Skip to content

Verify and rejections ​

Discovery is an estimate. A model an agent lists can still be one this account cannot use (a cached Claude catalog can be stale, or discovery may fall back to the evidence snapshot; a Codex model may be enterprise-only; a Cursor model may be off in the account). Orca does not check: orca orchestration worker-start reports the requested model as effective even when the agent cannot use it. orcacompass verify closes that gap. Code: src/verify.ts, src/routing/rejection.ts, src/routing/rejections.ts.

What goes wrong without it ​

AgentWhen launched with a model it cannot use
CodexPrints the refusal and idles at its prompt; no worker_done. Orca shows a live worker needing nothing
Claude CodeSame: prints the refusal and idles
Cursorcursor-agent exits 1; Orca fails the start after its 60 s readiness wait
Grok BuildExits 1
AntigravityExits 1
OpenCodePrints Error: Model unavailable: …

How verify works ​

sh
orcacompass verify --dispatch <dispatchId> [--wait 60] --json
  1. orca orchestration worker-show --dispatch <id> --json: the effective agent and model, and the Task id.
  2. With 1.5 s between polls, orca orchestration worker-read --dispatch <id> --source terminal --limit 80 --json: Orca's fleet projection (activity, outcome, liveness) and the terminal tail.
  3. Decide:
VerdictRule
rejectedA terminal line names the model and matches one of the known refusal phrases (below). Soft warnings that still run the model do not count
reportedOrca's outcome is succeeded: the worker already sent worker_done
failedThe terminal exited or the outcome is failed, with no recognized refusal
startedActivity has been working for 8 s (workingMs) with no refusal
stalledThe task header is in the terminal and activity is done while the outcome is still in_progress for 3 s: the agent ended its turn without worker_done (an unrecognized refusal, another error, or a question in its own prompt)
unknownNo verdict within --wait seconds (default and maximum 60), or Orca's commands kept failing. With unsent: true: the agent sat idle the whole time and Orca's === TASK === header never appeared in its terminal, so the prompt was probably not sent yet

The total timeout includes worker-show, reads, and polling delays. Each command uses at most 20 s and is shortened to the time remaining. Verification returns immediately when a verdict is available. --wait accepts 0–60 s; 0 returns unknown immediately. After a timeout, report it and continue normal Orca supervision instead of repeating startup verification.

verify reads worker state and records detected refusals in its local cache. It returns the cleanup argv lists for the orchestrator to run: worker-stop then worker-release, or worker-release alone when the worker already exited.

Refusal phrases recognized ​

Seen in November 2026 with Codex 0.166, Claude Code, cursor-agent, grok 1.0.70, agy 1.4.11 and opencode, launched with an unknown or unavailable model:

AgentRefusal (must be on a line that names the model)
CodexThe 'claude-opus-5-5' model is not supported when using Codex with a ChatGPT account.
Claude CodeThere's an issue with the selected model (gpt-6-terra). It may not exist or you may not have access to it.
CursorCannot use this model: claude-opus-9-9. Available models: …
Grok BuildCouldn't set model 'grok-9-fake': Invalid params: "unknown model id".
Antigravityinvalid model selection (--model "gemini-9-fake" …): model gemini-9-fake is not recognized as a known model …
OpenCodeError: Model unavailable: openrouter/fake/nonexistent-9

Matching is case-insensitive on these fragments: model is not supported, issue with the selected model, cannot use this model, couldn't set model, unknown model id, invalid model selection, not recognized as a known model, model unavailable. Long lines wrap in a terminal, so each line is also tried joined with the next. The saved message drops the agent's decoration (■, ERROR:), Cursor's list of available models, and a sentence cut off by wrapping; it is capped at 240 characters.

Not refusals: Codex's Model metadata for `x` not found and Claude Code's doesn't recognize the model "x", which still run the model. A changed wording in a newer agent version shows up as stalled or failed with the terminal tail instead of rejected.

What a rejection does ​

A rejected verdict is saved to rejected-models.json in the cache: agent, model (the launch id), the agent's words, and when. For rejectionTtlHours (168 hours, a week, by default):

  • every route skips that agent and model, in any tier, listing it under skipped as refused by codex 2h ago: "The 'claude-opus-5-5' model is not supported when using Codex with a ChatGPT account.";
  • status shows the same;
  • a rejection matches any of the model's launch ids, so Cursor's claude-opus-5-5-high and claude-opus-5-5-low are both skipped after one of them was refused;
  • a new refusal of the same agent and model replaces the older entry.
sh
orcacompass rejected           # what is skipped, with the agent's words and age
orcacompass rejected --clear   # forget everything (for example after upgrading a plan)

The replacement flow ​

After rejected, the orchestrator:

  1. runs each cleanup argv in order and checks it succeeded;
  2. routes again with the same flags (the refused model is now skipped, so the route differs);
  3. starts the replacement on the same Task: orca orchestration worker-start --task <taskId> --retry-of <dispatchId> … with the same placement and the new launchArgs;
  4. tells the user which model was refused and what replaced it.

For failed, the same, with --exclude-agent <agent> on the new route for this job. For stalled, read tail first: a model, account or limit error is treated like failed; a question is answered in the worker's terminal.

Example ​

sh
$ orcacompass verify --dispatch d-7f3a --json
{
  "status": "rejected",
  "dispatchId": "d-7f3a",
  "taskId": "t-19",
  "agent": "codex",
  "model": "claude-opus-5-5",
  "message": "The 'claude-opus-5-5' model is not supported when using Codex with a ChatGPT account.",
  "detail": "codex refused claude-opus-5-5",
  "cleanup": [
    ["orca", "orchestration", "worker-stop", "--dispatch", "d-7f3a", "--json"],
    ["orca", "orchestration", "worker-release", "--dispatch", "d-7f3a", "--json"]
  ],
  "skippedForHours": 168,
  "next": [
    "clean up: orca orchestration worker-stop --dispatch d-7f3a --json, then orca orchestration worker-release --dispatch d-7f3a --json",
    "route again with the same flags; codex claude-opus-5-5 is skipped for 168h",
    "start the replacement on the same Task: orca orchestration worker-start --task t-19 --retry-of d-7f3a --worktree <same placement> <new launchArgs> --json"
  ]
}

Text output has the same content: the verdict line, agent said: …, the terminal tail prefixed with |, then a numbered next: list.