Command Palette

Search for a command to run...

UnylyUnyly
Весь каталог

O Coding Navigation

БесплатноНе проверен

Local-first AI coding navigation CLI and MCP server with step gates, audit trail, and SOP-aware project state.

uncletimgzавтор: uncletimgz
GitHubEmbed

Описание

Local-first AI coding navigation CLI and MCP server with step gates, audit trail, and SOP-aware project state.

README

Local-first, MCP-first, state-machine-driven AI coding workflow operating system. CLI: ocn · MCP: ocn-mcp · License: Apache-2.0 Phase: SOP profile 0.9.0 mainline (Outcome Backbone, AM-017/DEC-043) + optional Auto Mode (AM-009) + optional Contract Backbone (AM-012) · Status: pre-GA beta · Public: on npm as latest and @beta → 0.9.0-beta.1 · GitHub pre-release: v0.9.0-beta.1

📑 This README has two parts: Part 1 — English (sections 1 – 11) · Part 2 — 中文版 (§§ A – K) 本 README 分两部分阅读：先看英文（§§ 1–11），中文从下方 §A 开始。

Part 1 · English

OCN turns AI coding from continuous-chat improvisation into a navigable, gated, auditable, reviewable systems-engineering process. It is a navigator, not an IDE, not a SaaS, not a project-management board.

Table of contents (English)

Understand OCN

  1. What OCN is
  2. Why OCN exists
  3. Current status

Use OCN 4. Install 5. First 5 minutes

  1. Core CLI commands
  2. MCP tools

Reference 8. Documentation map 9. Development 10. Roadmap 11. License

Understand OCN

1. What OCN is

OCN is the local discipline layer for AI coding. It runs on your machine, persists everything to plain files (Markdown + JSON + JSONL + YAML), exposes a small CLI for humans, and a small MCP server for agents. There is no cloud component.

OCN sells discipline — productized as:

  • a state machine (DISCOVERY → SPEC → DESIGN → PLAN → BUILD → VERIFY → SHIP → REFLECT) with forward-only transitions,
  • a Step Artifact Gate that blocks advancement when the current step's required sections are missing,
  • a dual-track audit trail (.ocoding/audit/audit-events.jsonl + docs/22-audit-trail.md),
  • a Logic Backbone (SOP 0.3.0) — a DESIGN-phase artifact whose computation/decision graph is machine-validated; ocn check blocks on orphan scores, dangling references, cycles, and unbound triggers, so the system's logic is wired before BUILD,
  • a Readiness Backbone (SOP 0.4.0) — a role-based cross-cutting readiness gate: 55 falsifiable checks derived from 54 curated IT roles (developer, QA, DevOps, CISO, service desk, …) run inside ocn check / ocn gate / ocn advance on every step; open-world semantics mean both FAIL and UNKNOWN block (silence is not a pass), catching role-blind completion before SHIP,
  • a Task Backbone (new in SOP 0.5.0) — build plans carry machine-parseable Task Spec blocks (goal / traces / touches / verify / DoD); the build-plan gate validates six hard defects and freezes each task's verify-command hash into .ocoding/task-ledger.json; a task is done only when ocn task check reruns the frozen command and it exits 0, and ocn advance out of BUILD is blocked while tasks are pending — closing the fourth false-completion class, receipt-only completion,
  • an Acceptance Backbone (new in SOP 0.8.0) — docs/03-acceptance-criteria.md carries a machine-parseable ## Acceptance Specs｜验收规格 section (### AC-<DOMAIN>-<n> blocks); the acceptance gate runs at step_acceptance_criteria, hard-blocks four structural defects (no_specs, duplicate_id, invalid_id, missing_field), and on pass freezes .ocoding/acceptance-specs.json — the canonical machine source of AC ids that build-plan traces bind to, closing the hole where AC written in a table silently didn't register,
  • a safe MCP surface that lets agents read, prepare, and create artifacts but never advance state, capture decisions, reset the project, or force-release the lock.

OCN is not a code generator, an IDE, a SaaS, a project-management tool, a notes app, or a scaffold-only doc factory.

2. Why OCN exists

Working with an AI coding agent for any non-trivial task tends to fail in four ways:

Failure Symptom
lost Nobody — human or AI — knows which step the project is on.
drift The agent keeps generating, but each new chunk is further from the original requirement.
amnesia A new chat starts and the agent re-derives what the previous chat already decided.
false-completion A document exists on disk, but it's missing a required section. The agent declares "done" anyway.

OCN treats these as the same problem: the AI coding loop has no rigorous notion of "where we are" and "what counts as done". OCN supplies both as code, not as exhortation.

3. Current status (SOP 0.8.0 (Acceptance Backbone) mainline)
Phase SOP 0.8.0 (Acceptance Backbone) mainline — Plan → Build → Verify, with the machine-verified DESIGN logic-backbone gate, a role-based cross-cutting readiness gate (55 checks) on every step, a frozen-verify task ledger that gates BUILD exit, and a SPEC-phase acceptance gate that freezes the canonical AC ids; plan-to-verify smoke covers all 20 steps
Tests full vitest suite green on Node 20 + Node 22
Coverage meets the publish-time gate
npm latest → 0.9.0-beta.1; beta → 0.9.0-beta.1; alpha → 0.1.0-alpha.2 (historical; preserved)
Maturity pre-GA beta — not stable, not GA, beta only (for controlled testing / dogfood)
External host validation Validated with Claude Desktop on Windows with WSL2. Cursor and Cline are not yet verified.
MCP transport stdio only (HTTP/SSE not started)

2026-05-04 republish note: Republished as 0.2.0-beta.2 to ship the Execution Navigator commands (exec status, github analyze-pr, evidence map, next-prompt, verify status, verdict draft) that were merged after 0.2.0-beta.1 was packaged. The README documentation and the published CLI binary now match.

3.1 Two-stage product model

OCN now works in two connected phases:

  1. Planning Gatekeeper — define scope, architecture, acceptance, and build plan before implementation. Drives state_discovery → state_plan, gates artifact completion, and is the surface covered by §6.1.
  2. Execution Evidence Navigator — read git, GitHub PR, acceptance evidence, and verification signals during implementation. Six new read-only commands (§6.2) summarise that evidence and draft the next agent prompt, the verification readiness signal, and the final verdict. They never mutate state, never call an LLM, never run npm/gh mutating commands.

The two stages are connected: planning artifacts (00–10) must pass their gates before the execution stage's evidence map, verify status, and verdict draft can read meaningful inputs. Authoritative reframe: DEC-024. Series closure report: docs/reports/2026-05-04-execution-navigator-verdict-draft.md.

Implemented

  • CLI: init, status, brief, doc create, check, gate, advance (full list in §6).
  • State machine: 8 states, forward-only transitions; DISCOVERY → PLAN have stable-ID steps wired (BUILD / VERIFY / SHIP / REFLECT have state IDs only — steps deferred).
  • Step Artifact Gate: required-section detection with NFKC-normalised heading match for bilingual Title｜标题 headings.
  • State safety: .ocoding/.lock (5 s timeout + stale recovery), state.json.bak rolling backup, atomic temp-rename writes; concurrent-advance race fixed (post-Codex P1).
  • Audit: dual-track persistence, 16 event types, correlationId threading across the entire ocn advance event chain.
  • MCP safe tools: 7 read/prepare/create/log tools over stdio; 4 forbidden tools never registered (full list in §7); projectRoot validator + threat model (docs/security/mcp-threat-model.md).
  • Real MCP Host validation: Claude Desktop on Windows with WSL2 validated end-to-end (DEC-017, report). Cursor and Cline remain unverified.
  • Executable example: examples/discovery-to-plan/ walks all 10 v1.0 SOP steps end-to-end via scripts/smoke.sh. Bundled fixtures derived verbatim from src/core/templates/*.ts so they cannot drift.
  • npm publish discipline: published as [email protected] (per DEC-039 the npm package version and the SOP profile version are unified and move in lockstep — both 0.8.0) under strict pre-publish checklists, prepublishOnly gate, and files allowlist. latest and beta both point to 0.9.0-beta.1; alpha is preserved at 0.1.0-alpha.2 for historical use. Annotated git tags and matching GitHub pre-releases published per beta.

Not implemented (deliberately deferred — see §10)

ocn doctor, ocn reset, ocn baseline, SOP versioning / upgrade, production / full tiers, mini-CRM dogfood, real-Host validation for Cursor / Cline (DEC-019), remote MCP transport, MCP auth.

Use OCN

4. Install

4.1 Recommended: install from npm

npm install -g o-coding-navigation

As of v0.9.0-beta.1, npm latest and beta both point to the SOP profile 0.8.0 release (Acceptance Backbone, AM-015/DEC-041) plus optional Auto Mode (AM-009) and optional Contract Backbone (AM-012). Note: per DEC-039 the npm package version and the SOP profile version are now unified and move in lockstep (both 0.8.0); SOP profile 0.8.0 adds the Acceptance Backbone on top of 0.7.0, and older profiles stay frozen + importable.

Verify:

ocn --version       # 0.9.0-beta.1
ocn --help
ocn-mcp             # starts the MCP stdio server; press Ctrl+C to exit

Use @beta when you want to pin the prerelease channel explicitly:

npm install -g o-coding-navigation@beta

To uninstall: npm uninstall -g o-coding-navigation.

4.2 Currently published channels

Channel Version npm tag Notes
latest (recommended) 0.9.0-beta.1 latest SOP profile 0.8.0 mainline (Acceptance Backbone) + optional Auto Mode (AM-009).
Beta (explicit prerelease pin) 0.9.0-beta.1 beta Same artifact as latest; use @beta when you want to pin the prerelease channel explicitly.
Alpha (still available) 0.1.0-alpha.2 alpha Prior pre-GA channel; preserved for historical use only.

Package home: https://www.npmjs.com/package/o-coding-navigation

Prerequisites: Node.js ≥ 20 (see engines in package.json).

Pre-GA caveat: this is a pre-GA beta release — beta only, intended for controlled testing / dogfood, not GA. Validated with Claude Desktop on Windows with WSL2. Cursor and Cline are not yet verified.

4.3 Alternative: local development from source

For contributing or local development:

git clone https://github.com/UncleTIM-GZ/O-CodingNavigation.git
cd O-CodingNavigation
npm install
npm run build
npm link              # exposes `ocn` and `ocn-mcp` on PATH

To uninstall the global links: cd O-CodingNavigation && npm unlink -g ocn ocn-mcp.

If you prefer not to npm link, run the CLI in-place via node /path/to/O-CodingNavigation/dist/cli/index.js ….

5. First 5 minutes

5.1 The minimal happy path

mkdir ocn-demo && cd ocn-demo

ocn init                        # creates .ocoding/ and docs/
ocn status                      # state_discovery / step_project_brief

ocn doc create project-brief    # creates docs/00-project-brief.md (template)
# Now edit docs/00-project-brief.md and fill in the 4 required sections:
#   Problem · Goal · Users · Success Criteria

ocn gate                        # read-only — confirms the gate now passes
ocn advance                     # gate + state mutation + audit trail
ocn status                      # state_discovery / step_scope (advanced)

ocn brief                       # session brief for an AI coding agent

5.2 What each command produces

  • init writes .ocoding/state.json, .ocoding/sop.yaml, the dual-track audit files, and a docs/ skeleton.
  • status reports currentStateId: state_discovery / currentStepId: step_project_brief.
  • doc create writes docs/00-project-brief.md with bilingual section headings already in place.
  • gate returns OK (exit 0) once the 4 required sections are filled in, or ERR_GATE_FAILED (exit 1) with a bilingual list of missing sections.
  • advance writes a correlationId-tagged chain of audit events: advance_started → artifact_gate_run → artifact_gate_passed → state_transitioned → state_write_succeeded → advance_succeeded.
  • brief prints the current-step required sections + AI governance reminders so a coding agent can resume without re-reading docs.

A more detailed walkthrough lives in docs/quickstart.md, including the expected file tree and common errors.

5.3 Try the executable example

For an end-to-end example that walks all 10 v1.0 SOP steps against a hermetic temp project, see examples/discovery-to-plan/ and run:

npm run build
bash examples/discovery-to-plan/scripts/smoke.sh

The smoke prints the final state and exits with Discovery-to-plan smoke completed. It does not modify your environment.

5.4 Full walkthrough: how it looks inside Claude Code

OCN is MCP-first by design. Inside Claude Code (or any other MCP host wired to ocn-mcp — Claude Desktop is the validated path per §7.1), you almost never need to remember an OCN command. You speak natural language; the AI calls the right OCN MCP tool; the flow runs itself. The one and only step that must be typed by you in a terminal is the state-advance step — that is OCN's design guardrail, not a bug.

Prerequisite: ocn-mcp wired into your MCP host as in §7.1. The validated host is Claude Desktop on Windows + WSL2 (DEC-017); other MCP hosts behave the same but are not yet validated (DEC-019).

Below is the full flow in 8 steps. For each step you get three things: what you say → what the AI does behind the scenes → what you see back.

Phase 1: one-time setup (do once per project)

Step 1. Initialise OCN

  • You say: "Initialise OCN in the current directory."
  • AI does: shells out via Bash to run ocn init (init is not an MCP tool — it must be a shell call).
  • You see: .ocoding/ and docs/ created; project initialised at state_discovery / step_project_brief.

Step 2. Confirm location

  • You say: "Which step is OCN on?"
  • AI does: calls MCP tool navigator.where_am_i.
  • You see: current state, current step, artifact path, suggested next action.
Phase 2: the per-step loop (repeat for every step)

Assume you're sitting on step_project_brief. Steps 3–8 close out one step; the next step uses the same loop.

Step 3. Pull this step's brief

  • You say: "Tell me what to do for the current step."
  • AI does: calls MCP tool navigator.brief.
  • You see: required sections, governance reminders, uncertainty policy. These enter the AI's context so it follows the rules when writing the artifact.

Step 4. Create the artifact template

  • You say: "Create the template for the current step's artifact."
  • AI does: calls MCP tool navigator.create_artifact with type project-brief.
  • You see: docs/00-project-brief.md written from the bundled template, with bilingual section headings in place.

Step 5. Fill out the artifact

  • You say: "Fill out the artifact according to the brief. Our project is [your description]."
  • AI does: based on the brief's required sections + your description, edits docs/00-project-brief.md directly via its Edit tool to fill Problem / Goal / Users / Success Criteria.
  • You see: a filled-in markdown. Iterate until you're satisfied.

Step 6. Verify the artifact

  • You say: "Check whether the current artifact passes the gate."
  • AI does: calls MCP tool navigator.run_gate.
  • You see: pass → AI says "gate passed"; fail → AI lists the missing sections and offers to fix them. On fail, return to Step 5.

Step 7. You type ocn advance yourself

⚠️ This is the one step in the entire flow that you must do yourself. The AI cannot do it for you.

After Step 6 passes, a disciplined AI will tell you:

"Gate passed. To advance, please run ocn advance in your terminal."

Why won't the AI just run it? Two reasons, both load-bearing:

  1. MCP does not expose advance. navigator.advance_phase is one of the four tools in §7.3 that is never registered. From the MCP path, the AI literally has no advance tool.
  2. OCN's AI governance rules reserve advance for humans. State advancement = project milestone = the sentence "we are now in the next phase" must be spoken by a human. Letting the AI say it would degrade OCN into one more autonomous scaffolder — exactly the failure mode OCN exists to prevent.

So open a terminal and run:

ocn advance

advance re-runs the gate internally, and on pass writes state.json under a lock with rolling backup and emits the full correlationId-tagged audit chain.

If you don't want to switch to a terminal, you can explicitly ask the AI to run it for you: "Run ocn advance for me." That still counts as your decision — the AI is just pressing the button on your behalf. The crucial difference: the AI cannot auto-run advance the moment Step 6 passes. It must wait until you say so.

Step 8. Loop back to Step 3

  • You say: "Where is OCN now?"
  • AI does: calls navigator.where_am_i.
  • You see: the new step (e.g., step_scope).

Then return to Step 3 for the new step: pull its brief, create its artifact, write, verify, advance. Every step uses this same loop until OCN reports the project has run out of steps.

Cheat sheet: what you say → which tool runs
What you say What the AI uses Mutates state.json?
"Where is OCN?" navigator.where_am_i (MCP, read-only) No
"Pull the current step's brief" navigator.brief (MCP, read-only) No
"Create the artifact template" navigator.create_artifact (MCP) No (writes docs/)
"Fill out the artifact" AI's own Edit tool on docs/ No
"Verify the artifact" navigator.run_gate (MCP, read-only) No
"I'll run ocn advance myself" ocn advance CLI (Bash, you trigger) ✅ Yes
Two hard rules

(1) navigator.run_gate passing ≠ state advancement. run_gate is read-only; a pass result does not move OCN to the next step. Only ocn advance (CLI) advances. If an AI runs run_gate and then claims "we're in the next step", it is lying — state.json is unchanged.

(2) The AI never advances on its own initiative. By design. After Step 6 passes, a disciplined Claude Code will say only "Gate passed; please run ocn advance". If it goes ahead and runs ocn advance without you asking, it has bypassed OCN's discipline guardrail — exactly the drift failure mode OCN exists to prevent.

This is the discipline OCN sells.

5.5 Adopting OCN mid-project

Scenario: your project is already running — half the code is written, the PRD lives on Notion, ACs are scattered across Slack threads, decisions exist only in your memory. You want to retrofit OCN's discipline now.

The fact you have to accept first: OCN does not start from "where you actually are"

OCN's state machine always starts at state_discovery / step_project_brief. The current release does not ship --skip-to, --start-at, or an --override flag on advance.

This is by design, not a limitation. OCN's value is forcing you to explicitly answer "what are we building?" and "what counts as done?". Skipping to BUILD makes brief, gate, and audit groundless — you might as well not install it.

So the only right path for mid-project adoption is backfill. The good news: you already have the content. You're just routing it into OCN's structure.

ocn init and existing files: hard guarantees

The most common worry when adopting OCN mid-project: "will ocn init overwrite my existing code or docs?"

Answer: no. Two source-level guarantees (verified against src/core/init.ts and src/core/artifact/template-writer.ts):

  1. ocn init only writes OCN's own state files under .ocoding/ (state.json, sop.yaml, gates.yaml, artifacts.yaml, config.yaml) and mkdirs docs/. It does not write any docs/XX-*.md artifact templates, and it does not touch src/, README.md, or any other path.
  2. If the project is already initialised (.ocoding/state.json exists), ocn init refuses to run and returns exit 4 (ERR_IO_OR_CONFIG) with the message "OCN is already initialized in this directory. / 当前目录已经初始化过 OCN。" — never overwrites your existing OCN state.

Similarly, ocn doc create <type> refuses to overwrite by default: on a name collision it returns exit 4 with "<type> already exists at <path>. Use --overwrite to replace it." You must explicitly pass --overwrite to replace.

So you can safely run them in an existing project root. Worst case: you already have a docs/00-project-brief.md — ocn doc create project-brief will refuse to write. Two options:

  • Manually merge your existing content into the OCN template (preserve OCN's bilingual section headings Title｜标题 — they are the gate's match anchors).
  • Rename your file as backup (e.g. docs/00-project-brief.legacy.md), let ocn doc create project-brief write the fresh template, then paste your content back into the OCN-shaped sections.
Recommended path: backfill early steps using material you already have

Step 1. Initialise in your project root

  • You say: "Initialise OCN in this directory."
  • AI does: runs ocn init via Bash.
  • You see: .ocoding/ created (docs/ is mkdir-ed); state at state_discovery / step_project_brief. Your existing code and docs are untouched.

Step 2. Feed the AI your project context in one shot

  • You say: "Here's the project context. The PRD is at [Notion URL / file path]. ACs pasted below. Code is in src/; [X module] is done, [Y] still pending. Architecture: [brief]. I want to backfill this into OCN's early steps."
  • AI does: ingests the context for use across all subsequent steps.
  • You see: the AI restates the situation to confirm it has the picture.

Step 3. Backfill the steps one at a time

Each step uses the §5.4 Steps 3–8 loop (pull brief → create template → AI fills using your material → run_gate → you type ocn advance). The difference from a greenfield project: because the content already exists, the AI mostly reorganises and aligns formats rather than inventing.

SOP 0.5.0 ships doc create templates for all 20 steps. The table below maps every step to its artifact path and where to mine the content from your existing project:

DISCOVERY phase

Step Artifact Where to mine it from your existing project
step_project_brief docs/00-project-brief.md Original project pitch, Notion landing page, the one-pager you sent investors / leadership
step_scope docs/01-scope.md Roadmap, Linear / Jira epic list, internal "what v1 / v2 will not do" doc

SPEC phase

Step Artifact Where to mine it from your existing project
step_prd docs/02-prd.md The PRD on Notion / Lark / Confluence — reorganise it into OCN's section structure
step_acceptance_criteria docs/03-acceptance-criteria.md Slack AC threads, "should…" lines in bug tickets, QA test-case descriptions — convert to Given/When/Then

DESIGN phase

Step Artifact Where to mine it from your existing project
step_technical_architecture docs/04-technical-architecture.md package.json / requirements.txt / go.mod, deployment diagrams, tech-selection notes
step_information_architecture docs/05-information-architecture.md Existing sitemap, URL routing tables, menu hierarchy, product screenshots
step_data_model docs/06-data-model.md schema.sql, Prisma schema, ORM models, ER diagrams, DB migration files
step_api_contract docs/07-api-contract.md OpenAPI / Swagger spec, Postman collections, router code, @RestController annotations
step_test_strategy docs/08-test-strategy.md tests/ directory layout, CI config (.github/workflows/), coverage targets, E2E framework choice
step_logic_backbone docs/07-logic-backbone.md Existing scoring/decision/rules logic, formula definitions, signal→action mappings

PLAN phase

Step Artifact Where to mine it from your existing project
step_mvp_plan docs/09-mvp-plan.md Existing roadmap, kanban, Linear/Jira epics, Phase 0/1/2 split
step_build_plan docs/10-build-plan.md Current sprint plan, todo list, near-term PR plan

BUILD phase (typically retroactive backfill)

Step Artifact Where to mine it from your existing project
step_implementation_log docs/11-implementation-log.md git log, merged PR titles and summaries
step_change_evidence docs/12-change-evidence.md Deployed changes, CHANGELOG.md, draft release notes
step_integration_notes docs/13-integration-notes.md Third-party integration notes, SDK call samples, integration test results

VERIFY phase

Step Artifact Where to mine it from your existing project
step_verification_report docs/14-verification-report.md Past QA reports, performance test results, user acceptance records
step_acceptance_mapping docs/15-acceptance-mapping.md AC-to-code/test mapping table (often needs fresh assembly)
step_failure_fix_log docs/16-failure-fix-log.md Bug fix records, incident reports, postmortems
step_regression_evidence docs/17-regression-evidence.md Regression test results, CI green-history records
step_final_build_verdict docs/18-final-build-verdict.md Final release verdict (typically newly written)

Step 4. Once you reach the step you're "actually" on, snap to it

When backfill catches up to your real position (say you've written half the code and you're at step_implementation_log), you don't start from a blank — your already-written code and merged PRs are the source material for that step. From here on, new work goes through OCN's discipline: every new feature gets a step, an artifact, then code.

The pre-OCN "anarchy period" is preserved as a faithful historical snapshot inside the backfilled artifacts.

Three common "can I…" questions

Q: Can I skip the early steps and start at BUILD? A: The current release does not expose a skip mechanism. Two options:

  • Honest backfill (recommended) — fill the early steps from material you already have.
  • Thin placeholder (not recommended) — write minimal content that just satisfies the required sections so the gate passes. This makes the gate rubber-stamp anything, which defeats the point of installing OCN.

Q: Will my existing docs/ numbering collide with OCN's? A: OCN reserves 00-project-brief.md through 18-final-build-verdict.md. On a name collision, ocn doc create refuses to write and returns exit 4 — your existing content is never overwritten. Recommended fix: rename your existing file as backup (e.g. docs/02-prd.legacy.md), run ocn doc create prd, then merge your content into the OCN template. The bilingual section headings (Title｜标题) in the OCN template are the gate's match anchors and must be preserved.

Q: Can I have the AI run all the early steps in one go? A: Technically yes. You can say:

"Backfill step_project_brief through step_technical_architecture from the material I provided. After each gate passes, tell me; I'll type ocn advance myself."

The AI will loop through doc create → fill → run_gate. After each step's run_gate passes, it stops and waits for you to type ocn advance (or to explicitly tell it to). Every advance remains your decision — by design, bulk backfill cannot replace the human "we are now in the next phase" call.

A minimal backfill example

Suppose you have:

  • A 3-month-old React + Node.js project with code in src/
  • A half-page Notion project pitch
  • Scattered AC discussions in Slack

Bringing this into OCN looks like (real conversation will be denser; this shows the skeleton):

You: "Initialise OCN in this directory." AI: runs ocn init. State: state_discovery / step_project_brief. src/ and your existing files are untouched.

You: "Project context: [paste the Notion pitch]. Fill docs/00-project-brief.md per the brief." AI: calls navigator.brief → gets required sections (Problem / Goal / Users / Success Criteria) → edits docs/00-project-brief.md.

