@clawplays/ospec-cli

OSpec.ai

English | 中文 | 日本語 | العربية

The official OSpec CLI package is @clawplays/ospec-cli, and the official command is ospec. OSpec is a spec-driven, agentic workflow framework for AI coding agents — it brings spec-driven development (SDD) and Loop Engineering (a verifiable plan → act → verify goal loop) to Claude Code, Codex, Gemini, OpenCode, MCP-based agents, and plain CLI workflows.

Prompt Guide | Usage | Overview | Installation | External Plugins | Plugin Release | Issues

Why OSpec?

AI coding assistants are powerful, but requirements that live only in chat history are hard to inspect, review, and close out cleanly. OSpec adds a lightweight workflow layer so the repository can hold the change context before code is written and after the work ships.

• Spec-driven work, saved to your repo — OSpec turns a request into files (proposal, design, plan, tasks, reviews, verification evidence) that live in your repo instead of in chat history, so any assistant (Codex/GPT, Claude Code, Gemini, OpenCode, or plain CLI) can pick up exactly where the last one stopped.
• ospec change — the everyday fast flow — one requirement becomes one active change on a short init -> change -> verify/finalize path, kept lightweight and easy to review.
• ospec goal for larger or riskier work — describe the result you need; the AI asks important questions, writes an inspectable plan, implements the work, runs tests, requests an independent review, updates project docs, and continues until the result is proven.
• You choose how much it may do automatically — L1 only checks, L2 may edit but pauses for important choices, and L3 may continue within limits you set. Progress is saved in the repository, so a later session can resume it.

Install With npm

npm install -g @clawplays/ospec-cli

Official package: @clawplays/ospec-cli
Command: ospec
Verify install: ospec --help

Quick Start

OSpec only takes 3 steps:

• initialize OSpec in your project directory
• create and advance one change for a requirement, document update, or bug fix
• archive the accepted change after deployment and validation are complete

1. Initialize OSpec In Your Project Directory

Recommended prompt:

OSpec, initialize this project.

Claude / Codex skill mode:

/ospec initialize this project.

Command line

ospec init .
ospec init . --summary "Internal admin portal for operations"
ospec init . --summary "Internal admin portal for operations" --tech-stack node,react,postgres
ospec init . --architecture "Single web app with API and shared auth" --document-language en-US

CLI notes:

• --summary: project overview text written into the generated docs
• --tech-stack: comma-separated stack list such as node,react,postgres
• --architecture: short architecture description
• --document-language: generated doc language, choose from en-US, zh-CN, ja-JP, or ar
• AI-first language resolution order: explicit language request in the conversation -> current conversation language -> persisted project language in .skillrc
• CLI language resolution order: explicit --document-language -> persisted project language in .skillrc -> existing project docs / managed for-ai/* guidance / asset manifest -> fallback en-US
• OSpec persists the chosen project document language in .skillrc and reuses it for for-ai guidance, ospec new, and ospec update
• new projects initialized by ospec init default to the nested layout: root .skillrc and README.md, with OSpec-managed files under .ospec/
• plain init does not create optional knowledge maps such as .ospec/knowledge/src/ or .ospec/knowledge/tests/; those appear only when a project already has legacy knowledge content to migrate or when future explicit knowledge-generation flows create them
• CLI commands still accept shorthand like changes/active/<change-name>, but the physical path in nested projects is .ospec/changes/active/<change-name>
• if you pass these values, OSpec uses them directly when generating project docs
• if you do not pass them, OSpec reuses existing docs when possible and otherwise creates placeholder docs first

2. Create And Advance A Change

Use this for requirement delivery, documentation updates, refactors, and bug fixes.

Recommended prompt:

OSpec, create and advance a change for this requirement.

Claude / Codex skill mode:

/ospec-change create and advance a change for this requirement.
/ospec-goal create and advance a full goal for this requirement.

Command line

ospec new docs-homepage-refresh .
ospec new fix-login-timeout .
ospec new update-billing-copy .

