discovery; make the changes in one agent that holds the whole picture.\nVisible cmux fleets: § Dispatch Boundaries › cmux adapter.\n\n## T3-native dispatch\n\nA lead in a T3 thread dispatches child work through the T3 MCP, not panes. The lifecycle, as rules: [references\u002Ft3-native.md](references\u002Ft3-native.md).\n\n1. **Discover:** call `orchestrator_capabilities` first; pin provider, model and effort in `target` from its live catalog (canon #5).\n   For ordinary Codex implementers, recon and reviewers, explicitly name `target.providerInstanceId: \"codex-worker\"` every time.\n   `codex` stays for: security (implementers + scans), UI \u002F computer-use recorders, browser \u002F web-QA workers and Etan (Codex has no lead seats).\n   A Codex child inherits its parent's Codex instance. A Claude lead naming only the codex driver gets whichever codex instance T3 lists first; always name the instance.\n   Remember: the role picks the model, the work type picks the instance. Codex reviewers on security or full-panel PRs stay on `codex`; ordinary Codex reviewers use `codex-worker`.\n   Otherwise each Codex worker thread starts codex-security, computer-history, node_repl and cua_repl (~24.5 MiB, 4 processes), which T3 keeps until settle\u002Fidle release.\n   Recreate, upgrade verification and missing-instance fallback: [Codex instances](references\u002Ft3-native.md#codex-instances).\n2. **Dispatch:** follow the [worktree launch guard](references\u002Fworktree-launch.md) for exact-base preflight and preparation checks; `delegate_task` with `mode: \"async\"`; keep the `taskId`; one stable `clientRequestId` per round. Edits get `isolation: \"worktree\"`, which starts from the parent's committed HEAD, so the brief names absolute paths.\n3. **Wait:** completion wakes the parent. End the turn; never poll or spawn a watcher for it.\n4. **Settle:** at the PR's terminal state (merged or handed off; never at the first DONE — a PR in review rounds keeps its implementer) (a child with no PR, such as a reviewer or recon: as soon as its report is verified), read the report, verify the artifact (PR URL and head, files) before relaying, then conclude the child, `task_cancel` its `taskId`, `unwatch_pull_request` its PR, then `t3_thread_organize {action: \"settle\", threadId: \u003CchildThreadId>}`. Never omit threadId: an omitted threadId settles the calling lead thread. See § Goal Contract for provider release timing and the worktree caveat.\n5. **Review rounds:** every round is a new `delegate_task` carrying prior findings; never `t3_thread_send` to the `childThreadId`.\n6. **Questions:** a child's pending question is answered by the lead, never forwarded to Etan as a routing question.\n7. **Workers:** `delegate_task` children of their lead, `isolation:\"worktree\"` for code, never top-level threads. `t3_thread_launch`\u002F`create_threads` are only for a separate conversation explicitly requested by Etan. Cross-project work goes through that repo’s lead: `delegate_task` runs only in the caller’s project. \"I thought codex workers were just like implementer threads ... we usually just delegate workers and not put them in threads, right?\" (Etan, 2026-10-28)\n   Messages to a RUNNING thread use `t3_thread_send` mode `auto`, never `queue`. Never settle, re-snooze or unsnooze a thread Etan snoozed himself.\n8. **PRs:** `link_pull_request` every PR; the parent owns `watch_pull_request`.\n9. **Who and how many:** Routing rules (SSOT) picks the vendors; fresh implementer per PR and the Codex cap live in § Lead Topology #5.\n10. **Scheduled tasks:** a thread that exists for a `schedule_task` run ends every run by snoozing itself to its next run time (`t3_thread_organize {action: \"snooze\", threadId: \u003Cits own threadId>, snoozedUntil: \u003Cnext run, ISO>}`), so it never sits idle in Etan's sidebar between runs. Write that as the last step of the scheduled prompt. If Etan snoozed the thread himself, his snooze stands (#7). (Etan, 2026-11-13, via orc.)\n\n## Dispatch Boundaries\n\n- Cursor model selection lives in Routing rules (SSOT).\n- Claude Workflow\u002FAgent-tool fan-out is read-only recon, verification, or synthesis except audio-dashboard builds.\n- T3 child work follows § T3-native dispatch.\n- Easy standalone read-only lanes go to Antigravity Flash-high (the 2026-11-02 rule in Routing rules (SSOT)); the 2026-10-28 tiers or Cursor are the fallback.