You: "Verify." AI: calls navigator.run_gate → pass. AI (disciplined): "Gate passed. Please run ocn advance."

You (terminal): ocn advance.

You: "Scope: v1 has shipped [X/Y/Z], [A] is in progress, explicitly out of scope: [B/C]. Fill docs/01-scope.md." AI: loop…

You: "Reorganise the Notion PRD into docs/02-prd.md: [link]." AI: loop…

You: "ACs from these Slack excerpts: [paste], converted to Given/When/Then." AI: loop…

(DESIGN phase: data-model from prisma/schema.prisma, api-contract from router code, technical-architecture from package.json + deployment diagram.)

(BUILD phase: implementation-log assembled from git log, change-evidence from merged PRs.)

Once the backfill catches up, all new work comes in through OCN's next step — every new feature has a brief, a gate, and an audit trail.

5.6 Post-plan execution flow

After your planning artifacts (00–10) pass their gates and the project enters implementation, the Execution Navigator commands take over. They are read-only summarisers — they help you and the agent see what is actually true in the repo, the PR, the acceptance evidence, and the verification signals.

  1. Build the planning artifacts (steps above) — your state_plan artifacts (00–10) gate the implementation.
  2. Start implementation in your normal repo workflow (your IDE, agents, etc.).
  3. During implementation, use the Execution Navigator commands (§6.2) to read evidence:
    ocn exec status                              # local git + OCN state snapshot
ocn github analyze-pr <n>                    # GitHub PR analysis (when a PR exists)
ocn evidence map --pr <n>                    # acceptance criteria coverage
ocn next-prompt --agent claude               # next agent brief
ocn verify status --mode combined --pr <n>   # verification readiness
ocn verdict draft --mode combined --pr <n>   # evidence-derived verdict
  4. Review the evidence and verdict draft before merge — the verdict is conservative; ready-to-merge requires every signal aligned (verification ready, PR clean, acceptance covered, no failed checks, no changes-requested).

6. Core CLI commands

All commands accept --json to emit a machine-readable CommandResult envelope. Exit codes are stable:

Exit code Meaning
0 OK
1 gate failed
2 artifact missing or invalid
3 state machine error
4 config / lock / IO error
5 SOP version incompatibility

6.1 Planning Gatekeeper commands

Command Purpose Reads / writes Audit emission
ocn init [--tier minimal] [--json] Initialise an OCN project in the current directory. Writes .ocoding/, docs/, the dual-track audit files. project_initialized + state-write events
ocn status [--json] Show current state, current step, the relative path of the current step's artifact, and the next-action hint. Read-only. None (avoids log spam — pull-mode)
ocn brief [--json] Print the current-step brief for an AI coding session: required sections, governance reminders, uncertainty policy. Read-only. None (pull-mode)
ocn doc create <type> [--overwrite] [--json] Create one of the 20 supported artifacts from its bundled template (refuses to overwrite an existing file unless --overwrite is passed). Writes the artifact under docs/. artifact_created
ocn check [--json] Check the current step's artifact against its required sections. Read-only. artifact_gate_run + artifact_gate_passed / artifact_gate_blocked
ocn gate [--json] Read-only artifact gate aggregation for the current step. Same emission as check; never mutates state. Read-only. artifact_gate_* (no correlationId)
ocn advance [--json] Run gate, then advance to the next step on pass. Lock-protected; never partial. Writes state.json (atomic). Full advance chain with shared correlationId
`ocn auto on --phase <1 2 all>/off [--phase]/resume/status/trace [--json]` Auto Mode switch (AM-009 / DEC-034) — delegate gate-green advances to the AI agent per phase (phase 1 = DISCOVERY→PLAN, phase 2 = BUILD→VERIFY incl. task check + the milestone rewind). The switch itself is human-only and refuses OCN_ACTOR=ai_agent; manual mode stays the default. status/trace are pull-mode (no audit); trace replays every ai_agent decision (rationale + engine-recorded context).
ocn sop upgrade [--target <version>] [--plan] [--json] Move the project's pinned SOP profile forward to a newer bundled version (forward-only; config.yaml preserved). --plan is a read-only dry-run. Apply rewrites .ocoding/ snapshots + state.json (atomic, lock-protected); --plan writes nothing. sop_upgraded (apply) / sop_version_diff_detected (plan)
ocn rewind --to <stepId> --reason <text> [--json] Move the cursor back to a strictly-earlier step in this round (DEC-033 / AM-008). --reason is mandatory; docs/ artifacts are untouched; no gate exemption — every advance after a rewind re-runs the full gate stack. Milestone loop (phased projects, e.g. P0→P4 with per-phase go/no-go): after each milestone's verdict, rewind to step_build_plan, APPEND the next phase's Task Specs (keep finished specs verbatim — verify-command text unchanged ⇒ hash unchanged ⇒ done preserved), re-gate — the ledger becomes the project's cumulative progress account. Human-only — never exposed over MCP. Writes state.json (atomic, lock-protected; latestGateResult cleared). cursor_rewind (success and failed)
ocn cycle new --yes [--json] Archive this round's runtime state to .ocoding/cycles/<n>-<ts>/ and restart at the first step (DEC-033 / AM-008). docs/ are kept so next-round gates fast-forward; the pin and user-owned config.yaml stay; the audit JSONL is never archived — one continuous log spans all rounds. --yes is mandatory. Use it when the round's completion boundary (01-scope) is reached — per-milestone iteration inside one round is ocn rewind's job. Human-only — never exposed over MCP. Moves .ocoding runtime files into the archive; rewrites snapshots + state.json. cycle_started (success and failed)
ocn stop --yes [--reason <text>] [--json] Terminate OCN for this project (DEC-042 / AM-016) — one-way. Marks state.json with a stoppedAt timestamp so every runtime surface goes quiet (brief / next-prompt / status stop driving; advance refuses; the Stop hook stops forcing the AI onward) AND uninstalls the injected Claude Code wiring (.claude/ocn.md, /ocn-next, OCN's hooks + OCN_ACTOR in .claude/settings.json, the @.claude/ocn.md import in CLAUDE.md) — user content is preserved. Use it once planning is done to move to free self-development (works from any state). No built-in re-attach — re-wire with ocn agent setup. --yes mandatory. Human-only — never exposed over MCP. Writes state.json (atomic, lock-protected); deletes/strips the wiring files. project_stopped (success and failed)
ocn readiness list [--json] Evaluate all readiness checks for the current state/tier and print the verdict table (PASS/WAIVED/FAIL/UNKNOWN/NA counts + blocking items). Read-only. None (pull-mode — avoids audit spam)
ocn readiness waive <checkId> --reason <r> --probe <cmd> [--json] Grant a conditional waiver ("waive-with-probe"): the probe must pass at grant time, is re-verified on every gate run, and the waiver expires when the project leaves the current state. Human-only — never exposed over MCP. Writes the waiver into .ocoding/readiness.json. readiness_waived
ocn task list [--json] Show the build-plan task ledger: per-task status (pending/done), the frozen verify command, and dependency/phase info. Read-only. None (pull-mode)
ocn task check [<id>] [--json] Run the task's frozen verify command — exit 0 marks the task done; a verify command that drifted from its frozen hash is refused. Human/agent CLI only — never exposed over MCP. Writes .ocoding/task-ledger.json. task_completed
ocn agent setup [--force] [--json] One-command Claude Code wiring (AM-006): Stop/PostToolUse hooks in .claude/settings.json (merge-only, command -v ocn guarded), .claude/ocn.md governance contract, CLAUDE.md import (append-once), /ocn-next slash command. Idempotent; human-only. Writes/merges the four .claude surfaces; never overwrites user entries. agent_setup_completed
ocn hook stop / ocn hook post-edit Machine-facing Claude Code hook handlers (called by the agent host, not humans): stop re-runs the gate when the agent ends a turn and blocks with fix hints; post-edit runs commands.lint/commands.typecheck for fast feedback. Fail-open by design. Raw hook-contract IO (stdin JSON in; block JSON / exit 2 + stderr out). gate-run events via the reused check engine

<type> for doc create (SOP 0.5.0 ships templates for all 20 steps): project-brief, scope, prd, acceptance-criteria, technical-architecture, information-architecture, data-model, api-contract, test-strategy, logic-backbone, mvp-plan, build-plan, implementation-log, change-evidence, integration-notes, verification-report, acceptance-mapping, failure-fix-log, regression-evidence, final-build-verdict.

--tier for init accepts minimal, production, full — only minimal is enforced today (production / full are accepted but their artifact sets are not yet differentiated).

6.2 Execution Navigator commands

Beyond the planning gate, OCN now reads execution-time evidence (git, GitHub PR, acceptance criteria, package scripts) and produces deterministic summaries — no LLM calls, no mutation, no npm run from inside the implementation. These commands are useful during implementation and review, after state_plan is closed. Authoritative reframe: DEC-024. Series closure: docs/reports/2026-05-04-execution-navigator-verdict-draft.md.

Command Purpose Typical use
ocn exec status [--json] Read local git evidence and OCN project state. Snapshot of branch, head, dirty/clean, changed files, recent commits, current state/step.
ocn github analyze-pr <n> [--json] Read-only GitHub PR evidence analysis via gh CLI. PR metadata, files changed, commits, checks, reviews — no mutation.
ocn evidence map [--json] [--pr <n>] Map acceptance criteria to available evidence. Per-criterion status: evidence-found / evidence-candidate / missing-evidence / needs-human-review.
ocn next-prompt [--json] [--agent ...] [--mode ...] [--pr <n>] [--issue ...] Generate the next deterministic agent task brief. 9 fixed sections — objective, evidence, allowed work, forbidden actions, verification commands, stop conditions.
ocn verify status [--json] [--mode ...] [--pr <n>] Summarise verification readiness and missing signals. Status ready / partial / blocked / pending / no-verification-data plus required commands and risk flags.
ocn verdict draft [--json] [--mode ...] [--pr <n>] Draft an evidence-derived final judgment. Category continue-work / request-changes / ready-for-review / ready-to-merge / hold-for-manual-review, with sorted supports / blocks / warnings.

All six commands are read-only by design: no writes to .ocoding/, docs/, or any project file; no git mutation; no gh mutation (only pr view / auth status); no LLM call. --mode accepts local | pr | combined where applicable; --pr <n> enables GitHub evidence in pr and combined modes. Per-MVP implementation reports live under docs/reports/2026-05-02-execution-navigator-*.md and the closure report linked above.

6.3 Claude Code integration — one command (AM-006 / DEC-031)

ocn agent setup      # idempotent; commit .claude/ + CLAUDE.md to share with the team

Wires the whole phase-2 runbook mechanically: a Stop hook re-runs ocn check whenever the agent tries to end its turn (blocked → fix hints are fed back and the agent keeps working; loop-protected, fail-open), a PostToolUse hook runs your configured commands.lint/commands.typecheck (from .ocoding/config.yaml) after every Edit/Write, the .claude/ocn.md governance contract loads into every session via a CLAUDE.md import, and /ocn-next pulls ocn brief + ocn next-prompt and starts the task. Hook commands are guarded with command -v ocn, so teammates without ocn installed are unaffected. After setup, the human workflow per task is two actions: /ocn-next in Claude Code → review + ocn advance in the terminal.

6.4 Auto Mode — optional, off by default (AM-009 / DEC-034)

ocn auto on --phase 1     # delegate the planning pipeline (DISCOVERY→PLAN)
ocn auto on --phase all   # full auto: planning + BUILD/VERIFY loop + milestone rewinds
ocn auto status           # current grant + circuit-breaker state
ocn auto trace            # replay every AI decision: rationale + machine context
ocn auto off              # back to fully manual (the default)

Auto Mode delegates the trigger, never the judgement: every advance still runs the full gate stack, task completion is still decided only by the frozen verify command's exit 0. The AI must sign each call (OCN_ACTOR=ai_agent, injected by ocn agent setup) and justify it (--rationale — background / evidence / action); the engine independently records machine context on the audit trail. Safety net: N consecutive gate failures on one step (default 5) trip a circuit breaker that suspends automation until a human runs ocn auto resume — humans are never affected. Hard human-only zones in every mode: readiness waive, cycle new, sop upgrade, overrides, any non-milestone rewind, ocn stop, and ocn auto itself. The MCP surface is unchanged (7 tools, test-pinned). With phase 2 on, multi-P build plans complete end-to-end: after each milestone's verdict the agent rewinds to step_build_plan, appends the next P's Task Specs, re-gates (done tasks carry over), and continues — stopping only when every milestone is done.

