Use this guide when a roster uses two or three seats while the operator pays for six or more model lanes. A subscription audit on one operator machine found rosters referencing 2 of roughly 15 usable Cursor model families, an activated-but-unused free reviewer lane, and a coding-plan quota of thousands of requests per week with zero dispatches. The gap is discoverability, not capability: every one of those lanes can hold a Brigade seat today.
This page maps common subscription lanes to seat roles, with copy-paste roster stanzas and the validation pattern to run before trusting a new seat.
Assign every seat one job. Mixed-purpose seats make outcome capture useless because a failure says nothing about which kind of work the seat is bad at.
| Role | What it does | What to route there |
|---|---|---|
| orchestrator | plans, dispatches, synthesizes | nothing directly. It owns the run |
| worker | implementation, patches, research | the bulk of dispatched tasks |
| scout | inventories, greps, triage, summaries | work that does not deserve a frontier model |
| reviewer | verify claims against the diff | exact-head review passes |
Model IDs drift. Every ID below was verified answering on 2026-07-17. Confirm against your harness's live inventory before wiring (see the validation pattern at the end, and issue #299 for making that check automatic).
The usual primary lane. Beyond the default frontier seats, the same OAuth serves cheap tiers that most rosters never touch:
[agents.scout]
cli = "cursor"
model = "gpt-5.4-mini-low"
role = "Cheap fast scout for file inventories and triage."The mini and nano tiers also answer through harnesses that front the ChatGPT OAuth directly.
Flat-sub plans carry far more than the headline models. Verified answering on one Ultra plan: composer-2.5, grok-4.5 tiers, kimi-k2.7-code, glm-5.2-high, gemini-3.5-flash, claude-sonnet-5 tiers, and the gpt-5.4-mini/nano ladder. An open-weight worker at zero marginal cost:
[agents.kimi27]
cli = "cursor"
model = "kimi-k2.7-code"
role = "Open-weight implementation worker on the flat sub; overflow relief for the primary worker."Known gap: Composer models return empty text in read-only plan mode (issue #206). Pin a non-composer model for read-only seats.
Mark direct Cursor Composer and Grok seats as incapable so the chef cannot route
--read-only work to them:
[agents.composer]
cli = "cursor"
model = "composer-2.5"
read_only_capable = false
role = "Fast implementation worker for writable runs."
[agents.grok]
cli = "cursor"
model = "grok-4.5-xhigh"
read_only_capable = false
role = "Alternate implementation worker for writable runs."Keep those direct seats set to false. To use either model for read-only findings,
create a separate ACP seat and set transport = "acpx",
transport_version = "0.12.0", and read_only_capable = true on that seat. ACP
model IDs differ from direct Cursor aliases, so copy the reviewed model ID from
the technical guide.
Two seats, two costs. The heavier model reviews. A lighter sibling takes routine passes at a fraction of the quota burn:
[agents.claude]
cli = "claude"
model = "claude-opus-4-8"
role = "Cross-model reviewer: verify claims against the diff."
[agents.sonnet5]
cli = "claude"
model = "claude-sonnet-5"
role = "Light reviewer and worker for routine passes."Keep the flagship interactive model out of headless seats entirely: one operator machine logged 502 API calls in an evening after review seats defaulted to the top tier with internal subagent fan-out. One dispatch should be one flat pass. Parallelism belongs in brigade run workers, where each lane is metered and receipted.
Often the most idle capacity an operator owns: two accounts can sit at 0-1% of weekly quota while paid lanes run hot. Route scans, summaries, research reads, and small changes here first.
[agents.flash]
cli = "antigravity"
model = "Gemini 3.5 Flash (Low)"
role = "Fast worker for research, summaries, scans, and small code changes."
[agents.reviewer2]
cli = "antigravity"
model = "Claude Sonnet 4.6 (Thinking)"
role = "Cross-model reviewer on the Google lane; verify claims without spending the Claude subscription."Remember: every seat's cli value must appear in limits.allow_models, in the same edit. Brigade validates at roster load, not at edit time, and a missing entry fails every run against that roster.
A direct Grok read-only worker can continue its exact session once when the CLI exits 0 without the required structured final. To allow the one terminal fallback after another invalid final, name a reviewed Cursor-Grok ACP seat on the direct seat:
[agents.grok-review]
cli = "grok"
model = "grok-4.5"
reasoning = "high"
role = "Review the requested change and report actionable findings."
invalid_final_fallback = "cursor-grok"
[agents.cursor-grok]
cli = "cursor"
model = "grok-4.5"
transport = "acpx"
transport_version = "0.12.0"
role = "Fallback review through the reviewed ACP transport."The reference is fail-closed: it must resolve to a non-orchestrator Cursor seat with a grok-* model and the reviewed ACPX version. Brigade does not search the roster for a similar model. The policy applies only to a typed malformed-final result from a direct Grok --worker read-only run; startup, authentication, network, timeout, permission, and nonzero failures are terminal.
Several open-weight providers expose Anthropic-compatible endpoints, which means the claude CLI can drive them and Brigade can seat them directly. Declare the lane's environment on the seat: plain values inline, secrets by reference with a _REF suffix naming the environment variable that holds the value (the roster never stores the secret itself):
[agents.k3]
cli = "claude"
model = "kimi-k3"
read_only_capable = false
role = "Open-weight worker on a coding-plan quota; relief for orchestrator-tier work."
env = { ANTHROPIC_BASE_URL = "https://api.moonshot.ai/anthropic", ANTHROPIC_AUTH_TOKEN_REF = "KIMI_API_KEY", CLAUDE_CONFIG_DIR = "/home/operator/.claude-lanes" }This proxy example opts out of read-only dispatch. Set the value to true only after
the validation smoke below returns a usable worker result in read-only mode.
When the proxy credential is rendered into a runtime systemd environment file, point the reference at that file rather than the parent process environment:
env = { ANTHROPIC_AUTH_TOKEN_REF = "env-file:/run/brigade/k3.env#CLIPROXY_API_KEY" }The path must be absolute. Brigade reads only KEY=VALUE entries from the file and does not
evaluate shell syntax.
Overrides apply to the spawned CLI process only, run.json records the override names and endpoint host (never values), and a missing referenced variable fails the worker before dispatch. If the CLI echoes any resolved override value, Brigade replaces the exact value with its target name in brackets before worker text, detail, stdout, or stderr can be stored. Direct CLI seats only: acpx and codex-cloud seats manage their own environment.
The isolated CLAUDE_CONFIG_DIR is load-bearing: with the default config directory, the claude CLI prefers its subscription OAuth over env auth, the upstream returns 401, and the CLI retries silently, which presents as an indefinite hang.
Brigade seats ollama models through a prefix reference: the model rides in the cli value, and the full reference must appear verbatim in limits.allow_models. Hosted free-tier models work too (kimi-k2.7-code:cloud verified as a seat):
[agents.k27cloud]
cli = "ollama:kimi-k2.7-code:cloud"
role = "Hosted open-weight worker on the ollama cloud free tier."Brigade never auto-pulls ollama models: dispatch fails unless the model is already present, which protects the disk from multi-GB surprise pulls. Hosted models can be retired upstream without notice, so validate before each wiring.
The only lane here that is neither an API nor a coding CLI. oracle
drives a real gemini.google.com session using the Chrome cookies Agent Pantry
syncs, so it costs no API key and no metered quota. It is a consult seat: one
shot, no tools, no file writes.
Requires a user-installed oracle (npm install -g @steipete/oracle) and a
pantry-synced cookie jar. Brigade installs neither.
[agents.researcher]
cli = "oracle"
model = "gemini-3.1-pro"
role = "researcher"
timeout_seconds = 300Models the browser engine accepts: gemini-3.5-flash, gemini-3.1-pro, and
gemini-3-deep-think (browser-only; oracle rejects it in API mode).
Read-only enforcement is hard by construction, since oracle has no filesystem
write path. When the seat fails with a login or expired-session message, run
brigade pantry expiry-alert and re-sync the source before retrying.
A browser round trip is much slower than a CLI seat. The research engine asks
for a 30 second planning call, which a browser session will not meet, so the
seat must declare its own timeout_seconds; research takes it as a floor that
raises short engine timings and never lowers generous ones.
Keep this to a single operator machine. Oracle's browser mode is experimental and cookie automation of a consumer session is grey against provider terms.
A seat that answers a smoke prompt is wired, not proven. The pattern:
- Smoke the seat:
brigade run --roster <roster> --read-only --worker <seat> --output-dir <dir> "Reply with exactly: SEAT-OK"and confirmrun.jsonrecordsstatus=okwith the expected text infinal.txt. - Bench it against an incumbent seat on a task with an objectively gradable answer (a planted bug with a known line and fix works well), and keep the receipts.
- Capture the outcome against the seat so
brigade outcome ranklearns something: a seat with no outcome history is a guess wearing a roster stanza.
Retire stale IDs on sight. Harness model inventories drift: one roster carried a model ID its harness no longer listed (it worked only through fuzzy mapping), and a hosted model was retired upstream two days before a run tried to use it. Run brigade roster doctor after changing direct Cursor, Grok, or Ollama seats. It reports exact, fuzzy-resolved, missing, or unavailable; the last three warn without rejecting the roster. A transient inventory failure is not evidence that a model disappeared.