\n- Skills, hooks, agent definitions, and global agent settings are skillCreator-domain work. If skillCreator is not already in the loop, stop and request its audit before patching.\n- **A lane blocked by a guard never changes that guard.** No edit to the gate, its allowlist, policy or tests to get its own action through. It reports the block (the exact command and the refusal) to its lead, and the gate's owner decides. (Etan, 2026-11-13, via orc, after a lane opened a PR widening the human-confirm gate to post its own approved comment.)\n\n- Superseded placement: \"**The M1 Pro becomes a TEST RIG only.** Development stays on the MBP. … Nothing changes on the M1 for this yet.\" (Etan 2026-10-18, via orc collab) The 2026-10-28 placement below now governs.\n- M1 placement (Etan, 2026-10-28): \"I want the m1 pro to be used for tests.\" Full suites, E2E and XCUITest default to M1 Pro: ≥10 GiB guard, max 3 heavy lanes, never during a sync window, PHX admission required. Quick focused unit tests stay local.\n- Dispatching to the other Mac (linked T3 environment, remote code writes, project ids, admission ledger)? Read [references\u002Fcross-machine.md](references\u002Fcross-machine.md). It replaces the m1-router launch transport.\n\n### cmux adapter\n\nApplies to visible cmux fleets only; T3 children follow § T3-native dispatch.\n\n| Task shape | Engine |\n|---|---|\n| Visible, multi-vendor, long-running, human-watchable workers | cmux fleet (`\u002Fcmux-agents`) |\n\n- Visible-worker launch law lives in fleet canon #6; `\u002Frepogolem` owns invocation mechanics and [delegation operations](references\u002Fdelegation-operations.md#launcher-boundary) owns the `--fast` prohibition.\n- Codex workers edit only their delegated worktree; their Claude lead owns acceptance.\n- Pane monitoring and reviewer handoff: `\u002Fcollab-monitor`.\n\n### M1 test-rig adapter\n\nActive placement follows the 2026-10-28 M1 rule in § Dispatch Boundaries.\n\n- Codex workers only, until a recorded Claude-in-T3 proof passes (real `brain_search`, one skill invoked, no hook errors). Full suites, E2E and XCUITest use M1 Pro by default under the guard, lane cap, sync-window exclusion and PHX admission above.\n- Each dispatch requires a same-turn BrainBar probe: the socket accepts and `brain_search` answers. A failed probe is reported to the lead; the work stays on the MBP.\n- Leads never move. Vendor routing, inner-loop order and PR-body records are unchanged.\n\n## Lead Topology\n\nDomain leads are orchestrators one tier below orc:\n\n1. Leads delegate implementation according to Routing rules (SSOT); fleet canon #7 owns their worker-monitor guard.\n2. Lead goals preserve orchestration duties: delegate, maintain health gates, synthesize, and verify.\n3. A lead is a managed `agent_id` with `role:\"orchestrator\"` and left-column placement.\n4. Tiny lead self-edits are capped at \u003C=20 changed lines, one single-purpose change, and zero new files. They require an isolated worktree plus same-post collab disclosure of what changed, why urgent, and line count. Urgency alone never qualifies; everything else follows Routing rules (SSOT).\n5. **One PR = one fresh implementer.** A new PR gets a new implementer (a new `delegate_task` or a new pane, fresh context), even when a healthy idle worker exists in the same repo\u002Fworkspace\u002Frole. Only that PR's own fixes (its review rounds) may reuse it. Dependent stacks are no exception: each stack layer is its own PR, so it gets its own implementer, handed a compact checkpoint; the 2026-10-18 rules supersede R6's dependent-stack exception. Why: \"it re-uses workers instead of settling and opening new ones, like that ship 2246 chat that had 25 prs, 24 merged\" (Etan 2026-10-18, via orc collab).\n   - **Lifetime cap:** \"No Codex session lives past one PR, or ~100 turns. After that, start a fresh delegate with a short brief and a compact checkpoint. Never re-inject full plans or logs.\" (Etan 2026-10-18, via orc collab) Tightened 2026-11-13: about 2 h or 40 turns (§ Codex seats, rule A).\n   - **Reviewers:** \"reviewers fresh per implementer (same PR may reuse them)\" (Etan 2026-10-18, via orc collab).\n   - **Concurrency:** \"at most 4 Codex workers running at once per lead\" (Etan 2026-10-18, via orc collab). A fifth queues until one settles; the Claude dispatch cap in § Model pins and dispatch is separate.