Skip to content

orchestration worker-start: accept --model (and --effort) for the pi and devin agents #26971

Open

Description

@Gabrielgvl

Summary

orca orchestration worker-start --agent pi --model <id> and --agent devin --model <id> are refused, although both CLIs accept a model at launch. Please add pi and devin entries to the agent session-option catalog, as #24796 did for OMP and #24624 / #24768 did for OpenCode.

Observed (Orca 1.4.221, Linux; pi 1.1.0; devin 3000.11.3)

$ orca orchestration worker-start --agent pi --model openai-codex/gpt-6-luna --effort low --task <task> --worktree new-child ...
Agent pi does not support launch-time model selection. Omit --model to run the model from its own config.

$ orca orchestration worker-start --agent devin --model swe-2-medium --effort low --task <task> --worktree new-child ...
Agent devin does not support launch-time model selection. Omit --model to run the model from its own config.

Without --model, every worker runs the machine-wide default (pi: defaultModel / defaultThinkingLevel in ~/.pi/agent/settings.json; devin: agent.model in ~/.config/devin/config.json), and the receipt reports launch.effective.model = null. Rewriting those global files before each dispatch races with other coordinators on the same machine (the same concern as #19980 and #22896).

Launch flags the agents already accept

  • pi: --model <provider/id> (optional :<thinking> suffix) and --thinking <off|minimal|low|medium|high|xhigh|max>. Model discovery: pi --list-models.
  • devin: --model <id>. Effort is part of the model id (swe-2-medium / swe-2-high / swe-2-max, claude-fable-5-1-low|medium|high, …). Model discovery: devin models list. A model-only entry, like OMP's, would cover it.

Evidence that the agents honour the flags under Orca supervision

On 1.4.221 we started each agent with the flag in an Orca terminal (orca terminal create --command "pi --model openai-codex/gpt-6-luna --thinking medium", and devin --model swe-2-medium), then bound a supervised worker to it with worker-start --terminal. pi ran gpt-6-luna • medium (its default is gpt-6.1-sol • low), and devin ran SWE-2 Medium (its default is Max). Both received the task and sent worker_done.

We would rather not rely on that path: it skips Orca's managed agent-first launch (for example, devin's first-run workspace trust prompt, #15860) and leaves launch.effective null.

Prior art

Expected

  • worker-start --agent pi --model X --effort Y launches pi --model X --thinking Y.
  • worker-start --agent devin --model X launches devin --model X.
  • The receipt's launch.effective reports the applied model and effort.

Activity

  1. AmethystLiang commented on Oct 12, 2026

    @AmethystLiang
    Contributor

    Thanks for reporting. Looking into it

  2. yoyofreeman commented on Oct 13, 2026

    @yoyofreeman

    Independent corroboration on a separate host (Orca 1.4.224, pi 1.1.0, Linux).

    orca-ide orchestration worker-start --agent pi --model <provider/id> ... is still refused in 1.4.224:

    Agent pi does not support launch-time model selection. Omit --model to run the model from its own config.
    

    Our coordinator dispatches each worker on a per-task model (a free gateway group), and the global-config race you describe is real for us too: several pi sessions share one machine, so rewriting ~/.pi/agent/settings.json before each dispatch is not viable.

    Workaround we run today (same shape as the issue): create the terminal with the flag, then bind the supervised worker to it:

    orca-ide terminal create --command "pi --model litellm/<group>"
    orca-ide orchestration worker-start --terminal <handle> ...
    

    It works, but as you note it leaves launch.effective.model null and skips the managed agent-first launch. A model-only catalog entry for pi (like OMP's) would let us drop the workaround.

  3. TomSpoct commented on Oct 23, 2026

    @TomSpoct

    Same refusal on the worktree/agent-launch path, Windows host:

    $ orca --version
    1.4.237 (win32)
    $ pi --version
    1.4.0
    
    $ orca worktree create --repo id:<rid> --name <task> --agent pi \
        --model "openai-codex/gpt-6.1-sol:medium" --prompt "..." --json
      "code": "invalid_argument",
      "message": "Agent pi does not support launch-time model selection. Omit --model to run the model from its own config."
    

    Use case: using Orca as the coordinator for implementation handoffs. Two constraints collide right now:

    1. The pi agent carries the operator's failover machinery (subscription rotation across accounts), but worktree create --agent pi resolves the model from pi's own config, so the coordinator can't pin the model the task needs.
    2. --agent codex --model <id> --effort <level> works and guarantees the model, but drops pi's failover entirely — the worker dies on one account's exhaustion instead of rotating.

    So the practical workaround today forces a choice between "wrong model" and "no failover".