7. MCP tools

OCN's MCP server (ocn-mcp) exposes 7 tools over stdio transport. Wire it into Claude Desktop on Windows with WSL2 — that is the validated Host (per DEC-017 and docs/reports/2026-04-30-mcp-external-host-validation-report.md). Cursor and Cline are MCP-aware but are not yet verified for OCN — see DEC-019 for the support boundary; do not treat them as supported until each has its own validation report.

7.1 Wire into Claude Desktop on Windows + WSL2

Add the OCN entry to %APPDATA%\Claude\claude_desktop_config.json:

{
  "mcpServers": {
    "ocn": {
      "command": "wsl.exe",
      "args": ["-e", "ocn-mcp"]
    }
  }
}

If ocn-mcp is not on the WSL2 PATH, replace "ocn-mcp" with the absolute path printed by which ocn-mcp inside WSL2 (typically the npm global bin, e.g. /home/<user>/.npm-global/bin/ocn-mcp). Save the config, fully quit Claude Desktop (system tray included), and reopen — the seven navigator.* tools should appear in the tools panel.

Full wiring guidance + per-tool envelopes: docs/mcp-usage.md. The validation transcript that proved this exact path works lives in docs/reports/2026-04-30-mcp-external-host-validation-report.md.

7.2 Allowed (7)

Tool Purpose Mutates?
navigator.where_am_i State snapshot. No
navigator.brief Current-step brief. No
navigator.run_gate Read-only gate aggregation. No
navigator.create_artifact Create from the full template registry (20 types: 00–19, logic-backbone at 07). Filesystem only
navigator.capture_log Append to docs/19-dev-log.md (type=dev) or docs/18-research-log.md (type=research). type=decision is hard-rejected. Filesystem only
navigator.detect_sop_version Drift between locked profile and bundled OCN SOP. No
navigator.generate_next_prompt Required sections + governance reminder + uncertainty policy + self-check rule. No

7.3 Forbidden (4) — NEVER exposed

Tool Why kept off MCP
navigator.advance_phase State advancement is human-only via the CLI.
navigator.capture_decision Decisions reflect human intent. The exposed capture_log rejects type=decision.
navigator.reset_project Destructive; twice-confirm flow is human-only.
navigator.force_release_lock Bypasses state-safety invariants; operator-only.

Enforced by tests/unit/mcp-tool-registry.test.ts (ALLOWED ∩ FORBIDDEN = ∅).

An MCP agent connected to OCN can read project state, render the next-step brief, prepare artifacts, run the read-only gate, create from the template registry, and capture dev / research logs. It cannot advance state, capture decisions, reset the project, or force-release the lock.

Full surface + wiring instructions: docs/mcp-usage.md.

Reference

8. Documentation map

OCN ships its own design baseline under docs/. Two governance points worth knowing before reading:

  • Canonical decision log: docs/20-decision-log.md. Some historical references say docs/19-decision-log.md — those refer to the same file before the path move recorded in AM-002.
  • Amendment index: docs/amendments/README.md. Active divergences from the frozen docs/00-08 design baseline are recorded as amendments rather than inline edits (DEC-004).
  • Frozen design baseline: docs/00-project-brief.md through docs/08-mvp-plan.md are Phase-2 design contracts, treated as historical artifacts. New projects initialised via ocn init get the SOP v1.1 step layout from the bundled default profile; the OCN repository itself runs against a project-level override per DEC-003.
  • Reports: docs/reports/ — Phase 2 closure, post-alpha P1 fix train (4 reports), Claude Desktop MCP Host validation, alpha.0 / alpha.1 / alpha.2 / beta.0 publish reports, examples F2/F3, beta release marker, bilingual install flow, doc audits, final Codex full-repo audit, post-Codex P1/P2 fix report. Phase 2 baseline lives in docs/reports/2026-04-28-phase2-completion-report.md.
  • Plans: docs/plans/ holds the planning artifacts for each PR. The active GA Prep plan is docs/plans/2026-04-28-ga-prep-gap-review-plan.md.
  • MCP usage: docs/mcp-usage.md.

9. Development

npm install
npm run lint           # ESLint (TypeScript-eslint)
npm run typecheck      # tsc --noEmit
npm run test           # vitest run — 459 tests, ~24 s
npm run test:coverage  # adds coverage report
npm run build          # tsc + chmod +x on bin entries

The pre-commit hook (Husky 9) runs lint + typecheck + test on every commit. CI runs the same checks plus build and reports coverage. Hard limits per CLAUDE.md: file ≤ 300 lines, function ≤ 50 lines, params ≤ 4, nesting ≤ 3, no raw any in exported APIs.

10. Roadmap

The GA Prep phase was a documentation, packaging, and operational-readiness audit that ran from Phase 2 closure through the beta candidate preparation track. Most GA Prep PRs are now complete; their evidence lives in docs/reports/ and docs/20-decision-log.md. Original plan: docs/plans/2026-04-28-ga-prep-gap-review-plan.md.

GA Prep PRs (status)

Execution Navigator MVP series (post-DEC-024)

  • MVP 1–6 complete — the six Execution Navigator commands (§6.2) shipped in PRs #63–#68 and merged to main. Cross-cutting review fixes landed in PR #69 (report). Series closure: docs/reports/2026-05-04-execution-navigator-verdict-draft.md.
  • Current external package: 0.9.0-beta.1 — Planning Gatekeeper (§6.1) + Execution Evidence Navigator (§6.2) + the Logic Backbone DESIGN gate + the Readiness Backbone cross-cutting gate + the Task Backbone BUILD task ledger + the Acceptance Backbone SPEC gate + optional Auto Mode (§6.4, AM-009) + optional Contract Backbone (AM-012) — a BUILD/VERIFY gate that cross-checks frontend call sites against the declared API contract.
  • Acceptance Backbone shipped in 0.9.0-beta.1 — SOP 0.8.0 (AM-015 / DEC-041): docs/03-acceptance-criteria.md now carries a machine-parseable ## Acceptance Specs｜验收规格 section (### AC-<DOMAIN>-<n> blocks with desc required + optional priority/given/when/then/trace). The acceptance gate runs at step_acceptance_criteria after the section gate, hard-blocks four structural defects (no_specs, duplicate_id, invalid_id, missing_field → ERR_ARTIFACT_INVALID, exit 2), and on pass freezes .ocoding/acceptance-specs.json — the canonical machine source of AC ids that build-plan traces bind to. Closes the false-completion hole where AC written in a table silently didn't register; reads are staleness-guarded (frozen projection while docs/03 is unchanged, else the live parse). Runtime default cut over to 0.8.0 per DEC-039 (npm/SOP lockstep → npm 0.9.0-beta.1); 0.7.0 and earlier stay frozen + importable; ocn sop upgrade migrates. See docs/acceptance-backbone-proposal.md.
  • Task Backbone shipped in 0.5.0-beta.0 — SOP 0.5.0 (AM-007 / DEC-032): build plans carry machine-parseable Task Spec blocks (## Task Specs｜任务规格: goal/traces/touches/verify/dod + optional depends/phase/timeout). The build-plan gate validates six hard defects (duplicate/invalid id, missing field, dangling traces → AC ids, dangling touches → logic-graph nodes, dangling/cyclic depends, zero tasks) and freezes each task's verify-command hash into .ocoding/task-ledger.json. Completion is decided only by ocn task check rerunning the frozen command (exit 0 → done + task_completed audit; drift → refused); /ocn-next dispatches the first pending task in state_build; ocn advance out of BUILD is blocked while tasks are pending. Closes the fourth false-completion class — receipt-only completion (honest-but-empty BUILD receipts that pass every gate; discovered via the Lattice dogfood). Runtime default cut over to 0.5.0 per DEC-032; ocn sop upgrade migrates existing projects.
  • Claude Code integration shipped in 0.4.0-beta.2 — AM-006 / DEC-031: ocn agent setup wires hooks + governance contract + /ocn-next in one idempotent command; ocn hook stop|post-edit carry the enforcement logic inside OCN (fail-open, loop-protected).
  • Readiness Backbone shipped in 0.4.0-beta.1 — SOP 0.4.0 (AM-004 / DEC-028) adds a role-based cross-cutting readiness gate: 55 falsifiable checks derived from 54 curated IT roles run inside ocn check / ocn gate / ocn advance after the section + logic gates. Open-world semantics (FAIL and UNKNOWN block); verdict ledger persisted to .ocoding/readiness.json; ocn readiness list / ocn readiness waive (waive-with-probe) commands; ocn sop upgrade migrates existing projects; runtime default cut over to 0.4.0 per DEC-030.
  • Logic Backbone shipped in 0.3.0-beta.1 — SOP 0.3.0 adds the DESIGN-phase docs/07-logic-backbone.md artifact; ocn check machine-validates its computation/decision graph (blocks on missing role, duplicate node id, dangling reference, dependency cycle, orphan node, unbound trigger) and on pass writes .ocoding/logic-graph.json, which ocn brief summarises as execution order + trigger bindings.
  • Still beta, not GA. Validated with Claude Desktop on Windows with WSL2; Cursor and Cline remain unverified.
  • Next gate: real external-repo dogfood pass before any further release sync; GA promotion remains gated on its own DEC.

Beyond GA Prep — not in any current plan; each requires its own DEC entry first

  • ocn doctor, ocn reset, ocn baseline
  • SOP versioning / upgrade tooling
  • Production / full tier artifact-set enforcement
  • Mini-CRM dogfood (Tier 2 GA success criterion)
  • HTTP / SSE MCP transport, MCP auth, MCP session management
  • Cursor / Cline real-Host validation (each Host needs its own DEC-017-style report)
  • GA promotion DEC (final gate tying together Host scope, multi-OS / Node 24+ CI, dogfood evidence)

11. License

Apache-2.0

Part 2 · 中文版

OCN（O'CodingNavigator）是一套本地优先、MCP 优先、状态机驱动的 AI 编程工作流操作系统。它把 AI 编程从"持续聊天式发挥"重塑成可导航、有门禁、可审计、可复盘的系统工程过程。它是一个 navigator，不是 IDE，不是 SaaS，不是项目管理工具。

CLI：ocn；MCP server：ocn-mcp；许可：Apache-2.0。

中文目录

理解 OCN

使用 OCN

参考资料

理解 OCN

A. OCN 是什么

OCN 是 AI 编程的本地纪律层。所有数据都跑在你自己的机器上，落到纯文本（Markdown + JSON + JSONL + YAML）。它对人提供一个小型 CLI，对 agent 提供一个小型 MCP server，没有任何云端组件。

OCN 把"纪律"产品化：

  • 状态机：DISCOVERY → SPEC → DESIGN → PLAN → BUILD → VERIFY → SHIP → REFLECT，单向推进，不可回退；
  • 步骤产物门禁（Step Artifact Gate）：当前 step 的必填章节缺失时，状态推进被阻断；
  • 双轨审计链：.ocoding/audit/audit-events.jsonl（机器源）+ docs/22-audit-trail.md（人类可读）；
  • 逻辑主干（Logic Backbone，SOP 0.3.0）：一个 DESIGN 阶段产物，其计算/决策图由机器校验；ocn check 会在出现孤儿分值、悬空引用、循环依赖、未绑定触发器时阻断，让系统逻辑在 BUILD 之前就连通；
  • 就绪主干（Readiness Backbone，SOP 0.4.0）：基于角色的横切就绪门禁——从 54 个精选 IT 角色（开发、QA、DevOps、CISO、服务台……）提炼出 55 条可证伪检查，在每个 step 的 ocn check / ocn gate / ocn advance 中运行；采用开放世界语义，FAIL 和 UNKNOWN 都会阻断（沉默不算通过），在 SHIP 之前拦住"角色盲区式完成"；
  • 任务主干（Task Backbone，SOP 0.5.0 新增）：build plan 携带机器可解析的任务规格块（goal / traces / touches / verify / DoD）；build-plan 门禁校验六类硬缺陷，并把每个任务的验收命令哈希冻结进 .ocoding/task-ledger.json；任务只有在 ocn task check 重跑冻结命令且退出码为 0 时才算完成，台账未清时 ocn advance 不准离开 BUILD——封堵第四类假完成："只有回执的完成"（receipt-only completion）；
  • 验收主干（Acceptance Backbone，SOP 0.8.0 新增）：docs/03-acceptance-criteria.md 携带机器可解析的 ## Acceptance Specs｜验收规格 章节（### AC-<DOMAIN>-<n> 块）；验收门禁在 step_acceptance_criteria 运行，硬拦四类结构缺陷（no_specs、duplicate_id、invalid_id、missing_field），通过后冻结 .ocoding/acceptance-specs.json——作为 build-plan traces 绑定的权威 AC id 机器源，封堵"AC 写在表格里却悄悄没登记"的假完成漏洞；
  • 安全的 MCP 工具面：agent 可读、可准备、可创建产物，但不能推进状态、记录决策、重置项目、强制释放锁。

