overdo-orchestrate

Run Overdo task graphs to completion using MCP as the only mutation boundary.

Use when

• User asks to run/continue an imported plan.
• User asks to recover after interruption or failure.

Required flow

1. Confirm Overdo MCP tools are available in this session.
2. Read next ready task(s) from Overdo.
3. Claim work with lease-safe dispatch.
4. Execute required gates (lint, unit, integration, e2e).
5. Record loop iteration and decision (retry, escalate, complete).
6. Apply task transitions and release leases.
7. Enqueue/process commit via commit coordinator.
8. Emit concise progress + evidence summary.

Tool availability gate (must pass first)

Before orchestration, verify these MCP tools are present:

• overdo_task_next_ready
• overdo_task_start
• overdo_task_complete
• overdo_task_progress
• overdo_orchestrate_autocontinue

If any are missing:

• Stop immediately.
• Report that Overdo MCP is not attached for this session.
• Do not orchestrate via ad-hoc shell commands.

MCP tools to call (required)

Use Overdo MCP tools directly for orchestration:

1. overdo_task_next_ready
2. overdo_task_start
3. overdo_task_complete
4. overdo_task_progress
5. overdo_orchestrate_autocontinue for continuous advancement without manual step-by-step prompts

When user asks for "autocontinue", prefer overdo_orchestrate_autocontinue with a bounded maxSteps and report progress after each run.

Guardrails

• Never mutate orchestration state outside MCP.
• Respect task blockers/dependency ordering.
• Persist every failed attempt with artifact/repro details.
• Resume from current SQLite state first; do not reinitialize existing DB.
• Keep commit processing serialized (single commit owner at a time).

Failure behavior

• On gate failure, record iteration and retry only within policy limits.
• On exhausted retries, escalate with actionable failure context.
• On crash/restart, continue from persisted state rather than replaying manual steps.

Output format

• Active task(s): claimed IDs and worker owner.
• Gate results: pass/fail per required gate.
• Loop decision: retry/escalate/complete with reason.
• State updates: transitions, lease release, commit queue status.
• Next action: exact command/prompt to continue.

Example invocation

/overdo-orchestrate --db .overdo/tasks.db --autocontinue