Skip to content

[Feature] automation cannot set params on agents #11043

Open

Description

@orca-discord-issues

hdsheena hdsheena via Discord

I want my automation to run as yolo, but i strictly DONT want this as default for this agent or project.

Activity

  1. orca-discord-issues commented on Jul 28, 2026

    @orca-discord-issues
    Author

    hdsheena hdsheena via Discord

    if i could add a custom agent, i could set up command code with yolo by default and use that in my automation?

  2. hdsheena commented on Jul 31, 2026

    @hdsheena
    Contributor

    [Feature] Automation runs cannot set per-run agent launch params (e.g. --yolo, --model)

    Related issues - this builds off the work done on these already:

    This issue is the focused slice of the "automation launch state" gap called out in #5700/#5708:
    #5708's definition of done includes run history recording the launch command and exposing
    clear "target, provider, schedule, precheck, setup, and launch state", but neither umbrella
    issue specifies how an automation owner sets per-run agent launch args (e.g. --yolo,
    --model). This issue narrows that to a concrete, shippable API: per-automation
    providerArgs passthrough.

    Description

    Orca automations can only select an agent by id (--provider <agent>). There is no way to
    pass launch arguments to that agent for a given automation run — e.g. Command Code's
    --yolo (skip permission prompts), --model <id>, -p/headless, --permission-mode,
    --add-dir, etc.

    Concretely:

    # today — cannot express "run this automation's command-code session with --yolo"
    orca automations create --name "nightly" \
      --provider command-code \
      --trigger daily --time 02:00 \
      --prompt "..." \
      # no way to add: --yolo / --model gpt-5.5 / --permission-mode auto-accept

    Motivation / Use case

    Automations that run unattended must not block on interactive permission prompts.

    Scheduled automations (cron, e.g. every 3 hours) run in new_per_run worktrees with a fresh
    terminal. The agent session does real work: it runs shell commands, edits files, commits, and
    rebase-pushes to origin/master. Without --yolo (or an equivalent auto-accept mode at the
    agent level), the session stalls at Command Code's "accept edits on [shift+tab]" permission
    prompt whenever the action isn't covered by the project's .commandcode/settings.json
    allow-list. Rebase-conflict resolution, edits to regenerated data files, and the initial
    worktree setup are the classic triggers.

    The only current workarounds are:

    1. Set "defaultMode": "auto-accept" in the project .commandcode/settings.json — but
      that applies to every session in the project, including interactive ones, which the user
      explicitly does not want.
    2. Hand-roll orca terminal create --command "command-code --yolo" inside the prompt — but
      the top-level automation session is launched by Orca itself and its launch command is
      not configurable, so the automation's own session still stalls.

    This is the exact gap: per-automation agent params, scoped to that automation's runs only,
    not the agent or project default.

    Desired behavior

    Add a way to set agent launch params on an automation, applied to every run of that
    automation:

    # proposed
    orca automations create ... \
      --provider command-code \
      --provider-arg "--yolo" \
      --provider-arg "--model gpt-5.5" \
      ...
    
    # or JSON form
    orca automations create ... --json '{
      "provider": "command-code",
      "providerArgs": ["--yolo", "--model", "gpt-5.5"]
    }'

    Requirements:

    • Scoped per automation. Setting args on automation A must not affect other automations,
      interactive sessions, or the agent's default.
    • Applied at launch. Each run's terminal session starts the agent with these args
      (equivalent to command-code --yolo ... for the command-code provider).
    • Repeatable (--provider-arg / array of strings) so multiple flags can be passed.
    • Persisted and visible in orca automations list/show --json so config-as-code
      workflows can audit them.
    • Editable via orca automations edit (replace the full arg list).
    • Optional passthrough validation: unknown/unsupported args should be surfaced (warn or
      error) rather than silently ignored, since they are provider-specific.

    Acceptance criteria

    • orca automations create accepts a repeatable --provider-arg flag (or equivalent) and
      persists it.
    • orca automations edit <id> --provider-arg ... replaces the stored args.
    • orca automations list --json / show <id> --json includes the args in the response.
    • When an automation with --provider-arg "--yolo" runs, the launched agent terminal
      session starts with --yolo (verify via run's terminal session — e.g. the session is in
      yolo/auto-accept mode and does not prompt on an action outside the project allow-list).
    • A second automation without args still launches with the provider's default (no args
      leaked across automations).
    • Empty/absent args behave as today (no regression for existing automations).

    Implementation notes (for the agent)

    • The command-code provider is registered under agentId: "command-code"; automation run
      records have sessionKind: "terminal" + terminalSessionId — the launch happens through
      the same path used by orca terminal create --command "command-code ...", so reuse that
      command-assembly code and inject providerArgs into the terminal launch command.
    • Persist providerArgs (list of strings) alongside the existing automation fields
      (prompt, precheck, rrule, workspaceMode, ...) visible in
      orca automations list --json today.
    • Consider a --provider-arg-append/--provider-arg-reset distinction if edit semantics get
      complex; simplest correct behavior is full replacement on edit.
    • The usage field on run records is frequently provider_unsupported — do not rely on
      it to detect launch params; prefer reading back the automation definition.

    Out of scope

    • Changing agent/project default permission modes (explicitly NOT wanted — must stay
      scoped per automation).
    • Per-run (ad-hoc) override for orca automations run <id> --provider-arg ... — nice
      future extension, not required here.
  3. pablontiv commented on Aug 3, 2026

    @pablontiv

    Related surface, same shape. I hit this on the orchestration side and filed #12360.

    orchestration worker-start has no per-dispatch agent parameter either. A supervised worker inherits whatever settings.agentDefaultArgs holds globally, and that record is GUI-only: git grep agentDefaultArgs over src/cli at v1.4.167 returns zero hits, and there is no settings command spec at all.

    The failure mode is the same one you describe, pointed at a different surface. When the global resolves to empty, which is exactly what Manual mode writes (src/shared/tui-agent-permissions.ts:152 at v1.4.167), every dispatched worker launches interactive and parks on its first approval prompt with nobody attached to answer it.

    Disposition: not a duplicate, and neither blocks the other. PR #11873 already models what worker-start needs, a per-invocation agentArgs field with a CLI flag that falls back to the global default when unset. If that lands first, extending the same field to WorkerStartParams is the smaller follow-up rather than a second design.

  4. AutomatorAlex commented on Nov 5, 2026

    @AutomatorAlex

    Adding a current OMP use case for automation-scoped launch settings (Orca 1.4.248, OMP 18.16.0 on Linux).

    --extra-agent-args now exists, but Orca's OMP allowlist still reports Allowed: --model. OMP itself supports --config <file>, --profile <name>, and --no-context-files.

    I want scheduled runs to disable OMP's advisor through a configuration overlay and omit automatically injected AGENTS.md instructions, while retaining each automation's saved task prompt. Interactive sessions should keep their normal settings. This needs configuration options beyond model selection; a global agent command override would affect interactive launches too.

    The current workaround is a Fish wrapper that recognizes and strips a marker from the automation's startup prompt, then applies --no-context-files --config <automation-overlay>. It works in a verified Orca launch, but depends on the shell and prompt-delivery convention.

    Please support automation-specific profiles/configuration overlays or additional supported launch arguments through both UI and CLI, applied consistently to scheduled and manual runs without changing other sessions. Clear validation errors are welcome; unrestricted shell execution is not required.

    Related: #34274 covers the argument editor, and #12930 covers reusable launch profiles.

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

Metadata

Metadata

Assignees

Labels

enhancementNew feature or request

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions