union-crew

# Union Crew Use Union Alpha for planning, task assignment, integration, and final acceptance. Use DeepSeek V4.1 Flash for delegated investigation, implementation, tests, and independent review. Apply this workflow to the user's actual task. ## Model setup - Orchestrator: `opencode-go-messages/union-alpha`. - Workers: `opencode-go/deepseek-v4.1-flash`. - Preferred worker role: `router_opencode_go_deepseek_v4_1_flash`. A skill supplies instructions; it cannot switch the current model. Check the active session's model information when available. A default in config.toml is not proof of the current task's model. If another model is active, explain that the user must select Union Alpha in the model picker and invoke this skill again. If identity is unavailable, disclose that limitation and ask the user to confirm the selection before claiming this model team is running. Do not present model roleplay as a switch or start a separate task to work around this requirement. Check the live spawn tool schema for the preferred worker role. Its configured model must be DeepSeek V4.1 Flash. If the role is unavailable, use another exposed role only when its configuration confirms that same model. The `agent_type` list and the optional `model` override list are separate. A DeepSeek role can be available even when DeepSeek is absent from the `model` override list. In that case, select the pinned `agent_type` and OMIT `model` and `reasoning_effort`; do not pass a display name, model slug, or placeholder override. The role selects the model. Absence from the override list alone is not a blocker. For the collaboration schema shown below, always set `fork_turns: "none"` for these worker calls and include the necessary context in `message`. Do not omit `fork_turns` or use `"all"`: a full-history fork inherits the parent's model and reasoning effort. An explicit model override is a fallback only when no suitable pinned role is exposed and the live tool schema accepts the exact worker model. Do not substitute DeepSeek V4 Flash, V4 Pro, an inherited parent model, or another provider route silently. Report unavailable routing as a blocker only after checking available pinned roles and permitted overrides. ### Spawn failures - Rejected `model` argument: reread the live schema and retry once using the selected pinned role, `fork_turns: "none"`, and no model or reasoning override. - Unknown `agent_type`: inspect the exposed roles and their model configurations. A file on disk is not proof that a running session has loaded the role. Refresh the Router model catalog and reopen the Codex session if the role is missing; do not invent a role name or retry the same unavailable role. - Provider HTTP 400, authentication, balance, or connection errors after an accepted spawn: report the exact error and investigate the provider/router. Changing the model argument cannot repair a provider tool-schema error. Verify that the worker completes a bounded task with a real tool call. When available, confirm the routed model in Router logs; worker self-identification and successful spawn creation alone do not establish successful execution. ## Orchestration 1. Identify the requested outcome, inspect the relevant project instructions, and define the checks that will establish completion. 2. Split substantial work into bounded assignments with clear file or module ownership. Start with one or two workers, increasing only when independent tasks and available concurrency justify it. Keep dependent work sequential. For a trivial task with no useful delegation, handle it directly and say so. 3. Spawn DeepSeek workers through the native collaboration tools. Use only argument names offered by the current tool schema. Prefer a fresh context with a self-contained assignment so model selection and scope are explicit. 4. While workers run, resolve decisions, inspect integration points, or prepare acceptance checks. Avoid duplicating the same investigation or editing files currently owned by a worker. 5. Read the returned evidence and inspect the actual changes. Reuse a worker for focused corrections. After two attempts with no useful progress, narrow the task or diagnose the blocker before trying again. Provider or balance failures are routing failures, not permission to change models. 6. Integrate the work and run checks proportionate to the change. For changes benefiting from independent review, assign a fresh DeepSeek worker to inspect the final diff and relevant evidence. Union Alpha decides acceptance. 7. Finish with the outcome, verification performed, and any remaining limitation. Distinguish configured routing from a successful live worker call. ## Worker assignment Include the objective, relevant context and paths, ownership, constraints, acceptance criteria, and the expected evidence. Tell every worker: - You share the workspace with other agents. Preserve existing changes, do not revert others' edits, and keep edits within your assigned ownership. - Follow applicable project instructions. Do not spawn further agents. - Perform the assigned work and verify it. Return changed files, findings with file and line references, checks and outcomes, and unresolved blockers. - If blocked, report the concrete cause rather than broadening scope. Example for the collaboration schema exposing `task_name`, `agent_type`, `fork_turns`, and `message`: ```json { "task_name": "inspect_failure", "agent_type": "router_opencode_go_deepseek_v4_1_flash", "fork_turns": "none", "message": "Investigate the reported failure in <project path>. Read applicable project instructions. Ownership: inspection only, no edits. Reproduce with <focused check> and return the cause with file:line evidence and a proposed fix. You share this workspace; preserve others' changes. Do not spawn agents." } ``` The role pins the model; omit a redundant model override. Adapt the assignment and schema to the live tools. Keep parallel writers on disjoint files; coordinate shared files explicitly. A worker's completion message is evidence to inspect, not a substitute for checking the final artifact.