coord-assign

Write assigned (agent) and/or model_* (per-task or per-subtask) on a coord task. Atomic: edit → commit → push. Never touches status — use /coord-requeue for that.

Project-root preflight:

• Resolve the canonical project checkout before reading, updating, patching, staging, committing, or pushing: COORD_MAIN=$("${COORD_TOOLS:-$HOME/Projects/coord-wright}/bin/coord-project-root") && cd "$COORD_MAIN"
• This is required when the skill is invoked from a sibling task worktree; assignment must update the main queue that the launchd worker reads.

Syntax

/coord-assign <task-id> [S1,S2,...] [--agent=claude|codex] [--model=<name>]

• <task-id>: required.
• [S1,S2,...]: optional subtask list. Only meaningful with --model.
• --agent=X: set task assigned: X. Subtask list (if any) is ignored when only --agent is given (agent is task-scoped).
• --model=Y: model override. Without subtask list → task-default model_<agent>. With subtask list → per-subtask model_<agent> on each named subtask.

Workflow

1. Parse args.
   • Extract <task-id>, optional S<n> list, --agent, --model.
   • Reject if neither --agent nor --model is supplied — there is nothing to do.
2. Resolve the agent for --model.
   • Infer from model name: opus|sonnet|haiku|*claude* → claude; gpt-*|*codex* → codex.
   • If --agent is also supplied, the inferred agent must match it. Reject mismatches.
   • If --agent is not supplied, read current task assigned via python3 "${COORD_TOOLS:-$HOME/Projects/coord-wright}/bin/coord" show <id> and require the inferred agent to match. Reject mismatches with a message pointing to /coord-split if the user wants per-subtask agent splits (not supported here).
3. Apply the agent change (if --agent supplied).
   • python3 "${COORD_TOOLS:-$HOME/Projects/coord-wright}/bin/coord" update <id> --assigned=<agent> --agents=<agent>
   • coord update --agents sets the agents: frontmatter field through the CLI; no direct file edits needed.
   • For roles.coder: read the current roles map from coord show <id>, build a merged spec preserving all existing keys (e.g. reviewer:claude), then apply: coord update <id> --roles=coder:<agent>,reviewer:<existing-reviewer>. Only include keys that were already present. Skip --roles entirely if the task has no roles: block.
   • roles.reviewer is intentionally left as-is (often stays claude for review quality even on codex-coder tasks).
   • Each coord update call auto-commits-and-pushes. Do not run manual git add, git commit, or git push.
4. Apply the model change (if --model supplied).
   • Task-level (no subtask list): python3 "${COORD_TOOLS:-$HOME/Projects/coord-wright}/bin/coord" update <id> --model_<agent>=<value> If it exits 3 as off-baseline for the task's complexity, report and stop; move the task to another rung only after the user agrees, by changing complexity in the same call (--complexity=simple --model_claude=sonnet).
   • Per-subtask:
     • Read current subtasks block from the task path returned by coord show <id> --path (the ### Subtasks checklist under ## Scope notes).
     • For each named S<n>, replace its model_<agent>: line with the new value (or insert one if missing).
     • Write the modified block to /tmp/coord-assign-subtasks-<id>.txt.
     • python3 "${COORD_TOOLS:-$HOME/Projects/coord-wright}/bin/coord" update <id> --set-subtasks=@/tmp/coord-assign-subtasks-<id>.txt
     • If any named S<n> does not exist in the current subtask list, stop without committing and report which ids were missing.
5. Report.
   • One line summarising the change (agent, model, or both).
   • Confirm with coord show <id> that assigned, agents, and roles reflect the new values.

Rules

• Never touch status. If the user wants to kick a stuck loop, that is /coord-requeue.
• Never use --force or --no-verify.
• Cross-agent subtask routing (S3 on a different agent than the task) is intentionally not supported — recommend /coord-split instead.