OCN 不是：代码生成器、IDE、SaaS、项目管理工具、笔记应用，也不是只会铺脚手架的"文档工厂"。

B. OCN 解决什么问题

任何稍微复杂的任务，AI 编程 agent 都倾向于踩四类坑：

失效模式 表现
迷路（lost） 没人——人或 AI——清楚项目当前在哪一步。
失控（drift） Agent 一直在产出，但每一段都离最初需求更远。
失忆（amnesia） 一开新对话，agent 又把上一个对话已经决定的东西重新推一遍。
假完成（false-completion） 文档存在，但缺关键章节；agent 还是宣布"完成"。

OCN 把这四个问题视为同一个问题：AI 编程闭环缺乏严肃的"我们在哪"和"什么算完成"。OCN 用代码而不是嘴上嘱咐去回答这两个问题。

C. 当前状态（SOP 0.8.0（验收主干）主干）

项目 状态
阶段 SOP 0.8.0（验收主干）主干——Plan → Build → Verify，含机器校验的 DESIGN 逻辑主干门禁、覆盖每个 step 的基于角色的横切就绪门禁（55 条检查）、把守 BUILD 出口的冻结验收任务台账，以及在 SPEC 阶段冻结权威 AC id 的验收门禁；plan-to-verify smoke 覆盖全部 20 步
测试 完整 vitest 套件在 Node 20 + Node 22 全部通过
覆盖率 满足发布门
npm latest → 0.9.0-beta.1；beta → 0.9.0-beta.1；alpha → 0.1.0-alpha.2（历史保留）
成熟度 pre-GA beta——非稳定、非 GA、仅 beta（用于受控测试 / dogfood）
已验证 Host 已在 Claude Desktop on Windows + WSL2 验证。Cursor 与 Cline 暂未验证。
MCP 传输 仅 stdio（HTTP/SSE 尚未启动）

2026-05-04 republish 说明：以 0.2.0-beta.2 重新发布，目的是把 0.2.0-beta.1 打包之后才合入的 Execution Navigator 命令（exec status、github analyze-pr、evidence map、next-prompt、verify status、verdict draft）真正包进 npm artifact。README 文档与已发布的 CLI 现在保持一致。

C.1 两阶段产品模型

OCN 现在分为两个衔接阶段：

  1. Planning Gatekeeper：在实现前锁定范围、架构、验收与 build plan。驱动 state_discovery → state_plan，把守 artifact 完整度，对应 §F.1 的命令面。
  2. Execution Evidence Navigator：在实现过程中读取 git、GitHub PR、验收证据与验证信号。新加 6 条只读命令（§F.2）汇总这些证据，并产出下一条 agent prompt、验证就绪信号与最终 verdict 草稿。这些命令永不写状态、永不调 LLM、永不跑 mutating 的 npm / gh 命令。

两个阶段是衔接的：planning 阶段的 artifact（00–10）必须先过门禁，execution 阶段的 evidence map、verify status、verdict draft 才能读到有意义的输入。权威 reframe：DEC-024。系列收口报告：docs/reports/2026-05-04-execution-navigator-verdict-draft.md。

已实现

  • CLI：init、status、brief、doc create、check、gate、advance（完整列表见 §F）。
  • 状态机：8 个状态，单向推进；DISCOVERY → PLAN 已挂上稳定 ID 的 step，BUILD/VERIFY/SHIP/REFLECT 仅有 state ID。
  • 步骤产物门禁：基于 NFKC 标准化的双语 Title｜标题 标题匹配。
  • 状态安全：.ocoding/.lock（5 秒超时 + 陈旧锁回收）、state.json.bak 滚动备份、临时文件 + atomic rename 写入；并发 advance 竞态已在 post-Codex P1 修复中关闭。
  • 审计：双轨持久化、16 种事件类型、correlationId 串联整条 ocn advance 事件链。
  • MCP 安全工具：stdio 上 7 个只读/准备/创建/日志类工具，4 个禁用工具不会被注册（详见 §G）；projectRoot 校验器 + 威胁模型（docs/security/mcp-threat-model.md）。
  • 真实 Host 验证：Claude Desktop on Windows + WSL2 已完成端到端验证（DEC-017、报告）。
  • 可执行示例：examples/discovery-to-plan/，scripts/smoke.sh 跑完 v1.0 SOP 全部 10 个 step；fixture 直接来源于 src/core/templates/*.ts，避免漂移。
  • npm 发布纪律：以 [email protected] 形式发布（按 DEC-039，npm 包版本与 SOP profile 版本统一、步调一致——均为 0.8.0），遵循严格的预发布清单与 prepublishOnly 门，files allowlist 收敛到 dist/ + LICENSE + README + docs/quickstart.md + docs/mcp-usage.md；latest 与 beta 都指向 0.9.0-beta.1，alpha 仍保留在 0.1.0-alpha.2；v0.9.0-beta.1 为带注释的 git tag + GitHub pre-release。

尚未实现（刻意延后，详见 §J）

ocn doctor、ocn reset、ocn baseline、SOP 版本/升级工具、production/full tier、mini-CRM dogfood、Cursor/Cline 真实 Host 验证（DEC-019）、远程 MCP 传输、MCP 鉴权。

使用 OCN

D. 安装

D.1 推荐：从 npm 安装

npm install -g o-coding-navigation

从 v0.9.0-beta.1 开始，npm latest 与 beta 均指向 SOP 0.8.0（验收主干）版本 + 可选自动模式（AM-009）。注：按 DEC-039，npm 包版本与 SOP profile 版本统一、步调一致——均为 0.8.0；SOP profile 0.8.0 在 0.7.0 基础上新增验收主干，旧 profile 保持冻结且可导入。

安装后验证：

ocn --version       # 0.9.0-beta.1
ocn --help
ocn-mcp             # 启动 MCP stdio server；按 Ctrl+C 退出

如果希望明确固定在 beta 预发布通道，可以使用 @beta：

npm install -g o-coding-navigation@beta

卸载：npm uninstall -g o-coding-navigation。

D.2 当前已发布的渠道

渠道 版本 npm tag 说明
latest（推荐） 0.9.0-beta.1 latest SOP 0.8.0（验收主干）主干 + 可选自动模式（AM-009）。
Beta（显式预发布通道） 0.9.0-beta.1 beta 与 latest 同一份产物；想明确固定在预发布通道时使用 @beta。
Alpha（仍可用） 0.1.0-alpha.2 alpha 之前的 pre-GA 通道；仅作历史保留。

包主页：https://www.npmjs.com/package/o-coding-navigation

前置依赖：Node.js ≥ 20（参见 package.json 的 engines）。

pre-GA 警告：当前是 pre-GA beta——仅 beta，用于受控测试 / dogfood，非 GA。已在 Claude Desktop on Windows + WSL2 验证。Cursor 与 Cline 暂未验证。

D.3 备选：从源码本地开发

git clone https://github.com/UncleTIM-GZ/O-CodingNavigation.git
cd O-CodingNavigation
npm install
npm run build
npm link              # 把 `ocn` 和 `ocn-mcp` 暴露到 PATH

卸载本地链接：cd O-CodingNavigation && npm unlink -g ocn ocn-mcp。

不想用 npm link 也可以直接 node /path/to/O-CodingNavigation/dist/cli/index.js …。

E. 5 分钟上手

E.1 最小可行流程

mkdir ocn-demo && cd ocn-demo

ocn init                        # 创建 .ocoding/ 与 docs/
ocn status                      # state_discovery / step_project_brief

ocn doc create project-brief    # 写出 docs/00-project-brief.md（模板）
# 然后编辑该文件，把 4 个必填章节填好：
#   Problem · Goal · Users · Success Criteria

ocn gate                        # 只读校验，确认门禁通过
ocn advance                     # 跑门禁 + 改状态 + 写审计链
ocn status                      # 已推进到 state_discovery / step_scope

ocn brief                       # 给接管的 AI agent 输出当前 step brief

E.2 每条命令的预期产出

  • init 写入 .ocoding/state.json、.ocoding/sop.yaml、双轨审计文件，以及 docs/ 骨架。
  • status 报告 currentStateId: state_discovery / currentStepId: step_project_brief。
  • doc create 写出 docs/00-project-brief.md，包含双语章节标题。
  • gate 在 4 个必填章节齐备后返回 OK（exit 0），否则返回 ERR_GATE_FAILED（exit 1）并列出缺的章节。
  • advance 会写一整条带 correlationId 的审计事件链：advance_started → artifact_gate_run → artifact_gate_passed → state_transitioned → state_write_succeeded → advance_succeeded。
  • brief 输出当前 step 的必填章节 + AI 治理提醒，让接管的 agent 不必重读全部文档。

更详细的 walkthrough 见 docs/quickstart.md，含期望文件树和常见报错。

E.3 跑可执行示例

仓库自带一个端到端可执行示例 examples/discovery-to-plan/：

npm run build
bash examples/discovery-to-plan/scripts/smoke.sh

该 smoke 在临时目录里走完 v1.0 SOP 全部 10 个 step，最后打印 Discovery-to-plan smoke completed.，不会污染你的环境。

E.4 完整操作流程：在 Claude Code 里走完一遍

OCN 是 MCP-first 设计。在 Claude Code（或其它接好 ocn-mcp 的 MCP host，如 Claude Desktop）里，你几乎不需要记命令——你说自然语言，AI 调对应的 OCN MCP 工具，流程自动跑通。整个流程里唯一必须你亲自敲终端的是状态推进那一步——这是 OCN 设计上的护栏，不是 bug。

前提：已按 §G.1 把 ocn-mcp 接入。当前经过端到端验证的 host 路径是 Claude Desktop on Windows + WSL2（DEC-017）；其它 MCP host 流程相同但暂未独立验证（DEC-019）。

下面把整个流程拆成 8 步。每一步给你三件事：你说什么 → AI 背后做什么 → 你看到什么。

阶段一：一次性初始化（每个项目做一次）

第 1 步：初始化 OCN

  • 你说：「帮我在当前目录初始化 OCN 项目。」
  • AI 做：用 Bash 工具跑 ocn init（init 不是 MCP 工具，必须 shell out）。
  • 你看到：.ocoding/ 与 docs/ 已创建；项目初始化在 state_discovery / step_project_brief。

第 2 步：确认位置

  • 你说：「OCN 现在在哪一步？」
  • AI 做：调 MCP 工具 navigator.where_am_i。
  • 你看到：当前 state、当前 step、产物路径、下一步建议动作。
阶段二：单 step 循环（每个 step 重复一遍）

假设当前停在 step_project_brief。第 3–8 步走完才进入下一个 step；下一 step 也是同一套循环。

第 3 步：拿这一步的 brief

  • 你说：「告诉我现在这一步该做什么。」
  • AI 做：调 MCP 工具 navigator.brief。
  • 你看到：当前 step 的必填章节、AI 治理提醒、不确定性策略——这些会进入 AI 的上下文，后面它写产物时会自觉遵守。

第 4 步：建产物模板

  • 你说：「帮我建当前 step 的产物模板。」
  • AI 做：调 MCP 工具 navigator.create_artifact，type=project-brief。
  • 你看到：docs/00-project-brief.md 已写出，含双语章节占位。

第 5 步：把产物写完

  • 你说：「按 brief 要求，把这一步的产物写好。我们项目是 [简单介绍]。」
  • AI 做：根据 brief 的必填章节 + 你的描述，直接用 Edit 工具改 docs/00-project-brief.md，把 Problem / Goal / Users / Success Criteria 四个章节填满。
  • 你看到：填好的 markdown。让它继续改，直到你满意。

第 6 步：验产物合不合格

  • 你说：「验证一下这一步的产物。」
  • AI 做：调 MCP 工具 navigator.run_gate。
  • 你看到：通过 → AI 告诉你「门禁通过」；不通过 → AI 列出缺哪些章节，问你要不要它补。不通过就回第 5 步，直到通过。

第 7 步：你亲自敲 ocn advance

⚠️ 这是整套流程里 唯一必须你自己做 的事，AI 帮不了你。

第 6 步通过后，一个守纪律的 AI 会告诉你：

「门禁通过。要推进，请你在终端运行：ocn advance」

为什么 AI 不直接帮你跑？两条原因，缺一不可：

  1. MCP 不暴露 advance。 navigator.advance_phase 是 §G.3 列出的 4 个永不暴露的工具之一。从 MCP 这条路，AI 没有这个工具。
  2. OCN 的 AI 治理规则把 advance 锁成人类专属。 状态推进 = 项目里程碑 = 「我们正式进入下一阶段」——这种话只能由人来说。让 AI 自己说，OCN 就退化成另一个会幻觉的脚手架。