\n   - **Settle, then open new:** keep the implementer through its own PR's review rounds, not past them. At the PR's terminal state (merged or handed off, artifacts harvested) settle it, never at its first DONE; the next PR opens a new one.\n   - **cmux adapter:** canon #7's \"minimize pane churn\" is cmux fd-leak hygiene. It means no needless pane open\u002Fclose, never keeping one worker across PRs.\n\n6. A lead's model is written in its lane brief, signature line and claim post; a mid-sprint tier change names its reason (budget | quality | stall).\n7. **Leads never do ops themselves:** \"Leads never do ops themselves: deploy hops go to the `deploy-local-prod` sub-agent; mechanical checks and waits go to Haiku or gpt-6-luna.\" (Etan 2026-10-19, via orc) A deploy, install, restart or reconnect hop is a `deploy-local-prod` dispatch; a sit\u002Fpoll\u002Fwait task is a `claude.subagent.mechanical` or `codex.subagent.mechanical --use watcher` dispatch (§ Cheap sub-agents). Resolve both through `scripts\u002Fmodel-roles.mjs`; never write the model names. The lead's part is routing the hop and verifying the artifact it reports; the lead never runs the hop, the probe or the wait in its own session.\n8. **The lead shape (default, 2026-11-13).** Fable orchestrates a few Opus per domain; Opus hands reading and mechanical work to Sonnet or Haiku. A domain lead is a Fable seat that dispatches a few Opus (`claude.judgment`) workers for its domain's judgment and holistic work. Each Opus hands its reading to Sonnet (`claude.subagent.cheap`) and its mechanical work (sit, poll, wait, bounded checks) to Haiku (`claude.subagent.mechanical`), per § Cheap sub-agents. Fable is selected explicitly at each lead launch (§ Model pins and dispatch); the bare boot pin stays on `claude.judgment`, and #6 still records the lead's model. orc launches Fable leads: `model-pin-gate` allows a Fable target only from an apex seat, so a lead does not spawn Fable workers. The Routing rules (SSOT) table is unchanged: Codex keeps narrow technical work and other-vendor review, and gathers keep their routes.\n   - **Concurrency:** fleet canon #5 owns the numbers; this shape changes none. A Fable lead takes one of canon #5's Claude lead tracks and its place in the shared ceiling with leads. \"A few Opus\" means at most canon #5's `claude.judgment` worker cap running at once; the rest queue. Sonnet and Haiku draw on their own canon #5 caps.\n   - **Caveat:** it depends on the lead, and Saturday's weave should say whether Fables are better as advisers to Opus leads (orc's routing of card 5). Etan's words: \"Yeah, I think it depends on the lead, but I think some leads could be opposes [Opuses]. Maybe most of them should be fable but I think the weave should help us figure out if maybe fables should be more of advisors that come in to correct or give extra insight to opus leads before they make opus workers or soul [Sol] workers as well as maybe sometimes they should utilize fable for workers so we’ll need to see, but for now I think we should adopt that default\" (Etan, 2026-11-13, dashboard card 5, via orc). Until that weave (2026-11-14) rules, a lead may stay on Opus when its domain calls for it; the brief names why.\n\n## Goal Contract\n\nEvery complex or multi-hour dispatch uses one absolute file-backed goal containing the full user\nmission, constraints, green\u002Fno-green criteria, report path, and exact DONE marker. Name the CLI,\nexecution profile and required tool access, matched to the task under § Execution capability. Supersede an\nimplementer's goal only within its own PR (its review rounds); a new PR gets a fresh implementer (§ Lead Topology #5). A lane may\nclose only as verified `DONE`, file-backed `BLOCKED`\u002F`NOT_GREEN`, or `TRANSFERRED` with successor\nevidence. A DONE marker is only a prompt to verify the contracted artifact.\n\n**Stay on task:** a worker reads only its brief, the files it names, and repo instruction files. Never read other agents' reports, inboxes, briefs or collab threads not addressed to it unless the brief explicitly allows it.\n\n**Capability mismatch:** if the worker's profile\u002Ftools cannot do the task, stop and report one line to its lead: `BLOCKED: \u003Ctask> needs \u003Ccapability>; profile \u003Cx> lacks it`. Never substitute other work or explore.