Other

/agent-routing

Invoke before any git/gh and before starting another agent: dispatch a worker/reviewer, delegate_task, hand off; who gathers, gemini gatherer, Codex model/effort, /multitask, fan out in parallel. Pick provider/worktree. NOT for one edit or dependent steps.

$ golems-cli skills install agent-routing
86 evals

Updated today

This skill owns fleet routing rules, role selection, delegation checks, review ownership, and model/effort dispatch policy. standards/model-roles.json owns role → model; /repogolem owns launcher mechanics and canon #6 owns launcher law.

Auto-dispatch triggers (canonical in /orc C4): batch reads >=3, transcription >=2, web research >=1, or any "in parallel" / "all of these" phrasing -> fan out sub-agents in the SAME message before asking permission.

Kickoff contract

  • Claude seats: invoke /agent-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/model choice. Make it the first tool call, before any git/gh/grep orientation: live misses open with git status/gh pr and never load the skill (PR15 trigger eval, 2026-10-19).
  • Codex seats: read this SKILL.md body (a file read) before that action; Codex has no Skill tool.
  • A lead's kickoff brief to a worker names the skills the worker invokes (/large-plan, /agent-routing, /pr-loop).
  • 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)
  • 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).

Service tier and placement

  • Lead launch prompts say: "Read and follow /LEAD.md and the global LEAD.md" (~/.codex/LEAD.md, installed from the private global instruction source). Neither shared instruction body imports the lead layer; delegates' briefs never name it.
  • Every delegate/sub-agent explicitly passes Claude fastMode:false or Codex serviceTier:"default". Elevated service requires Etan's explicit request for that task.
  • Mechanics (2026-11-19): pass Claude fastMode:false only when that model's orchestrator_capabilities options list fastMode (Opus today). Never pass it to Haiku or Sonnet: T3 refuses the child ("Unknown option fastMode"). Each Claude role's t3_delegate.options in standards/model-roles.json holds what to pass.
  • Harness examples do not authorize elevated service.
  • Lead threads use the repo root checkout (t3_thread_launch with workspaceStrategy:{type:"root"}); only worker threads bind worktrees.
  • Use Theo's official T3 Code nightly; never build or install T3 from our fork.

Attribution: Etan, 2026-10-26.

Routing rules (SSOT)

This 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/model-roles.json.

Gathering

Antigravity is an underused pool (2026-11-08, Etan, via orc; corrected ruling).

"we gotta be using agy much more, … y'all arent doing what I asked for."

"ummm I never said all these need to go there??"

"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?)"

Draw on Antigravity much more where it fits: Gemini Flash/Argon AND Claude via agy, now natively supported in T3 Code. There is no blanket Gemini-first routing and no ban on agy Claude/GPT. The 2026-11-01 agy-Claude rule and all existing routing guardrails stand unchanged. The quoted quota/window is the dated snapshot. A same-model bench of agy-Claude versus native Claude Code is pending, owned by orc; no result is claimed.

Easy gathers go to Antigravity Flash-high at any depth (the 2026-11-02 rule below); Codex Luna/Terra 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. Older Gemini models keep that gather-only rule; Gemini 4 Neon and Hydrogen may build under § Execution capability (2026-11-20 ruling; Argon only as a named trial with a mechanical verifier). 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 <absolute brief path>. Each brief says Report to <collab 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.

Execution capability

A task that must RUN COMMANDS (shell, ffmpeg/media 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.

Gemini 4 Neon and Hydrogen may build (2026-11-20, Etan).

"You have the exception." / "just test it for fuck's sake." / "neon and hydrogen too tbh and MAYBE argon, and all their versions, orchestrator and none"

The change for Orchestrator (Neon) is permanent.

  • In force: a lead may give Gemini 4 Neon or Hydrogen (plain or Orchestrator, provider antigravity) a task that writes files and runs commands, in a worktree, in full-access mode. Argon is allowed only as a named trial with a mechanical verifier, never by default.
  • Older Gemini models: keep the 2026-11-02 gather-only rule — use the gatherer profile, or video-qa only to describe what frames show, with the QA verdict staying with Claude; never send older Gemini models command execution, builds, tests, installs, fixes or file writes (the shell-worker profile included).
  • Unchanged: no Gemini model reviews, merges on its own verdict, or decides a deletion or a test edit. Neon-built code (and any Hydrogen or trial-Argon build) is reviewed by Sol or Claude. No Gemini holds a lead seat except the one thread the owner named.
  • Brief rule: a brief to a Gemini builder child must say what it may write and carry the hard rules itself (worktree only, no rm -r, no secrets, no live store, no host), because those threads run without guard hooks.

Implementation and review

Implementation default (Etan, 2026-10-29)

Historical quote (superseded by the 2026-10-28 default):

"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)

  • 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) Etan (2026-10-28, via orc/PHX): use Sol/Terra/Luna for more of the work so the Claude 5-hour window survives; not all work needs Claude.

  • Default implementers (2026-10-29, replaces the 2026-10-28 Sol/Terra default; Etan, via fleet control: Claude quality is generally higher on holistic work, Codex stays heavy on recon): holistic/product work (skill and doc text, UX/UI, product behaviour, multi-file feature work, anything where judgment and coherence matter) → Claude Opus implements. Sonnet/Haiku never implement; they are sub-agents only for trivial edits (changing some words). Narrow technical work (a bounded fix, a parser/guard/CI 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".

  • Reviewers (2026-10-29) are always the other vendor: Sol reviews a Claude implementation, claude.judgment reviews a Codex implementation, with no exception. Gemini/Antigravity gather and never review.

  • UI/UX (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.

  • 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)

  • 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" / "either leads, subagents or subagents of subagents should use them, not lead being agy". Easy work is a delegate_task on provider antigravity at high thinking, from a lead, its sub-agents and theirs. Per Mac (2026-11-19, orc ruling): check with orchestrator_capabilities first; use gemini-3.8-flash if that Mac's catalog lists it, otherwise gemini-4-hydrogen; both at high thinking. The thinking option's name comes from that Mac's catalog (thinking_level on one build, thought_level on another); never hard-code one. Hydrogen is not benchmarked: treat its output as a pointer to verify. Never Argon for easy reads. The list lives in t3_delegate of the gemini.gather.* roles in standards/model-roles.json. Easy work means ONLY gathers, reads, greps, inventories, log/screen 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 20/20 on mixed gathers and 44/44 on visual reads with 0 fabrications, but 0/3 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.

  • Luna/Terra/Haiku (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.

  • 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.

  • Keep the Claude 5h window alive (2026-10-28). Never fast/priority: explicitly pass fastMode:false (only where the model's catalog offers fastMode) / serviceTier:"default". This supersedes the dated budget-pacing note; it does not revert at a window reset.

WorkImplementsReviewsPlus
UX/UIclaude.judgmentCodex (codex.implement)—
Securitycodex.security at effort highclaude.judgmenta codex-security deep scan per security PR
Holistic/product work (skill and doc text, product behaviour, multi-file features, anything where judgment and coherence matter)claude.judgment (Opus; Sonnet/Haiku only as sub-agents for trivial wording edits), 2026-10-29Codex Sol (codex.implement)a deletion/test-edit decision by claude.judgment first (decision rule 2)
Narrow technical work (a bounded fix, a parser/guard/CI script, a mechanical refactor, a dependency bump, tests, deletions)Codex Sol at medium by default (high with one line of why), § Codex seats D, 2026-11-13claude.judgmenta deletion/test-edit decision by claude.judgment first (decision rule 2)
Ratchets and TLC/TLA+ 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 conformancethis 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

The reviewer is always the other vendor.

Interim route (Etan 2026-09-30/10-01): 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.