所以你打开终端，打：

ocn advance

advance 内部会再跑一次门禁、通过则带锁 + 备份 + 原子写更新 state.json、写入完整带 correlationId 的审计事件链。

如果你不想切到终端，显式让 AI 替你按按钮也可以： 「替我跑 ocn advance。」 这仍然是 你的决策，AI 只是替你执行 Bash。关键差异：AI 不能在第 6 步通过后自己接着跑——它必须等你明确开口。

第 8 步：回到第 3 步

  • 你说：「OCN 现在在哪一步？」
  • AI 做：调 navigator.where_am_i。
  • 你看到：新 step（如 step_scope）。

然后从 第 3 步 重复：拿新 step 的 brief、建产物、写、验、advance。每个 step 都是这一套，直到 OCN 报告项目走完。

速查表：你说的话对应哪个工具
你说什么 AI 背后用 是否改 state.json
"OCN 在哪一步？" navigator.where_am_i（MCP，只读） 否
"拿一下当前 step 的 brief" navigator.brief（MCP，只读） 否
"建产物模板" navigator.create_artifact（MCP） 否（只写 docs/）
"把这一步的产物写好" AI 自己的 Edit 工具改 docs/ 否
"验证一下产物" navigator.run_gate（MCP，只读） 否
"我自己跑 ocn advance" ocn advance CLI（Bash，你触发） ✅ 是
两条硬规则

(1) navigator.run_gate pass ≠ 状态推进。 run_gate 只读，跑出 pass 不会让 OCN 自动进入下一 step。推进只能靠你敲 ocn advance。如果 AI 跑完 run_gate 就宣称「我们进入下一步了」——它在骗你，state.json 根本没动。

(2) AI 永远不会主动 advance。 这是设计。一个守纪律的 Claude Code 在第 6 步通过后只会说：「门禁通过了，请你跑 ocn advance。」如果它没说就自己跑了——它绕过了 OCN 的纪律护栏。这种「自己往前冲」的行为，恰好就是 OCN 想消灭的「失控（drift）」失效模式。

这就是 OCN 卖的东西——纪律。

E.5 把 OCN 加入一个进行中的项目

适用场景：项目已经在跑——代码写了一半、PRD 在 Notion 上、AC 散在 Slack 里、决策记录靠回忆。现在你想把 OCN 的纪律层补上。

一个必须先承认的事实：OCN 不从你"现在的真实进度"开始

OCN 的状态机起点永远是 state_discovery / step_project_brief。当前版本没有 --skip-to 或 --start-at 选项，也不暴露 advance 时的 --override。

这是设计，不是限制。OCN 的价值在于强迫你显式回答「我们到底要做什么」「什么算完成」——如果让你直接跳到 BUILD，那 OCN 的 brief、门禁、审计就全是无源之水，跟没装一样。

所以加入中途的项目，唯一正路是回填。好消息是：内容你都有，只是搬到 OCN 的结构里。

ocn init 与已有文件：硬保证

加入中途项目最容易担心的一件事：ocn init 会不会覆盖我现有的代码或文档？

答：不会。源码层面的两条硬保证（已对照 src/core/init.ts 与 src/core/artifact/template-writer.ts 验证）：

  1. ocn init 只在 .ocoding/ 下写 OCN 自己的状态文件（state.json、sop.yaml、gates.yaml、artifacts.yaml、config.yaml），并 mkdir docs/。它不写任何 docs/XX-*.md 产物文件，也不动你的 src/、README.md、其它任何路径。
  2. 如果项目已经初始化过（.ocoding/state.json 已存在），ocn init 会拒绝执行并返回 exit 4（ERR_IO_OR_CONFIG），消息：「OCN is already initialized in this directory. / 当前目录已经初始化过 OCN。」——绝不覆盖你已有的 OCN 状态。

类似地，ocn doc create <type> 也默认拒绝覆盖：撞名时返回 exit 4，消息：「<type> already exists at <path>. Use --overwrite to replace it.」要替换必须显式加 --overwrite。

所以你可以放心在已有项目根目录直接跑。最坏情况是你已经有 docs/00-project-brief.md——这时 ocn doc create project-brief 会拒绝写，你有两个选择：

  • 把已有内容手动合并进 OCN 模板（保留 OCN 模板的双语章节标题 Title｜标题——它们是门禁的匹配锚点）。
  • 给已有文件改名备份（如 docs/00-project-brief.legacy.md），再跑 ocn doc create project-brief 让 OCN 写新模板，然后把原内容粘回 OCN 模板章节里。
推荐路径：把已有材料"灌入" OCN 的早期 step

第 1 步：在项目根目录初始化

  • 你说：「在当前目录初始化 OCN。」
  • AI 做：用 Bash 跑 ocn init。
  • 你看到：.ocoding/ 创建（docs/ 也被 mkdir）；状态在 state_discovery / step_project_brief。你现有的代码和文档原封不动。

第 2 步：把项目上下文一次性喂给 AI

  • 你说：「这是我们项目现状。PRD 在 [Notion 链接 / 文件路径]。AC 我贴在下面。当前代码在 src/，已经做完 [X 模块] 还要做 [Y]。架构是 [简述]。我现在要把这些灌进 OCN 的早期 step。」
  • AI 做：吸收上下文，准备后面每个 step 直接复用。
  • 你看到：AI 复述项目状况，确认它理解到位。

第 3 步：逐 step 回填

每个 step 都走 §E.4 第 3–8 步循环（拿 brief → 建模板 → AI 据已有材料填 → run_gate → 你敲 ocn advance）。区别是：因为内容已存在，AI 主要做"重组"和"格式对齐"，不是"创造"。

SOP 0.5.0 的 20 个 step 全有 doc create 模板。下表把每个 step 对应的产物路径，以及在你已有项目里去哪找这部分内容列出来：

DISCOVERY 阶段

step 产物 在你已有项目里去哪找
step_project_brief docs/00-project-brief.md 项目立项 PRD、Notion 项目主页、给投资人 / 老板的一页纸介绍
step_scope docs/01-scope.md 路线图、Linear / Jira 的 epic 列表、"v1 / v2 不做什么"的内部对齐文档

SPEC 阶段

step 产物 在你已有项目里去哪找
step_prd docs/02-prd.md Notion / 飞书 / Confluence 上的 PRD，重新组织进 OCN 章节结构
step_acceptance_criteria docs/03-acceptance-criteria.md Slack 上的 AC 讨论、bug ticket 里的"应该…"、QA 测试用例描述，整理成 Given/When/Then

DESIGN 阶段

step 产物 在你已有项目里去哪找
step_technical_architecture docs/04-technical-architecture.md package.json / requirements.txt / go.mod、部署架构图、技术选型记录
step_information_architecture docs/05-information-architecture.md 已有 sitemap、URL 路由表、菜单层级、产品截图
step_data_model docs/06-data-model.md schema.sql、Prisma schema、ORM models、ER 图、数据库迁移文件
step_api_contract docs/07-api-contract.md OpenAPI / Swagger spec、Postman collection、router 代码、@RestController 注解
step_test_strategy docs/08-test-strategy.md tests/ 目录结构、CI 配置（.github/workflows/）、覆盖率目标、E2E 框架选择
step_logic_backbone docs/07-logic-backbone.md 已有的打分/决策/规则逻辑、公式定义、信号→动作映射

PLAN 阶段

step 产物 在你已有项目里去哪找
step_mvp_plan docs/09-mvp-plan.md 已有路线图、kanban、Linear/Jira epic、Phase 0/1/2 划分
step_build_plan docs/10-build-plan.md 当前 sprint 计划、待办列表、近期 PR 计划

BUILD 阶段（通常是回顾性回填）

step 产物 在你已有项目里去哪找
step_implementation_log docs/11-implementation-log.md git log、已合并的 PR 标题与摘要
step_change_evidence docs/12-change-evidence.md 已部署的变更、CHANGELOG.md、release notes 草稿
step_integration_notes docs/13-integration-notes.md 第三方 API 接入笔记、SDK 调用样例、集成测试结果

VERIFY 阶段

step 产物 在你已有项目里去哪找
step_verification_report docs/14-verification-report.md 已跑过的 QA 报告、性能测试结果、用户验收记录
step_acceptance_mapping docs/15-acceptance-mapping.md AC 与代码 / 测试的对应表（往往要新整理）
step_failure_fix_log docs/16-failure-fix-log.md bug 修复记录、incident 报告、postmortem
step_regression_evidence docs/17-regression-evidence.md 回归测试结果、CI 历史绿灯记录
step_final_build_verdict docs/18-final-build-verdict.md 最终发布判定（一般要新写）

第 4 步：到了你"真实的"当前位置

回填到你"真正在做"的 step 时（比如你已经写了一半代码，正在 step_implementation_log），你不是从空白开始——已写的代码、已合并的 PR 就是这一步的产物来源。从这里开始，新的工作全部走 OCN 的纪律：每个新功能先有 step、再有产物、再写代码。

之前的"无政府期"作为历史快照，被忠实地保留在回填的 artifact 里。

三个常见的「我能不能…」

Q：能不能跳过早期 step，直接从 BUILD 开始？ A：当前版本不暴露这种跳跃。两个选择：

  • 诚实回填（推荐）——把前面 step 用已有材料填好。
  • 薄占位（不推荐）——写仅满足必填章节的最小占位让门禁过。这等于让 OCN 闭眼放行，等于你白装。

Q：我已有的 docs/ 编号会不会跟 OCN 撞？ A：OCN 占用 00-project-brief.md 到 18-final-build-verdict.md 这一段编号。撞名时 ocn doc create 会拒绝写并返回 exit 4，不会覆盖你已有内容。处理方式：把已有文件改名备份（如 docs/02-prd.legacy.md），再 ocn doc create prd，把原内容合并进 OCN 模板。OCN 模板里的双语章节标题（Title｜标题）必须保留——它们是门禁的匹配锚点。

Q：能不能让 AI 一次性把全部早期 step 跑完？ A：技术上可以。你说一句：

「按我提供的材料，把 step_project_brief 到 step_technical_architecture 全部 artifact 填好并验证。每个 step 验证通过后告诉我，我自己敲 advance。」

AI 会循环走 doc create → 编辑 → run_gate，每个 step 跑完 run_gate pass 后停下来等你敲 ocn advance。每一次 advance 仍是你的决策——这是设计，批量回填不能让 AI 替代人的"我们进入下一阶段"决策。

一个最小回填例子

假设你已有：

  • 一个跑了 3 个月的 React + Node.js 项目，代码在 src/
  • 半页 Notion 立项文档
  • Slack 里散落的 AC 讨论

灌入 OCN 大概这样跑（实际对话节奏会更密，下面只展示骨架）：

你：「把当前目录初始化为 OCN 项目。」 AI：跑 ocn init。状态：state_discovery / step_project_brief。src/ 和已有文件没动。

你：「项目背景：[贴 Notion 立项文档]。请按 brief 填 docs/00-project-brief.md。」 AI：调 navigator.brief → 拿必填章节（Problem / Goal / Users / Success Criteria）→ 用 Edit 改 docs/00-project-brief.md。

你：「验证。」 AI：调 navigator.run_gate → pass。 AI（守纪律）：「门禁通过。请你跑 ocn advance。」

你（终端）：ocn advance。

你：「scope 这步：v1 已经做完 [X/Y/Z]，还在做 [A]，明确不做 [B/C]。填 docs/01-scope.md。」 AI：循环……

你：「prd 用 Notion 那份重组进来：[贴链接]。」 AI：循环……

你：「ac 用这些 Slack 摘录：[贴]。Given/When/Then 化。」 AI：循环……

（DESIGN 阶段：data-model 用 prisma/schema.prisma、api-contract 用 router 代码、technical-architecture 用 package.json + 部署图。）

（BUILD 阶段：implementation-log 让 AI 从 git log 摘。change-evidence 从 PR 摘。）

走完后，新的工作全部从 OCN 的下一 step 开始——每个新功能都有 brief、有门禁、有审计。

E.6 实现阶段：执行证据流

planning 阶段的 artifact（00–10）跑过门禁、项目进入实现阶段后，Execution Navigator 命令接力。它们都是只读的"汇总器"——帮你和 agent 看清 repo、PR、验收证据、验证信号到底是什么状态。