Agent Execution (Goal Workflow)

The classic change flow above stays simple: proposal.md → tasks.md → implement → verification.md → review.md, with no controller layer. The agent controller layer — parallel worker dispatch, reviewer gates, and durable evidence — belongs to the full goal workflow. Use it with ospec goal, or on a single change only when you explicitly opt into agent/worker execution. OSpec keeps the controller state in repo artifacts, and the current AI harness starts native worker agents when available.

ospec session .
ospec execute bootstrap changes/active/<goal-name>
ospec execute workspace changes/active/<goal-name>
ospec execute status changes/active/<goal-name>
ospec execute dispatch changes/active/<goal-name> --limit 2
ospec execute launch changes/active/<goal-name> --task <task-id> --target codex
ospec execute complete <task-id> changes/active/<goal-name> --status DONE --summary "..."
ospec execute review changes/active/<goal-name> --task <task-id>
ospec execute verify changes/active/<goal-name> --command "npm test" --status PASSED --exit-code 0

launch writes artifacts/agents/launch-plan.md; it does not start workers by itself. Codex/GPT use spawn_agent plus the harness wait lifecycle (and close only when exposed), Claude Code uses Task, Gemini uses @generalist, and OpenCode uses @mention. Use launch --run --command or orchestrate --command only when the current harness cannot start native subagents and the user explicitly accepts CLI fallback.

3. Archive After Acceptance

After the requirement has passed deployment, testing, QA, or other acceptance checks, archive the validated change.

Recommended prompt:

OSpec, archive this accepted change.

Claude / Codex skill mode:

/ospec archive this accepted change.

Command line

ospec verify changes/active/<change-name>
ospec finalize changes/active/<change-name>

Archive notes:

• run your project-specific deploy, test, and QA flow first
• use ospec verify to confirm the active change is ready
• use ospec finalize to rebuild indexes and archive the accepted change
• new nested projects archive under .ospec/changes/archived/YYYY-MM/YYYY-MM-DD/<change-name>; CLI shorthand under changes/archived/... still works
• existing flat archives are reorganized by ospec update

Goal: Use It For Work That Needs More Care

Use a goal when the work touches several parts of the project, has important design choices, changes an API or data, carries security or migration risk, or will take several rounds to finish. For a small, well-defined edit, use ospec new instead.

Start from a terminal:

ospec goal improve-checkout --level L2 --target codex --execution-model controller --harness-interactive true --native-subagents supported

Then tell the AI what outcome you need in ordinary language. You can also skip the terminal command and say: "Use OSpec goal for this requirement and carry it through to completion."

That is all a normal user needs to operate. The AI will:

• Ask only the choices that materially change the result.
• Write down the agreed approach so you can inspect it before code changes begin.
• Split the work into safe pieces and run independent pieces in parallel when useful.
• Test the implementation, ask a separate reviewer to check it, and fix the findings.
• Update the relevant project documentation and indexes before archiving the completed work.

You remain in control. The AI explains what it is about to do, pauses when it needs a decision, and records task, review, repair, verification, and loop progress in the repository so a fresh worker or later session can continue without replaying the conversation. You do not need to run the internal ospec execute commands yourself.

Choose how much the goal may do on its own with --level L1|L2|L3 (default L1):

• L1 checks and reports, but does not change project files.
• L2 may make changes and pauses for important decisions.
• L3 may continue without waiting for routine steps, but only inside non-empty path and command allowlists.

Required user decisions block every level. The integrated goal loop reads task-graph.json, emits a bounded parallel batch, and observes durable completion or review evidence before advancing. When the ospec-goal skill is active and the IDE exposes a native subagent API, controller mode requires the current IDE AI to launch one fresh native subagent per referenced packet, wait for the batch, record evidence, and continue ticking without another user prompt; it does not stop at Loop initialization or require loop watch. L1 remains report-only, so executable work requires the user to select L2 or L3. If the IDE lacks native subagents, controller auto-dispatch fails clearly instead of silently pretending the CLI started an IDE agent.

