scrum-agents

Files (1)

• SKILL.md

SKILL.md (first lines)

Optional `model` on create accepts an explicit provider model ID or alias (max256 characters, no whitespace or command options). Blank/omitted preserves the provider default. list/get display the saved value. Use `agent_manage(project,actor=<human>,action=update,id,model)` to change or clear it for the next authorized run; it never restarts active work. Creation and update validate syntax; provider availability is checked only on the authorized run. Errors appear in Agents/Chat; no replacement model is chosen. No catalog or account probe is needed.

Both providers use their documented `--model` argument: [Codex CLI](https://developers.openai.com/codex/cli/reference/) and [Claude CLI](https://code.claude.com/docs/en/cli-reference). Enter IDs/aliases supported by the configured account; this skill does not maintain a current catalog.

Use agent_manage(project,action=list|get|create|update|start|stop). Match the checkout project; reuse a matching profile after uncertain writes. Lists omit full role text; get(id,include_instructions=true) loads it only when needed. Profiles and launch configuration stay in this checkout's ignored .scrum/local; snapshots never restore running agents.

Creation/control requires an active human actor and the human's authorization to run that provider. A request to implement this feature or merely inspect profiles is not permission to start real agents. create requires name (1..80) and defaults client=codex, instructions='', start_immediately=true. Empty instructions use the standard Scrum worker. To prepare without execution, explicitly pass start_immediately=false.

For a user-selected .md or .toml, read its UTF-8 text and pass instructions plus source=<filename.md|filename.toml>; the MCP does not open arbitrary file paths. TOML is preserved as role text, not parsed into executable/client/model settings. Otherwise pass a role prompt or leave instructions empty. Max 8000 characters; concise role/specialty/preferences only. Role instructions do not authorize additional scope or weaker ownership, quota, Stop or verification rules. There is no custom executable/shell-command field.

Claude start additionally requires confirm_client_account=true for an explicitly authorized configured account and usage_file=<absolute telemetry JSON path>. Do not run, probe or configure a shared Claude account merely to test compatibility. Existing CLI permissions stay in force; the site does not install or sign in to providers.

Inspect request and error: launch_failed means a profile was created but not started; do not duplicate it. starting is not proof of a running model. Initial start waits for fresh usage and respects project Stop; unknown/high quota holds work. Existing connected idle profiles receive a new individual start request; a new process is not spawned for the same connected profile. Wait for a stopped bridge to disconnect before restarting.

stop(id) disables only that managed profile. Its bridge interrupts its own worker, then uses the standard recorded-progress handoff and releases own reservations after process exit. Project Stop/Wakeup still affects participating bridges. Chat records continuation instructions, final outputs or interrupted/held runs. Do not poll with model calls; use status when requested or needed for a concrete operation. Never use process-name/port kills to replace control errors.

The website Agents cards offer Delete agent, using user_manage(actor=<human>,id=<profile actor>,action=delete). This preserves history and recovers unfinished work through the existing deletion flow; pending=true means Stop requested, so wait for exit and retry Delete. Lists mark deleted identities with deleted=true; omit those from active cards. Restore the identity from Users / Show deleted profiles before starting it again. Do not directly delete profile rows or force another worker's reservations.

Project Wakeup now restores the previous enabled managed-profile set, including disconnected bridges. resume_on_wakeup records membership; explicit profile Stop removes it, Start adds it, and launch failure does not erase the intended roster. Existing roles/client/actor are reused. If no profiles or connected manual bridges exist, Wakeup creates/reuses one default Codex agent. Prepared profiles with start_immediately=false are not implicitly enabled. Claude restoration still requires explicit configured-account authorization via confirm_client_account=true; web users can use the profile's existing account-confirmation Start dialog. Wakeup starts managed bridges with an immediate request, while quota checks gate provider execution. For signal-only behavior pass start_agents=false to project_control.

Task model routing: fixed task model (human-only) overrides the human agent-profile model; otherwise automatic model_tier selects economy/standard/advanced. AI may set model_tier with concise model_reason during existing planning or task_update with normal ownership. Auto defaults to standard regardless of scope size. Size is not difficulty: choose economy only for mechanical low-risk work, standard for implementation/debugging/Unity integration, and advanced for demonstrated difficult reasoning. Set the tier and concise reason during existing planning or a necessary task update; never launch a separate selector model. No separate model call makes this decision. Manual model IDs are never replaced automatically. Human task_update(id,actor,model) alone can queue an override during an active claim, applying on the next launch without editing another worker’s code. Managed workers execute only their offered task, close, then exit; the supervisor re-evaluates the next task. For an automatic tier change during work, preserve task continuity and close safely; at most two routing relaunches occur. The desktop chat itself cannot change its own model through this field. No Claude account probe or paid test run is authorized by implementation.

For a human-requested idle supervisor upgrade, start(start_immediately=false) reconnects local control without requesting a model task. It preserves normal Stop/quota/autoresume policy and does not signal Wakeup. Default start behavior is unchanged. Never stop an active worker merely to upgrade instructions.

From martiin3d/scrumstudio, MIT.

More from martiin3d/scrumstudio