kurihada/ai-workflow-skill

Fork 0

Files

T

kurihada 7840b2767f Add orch Codex launch bridge workflow

2026-03-20 18:28:48 +08:00

16 KiB

Raw Blame History

Orch CLI

Purpose

orch is the leader-facing scheduler and control plane. It owns the run, task graph, dependencies, ready queue, dispatch decisions, retries, and reassignment logic.

orch does not replace inbox. It uses inbox as the durable transport and execution record.

In normal operation:

leaders use orch
orch creates and monitors inbox threads
workers continue using inbox
a separate worker runtime or worker agent must still consume the assigned inbox thread after dispatch
Codex-specific launch bridges may sit above orch, but they should consume dispatch output rather than change the CLI contract

Responsibilities

orch is responsible for:

creating a run for one user request or project
defining tasks and dependencies
calculating which tasks are ready
dispatching ready tasks to workers
tracking attempts and mapping them to inbox threads
allocating attempt worktrees for code tasks
surfacing blocked tasks to the leader
sending answers back into the active inbox thread
reconciling thread state into task state
blocking until actionable events arrive for the leader
retrying, reassigning, cancelling, or adding follow-up tasks

Non-Responsibilities

orch should not implement:

worker claiming
direct worker polling
automatic worker-runtime launch
raw message append storage
low-level thread history management

Those belong to inbox.

Core Objects

run: one coordinated execution for a user request
task: one schedulable unit of work
dependency: an edge between tasks
attempt: one execution try for a task
dispatch: the act of materializing a task into an inbox thread
workspace: the branch and worktree assigned to one code-writing attempt

Workspace Model

For code-writing tasks, orch should allocate one Git worktree per attempt.

Strict policy:

dispatch from a concrete committed base_ref
fail dispatch if strict mode is enabled and the leader is implicitly relying on uncommitted state
create a fresh worktree for every retry
do not let workers edit the user's primary checkout

See worktree-execution.md for the full execution model.

Task State Model

planned: task exists but is not yet eligible for dispatch
ready: dependencies are satisfied and it can be dispatched
dispatched: an inbox thread exists but the worker has not started yet
running: the task has been claimed and is actively executing
blocked: the active attempt needs clarification or an external dependency
done: task completed and passed its current acceptance gate
failed: task completed unsuccessfully
cancelled: task was cancelled and should not continue

Suggested transitions:

planned -> ready
ready -> dispatched
dispatched -> running
running -> blocked
blocked -> running
running -> done
running -> failed
failed -> ready through explicit retry
* -> cancelled by leader action

Leader Workflow

The normal leader loop is:

create a run
add tasks
add dependencies
inspect ready
dispatch tasks
arrange or launch a separate worker runtime that consumes the assigned inbox threads
use status for the current operational view; it reconciles first and includes latest attempt and message context
inspect blocked
answer blocked questions
if nothing is actionable, call wait
retry or reassign failures when needed
finish when all required tasks are done

The leader should block on orch wait, not on ad hoc sleep.

CLI Surface

The binary name is orch.

Global Flags

--db PATH
--json

`orch run init`

Create a new run.

Suggested flags:

--run RUN_ID
--goal TEXT
--summary TEXT

Example:

orch run init --db .agents/coord.db --run blog_mvp_001 --goal "Build blog MVP" --summary "Public blog plus admin CRUD"

`orch run show`

Show run metadata and current aggregate status.

Suggested flags:

--run RUN_ID

`orch task add`

Add a task to a run.

Suggested flags:

--run RUN_ID
--task TASK_ID
--title TEXT
--summary TEXT
--default-to AGENT
--acceptance-json STRING
--priority low|normal|high

`orch dep add`

Add a dependency edge.

Suggested flags:

--run RUN_ID
--task TASK_ID
--depends-on TASK_ID

`orch ready`

List tasks ready for dispatch.

Suggested flags:

--run RUN_ID
--limit N

`orch dispatch`

Dispatch a ready task to a worker by creating an inbox thread and the first task message.