\n\n**Questions go to the lead:** \"workers never ask Etan routing questions\" (Etan 2026-10-18, via orc collab). A worker\nunsure about routing, scope, model or reviewer asks its lead. T3: the child's pending question, which the lead reads with\n`t3_pending_request_read` and answers with `t3_pending_request_respond` ([t3-native § 6](references\u002Ft3-native.md#6-pending-questions)).\ncmux adapter: one addressed collab line to the lead. The lead answers from this skill and escalates only a decision Etan\nalone can make, as an `@etan` ask in the collab. Etan alone decides only true one-way doors (deleting\ndata\u002Fcreds\u002Faccounts, money, messaging people); a tag, release, install, deploy or upgrade ships without asking (Etan\n2026-10-19, via orc; `\u002Fpr-loop` § Hard Gates).\n\n**Settle at the terminal state:** \"settle idle workers\" (Etan 2026-10-18, via orc collab). When a lane closes (verified\n`DONE`, filed `BLOCKED`\u002F`NOT_GREEN`, or `TRANSFERRED`), the lead settles its worker. T3: conclude the task (no new\nsends), `task_cancel` its `taskId` so queued runs stop, and `unwatch_pull_request` the lane's PR, then `t3_thread_organize {action: \"settle\", threadId: \u003CchildThreadId>}`. Never omit threadId: an omitted threadId settles the calling lead thread.\nThis releases its provider session (Claude: MCP children exit at once; shared Codex: the thread unloads and its MCPs close after Codex's unload delay); `task_cancel` alone does not detach.\nSettle once nothing still needs the child's worktree (a pending reviewer or fix round): T3's hourly cleanup removes settled threads' worktree checkouts (it keeps the branch).\nSource: T3 `storageCleanup.ts:85-90,1017-1029,1236-1261`. cmux adapter: harvest,\nthen close the pane through the cmuxlayer MCP (`close_surface`). Idle with no artifact pending = settle now, not later.\nA PR still in its review rounds is not idle; its implementer stays for those fixes (§ Lead Topology #5). The implementer\nremoves its own worktree at DONE (`\u002Fpr-loop` § Cleanup Receipt): `git worktree remove` (no `--force`),\nthen `git worktree prune`. Before lane close or settle, the lead verifies removal proof: the path is absent\nfrom `git worktree list --porcelain`. Dirty files or local-only commits stay; name the path and reason\nin the receipt and escalate to the lead. No proof or escalated exception = invalid receipt.\n\nIf the user questions why work is happening or corrects the route, pause spawning and patching,\nexplain the evidence-backed state, and store the correction separately.\n\n## Cross-Skill Ownership\n\n- `\u002Frepogolem`: launcher flags, defaults, resume behavior, and raw escape hatches.\n- `\u002Fpr-loop`: branch through draft PR to ready-for-review; Routing rules (SSOT) above owns the inner-loop pair, which runs on the draft.\n- `\u002Fcollab-monitor`: cmux\u002Fcollab-mail only: pane worker monitoring and reviewer handoff mechanics. T3 children use completion notifications instead (§ T3-native dispatch).\n- `\u002Fqa-video`: video QA\u002Fdebrief pipeline; one `claude.judgment` judge child runs where the recording lives, `gemini.gather.visual` gatherers (delegate_task \u002F visual-batch.py) read the frames, the judge judges their text and spot-checks.\n- `\u002Fwhats-new`: tool-surface changes; they do not revise Routing rules (SSOT) by implication.\n\n## Installed design skill constraints\n\nThese fleet constraints apply independently of the usage proposal below:\n\n- `improve-animations` is audit mode only (`quick` or `standard`). Its `execute`\n  mode is **never** used in the fleet: it self-dispatches and self-reviews.\n  Write plans only under `docs.local\u002Fanimation-plans\u002F`, never a tracked `plans\u002F`\n  directory. Send the plans to the lead, who routes implementation and review.\n  Any authorized read-only helpers are capped at two and follow normal model\u002Ftier\n  rules (default service, no fast; Cursor stays Auto-only).\n- `animate` must use the repo's existing dependencies. `pick-ui-library` is not\n  installed: never fetch it or a replacement library to satisfy that reference.\n\n## Design skills usage — FRONTEND RULE, 2026-10-28\n\nEtan: \"Split: Claude for visual slices\" (2026-10-28, via orc\u002FPHX).