To see progress, child executor ids, heartbeats, leases, token sources, and active guards, run ospec loop status. Use ospec loop configure for concurrency, allowlists, test commands, deadlines, token/time/iteration budgets, no-progress limits, comprehension checkpoints, and prompt bounds. IDE controllers persist child ownership with ospec loop heartbeat and each result with ospec loop result; after a confirmed session/child loss, ospec loop recover --force expires only unfinished items so they can be requeued without duplicating completed siblings. Advanced loop and triage commands are documented in docs/loop-engineering.md.

Internally, OSpec keeps implementation and independent review separate, feeds blocked work and review findings into retry or grouped repair, and requires current test evidence before the goal is considered complete.

Claude Code hard enforcement (one-time; the AI runs this for you automatically in a Claude Code harness):

ospec session hook --target claude --apply

This writes a hook bundle under .ospec/hooks/claude/ and merges it into .claude/settings.json (idempotent and reversible). The hooks:

• announce every subagent dispatch and every ospec command at the tool level,
• hard-block subagent dispatch while a required decision is still pending,
• inject the static Announce-Before-Act / Brainstorm-First contract only on startup, clear, or compact, while prompt submission stays silent unless a required decision is pending.

Hooks load at session start, so they take effect from the next Claude Code session.

Update With npm

For an existing OSpec project, after upgrading the CLI with npm, run this in the project directory to refresh the project's OSpec files:

ospec update

ospec update also migrates legacy root-level build-index-auto.cjs / build-index-auto.js tooling into .ospec/tools/build-index-auto.cjs and refreshes OSpec-managed hook entrypoints to use the new location. It also repairs older OSpec projects that still have an OSpec footprint but are missing newer core runtime directories, refreshes managed skills and archive layout metadata, and syncs project assets for already-enabled plugins. For nested projects that still carry legacy knowledge under .ospec/src/ or .ospec/tests/, ospec update migrates those paths into .ospec/knowledge/src/ and .ospec/knowledge/tests/. When an already-enabled plugin has a newer compatible npm package version available, ospec update upgrades that global plugin package automatically and prints the version transition. It does not upgrade the CLI itself, and it does not enable plugins or migrate active / queued changes automatically. It also does not switch a classic project layout to nested automatically. If you want to convert an older classic project to the new layout, run ospec layout migrate --to nested explicitly.

How The OSpec Workflow Works

┌─────────────────────────────────────────────────────────────────┐
│  1. USER REQUEST                                               │
│     "OSpec, create and advance a change for this task."       │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│  2. INIT TO CHANGE-READY                                       │
│     ospec init                                                 │
│     - .skillrc                                                 │
│     - README.md                                                │
│     - .ospec/                                                  │
│     - .ospec/changes/active + .ospec/changes/archived          │
│     - .ospec/SKILL.md + .ospec/SKILL.index.json + .ospec/for-ai│
│     - .ospec/docs/project/* baseline knowledge docs            │
│     - reuse docs or fall back to placeholders                  │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│  3. EXECUTION                                                  │
│     ospec new <change-name>      # classic fast change          │
│     ospec goal <goal-name>       # full goal workflow           │
│     ospec brainstorm / plan (optional pre-change aids)         │
│     ospec session                                              │
│     ospec session hook                                         │
│     ospec progress                                             │
│     ospec execute bootstrap / handoff / doc-review / status    │
│     ospec execute next                                         │
│     ospec execute workspace / worktree / worktree --create     │
│     ospec execute worktree --cleanup / finish                  │
│     ospec execute dispatch / launch / collect / retry / review │
│     ospec execute debug                                        │
│     ospec execute tdd                                          │
│     ospec execute verify                                       │
│     ospec execute sync                                         │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│  4. DEPLOY + VALIDATE                                          │
│     project deploy / test / QA                                 │
│     ospec verify                                               │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│  5. ARCHIVE                                                    │
│     ospec finalize                                             │
│     rebuild index + archive                                    │
└─────────────────────────────────────────────────────────────────┘

Core Concepts

Concept	What It Means
Protocol Shell	The minimum collaboration skeleton: root .skillrc and README.md, plus managed OSpec files under .ospec/ for change state, SKILL docs, index state, for-ai/ guidance, and project docs.
Project Knowledge Layer	Explicit project context such as docs/project/*, layered skill files, and index state that AI can read consistently.
Active Change	A dedicated execution container for one small or routine requirement, using the classic fast files: proposal.md, tasks.md, state.json, verification.md, and review.md, plus plugin artifacts when activated.
Active Goal	A full-workflow execution container created with ospec goal, adding design.md, implementation-plan.md, artifacts/agents/task-graph.json, handoff artifacts, document-review artifacts, launch-plan artifacts, worker-run artifacts, reviewer-run artifacts, retry artifacts, review artifacts, artifacts/agents/worker-status.md, and evidence artifacts.

Features

• Change-ready initialization: ospec init creates the protocol shell and baseline project knowledge docs in one pass.
• Guided initialization: AI-assisted init can ask once for missing summary or tech stack; direct CLI init falls back to placeholder docs when context is missing.
• Stable project language: the chosen document language is stored in .skillrc so later guidance and generated change docs stay consistent unless you explicitly change it.
• One indexed document per archived change: every successful finalize or archive writes docs/project/changes/<archive-path>.md, then links it from docs/project/feature-index.md, SKILL.index.json.documents, and SKILL.index.json.archived_changes. Before moving the active change, archive preflight refuses to overwrite a human-owned file at that path and verifies the managed output directories are writable. Goals still update the relevant human-maintained architecture, API, module, or operational docs; the generated change summary does not replace them.
• Scoped review evidence and cost metrics: task and final review dispatches write artifacts/agents/review-packages/*.diff with scoped Git evidence, while artifacts/agents/execution-metrics.json records packet/report/package bytes and task duration. Goal task graphs can enable documentation_updates so missing or undeclared project docs block archive.
• Tracked requirement execution: small changes keep proposal, tasks, state, verification, and review files aligned; full goals also keep design, implementation plan, task graph, handoff, review, worker status, and evidence artifacts aligned.
• Goal experience contracts: every goal runs with Announce-Before-Act (the AI announces its skill and stage, the ospec execute … command and the artifact it writes, and each subagent dispatch), Brainstorm-First (open direction, architecture, API, UI, risk, and scope decisions are asked one at a time through the native question UI before design is locked), and Zero-Setup (you only start a goal and describe the requirement — the AI runs every ospec command itself). In Claude Code, ospec session hook --target claude --apply adds hooks that announce every dispatch and hard-block subagent dispatch while a required decision is still pending.
• Optional pre-change aids: ospec brainstorm writes durable exploration artifacts under .ospec/brainstorms/, with an optional static visual companion; ospec plan writes plan drafts under .ospec/plans/ and only updates implementation-plan.md when --apply is passed. The default small-change flow starts with ospec new; the full workflow starts with ospec goal.
• Session brief and hooks: ospec session writes .ospec/session-brief.json and .ospec/session-brief.md so agents or humans entering an existing project can see active changes, queued changes, queue-run state, indexed document and archived-feature counts, a cache fingerprint, and the next safe command before touching a change; ospec session hook --target claude writes opt-in harness startup artifacts plus a Claude Code hook bundle under .ospec/hooks/, and --apply idempotently merges it into .claude/settings.json.
• Integrated goal loop: ospec loop run --once emits token-bounded task/review/verification action batches for fresh native subagents, while CLI-driven ospec loop watch executes supported external agents in fresh processes. Both use packet paths instead of duplicating the whole goal, persist pending actions and feedback, route review failures through retry or one grouped repair wave, and enforce required decisions, L3 allowlists, budgets, no-progress stops, and comprehension-review pauses.
• Task graph controller: ospec execute bootstrap writes a one-change startup/resume snapshot with the project session brief snapshot and next safe action; handoff writes a cross-tool worker handoff guide with the project session brief snapshot; doc-review creates design and implementation-plan reviewer packets before task execution; status and next report controller state and safe next task candidates; workspace records git workspace safety before worker handoff; worktree records an isolated-worktree preparation plan by default, while explicit --create or --cleanup runs the matching git worktree command and captures artifacts/agents/worktree-runs/; finish records closeout readiness before finalize, archive, push, merge, or worktree cleanup; dispatch and complete create parallel-safe worker packets with worker profiles and target tool mapping, then record task results as OSpec artifacts; review --task creates one combined per-task code review packet (spec compliance + code quality in a single pass) that blocks dependent tasks until approved, while final review creates one combined whole-change code review packet; launch writes the native agent launch plan for the current AI harness, including Codex/GPT spawn_agent, Claude Code Task, Gemini @generalist, and OpenCode @mention guidance; orchestrate is the final CLI fallback for harnesses without native subagents and runs explicit command templates only; launch --run --command is the single-worker CLI fallback; collect turns a fallback worker run into task completion state; retry reopens blocked, needs-context, or failed task work; explicit review --run --command captures artifacts/agents/review-runs/; debug, tdd, and verify record durable evidence; sync rebuilds worker-status.md from execution and review artifacts.
• Adaptive review and model routing: .skillrc.workflow.document_review_policy keeps independent document review by default. With adaptive, inline preflight still requires an explicit risk_level: low (or none) declaration and no detected risk signal; .skillrc.workflow.model_profiles maps logical roles without provider model names in OSpec defaults.
• Measured execution and grouped repair: command runners can write authoritative usage to OSPEC_USAGE_FILE for automatic ingestion, while --usage-file remains a manual input. Metrics distinguish complete, partial, and missing coverage. ospec execute repair turns all structured NEEDS_CHANGES findings into one repair task.
• Verified durable documentation: declared documentation targets capture before/after normalized content hashes, so an unchanged file cannot satisfy a new run. Feature indexes link completed work directly to the durable project documents it updated.
• Queue helpers: queue and run support explicit multi-change execution when one active change is not enough.
• Plugin workflow gates: plugin commands support Stitch design review and Checkpoint automation through npm-installed official plugins.
• Skill management: install and inspect OSpec skills for Codex and Claude Code.
• Standard closeout: finalize verifies and archives the change, then refreshes the generated feature locator and knowledge index before manual Git commit. It never overwrites human-maintained architecture, module, or API prose.

Plugin Installation

OSpec supports plugins for UI review and runtime validation. Keep the public flow simple:

/ospec open Stitch for this project.
/ospec open Checkpoint for this project.

In AI / /ospec flows, requests like "open Stitch" or "open Checkpoint" should be handled as: check whether the plugin is already installed globally, install only when missing, then enable it in the current project.

Command line fallback:

ospec plugins list
ospec plugins install stitch
ospec plugins enable stitch .
ospec plugins install checkpoint
ospec plugins enable checkpoint . --base-url http://127.0.0.1:3000

Official npm plugin packages:

• @clawplays/ospec-plugin-stitch
• @clawplays/ospec-plugin-checkpoint

After a plugin is enabled, its detailed setup docs are synced into .ospec/plugins/<plugin>/docs/.

Maintainers can find plugin publishing and automation details in docs/plugin-release.md.

Documentation

Core Docs

• Prompt Guide
• Usage
• Project Overview
• Installation
• Skills Installation
• External Plugins
• Plugin Release

Repository Structure

dist/                       Compiled CLI runtime
assets/                     Managed protocol assets, hooks, and skill payloads
docs/                       Public documentation
scripts/                    Release and installation helpers
.ospec/templates/hooks/     Hook templates shipped with the package

License

This project is licensed under the MIT License.