Suggested flags:

--run RUN_ID
--task TASK_ID
--to AGENT
--repo-path PATH
--base-ref REF
--workspace-root PATH
--strict-worktree
--body TEXT
--body-file PATH

Behavior:

creates a new attempt
automatically enables strict worktree mode for code-like tasks inferred from task metadata when worktree flags are omitted
resolves the source repository from --repo-path or the current working directory
resolves a committed base revision
creates a branch and worktree for the attempt when the task writes code
creates or links an inbox thread
writes workspace metadata into attempt storage and task payload
moves the task to dispatched
does not start a worker runtime on its own

Integration note:

a higher-level Codex bridge may save this JSON output, render a worker brief, and then spawn a worker sub-agent
that bridge should remain outside the core orch runtime so the scheduling contract stays portable

Strict-mode recommendation:

if --base-ref is omitted and the repository is clean, default to HEAD
if --base-ref is omitted and the repository is dirty, fail dispatch
if --base-ref is provided, resolve it to a commit and use it exactly
if --workspace-root is omitted in worktree mode, default to .orch/worktrees under the source repository

`orch reconcile`

Read inbox state and update run/task state.

Suggested flags:

--run RUN_ID

Behavior:

maps inbox claimed or in_progress to running
maps inbox blocked to blocked
maps inbox done to done
maps inbox failed to failed

`orch blocked`

List blocked tasks and their latest question.

Suggested flags:

--run RUN_ID

`orch wait`

Block until one or more run-scoped events become available.

This is the normal wait primitive for the interactive leader.

Suggested flags:

--run RUN_ID
--for task_ready,task_blocked,task_done,task_failed
--after-event EVENT_ID
--timeout-seconds N

Behavior:

blocks until a later matching event exists
reconciles inbox state while polling so worker thread transitions can surface as task_* events
returns a cursor for the next wait
lets the leader wait for worker activity without manual sleep loops

`orch answer`

Answer the active blocked question for a task by writing into the mapped inbox thread.

Suggested flags:

--run RUN_ID
--task TASK_ID
--body TEXT
--body-file PATH
--payload-json STRING

`orch retry`

Explicitly retry a failed task.

Suggested flags:

--run RUN_ID
--task TASK_ID
--to AGENT
--body TEXT
--body-file PATH

Behavior:

creates a new attempt
links the retry to the prior failed attempt
dispatches a new inbox thread or fresh task message

`orch reassign`

Move a blocked or failed task to another worker.

Suggested flags:

--run RUN_ID
--task TASK_ID
--to AGENT
--reason TEXT

`orch cancel`

Cancel a task or an entire run.

Suggested flags:

--run RUN_ID
--task TASK_ID
--reason TEXT

`orch cleanup`

Remove completed or abandoned attempt worktrees that are no longer needed.

Suggested flags:

--run RUN_ID
--task TASK_ID
--attempt N
--all-completed
--force

`orch status`

Show task state summary for the run.

Suggested flags:

--run RUN_ID

Behavior:

reconciles inbox thread state before returning the view
returns run aggregate counts plus per-task detail
includes the latest attempt for each task when one exists
includes the latest thread message for each task when one exists
includes the latest blocked question for blocked tasks so the leader can inspect the current issue without a separate blocked call in the common case

`orch show`

Show one task with dependencies, attempts, and inbox mapping.

Suggested flags:

--run RUN_ID
--task TASK_ID

`orch council start`

Start a three-reviewer council workflow for one target.

Suggested flags:

--run RUN_ID
--target TEXT
--target-file PATH
--repo-path PATH
--task-id TASK_ID
--target-type text|repo|mixed
--mode brainstorm|review
--output markdown|json|both
--only-unanimous

Default behavior:

fixed reviewer roles: architecture-reviewer, implementation-reviewer, risk-reviewer
analysis only
--target-type mixed
--output both
unanimous-only disabled unless requested
reviewer count fixed at 3 in v1

`orch council wait`

Block until the council has enough reviewer responses to continue.

Suggested flags:

--run RUN_ID
--timeout-seconds N

`orch council tally`

Group similar reviewer suggestions and compute support counts.

Suggested flags:

--run RUN_ID
--similarity strict|normal

Behavior:

groups semantically similar reviewer proposals
assigns consensus, majority, or minority
persists grouped recommendations in orch storage

Default behavior:

--similarity normal

`orch council report`

Render the final grouped council output.

Suggested flags:

--run RUN_ID
--show consensus|majority|minority|all|consensus,majority

Default behavior:

show consensus,majority
preserve minority in persisted storage even if omitted from the main report
support both markdown artifacts and JSON output

Relationship To Inbox

orch should be implemented as a control plane on top of inbox.

orch dispatch writes the first task message into inbox
orch dispatch also writes worktree metadata for code tasks into the attempt record and inbox payload
workers claim and update status through inbox
orch reconcile reads thread state and converts it into task state
orch answer writes an inbox answer message to the active thread

The leader should not need to hand-write inbox send during normal dispatch.

Higher-level workflows such as council review should also run on top of orch, not as a separate infrastructure layer. See council-review.md.

Waiting Model

The leader does not receive worker output as an in-memory push. Instead:

workers write updates into inbox
inbox appends events
orch reconcile converts thread state into task state
orch wait blocks on the run-scoped event stream

This is still a single leader model. orch wait is just the leader's blocking read primitive.

JSON Contract

Every command should support --json.

Suggested success shape:

{
  "ok": true,
  "command": "dispatch",
  "run_id": "blog_mvp_001",
  "task": {
    "task_id": "T4",
    "status": "dispatched",
    "assigned_to": "backend-worker"
  },
  "attempt": {
    "attempt_no": 1,
    "thread_id": "thr_987",
    "base_ref": "main",
    "base_commit": "abc1234",
    "branch_name": "orch/blog_mvp_001/T4/attempt-1",
    "worktree_path": ".orch/worktrees/blog_mvp_001/T4/attempt-1"
  }
}

Suggested wait wake shape:

{
  "ok": true,
  "command": "wait",
  "woke": true,
  "next_event_id": 127,
  "events": [
    {
      "event_id": 127,
      "type": "task_blocked",
      "run_id": "blog_mvp_001",
      "task_id": "T5",
      "thread_id": "thr_t5_attempt1",
      "summary": "Editor choice undecided",
      "payload": {
        "question": "Should the admin editor use a rich text editor or plain textarea in MVP?"
      }
    }
  ]
}

Exit Codes

0: success
10: no ready or matching tasks
20: conflict
30: invalid input or invalid state transition
40: not found
50: storage or internal error

SQLite Schema Draft

These tables should live in the same SQLite file as the inbox tables.

CREATE TABLE IF NOT EXISTS runs (
  run_id TEXT PRIMARY KEY,
  goal TEXT NOT NULL,
  summary TEXT NOT NULL DEFAULT '',
  status TEXT NOT NULL DEFAULT 'active',
  created_at TEXT NOT NULL,
  updated_at TEXT NOT NULL
);

CREATE TABLE IF NOT EXISTS tasks (
  run_id TEXT NOT NULL,
  task_id TEXT NOT NULL,
  title TEXT NOT NULL,
  summary TEXT NOT NULL DEFAULT '',
  status TEXT NOT NULL,
  default_to TEXT,
  priority TEXT NOT NULL DEFAULT 'normal',
  acceptance_json TEXT NOT NULL DEFAULT '[]',
  latest_attempt_no INTEGER,
  created_at TEXT NOT NULL,
  updated_at TEXT NOT NULL,
  PRIMARY KEY (run_id, task_id),
  FOREIGN KEY(run_id) REFERENCES runs(run_id)
);

CREATE TABLE IF NOT EXISTS task_dependencies (
  run_id TEXT NOT NULL,
  task_id TEXT NOT NULL,
  depends_on_task_id TEXT NOT NULL,
  PRIMARY KEY (run_id, task_id, depends_on_task_id)
);

CREATE TABLE IF NOT EXISTS task_attempts (
  run_id TEXT NOT NULL,
  task_id TEXT NOT NULL,
  attempt_no INTEGER NOT NULL,
  assigned_to TEXT NOT NULL,
  thread_id TEXT NOT NULL,
  base_ref TEXT,
  base_commit TEXT,
  branch_name TEXT,
  worktree_path TEXT,
  workspace_status TEXT,
  result_commit TEXT,
  status TEXT NOT NULL,
  created_at TEXT NOT NULL,
  updated_at TEXT NOT NULL,
  PRIMARY KEY (run_id, task_id, attempt_no)
);