    Since pi's --model already accepts a full reference with an optional thinking suffix (provider/id:medium), the mapping for this agent could be: --model → --model provider/id, --effort → the :level suffix (or pi's --thinking flag, if that's cleaner than suffix parsing in the launcher).

  4. floriandotorg commented on Oct 30, 2026

    @floriandotorg

    Would be so nice to fix this, it blocks using cheap subagents.

  5. farrrr commented on Nov 7, 2026

    @farrrr

    Still reproducible on Orca 1.4.252 (Linux) with Pi 1.5.0:

    $ orca orchestration worker-start --spec-file spec.md --agent pi \
        --model litellm/zai/glm-5.3 --effort medium --worktree new-child --name pi-probe --json
    invalid_argument: Agent pi does not support launch-time model selection.
    Omit --model to run the model from its own config.
    

    In the shipped build, out/shared/agent-session/options/agent-session-option-catalog.js registers launch catalogs for antigravity, claude, codebuddy, codex, gemini, cursor, grok, muse and omp, with no entry for pi. Pi's discovered catalog (agent-model-catalog.json, discovered.origin = "probe") feeds the model picker but not launch-time validation.

    Two related gaps make this harder to work around:

    • For Pi dispatches, the start receipt's launch.effective.model is null, and worker-read --source transcript returns provider_unsupported, so a coordinator cannot confirm from Orca which model actually ran.
    • The workaround we use is to write .pi/settings.json (defaultProvider, defaultModel, defaultThinkingLevel) into an existing worktree, then start with --worktree path:<it> and no --model. It only works when Pi trusts project settings, and that same trust also lets a worktree auto-load project extensions and MCP config. A native launch mapping (--model to pi --model, --effort to pi --thinking) would let us drop it.

    Related: #36354, #36544.

  6. Akindu23 commented on Nov 9, 2026

    @Akindu23

    I can reproduce the Devin half of this on macOS, Orca 1.4.251, Devin CLI 3000.12.32:

    $ orca orchestration worker-start --agent devin --model swe-2-flash-high --worktree path:<folder> --spec "..." --json
    invalid_argument: Agent devin does not support launch-time model selection. Omit --model to run the model from its own config.
    

    The terminal create --command "devin --model …" + worker-start --terminal workaround from the original report works here too. The worker ran on the model we passed (SWE-2 Flash High; the default in our config is SWE-2 Max) and sent worker_done normally. A model-only catalog entry for Devin, passing --model <id>, would cover our use case. Effort is already part of Devin's model ids.

    A side note in case it's useful: launched the default way (--agent devin, no flags), the worker runs in Devin's auto permission mode. In that mode every shell command waits for approval in the worker's tab, including the orca orchestration worker_done report, so a supervised Devin worker sits at "needs input" until someone clicks through. Starting Devin with --permission-mode smart in the workaround avoided that for read-only tasks. If Orca ends up building Devin's launch arguments, it might be worth letting the permission mode be set too, or allowlisting orca orchestration * for supervised Devin workers.

    Thanks for looking into this. SWE-2 is a great fit for cheap exploration workers, so first-class model selection would be really nice.

  7. maxsroda commented on Nov 10, 2026

    @maxsroda

    Still present on Orca 1.4.254 (Linux), and one more data point.

    $ orca orchestration worker-start --spec-file <file> --worktree current --agent pi --model anthropic/claude-opus-5-5 --effort low
    error: invalid_argument: Agent pi does not support launch-time model selection. Omit --model to run the model from its own config.
    

    Two details the refusal hides:

    • The agent model catalog in Orca's user data directory already has a pi entry that lists pi's models with their effort levels (for example a Mistral-provider model with low, high, max), so Orca knows the choices and still cannot launch one.
    • Omitting --model is not always a way out. If the default model in pi's own config belongs to a provider that is not authenticated (pi auth check --provider <p> returns not_ready) while another provider is ready, the worker can only be started on the unusable default. The caller then has to edit pi's global config for one worker, or leave Orca orchestration and run pi -p --model <id> directly.

    Use case: a coordinator that picks a cheaper or a stronger model per task for its workers.

  8. jondmarien commented on Nov 12, 2026

    @jondmarien

    Still reproducible on Windows, Orca 1.4.256 (2026-11-11).

    orca orchestration worker-start with --agent devin --model … is rejected at validation. There is no dispatch, no new task ID, and no successful creation receipt.

    Request id: 1e1d781d-8dad-4b77-92e8-2c99789c8be9

    On this machine the Devin CLI accepts --model, and Devin's config is agent_model: swe-2-flash-high with default_model: null. Omitting --model would only run that machine-wide default, which is the same limitation already described here.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions