Skip to content

Orca agents and OrcaCompass support ​

Orca's observed agent support, and what OrcaCompass does with each agent.

The Orca launch and trust behavior below was inspected on 2026-11-03 with Orca IDE 1.4.246. It is a compatibility reference, not a guarantee for newer Orca versions. The OrcaCompass adapter descriptions reflect this checkout.

Every agent id below can be named in orca orchestration worker-start --agent <id>. The list comes from Orca's built-in agent table (resources/app.asar.unpacked/out/main/chunks/tui-agent-config-*.js in the Orca install).

Columns:

  • --model at launch: from Orca's per-agent catalogs (resources/app.asar.unpacked/out/shared/agent-session/options/agent-session-option-catalog-*.js). ✅ Orca passes --model (and --effort where listed). ❌ Orca has a catalog but refuses --model at launch (Agent <id> does not support launch-time model selection). — no catalog. With ❌ or —, the agent runs the model from its own config.
  • Auto-trust: Orca answers the agent's "trust this folder" prompt before launch (preflightTrust), or the launch command carries a trust flag. Without it, a worker in a folder the agent has not trusted can stop at that prompt; Orca then fails the start with Agent startup blocked: agent-trust-workspace.
  • OrcaCompass: what OrcaCompass does with the agent today.
Agent idCommand--model at launchAuto-trustOrcaCompass
claudeclaude✅✅Supported: routed
claude-agent-teamsorca claude-teams (not on Windows or WSL)not checked✅Not supported
codexcodex✅✅Supported: routed
cursorcursor-agent✅ (effort is part of the model id)✅Supported: routed; also carries xAI, Google, Meta and other models through Cursor
antigravityagy✅✅Supported: routed
musemuse --trust-workspace✅✅ (flag)Orca accepts any --model; neither Muse nor Orca lists Muse models, so OrcaCompass routes a built-in list
grokgrok✅ (since 1.4.246; passed as -m / --reasoning-effort)❌Supported: routed. OrcaCompass reads Grok's models (~/.grok/models_cache.json, then grok models), tiers them and knows their effort limits. Orca does not auto-trust folders for Grok, so a worker in a folder Grok has not trusted stops at its trust prompt. xAI models also route through Cursor
hermeshermes --tui——Not supported: Orca passes no model to Hermes
geminigemini❌—Not an agent here: Gemini models route through Antigravity or Cursor
codebuddycodebuddy / cbc✅—Not supported yet (Orca side allows it)
ompomp✅ (no effort)—Not supported yet (Orca side allows it)
opencodeopencode✅ (existing worktree only, no effort)—Supported: routed as a last resort by default (lastResortAgents). Models from a live opencode models call: API-key providers such as OpenRouter (pay per token), OpenCode's own models, and local servers such as Ollama. No effort is passed. Its routes need --worktree current or another existing worktree, because Orca refuses --model for OpenCode in a new one
qoderqodercli / qoder—✅Not supported
qoder-cnqoderclicn / qodercn—✅Not supported
kimikimi / kimi-code—✅Not supported
copilotcopilot—✅Not supported
command-codecommand-code --trust—✅ (flag)Not supported
opencode2opencode2 --standalone——Not supported
openclaudeopenclaude——Not supported
autohandautohand——Not supported
anteante——Not supported
traetraecli——Not supported
mimo-codemimo——Not supported
pipi——Not supported
prime-agentprime-agent——Not supported
aideraider——Not supported
goosegoose——Not supported
ampamp——Not supported
kilokilo——Not supported
kirokiro-cli chat --tui——Not supported
crushcrush——Not supported
augauggie——Not supported
clinecline——Not supported
freebufffreebuff——Not supported
codebuffcodebuff——Not supported
continuecn——Not supported
droiddroid——Not supported
mistral-vibevibe / mistral-vibe——Not supported
qwen-codeqwen——Not supported
rovoacli rovodev run——Not supported
openclawopenclaw——Not supported
dshdsh-tui .——Not supported
zcodezcode——Not supported
devindevin——Not supported
jcodejcode——Not supported

How OrcaCompass handles each agent ​

Source: src/routing/environment.ts and src/placements/agents/*.ts, src/routing/router.ts, src/routing/rejection.ts, src/verify.ts.