\nClaude (`claude-opus-5-5`, `fastMode:false`), loaded with BOTH\n`frontend-design:frontend-design` and `emil-design-eng`, implements taste-heavy\nvisual slices; Codex does the plumbing. This is the taste-critical exception to\nthe Claude budget hold; keep visual slices small. A fresh red-team Claude\n(`claude.judgment`, never the implementer) with BOTH skills reviews every visual\nPR and cites [pr-loop's required proof](..\u002Fpr-loop\u002FSKILL.md#visual-proof).\nCodex computer-use recorders may produce the artifacts and functional checks.\nClaude uses the official plugin; Codex\u002FCursor use vendored `frontend-design`.\nPlugin namespacing denotes one route, not a duplicate.\n\nSequence (lead guidance): `frontend-design` sets the holistic direction before planning\nand reviews the assembled full page at the end. `emil-design-eng` handles each\ninteractive component\u002Fsection inside that direction: feedback, keyboard\u002Ffocus,\nloading\u002Fempty\u002Ferror states, gestures, interruption and reduced motion. Holistic\nlayout\u002Ftokens win on visual cohesion; Emil governs local interaction craft.\n\n- VoiceBar and app UIs: holistic → every page\u002Fsection with Emil → holistic review\n  of the complete navigation\u002Fsettings flow. Adapt web recipes to native conventions.\n- BrainLayer, cmuxlayer, VoiceLayer, Golems and etanheyman.com websites: holistic\n  frontend-design pass → `\u002Flarge-plan` → each interactive row's implementer gets\n  both skills, the shared tokens and required states → holistic assembled-page pass.\n- Use `animate` to implement justified motion, `review-animations` explicitly for\n  a motion diff, `improve-animations` to audit\u002Fwrite plans, and\n  `find-animation-opportunities` for restrained proposals. Do not load the full suite\n  into every worker; no animation is a valid decision. Give Emil a concrete task\n  because a bare invocation only greets. `animate` references the uninstalled\n  `pick-ui-library`; use existing repo dependencies and never fetch it. The\n  audit-only, docs.local output and never-execute constraints above always apply.\n\nClaude keeps the official auto-updating\nfrontend-design plugin and the five vendored Emil skills. Codex uses all six\nvendored skills plus the upstream drift\u002Fupdater path. Cursor has the six skills\nfor routed read-only review only, never implementation. Gemini\u002Fagy receive none.\nAdditional Emil skills await Etan's decision.\n",{"name":5,"description":7},"master","3400bc971628f00ab7bf1063993477e211a5508a","skills\u002Fgolem-powers\u002Fagent-routing","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems","https:\u002F\u002Fgithub.com\u002FEtanHey","etanheyman.com","2026-11-15 03:05:08",{"source":8,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add EtanHey\u002Fgolems --skill agent-routing",{"tool":37,"command":38},"bunx","bunx skills add EtanHey\u002Fgolems --skill agent-routing",{"tool":40,"command":41},"pnpm","pnpm dlx skills add EtanHey\u002Fgolems --skill agent-routing","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002FSKILL.md",true,[45,51,55,61,65,69,73,77,81,85,89,93,97,101,105,109,113,117,121,125,129,133,137,141,145,149,153,157,163],{"path":46,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":49,"sourceHtmlUrl":50},"evals\u002Fevals.json","REFERENCE","参考文档",116901,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fevals.json",{"path":52,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":53,"sourceHtmlUrl":54},"evals\u002Fresults\u002F2026-09-30-model-role-routing-smoke.json",3792,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002F2026-09-30-model-role-routing-smoke.json",{"path":56,"resourceType":57,"resourceTypeDesc":58,"sizeBytes":59,"sourceHtmlUrl":60},"evals\u002Fresults\u002F2026-09-30-model-role-sweep-classification.csv","ASSET","资产文件",16708,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002F2026-09-30-model-role-sweep-classification.csv",{"path":62,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":63,"sourceHtmlUrl":64},"evals\u002Fresults\u002F2026-09-30-model-role-sweep-rest-smoke.json",6294,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002F2026-09-30-model-role-sweep-rest-smoke.json",{"path":66,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":67,"sourceHtmlUrl":68},"evals\u002Fresults\u002F2026-09-30-model-role-sweep-smoke.json",6078,