Security 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. Deep scans run from <repo>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/S security PRs get a diff-scoped Standard scan, M/L or anything touching auth/guards a full Deep scan; the PR body records the scan's reported token use, and cost is never a reason to skip it.

Proven 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.

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/guard 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.

When 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).

Inner loop (sequential)

  • 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.
  • 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 <N>) and runs the rest of /pr-loop. Merge follows the lane's merge authority after the PR loop is happy.
  • 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.

This inner-loop pair review happens while the PR is a draft, before it is marked ready. /pr-loop bot and PR reviewers are separate and start at the ready flip. Pane mechanics live in /collab-monitor section "Completion -> Reviewer Handoff".

Review sizing

Review is sized to risk; the PR's size:* label sets the tier.

  • "XS/S: 1 reviewer, 1 round." (Etan 2026-10-18, via orc collab) That reviewer is the other vendor, per the table above; UI/style PRs follow the UI flow below.
  • M: up to 2 rounds, and round 2 only for BLOCKING findings (canon #9).
  • "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/product or UX/UI work), the table's Codex reviewer stays on, so cross-vendor review still holds. No Gemini model ever reviews, at any size.
  • UI/style 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/UI." (Etan 2026-10-18, via orc collab) The recorder resolves codex.implement and reports functional defects; proof requirements now follow pr-loop Visual proof.
  • 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. 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.

Claude keeps the official auto-updating frontend-design:frontend-design plugin; Codex/Cursor use the pinned frontend-design skill. These names each denote one provider route (Etan, 2026-10-27).

Records

Each PR body records its implementer, review rounds, and bot/reviewer defects. Claude leads orchestrate and route work through visible panes.

Model pins and dispatch

  • 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.
  • Fable only via explicit per-invocation selection; a prior session's model never persists into the next.
  • Every non-Cursor Agent/Workflow/Task spawn pins its model explicitly, resolved from roles. Cursor is the Auto-only exception.
  • 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."
    • Resolve role defaults through § Model roles; the 2026-10-28 Luna/Terra fan-out permission is policy. Dispatch still respects configured candidate gates until the separate RT-2 model-role evaluation changes them.
    • 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).
    • Astra needs Etan's own explicit ask; a lead's or worker's suggestion does not count (2026-10-18 ruling).
    • 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).
  • Effort is per /large-plan phase; declare effort + why and pass it explicitly at dispatch. Codex effort caps: § Codex seats, rule D.
  • Claude concurrency (lead/worker caps): fleet canon #5 owns the numbers.
  • Usage is managed by default-pinning and dispatch-counting, not by usage-blocking buckets.

Codex seats (2026-11-13)

Every Codex dispatch follows this section. It is the only home of these rules; other skills point here.

Etan (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."

Evidence (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/docs.local/collab/CONVERGENCE.SYNCED.md; report orchestrator/docs.local/reports/2026-11-13/CODEX-USAGE-AUDIT.md.

The rules, verbatim (orc, 2026-11-13 03:05; they replace rule 3 of orc's 02:46 post):

  • 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.
  • 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.)
  • 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.
  • 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.
  • E. Review stays on Codex untouched: it was 7% and it is what the subscription is for.
  • 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.

How they bind the rest of this skill:

  • Task-scoped precedence (orc, 2026-11-17): Explicit models named in a lead's or orc's brief win over routing defaults for that task, including C; model-permission, effort, execution and role/review safeguards still apply. If the named model is unavailable, report to the lead rather than silently substitute. Without an explicit model, quota-exhausted gathers try the remaining eligible Cursor/Gemini/Antigravity pools under § Gathering, then Sonnet, then Haiku for easy/mechanical gathers only; skip exhausted pools.
  • 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.
  • 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/model-and-effort.md and the Codex gather fallbacks in § Gathering and the 2026-10-28 tiers (Luna/Terra for easy work, a Codex gather when Antigravity is out): those fall back to Cursor, Sonnet or Haiku instead.
  • D wins over /large-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.
  • 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).