Shared steps (every agent) ​

  1. Find the CLI with find(<binary>, extra dirs). Not found → installed: false, never routed.
  2. Read the model list when the placement table is built (every placementsTtlHours, on refresh, or when the table is stale), never on a route that reuses a fresh table. For Claude Code and Codex, Orca's model catalog comes first (src/placements/agents/orca-catalog.ts): the newest probe in agent-model-catalog.db that found models for the agent, opened read-only once per build, then the agent-model-catalog.json older Orca versions wrote. Orca saves no other agent's list there. Live model-list commands run through listModels (src/placements/agents/list-cache.ts): a good listing is saved to the OrcaCompass cache; if the command later fails, the saved copy is used with its age ("using the list from 3h ago"). A sign-in prompt in the output marks the agent signedOut and never falls back.
  3. Give each model its profile: coding, agentic and reasoning scores from the evidence snapshot, and a cost tier from your tierOverrides, the snapshot price, the family or lineup hint, then the evidence class (src/placements/profiles.ts), stored in placements.json (src/placements/build.ts).
  4. Check usage with orca account list --json (src/routing/orca-usage.ts): each model draws on named usage pools; a missing number is skipped, not read as 0 %.
  5. Skip the agent when it is turned off in Orca's Settings → Agents (disabledTuiAgents, read-only from Orca's profile state, src/routing/orca-settings.ts), when Orca reports it unavailable, or when it is signed out.
  6. Skip models the agent refused at an earlier launch (orcacompass verify saves them for rejectionTtlHours).
  7. Build launchArgs (resolve in src/routing/router.ts):
    • flag models: --agent <id> --model <id> [--effort <e>], effort snapped to the nearest one the model offers (applied: flag, or provider-default when none).
    • variant models (effort inside the id): --agent <id> --model <id-for-effort>, no --effort (applied: model-id). Explicit --effort requests are never snapped: models that cannot apply that exact level are excluded.
  8. After launch, orcacompass verify --dispatch <id> reads the worker's output and matches the agent's own refusal text (below).

Per agent ​

ClaudeCodexCursorAntigravityGrokOpenCodeMuse
Adapterclaude.tscodex.tscursor.tsantigravity.tsgrok.tsopencode.tsmuse.ts
Findsclaude on PATHcodex on PATHcursor-agent on PATH or %LOCALAPPDATA%\cursor-agentagy on PATH, %LOCALAPPDATA%\agy\bin or ~/.local/bingrok on PATH, $GROK_HOME/bin or ~/.grok/binopencode on PATH or ~/.opencode/binmuse on PATH
Model listOrca's agent-model-catalog.db, then Claude Code's cached account catalog, then Anthropic families in the evidence snapshotOrca's agent-model-catalog.db, then $CODEX_HOME/models_cache.json (visibility: list), then the evidence snapshot's Codex ladder families without -latest aliasescursor-agent models, liveagy models, live (id<TAB>label rows)~/.grok/models_cache.json (Grok's own, hidden models dropped); else grok models, liveopencode models, live (provider/model rows; OpenRouter ~ aliases skipped)Built-in: muse-wonder-1.0, muse-spark-1.1, muse-ember-1.0, muse-spark-1.3 (MUSE_FALLBACK_MODELS)
Grouping——Variants grouped by base id: effort, -thinking and -fast are stripped; cursor- prefix and old claude-4.6-sonnet ids normalized-low/-medium/-high ids grouped; -thinking dropped to match the evidence snapshot—A model listed by several providers kept once, from the cheapest (local, then free, then paid)—
EffortOrca's catalog levels + default, else the account catalog; on the evidence snapshot fallback, assumed low to maxOrca's catalog levels + default, else the Codex cache's; on the evidence snapshot fallback, assumed low to maxInside the model id (…-high); the plain id stands for medium; fast-only levels droppedGemini: inside the id. Others: Orca's low/medium/high via --effortGrok cache levels + default (none for multi-agent models); from grok models: low to xhigh, capped as in Orca's seed (4.5 up to high)None (Orca passes no effort to OpenCode)—
launchArgs--model --effort--model --effort--model <variant id>--model <variant id> (Gemini) or --model --effort--model --effort--model <provider/model>—
Usage poolssession, weekly, plus fableWeekly for Fablesession, weeklymonthly + Cursor Models or Other Models bucket (which models draw on which is inferred)Per quota group (Gemini Models, Claude and GPT models): 5-hour and long; else Orca's session/weeklyweekly, monthlyNone: pay per use or local; the route says how it is paid—
Refusal verify matchesThere's an issue with the selected model (idles)model is not supported when using Codex… (idles)Cannot use this model (exits)invalid model selection / not recognized as a known model (exits)Couldn't set model … unknown model id (exits)Model unavailable—
Special rule——Cursor auto placed only in Light and Daily——Last resort by default (lastResortAgents); routes note that Orca needs an existing worktree—

Where OrcaCompass and Orca disagree today ​

  • Grok: Orca 1.4.245 refused --model for Grok workers; 1.4.246 accepts --agent grok --model <id> --effort <e> (stablyai/orca#36480) and still refuses effort levels above a model's ceiling (Grok 4.5 above high), which matches the ceilings in grok.ts. End-to-end validation remains incomplete.
  • Effort lists for Claude and Codex come from their local catalogs, Orca's catalog, or the the evidence snapshot fallback. Orca checks effort against its own catalog for models it knows (for example gpt-5.5 up to xhigh) and refuses others. Whether the two lists always agree has not been checked.
  • OpenCode: the worktree rule is only a note in reasons; launchArgs do not add --worktree, so the caller must pick --worktree current or an existing one.
  • Folder trust: OrcaCompass does not check it. Orca handles it for Claude, Codex, Cursor and Antigravity (and Muse through its launch flag). Grok can stop at its trust prompt; OpenCode has no auto-trust in Orca, but whether it asks has not been checked.