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002F2026-09-30-model-role-sweep-smoke.json",{"path":70,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":71,"sourceHtmlUrl":72},"evals\u002Fresults\u002F2026-10-01-routing-ssot-smoke.json",5665,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002F2026-10-01-routing-ssot-smoke.json",{"path":74,"resourceType":57,"resourceTypeDesc":58,"sizeBytes":75,"sourceHtmlUrl":76},"evals\u002Fresults\u002F2026-10-04-routing-restatement-classification.csv",63213,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002F2026-10-04-routing-restatement-classification.csv",{"path":78,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":79,"sourceHtmlUrl":80},"evals\u002Fresults\u002F2026-10-04-routing-restatement-proof.json",1853,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002F2026-10-04-routing-restatement-proof.json",{"path":82,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":83,"sourceHtmlUrl":84},"evals\u002Fresults\u002F2026-10-04-routing-restatement-sweep.md",5555,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002F2026-10-04-routing-restatement-sweep.md",{"path":86,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":87,"sourceHtmlUrl":88},"evals\u002Fresults\u002F2026-11-13-codex-seats-red-green.json",40805,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002F2026-11-13-codex-seats-red-green.json",{"path":90,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":91,"sourceHtmlUrl":92},"evals\u002Fresults\u002Fcursor-multitask-headless-ab-2026-06-05.json",3839,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002Fcursor-multitask-headless-ab-2026-06-05.json",{"path":94,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":95,"sourceHtmlUrl":96},"evals\u002Fresults\u002Flive-2026-08-02-effort-routing-baseline.json",2527,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002Flive-2026-08-02-effort-routing-baseline.json",{"path":98,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":99,"sourceHtmlUrl":100},"evals\u002Fresults\u002Flive-2026-08-02-effort-routing-with-skill.json",2316,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002Flive-2026-08-02-effort-routing-with-skill.json",{"path":102,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":103,"sourceHtmlUrl":104},"evals\u002Fresults\u002Flive-2026-08-03-effort-consciousness-baseline.json",2413,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002Flive-2026-08-03-effort-consciousness-baseline.json",{"path":106,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":107,"sourceHtmlUrl":108},"evals\u002Fresults\u002Flive-2026-08-03-effort-consciousness-with-skill.json",3333,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002Flive-2026-08-03-effort-consciousness-with-skill.json",{"path":110,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":111,"sourceHtmlUrl":112},"evals\u002Fresults\u002Flive-2026-08-03-effort-low-baseline.json",2518,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002Flive-2026-08-03-effort-low-baseline.json",{"path":114,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":115,"sourceHtmlUrl":116},"evals\u002Fresults\u002Flive-2026-08-03-effort-low-with-skill.json",3175,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002Flive-2026-08-03-effort-low-with-skill.json",{"path":118,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":119,"sourceHtmlUrl":120},"evals\u002Fresults\u002FSUPERSEDED.md",998,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Fresults\u002FSUPERSEDED.md",{"path":122,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":123,"sourceHtmlUrl":124},"evals\u002Ftrigger-evals.json",1649,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fevals\u002Ftrigger-evals.json",{"path":126,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":127,"sourceHtmlUrl":128},"references\u002Fcross-machine.md",3498,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Freferences\u002Fcross-machine.md",{"path":130,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":131,"sourceHtmlUrl":132},"references\u002Fdelegation-operations.md",6695,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Freferences\u002Fdelegation-operations.md",{"path":134,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":135,"sourceHtmlUrl":136},"references\u002Ffan-out-engines.md",8505,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Freferences\u002Ffan-out-engines.md",{"path":138,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":139,"sourceHtmlUrl":140},"references\u002Ffan-out-head-to-head.md",6626,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Freferences\u002Ffan-out-head-to-head.md",{"path":142,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":143,"sourceHtmlUrl":144},"references\u002Fmodel-and-effort.md",6698,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Freferences\u002Fmodel-and-effort.md",{"path":146,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":147,"sourceHtmlUrl":148},"references\u002Ft3-native.md",13732,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Freferences\u002Ft3-native.md",{"path":150,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":151,"sourceHtmlUrl":152},"references\u002Fverification-and-incidents.md",5224,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Freferences\u002Fverification-and-incidents.md",{"path":154,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":155,"sourceHtmlUrl":156},"references\u002Fworktree-launch.md",5162,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Freferences\u002Fworktree-launch.md",{"path":158,"resourceType":159,"resourceTypeDesc":160,"sizeBytes":161,"sourceHtmlUrl":162},"scripts\u002Flaunch-guard.mjs","SCRIPT","脚本",6173,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fscripts\u002Flaunch-guard.mjs",{"path":164,"resourceType":159,"resourceTypeDesc":160,"sizeBytes":165,"sourceHtmlUrl":166},"scripts\u002Flaunch-guard.test.mjs",12018,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagent-routing\u002Fscripts\u002Flaunch-guard.test.mjs",[168,175,182,189,196,203],{"skillNo":169,"name":170,"nameZh":6,"description":171,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":172,"categories":173,"keywords":174,"publishedAt":19},"01M7C53JEF6A6SS20Q0D2YJYM8","yash","Push dependency bugs upstream. Triggers: dependency bug, workaround, monkey-patch, fork\u002Ffix. NOT app bugs.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fyash\u002FSKILL.md",[],[],{"skillNo":176,"name":177,"nameZh":6,"description":178,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":179,"categories":180,"keywords":181,"publishedAt":19},"01M7C53EPMNSZKHK47VY7WQDP0","whats-new","Diff Claude\u002FCodex\u002FCursor\u002FWispr\u002Fagy changelogs vs local setup. Triggers: whats new, changelog, version bump.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fwhats-new\u002FSKILL.md",[],[],{"skillNo":183,"name":184,"nameZh":6,"description":185,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":186,"categories":187,"keywords":188,"publishedAt":19},"01M7C53C4TVQ9G7AB3X57DACPB","weave","Orchestrator convergence: mine JSONLs, cite findings, route actions. Triggers: weave, run weave.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fweave\u002FSKILL.md",[],[],{"skillNo":190,"name":191,"nameZh":6,"description":192,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":193,"categories":194,"keywords":195,"publishedAt":19},"01M7C5342EDDN7YH53KKW511V2","unslop","Cut AI tells while keeping every receipt, number, path, and hedge intact. Triggers: unslop, cut the slop, wall of text, too verbose, rambly, make this readable. NOT for delivery cadence (see i-have-adhd), channel choice, or artifact length governance.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Funslop\u002FSKILL.md",[],[],{"skillNo":197,"name":198,"nameZh":6,"description":199,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":200,"categories":201,"keywords":202,"publishedAt":19},"01M7C52M0CTDW5DDP4452FJV0G","tmp-block","Guard denying durable writes to \u002Ftmp, \u002Fprivate\u002Ftmp, \u002Fvar\u002Ffolders, $TMPDIR; enforces \u003Crepo>\u002F.worktrees\u002F\u003Cname>. Triggers: tmp guard, TMP-BLOCK deny, WORKTREE-CONVENTION deny, bypass ledger. NOT for storage policy.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002FSKILL.md",[],[],{"skillNo":204,"name":205,"nameZh":6,"description":206,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":207,"categories":208,"keywords":209,"publishedAt":19},"01M7C52HN4Q9E1Z74Z11K23E2H","t3-visuals","Show\u002Fvisualize\u002Fchart\u002Fdiagram\u002Fcompare\u002Fstatus board in T3 with compact interactive HTML. Not standalone dashboards.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ft3-visuals\u002FSKILL.md",[],[]]