CREATE INDEX IF NOT EXISTS idx_tasks_run_status
  ON tasks(run_id, status, priority, updated_at);

CREATE TABLE IF NOT EXISTS events (
  event_id INTEGER PRIMARY KEY AUTOINCREMENT,
  run_id TEXT NOT NULL,
  task_id TEXT NOT NULL,
  thread_id TEXT,
  source TEXT NOT NULL,
  event_type TEXT NOT NULL,
  message_id TEXT,
  summary TEXT NOT NULL DEFAULT '',
  payload_json TEXT NOT NULL DEFAULT '{}',
  created_at TEXT NOT NULL
);

CREATE INDEX IF NOT EXISTS idx_events_run_event
  ON events(run_id, event_id);

Embedded Skill Draft

The following block is a draft SKILL.md for the leader-facing orch skill.

```markdown
---
name: orch
description: Use this skill when the leader needs to plan and schedule work through the orch CLI. It is for creating runs, adding tasks and dependencies, finding ready work, dispatching tasks to workers, allocating task worktrees, reconciling inbox state, waiting for worker events, reviewing blocked tasks, answering them, retrying failures, reassigning work, and cleaning up attempt worktrees. Do not use this skill for worker-side claim or progress updates; use inbox for that.
---

# Orch

Use this skill when you are the leader and need to control the task graph through the `orch` CLI.

## When To Use

- you need to decompose a goal into tasks
- you need to record dependencies
- you need to know which tasks are ready
- you need to dispatch work to workers
- you need to allocate isolated worktrees for code-writing tasks
- you need to inspect blocked tasks and answer them
- you need to retry or reassign a failed task

## Rules

- Prefer `orch` over hand-written `inbox send` for normal leader operations.
- Reconcile inbox state before making new dispatch decisions.
- If nothing is actionable, use `orch wait` instead of manual sleep loops.
- For code tasks, dispatch from a committed base and allocate a fresh worktree per attempt.
- Keep tasks small enough to be checkable and to minimize clarification loops.
- Use `inbox` directly only for inspection or manual repair.
- Keep user-facing discussion in the leader.

## Typical Commands

```bash
orch run init --run blog_mvp_001 --goal "Build blog MVP" --summary "Public blog plus admin CRUD" --json
orch task add --run blog_mvp_001 --task T1 --title "Project skeleton" --summary "Initialize app structure and database wiring" --default-to foundation-worker --json
orch dep add --run blog_mvp_001 --task T2 --depends-on T1 --json
orch ready --run blog_mvp_001 --json
orch dispatch --run blog_mvp_001 --task T1 --to foundation-worker --base-ref main --workspace-root .orch/worktrees --strict-worktree --body-file tasks/t1.md --json
orch reconcile --run blog_mvp_001 --json
orch wait --run blog_mvp_001 --for task_blocked,task_done,task_failed --after-event 0 --timeout-seconds 900 --json
orch blocked --run blog_mvp_001 --json
orch answer --run blog_mvp_001 --task T2 --body "MVP supports draft and published only." --json
orch retry --run blog_mvp_001 --task T7a --to backend-worker --body "Retry after fixing the contract mismatch." --json
orch cleanup --run blog_mvp_001 --all-completed --json
```
```

16 KiB Raw Blame History

Orch CLI

Purpose

Responsibilities

Non-Responsibilities

Core Objects

Workspace Model

Task State Model

Leader Workflow

CLI Surface

Global Flags

orch run init

orch run show

orch task add

orch dep add

orch ready

orch dispatch

orch reconcile

orch blocked

orch wait

orch answer

orch retry

orch reassign

orch cancel

orch cleanup

orch status

orch show

orch council start

orch council wait

orch council tally

orch council report

Relationship To Inbox

Waiting Model

JSON Contract

Exit Codes

SQLite Schema Draft

Embedded Skill Draft

16 KiB

Raw Blame History

`orch run init`

`orch run show`

`orch task add`

`orch dep add`

`orch ready`

`orch dispatch`

`orch reconcile`

`orch blocked`

`orch wait`

`orch answer`

`orch retry`

`orch reassign`

`orch cancel`

`orch cleanup`

`orch status`

`orch show`

`orch council start`

`orch council wait`

`orch council tally`

`orch council report`