  1. 先把 planning 阶段的 artifact 写完（上一节流程）——state_plan 的 artifact（00–10）是实现的门禁。
  2. 在你日常的 repo 流程里推进实现（你的 IDE、agent 等）。
  3. 实现过程中，用 §F.2 的 Execution Navigator 命令读证据：
    ocn exec status                              # 本地 git + OCN 状态快照
ocn github analyze-pr <n>                    # GitHub PR 分析（已开 PR 时）
ocn evidence map --pr <n>                    # AC 覆盖度
ocn next-prompt --agent claude               # 下一条 agent brief
ocn verify status --mode combined --pr <n>   # 验证就绪度
ocn verdict draft --mode combined --pr <n>   # 证据驱动的 verdict 草稿
  4. 合入前看一眼证据和 verdict 草稿——verdict 是保守的：ready-to-merge 要求所有信号一致（验证 ready、PR 干净、AC 覆盖、无失败 check、无 changes-requested）。

F. 核心 CLI 命令

所有命令都接受 --json，输出机器可读的 CommandResult。退出码是稳定契约：

退出码 含义
0 OK
1 门禁未通过
2 产物缺失或非法
3 状态机错误
4 配置/锁/IO 错误
5 SOP 版本不兼容

F.1 Planning Gatekeeper 命令

命令 用途 读/写 审计输出
ocn init [--tier minimal] [--json] 在当前目录初始化 OCN 项目。 写 .ocoding/、docs/ 与双轨审计文件。 project_initialized + state-write 事件
ocn status [--json] 显示当前 state、当前 step、当前 step 产物的相对路径，以及下一步动作提示。 只读。 无（避免 pull 模式刷日志）
ocn brief [--json] 输出当前 step 的 brief：必填章节、治理提醒、不确定性策略。 只读。 无（pull 模式）
ocn doc create <type> [--overwrite] [--json] 用内置模板生成 20 类产物之一（撞名时拒绝写入，需显式 --overwrite 才覆盖）。 在 docs/ 写产物。 artifact_created
ocn check [--json] 检查当前 step 的产物是否满足必填章节。 只读。 artifact_gate_run + artifact_gate_passed / artifact_gate_blocked
ocn gate [--json] 当前 step 的只读门禁聚合。和 check 输出相同；不改状态。 只读。 artifact_gate_*（无 correlationId）
ocn advance [--json] 跑门禁，通过后推进到下一 step。带锁保护，永不部分写入。 原子写 state.json。 完整 advance 事件链，共享 correlationId
`ocn auto on --phase <1 2 all>/off [--phase]/resume/status/trace [--json]` 自动模式开关（AM-009 / DEC-034）——按阶段把「门禁全绿后的推进」委托给 AI（第一阶段 = DISCOVERY→PLAN；第二阶段 = BUILD→VERIFY，含 task check 与里程碑回拨）。开关本身人工专属（拒绝 OCN_ACTOR=ai_agent）；手动模式仍是默认。status/trace 为 pull 模式不写审计；trace 按时间线重放每条 AI 决策（理由 + 引擎记录的机器上下文）。
ocn sop upgrade [--target <version>] [--plan] [--json] 把项目锁定的 SOP profile 向前迁移到更新的内置版本（只能向前；保留 config.yaml）。--plan 为只读 dry-run。 apply 重写 .ocoding/ 快照 + state.json（原子、带锁保护）；--plan 不写任何文件。 sop_upgraded（apply）/ sop_version_diff_detected（plan）
ocn rewind --to <stepId> --reason <text> [--json] 把游标拨回本轮中严格更早的一步（DEC-033 / AM-008）。--reason 必填；docs/ 产物一律不动；回拨零豁免——之后每次 advance 重过完整门禁。里程碑循环（分阶段项目，如 P0→P4 各配 go/no-go）：每个里程碑裁定后回拨到 step_build_plan，追加下一阶段 Task Specs（已完成任务规格原文保留——verify 命令一字不动 ⇒ 哈希不变 ⇒ done 保留），重过门禁——台账即成为项目级累积进度账。人类专属——不暴露 MCP。 写 state.json（原子、带锁；latestGateResult 置空）。 cursor_rewind（成功与失败）
ocn cycle new --yes [--json] 把本轮运行时状态归档至 .ocoding/cycles/<n>-<ts>/ 并从首步重开新一轮（DEC-033 / AM-008）。docs/ 保留供下一轮门禁快进；pin 与用户所有的 config.yaml 不变；审计 JSONL 不随轮归档——单链贯穿所有轮次。--yes 强制。在本轮完成边界（01 范围）达成时使用——轮内的逐里程碑迭代属于 ocn rewind 的职权。人类专属——不暴露 MCP。 运行时文件移入归档；重写快照 + state.json。 cycle_started（成功与失败）
ocn readiness list [--json] 按当前 state/tier 评估全部就绪检查，输出判定表（PASS/WAIVED/FAIL/UNKNOWN/NA 计数 + 阻断项）。 只读。 无（pull 模式——避免刷审计日志）
ocn readiness waive <checkId> --reason <r> --probe <cmd> [--json] 授予条件豁免（"带探针豁免"）：探针在授予时必须通过，每次跑门禁都会重新验证，项目离开当前 state 时豁免失效。仅限人工操作——永不通过 MCP 暴露。 把豁免写入 .ocoding/readiness.json。 readiness_waived
ocn task list [--json] 列出 build plan 任务台账：每个任务的状态（pending/done）、冻结的验收命令、依赖/阶段信息。 只读。 无（pull 模式）
ocn task check [<id>] [--json] 跑该任务冻结的验收命令——退出码 0 才标记完成；验收命令与冻结哈希不一致时拒绝执行。仅限人/agent 经 CLI 调用——永不通过 MCP 暴露。 写 .ocoding/task-ledger.json。 task_completed
ocn agent setup [--force] [--json] 一条命令完成 Claude Code 接线（AM-006）：.claude/settings.json 写入 Stop/PostToolUse 钩子（仅合并、带 command -v ocn 守卫）、.claude/ocn.md 治理契约、CLAUDE.md 追加导入（仅一次）、/ocn-next 斜杠命令。幂等；仅限人工。 写入/合并四个 .claude 文件面；永不覆盖用户已有条目。 agent_setup_completed
ocn hook stop / ocn hook post-edit 机器侧 Claude Code 钩子处理器（由 agent 宿主调用，非人工）：stop 在 agent 想结束回合时重跑门禁、不过则带 fix hints 顶回；post-edit 跑 commands.lint/typecheck 给即时反馈。设计为 fail-open。 原始钩子契约 IO（stdin JSON 入；block JSON / exit 2 + stderr 出）。 经复用的 check 引擎产生门禁审计事件

doc create 的 <type>（SOP 0.5.0 为全部 20 个 step 都提供了模板）： project-brief、scope、prd、acceptance-criteria、technical-architecture、 information-architecture、data-model、api-contract、test-strategy、logic-backbone、 mvp-plan、build-plan、implementation-log、change-evidence、integration-notes、 verification-report、acceptance-mapping、failure-fix-log、regression-evidence、 final-build-verdict。

init 的 --tier：minimal、production、full，目前只对 minimal 强制约束（production / full 接受参数但产物集合尚未差异化）。

F.2 Execution Navigator 命令

planning 门禁之外，OCN 现在还能读"实现期证据"（git、GitHub PR、AC、package scripts），并产出确定性汇总——不调 LLM、不写状态、实现内部不跑 npm run。这些命令在实现与评审阶段（state_plan 关闭之后）有用。权威 reframe：DEC-024。系列收口：docs/reports/2026-05-04-execution-navigator-verdict-draft.md。

命令 用途 典型用法
ocn exec status [--json] 读取本地 git 证据 + OCN 项目状态。 branch、head、dirty/clean、变更文件、近期 commit、当前 state/step 的快照。
ocn github analyze-pr <n> [--json] 通过 gh CLI 做只读 GitHub PR 证据分析。 PR metadata、变更文件、commit、checks、reviews——不写。
ocn evidence map [--json] [--pr <n>] 把 AC 映射到可用证据。 每条 AC 状态：evidence-found / evidence-candidate / missing-evidence / needs-human-review。
ocn next-prompt [--json] [--agent ...] [--mode ...] [--pr <n>] [--issue ...] 生成下一条确定性 agent 任务 brief。 9 个固定段：objective、evidence、allowed work、forbidden actions、verification commands、stop conditions。
ocn verify status [--json] [--mode ...] [--pr <n>] 汇总验证就绪度与缺失信号。 状态 ready / partial / blocked / pending / no-verification-data，含必跑命令与风险旗标。
ocn verdict draft [--json] [--mode ...] [--pr <n>] 输出基于证据的最终判定草稿。 类别 continue-work / request-changes / ready-for-review / ready-to-merge / hold-for-manual-review，含按字典序排序的 supports / blocks / warnings。

6 条命令按设计都是只读的：不写 .ocoding/、docs/、任何项目文件；不做 git 写操作；gh 仅 pr view / auth status；不调 LLM。--mode 可选 local | pr | combined；--pr <n> 在 pr / combined 模式下启用 GitHub 证据。各 MVP 实现报告见 docs/reports/2026-05-02-execution-navigator-*.md 与上面的收口报告。

F.3 Claude Code 集成——一条命令（AM-006 / DEC-031）

ocn agent setup      # 幂等；提交 .claude/ 与 CLAUDE.md 即可全队共享

把整套第二阶段 runbook 机械化接线：Stop 钩子在 agent 想结束回合时重跑 ocn check（不过则把 fix hints 顶回去继续修；带防环旗标、fail-open），PostToolUse 钩子在每次 Edit/Write 后跑你在 .ocoding/config.yaml 配置的 commands.lint/commands.typecheck，.claude/ocn.md 治理契约经 CLAUDE.md 导入随每个会话加载，/ocn-next 自动注入 ocn brief + ocn next-prompt 并开始任务。钩子命令带 command -v ocn 守卫，没装 ocn 的队友完全不受影响。接线后每个任务人只剩两个动作：Claude Code 里 /ocn-next → 终端里 review + ocn advance。

F.4 自动模式——可选，默认关闭（AM-009 / DEC-034）

ocn auto on --phase 1     # 委托第一阶段（DISCOVERY→PLAN 规划流水线）
ocn auto on --phase all   # 全自动：规划 + BUILD/VERIFY 循环 + 里程碑回拨
ocn auto status           # 当前授权 + 熔断器状态
ocn auto trace            # 复盘每条 AI 决策：理由 + 机器上下文
ocn auto off              # 回到全手动（默认）

自动模式委托的是触发权，永不委托裁决权：每次推进仍跑完整门禁栈，任务完成仍只认冻结验收命令的 exit 0。AI 每次调用必须签名（OCN_ACTOR=ai_agent，由 ocn agent setup 注入）并说明理由（--rationale——背景/依据/操作）；引擎另行独立记录机器上下文入审计。兜底：同一步骤连续 N 次（默认 5）门禁失败即熔断暂停，需人工 ocn auto resume 恢复——人工操作不受任何影响。任何模式下的人工专属硬禁区：readiness waive、cycle new、sop upgrade、override、非里程碑回拨、ocn auto 本身。MCP 面零变化（7 个工具，测试钉死）。开启第二阶段后，多 P 构建计划可一跑到底：每个里程碑裁定后 AI 自动回拨到 step_build_plan、追加下一个 P 的任务规格、重过门禁（done 任务保留）继续，直至全部里程碑完成才停。

G. MCP 工具

OCN 的 MCP server（ocn-mcp）通过 stdio 暴露 7 个工具。唯一已验证的 Host 路径是 Windows + WSL2 + Claude Desktop（DEC-017、报告）。Cursor 与 Cline 都支持 MCP，但暂未对 OCN 完成验证（DEC-019）；在它们各自的验证报告落地之前，不要把它们当作正式支持路径。

G.1 接入 Windows + WSL2 中的 Claude Desktop

在 Windows 中编辑 %APPDATA%\Claude\claude_desktop_config.json，并入：

{
  "mcpServers": {
    "ocn": {
      "command": "wsl.exe",
      "args": ["-e", "ocn-mcp"]
    }
  }
}

如果 WSL2 里 ocn-mcp 不在 PATH 上，把 "ocn-mcp" 替换成 which ocn-mcp 输出的绝对路径（通常是 npm 全局 bin，如 /home/<user>/.npm-global/bin/ocn-mcp）。保存后完全退出（含系统托盘）再重启 Claude Desktop，工具面板里应当看到 7 个 navigator.* 工具。

完整接入说明与每个工具的输入/输出 envelope：docs/mcp-usage.md。验证当时的实际记录见 docs/reports/2026-04-30-mcp-external-host-validation-report.md。

G.2 暴露的 7 个工具

工具 用途 是否写入
navigator.where_am_i 状态快照。 否
navigator.brief 当前 step 的 brief。 否
navigator.run_gate 只读门禁聚合。 否
navigator.create_artifact 用完整模板库创建产物（20 类：00–19，logic-backbone 在 07）。 仅文件系统
navigator.capture_log 追加到 docs/19-dev-log.md（type=dev）或 docs/18-research-log.md（type=research）；type=decision 强拒。 仅文件系统
navigator.detect_sop_version 比较已落盘 SOP profile 与内置 OCN SOP，输出漂移状态。 否
navigator.generate_next_prompt 当前 step 的必填章节 + 治理提醒 + 不确定性策略 + 自检规则。 否

G.3 4 个被禁工具——永远不会暴露

工具 原因
navigator.advance_phase 状态推进只能由人通过 CLI 触发。
navigator.capture_decision 决策必须反映人的意图；MCP 上的 capture_log 已硬拒 type=decision。
navigator.reset_project 破坏性操作，必须由人二次确认。
navigator.force_release_lock 会绕过状态安全不变量，仅运维使用。

由 tests/unit/mcp-tool-registry.test.ts 强制（ALLOWED ∩ FORBIDDEN = ∅）。

接入 OCN 的 MCP agent 可以读项目状态、渲染下一步 brief、准备产物、跑只读门禁、用模板创建产物、捕获 dev / research 日志。不可以推进状态、记录决策、重置项目、强制释放锁。

完整工具面 + 接入说明：docs/mcp-usage.md。

参考资料

H. 文档地图

OCN 把自己的设计基线放在 docs/ 下。开始阅读前两个治理点：

  • 决策日志（唯一权威）：docs/20-decision-log.md。早期文档里出现的 docs/19-decision-log.md 是同一份文件的旧路径，迁移记录见 AM-002。
  • 修订（amendment）索引：docs/amendments/README.md。冻结的 docs/00-08 设计基线之上的所有有效偏离都以 amendment 形式记录，不就地改动 (DEC-004)。
  • 冻结的设计基线：docs/00-project-brief.md 至 docs/08-mvp-plan.md 是 Phase 2 设计契约，作为历史档案对待。新项目通过 ocn init 拿到的是内置 SOP v1.1 的 step 布局；OCN 仓库自身按 DEC-003 跑项目级 override。
  • 报告：docs/reports/ ——Phase 2 收口、post-alpha P1 修复（4 篇）、Claude Desktop MCP Host 验证、alpha.0 / alpha.1 / alpha.2 / beta.0 发布报告、examples F2/F3、beta release marker、双语安装流程、文档审计、最终 Codex 全 repo 审计、post-Codex P1/P2 修复报告。Phase 2 基线见 docs/reports/2026-04-28-phase2-completion-report.md。
  • 计划：docs/plans/。当前活动的 GA Prep 计划是 docs/plans/2026-04-28-ga-prep-gap-review-plan.md。
  • MCP 用法：docs/mcp-usage.md。

I. 开发

npm install
npm run lint           # ESLint（TypeScript-eslint）
npm run typecheck      # tsc --noEmit
npm run test           # vitest run —— 459 个用例，约 24 秒
npm run test:coverage  # 加上覆盖率报告
npm run build          # tsc + 给 bin 入口加 chmod +x

pre-commit 钩子（Husky 9）会在每次 commit 前跑 lint + typecheck + test。CI 跑同样的检查再加 build 与覆盖率上传。CLAUDE.md 里的硬约束：单文件 ≤ 300 行、单函数 ≤ 50 行、参数 ≤ 4、嵌套 ≤ 3、对外 API 不许 raw any。

J. 路线图

GA Prep 阶段是从 Phase 2 收口走到 beta candidate 准备的一段文档/打包/运营就绪审计。大部分 GA Prep PR 已完成，证据全在 docs/reports/ 与 docs/20-decision-log.md。原始计划见 docs/plans/2026-04-28-ga-prep-gap-review-plan.md。

GA Prep 各 PR 状态

  • PR A——文档编号修复 + amendments 索引。已完成。
  • PR B——README first-5-minutes 与 CLI help 文案审计；最近一次双语安装流程刷新见 docs/reports/2026-05-02-readme-install-flow-completion.md。
  • PR C——MCP projectRoot 路径穿越审计 + 威胁模型。docs/security/mcp-threat-model.md 已合入；projectRoot 校验器在 P1-001 中由 validateInitializedProjectRoot 加固。
  • PR D——外部 MCP Host 验证。Claude Desktop on Windows + WSL2 已端到端验证（见 报告 与 DEC-017）；Cursor 与 Cline 留给后续工作。
  • PR E——npm 发布门禁计划 + CI 稳定性审计。发布纪律：DEC-008 / DEC-012 / DEC-015 / DEC-016 / DEC-021 / DEC-022。CI 矩阵已扩展到 Node 20 + Node 22（报告）。锁观察 flake 已加固（报告）。
  • PR F——examples/ 计划。F1 + F2 + F3 完成：examples/discovery-to-plan/ 是端到端可执行 smoke，覆盖 v1.0 SOP 全部 10 个 step（报告）。F4（README 顶层 "Try the example" 链接）即 §E 中的示例引用。
  • post-Codex 审计修复 train——最终 Codex 全 repo 审计（报告）与 P1/P2 修复（报告）；剩两个 P3 polish 项。

Execution Navigator MVP 系列（DEC-024 之后）

  • MVP 1–6 已完成——§F.2 的 6 条 Execution Navigator 命令在 PR #63–#68 中陆续 ship 并合入 main；横切 review 修复在 PR #69 落地（报告）。系列收口：docs/reports/2026-05-04-execution-navigator-verdict-draft.md。
  • 当前外部包：0.9.0-beta.1——Planning Gatekeeper（§F.1）+ Execution Evidence Navigator（§F.2）+ Logic Backbone DESIGN 门禁 + Readiness Backbone 横切门禁 + Task Backbone BUILD 任务台账 + Acceptance Backbone SPEC 门禁 + 可选自动模式（§F.4，AM-009）+ 可选契约主干（AM-012）——在 BUILD/VERIFY 把前端调用点与声明的 API 契约交叉校验的门禁。
  • Acceptance Backbone 在 0.9.0-beta.1 ship——SOP 0.8.0（AM-015 / DEC-041）：docs/03-acceptance-criteria.md 携带机器可解析的 ## Acceptance Specs｜验收规格 章节（### AC-<DOMAIN>-<n> 块，desc 必填，priority/given/when/then/trace 可选）。验收门禁在 step_acceptance_criteria 于 section 门禁之后运行，硬拦四类结构缺陷（no_specs、duplicate_id、invalid_id、missing_field → ERR_ARTIFACT_INVALID，exit 2），通过则冻结 .ocoding/acceptance-specs.json——作为 build-plan traces 绑定的权威 AC id 机器源。封堵"AC 写在表格里却悄悄没登记"的假完成漏洞；读取有陈旧性保护（docs/03 自门禁后未改则用冻结投影，否则走实时解析）。运行时默认 profile 按 DEC-039 切到 0.8.0（npm/SOP 步调一致 → npm 0.9.0-beta.1）；0.7.0 及更早保持冻结且可导入；ocn sop upgrade 迁移。见 docs/acceptance-backbone-proposal.md。
  • Task Backbone 在 0.5.0-beta.0 ship——SOP 0.5.0（AM-007 / DEC-032）：build plan 携带机器可解析的任务规格块（## Task Specs｜任务规格：goal/traces/touches/verify/dod + 可选 depends/phase/timeout）。build-plan 门禁校验六类硬缺陷（id 重复/非法、字段缺失、traces 悬空指向 AC、touches 悬空指向逻辑图节点、depends 悬空或成环、零任务），并把每个任务的验收命令哈希冻结进 .ocoding/task-ledger.json。完成与否只由 ocn task check 重跑冻结命令裁决（退出码 0 → done + task_completed 审计；命令漂移 → 拒绝执行）；/ocn-next 在 state_build 中派发第一个待办任务；台账未清时 ocn advance 不准离开 BUILD。封堵第四类假完成——"只有回执的完成"（receipt-only completion：BUILD 回执诚实但空洞，却能通过每道门禁；在 Lattice dogfood 中发现）。运行时默认 profile 按 DEC-032 切到 0.5.0；旧项目用 ocn sop upgrade 迁移。
  • Claude Code 集成在 0.4.0-beta.2 ship——AM-006 / DEC-031：ocn agent setup 一条幂等命令接线钩子 + 治理契约 + /ocn-next；执行逻辑以 ocn hook stop|post-edit 内置在 OCN 中（fail-open、防环）。
  • Readiness Backbone 在 0.4.0-beta.1 ship——SOP 0.4.0（AM-004 / DEC-028）新增基于角色的横切就绪门禁：从 54 个精选 IT 角色提炼出 55 条可证伪检查，在 section + logic 门禁之后于 ocn check / ocn gate / ocn advance 中运行。开放世界语义（FAIL 和 UNKNOWN 都阻断）；判定账本持久化到 .ocoding/readiness.json；新增 ocn readiness list / ocn readiness waive（带探针豁免）命令；旧项目用 ocn sop upgrade 迁移；运行时默认 profile 按 DEC-030 切到 0.4.0。
  • Logic Backbone 在 0.3.0-beta.1 ship——SOP 0.3.0 新增 DESIGN 阶段产物 docs/07-logic-backbone.md；ocn check 机器校验其计算/决策图（在缺角色、节点 id 重复、悬空引用、依赖循环、孤儿节点、未绑定触发器时阻断），通过则写出 .ocoding/logic-graph.json，ocn brief 把它汇总成执行顺序 + 触发器绑定。
  • 仍是 beta，非 GA。已在 Claude Desktop on Windows + WSL2 验证；Cursor 与 Cline 仍未验证。
  • 下一道门：在真实外部 repo 完成一次 dogfood，再考虑后续 release sync；GA promotion 仍由独立 DEC 把守。

GA Prep 之外——目前不在任何活动计划里，落地前必须先开 DEC

  • ocn doctor、ocn reset、ocn baseline
  • SOP 版本/升级工具
  • production / full tier 的产物集合差异化
  • mini-CRM dogfood（Tier 2 GA 成功条件）
  • HTTP / SSE MCP 传输、MCP 鉴权、MCP 会话管理
  • Cursor / Cline 真实 Host 验证（每个 Host 都需要 DEC-017 模式的独立验证报告）
  • GA promotion DEC（最终把 Host 范围、多 OS / Node 24+ CI 矩阵、dogfood 证据捆在一起的门）

K. 许可

Apache-2.0

from github.com/UncleTIM-GZ/O-CodingNavigation

Установить O Coding Navigation в Claude Desktop, Claude Code, Cursor

Рекомендуется · одна команда, все IDE
unyly install o-coding-navigation

Ставит в Claude Desktop, Claude Code, Cursor и VS Code — сам разбирается с npx, uvx и сборкой из исходников.

Впервые? Поставь CLI: curl -fsSL https://unyly.org/install | sh

Или настроить вручную

Выполни в терминале:

claude mcp add o-coding-navigation -- npx -y o-coding-navigation

FAQ

O Coding Navigation MCP бесплатный?

Да, O Coding Navigation MCP бесплатный — установка в пару кликов через Unyly без оплаты.

Нужен ли API-ключ для O Coding Navigation?

Нет, O Coding Navigation работает без API-ключей и переменных окружения.

O Coding Navigation — hosted или self-hosted?

Self-hosted: сервер запускается локально на твоей машине командой из раздела установки.

Как установить O Coding Navigation в Claude Desktop, Claude Code или Cursor?

Открой O Coding Navigation на unyly.org, выбери вкладку своего клиента (Claude Desktop, Claude Code, Cursor) и нажми Install — конфиг сгенерируется автоматически, без правки JSON.

Похожие MCP

Fetch

Web content fetching and conversion for efficient LLM usage.

автор: Community

AWS KB Retrieval

Retrieval from AWS Knowledge Base using Bedrock Agent Runtime.

modelcontextprotocolавтор: modelcontextprotocol

Spring AI MCP Server

Provides auto-configuration for setting up an MCP server in Spring Boot applications.

автор: Community

llm-analysis-assistant

A very streamlined mcp client that supports calling and monitoring stdio/sse/streamableHttp, and can also view request responses through the /logs page. It also

xuzexin-hzавтор: xuzexin-hz

Compare O Coding Navigation with

O Coding NavigationvsFetchO Coding NavigationvsAWS KB RetrievalO Coding NavigationvsSpring AI MCP ServerO Coding Navigationvsllm-analysis-assistant

Не уверен что выбрать?

Найди свой стек за 60 секунд

Автор?

Embed-бейдж для README

Похожее

Все в категории ai