Execute task per plan

Act as the orchestrator. Detect the input shape, pick a mode, and run that mode's procedure from its reference file. The flow is grill → do → PR: every run ends by finishing per reference/finish.md — the pull request(s) are the deliverable.

Principle: the developer starts the run and returns on notification.

Input

$ARGUMENTS — one of:

• empty — no input; work from the current conversation (inline)
• a plain task description — chat text, no ticket (inline)
• a single issue URL, a bare <slug>, or a *-task.md path (sub-agents)
• a *-plan.md path — a pre-built plan (sub-agents, back-compat)
• a PRD issue that has GitHub sub-issues (team)
• a Draft PR URL, or a bare <slug> whose .yoke/ai/<slug>/<slug>-draft.md exists (draft execution)

Optional flag --update-docs enables the documentation step. Without it, /do skips docs unless the plan's frontmatter sets update_docs: true.

Router

Detect the mode from $ARGUMENTS. First match wins:

1. $ARGUMENTS is a path ending -plan.md → sub-agents mode. A plan already exists, so skip planning. Read the file's **Mode:** header for back-compat — an agent-team header maps to team mode instead. Then read and follow reference/mode-sub-agents.md (or reference/mode-team.md when the header says agent-team).
2. $ARGUMENTS is a GitHub pull-request URL (path contains /pull/), or a bare <slug> for which .yoke/ai/<slug>/<slug>-draft.md exists → draft execution mode: a /draft run marked the code and opened the Draft PR; implement it per the review. Read and follow reference/mode-draft-execution.md.
3. $ARGUMENTS is a single issue URL, a bare <slug>, or a *-task.md path → sub-agents mode (investigate → plan → pause → execute). Read and follow reference/mode-sub-agents.md.
4. $ARGUMENTS points at a PRD issue that HAS sub-issues (detect via gh api) → team mode. Read and follow reference/mode-team.md.
5. empty / a plain chat description / no ticket → inline mode. Read and follow reference/mode-inline.md.

Each mode's full procedure lives in its reference file. Read only the one this ladder selects, then execute it.

Modes

Principles

These hold for every mode run:

• Worktree on default branch. Before touching code (after the cold-start pause, if any), enter a worktree per reference/finish.md §2 when the run starts on the default branch.
• Finish at PR. Once the tasks are done, hand off to reference/finish.md §3-§7: finish each repo by its finish policy from the flow map, comment on the ticket, and send the notify carrying the PR link(s).
• Pause only on cold start. The plan-confirmation pause fires only when no grill preceded in the session and no approved plan was handed in. A plan already agreed in grill runs straight through.
• Never merge, except one signal. An explicit "straight to main" from the user is the single merge exception (reference/finish.md §6). Otherwise do finishes at the PR and never merges; it never runs deploy or release.
• Artifacts under .yoke/. Every file do writes lives under .yoke/.
• Commits by convention. Format, ticket ID, and git initiative — from ${CLAUDE_PLUGIN_ROOT}/skills/gca/reference/commit-convention.md (see its "Git initiative and defaults"). Do not ask "commit?" mid-run.
• Read the git memory first. Before modifying a file, read its recent commit history per ${CLAUDE_PLUGIN_ROOT}/skills/gca/reference/history-reading.md. A Constraint: or Directive: that conflicts with the plan is surfaced, not silently overridden; a Rejected: approach is not re-proposed without new evidence.
• Review after each task. Spec compliance → code quality. Mandatory.
• Context isolation. A sub-agent receives only its own task text, not the whole plan.
• CLI output. Run commands with long output (formatter, lint, build, test) with 2>&1 | tail -20.
• Language: match the ticket/input language, or follow the project-level definition in CLAUDE.md / AGENTS.md.

Report

Sub-agents and team modes write .yoke/ai/<slug>/<slug>-report.md and append the finish block from reference/finish.md §7; inline prints its summary in chat. Full template: reference/report-format.md.