cc-devflow

Agent-first development workflow for planning, diagnosis, implementation, verification, and shipping.

中文文档 | English | Getting Started | Contributing | Security

CC-DevFlow is a small, explicit workflow system for agent coding. It gives an AI agent a PDCA feature loop, a PR Harness lane, and a lightweight cc-diagnose hotfix/debugging path.

Why cc-devflow

• Small public surface: core workflow skills, a PR Harness lane, one optional deep review skill, plus a CLI for installation and platform adaptation.
• Evidence before done: implementation must pass through verification proof before shipping or handoff.
• Skill-first distribution: the public contract lives in .claude/skills/<skill>/SKILL.md and PLAYBOOK.md, not in hidden runtime behavior.
• Multi-platform output: install once, then adapt for Codex, Cursor, Qwen, Antigravity, and related agent environments.
• Durable project memory: specs, planning, review, and handoff artifacts stay in devflow/; temporary worker scratch stays outside durable truth.

Quick Start

Prerequisites:

• Node.js 18+
• npm or a compatible package runner
• A Git repository
• Claude Code or another supported agent environment

Install the whole skill pack:

npx cc-devflow@latest init --dir /path/to/your/project

Generate platform outputs:

npx cc-devflow@latest adapt --cwd /path/to/your/project --platform codex
npx cc-devflow@latest adapt --cwd /path/to/your/project --platform cursor
npx cc-devflow@latest adapt --cwd /path/to/your/project --platform qwen
npx cc-devflow@latest adapt --cwd /path/to/your/project --platform antigravity

Refresh every supported platform output:

npx cc-devflow@latest adapt --cwd /path/to/your/project --all

After installation, ask your agent to use the workflow skills directly. Start with cc-plan for planned work or cc-diagnose for bugs. Use cc-next when you want the agent to rank local changes and issues, cc-dev to drive selected work until cc-act selects delivery, cc-pr-review to review remote PRs in a separate session, and cc-pr-land to land reviewed PRs into main.

Workflow Map

PR Harness: cc-next -> cc-dev -> cc-pr-review -> cc-pr-land

PDCA: cc-plan -> cc-do -> cc-check(review convergence) -> cc-act
Parallel PDCA: cc-plan -> cc-dev dispatch loop -> child cc-* environments -> integrate -> cc-check -> cc-act
Hotfix: cc-diagnose -> focused fix -> regression proof

flowchart TD
  Next --> Dev["cc-dev\nDrive current worktree to PR"]

  Dev --> Route{"Route"}
  Route -->|Feature or change| Plan["cc-plan\nFreeze scope and tasks"]
  Route -->|Bug or regression| Diagnose["cc-diagnose\nBuild feedback loop and fix"]

  Plan --> Do
  Diagnose --> Check

  Do --> Check["cc-check\nFresh verification + review convergence"]
  Check --> Act["cc-act\nCreate or update remote PR"]
  Act --> PRReview["cc-pr-review\nSeparate PR review session"]
  PRReview --> PRLand["cc-pr-land\nRebase, land, prove main parity"]
  PRReview -->|Fixes required| Dev
  PRLand --> Main["main\nLocal and remote parity"]

Workflow Skills

Maintenance skills are shipped with the pack:

• cc-simplify: review changed code for reuse, quality, efficiency, and spec drift

Planning Quality Gates

cc-plan freezes implementation decisions before cc-do starts. Non-trivial plans compare minimal viable and ideal architecture options, full designs include decision horizon plus error/rescue mapping, and test-first plans record test framework evidence, public test seams, behavior assertions, mock boundaries, coverage quality, mandatory regression tests, refactor candidates, vertical tracer-bullet slices, and confidence-per-minute test strategy when existing behavior changes. It records the expected final cc-check review convergence gate instead of spawning default cc-review child threads. cc-diagnose is deliberately lighter: reproduce with the sharpest loop available, rank falsifiable hypotheses, instrument narrowly, fix, prove the original repro is gone, and keep debug probes out of the final tree.

When large work needs parallel execution, cc-plan first freezes an execution environment graph in task.md#Execution Environments: dependencies, touched surfaces, routed skill, verification commands, and merge gates. cc-dev then uses that graph to create sibling worktrees / child sessions, dispatch cc-do, explicit standalone cc-review, cc-check, cc-diagnose, or bounded cc-act, and keep integration authority in the orchestrator thread through serial cherry-picks, phase gates, and final cc-check. Child sessions own only their assigned environment; they do not unlock phases, merge main, or make final delivery decisions. Codex child dispatch uses a fixed dispatch packet, real thread-tool preflight, and read-only integration audit evidence before cherry-pick.

Canonical language and durable decisions stay inside cc-devflow-native sources: task.md, Git history, PR truth, and handoff artifacts. Legacy planning artifacts are readable fallback inputs only.

Every shipped skill carries its own references/checklist-contract.md; there is no shared checklist reference. The checklist is a pause-point contract for that skill only, and its proof must land in the skill's normal evidence sink such as task.md, the response, PR truth, Git history, or release verification.

Every post-planning stage starts from task.md, current Git history/status, and PR or handoff truth when present. There is no runtime context query layer; disputed facts must be re-read from source artifacts. Use npm run benchmark:skills to keep public skill entrypoints thin; deeper planning rules should live behind conditional references instead of default context.

Verification And Ship Gates

cc-check now treats QA as a feedback-loop problem, not only a green-test problem. Bugfix and behavior work records the loop used to prove reality, expected versus actual behavior, reproduction steps, confidence-per-minute proof value, test boundary quality, and architecture follow-ups when no clean public test seam exists. It also launches review subAgents to run repeated cc-review against task.md, the current diff, and fresh verification evidence until no P0/P1/P2 finding remains; any unresolved P0/P1/P2 routes back to cc-plan, cc-do, or cc-diagnose instead of passing. Green tests with no meaningful proof value, broad snapshots, duplicate happy paths, no-op smoke tests, brittle internal assertions, or overmocked internals route back to cc-do or cc-plan instead of supporting a pass. It also classifies task.md#Failure Ledger entries, including eligible review escape candidates, as confirmed lessons, noise, or unresolved risks before closeout.

cc-act carries that evidence into PR briefs, handoffs, or incident postmortems when needed. PR and handoff output carries a release-readiness gate ledger for local checks, config/env, migrations/data, deploy/health, smoke/cleanup, rollback, and watch items; skipped, blocked, and not-applicable gates must say why.

Installation Modes

Whole-pack install

Use this when you want the complete .claude skill pack:

npx cc-devflow@latest init --dir /path/to/your/project
npx cc-devflow@latest init --dir /path/to/your/project --force

--force upgrades only cc-devflow-managed distributed skills and preserves unrelated project files under .claude.

Source checkout

When developing this repository locally:

node bin/cc-devflow-cli.js --help
node bin/cc-devflow-cli.js init --dir /tmp/example-project
node bin/cc-devflow-cli.js adapt --cwd /tmp/example-project --platform codex

Single-skill install with skills.sh

skills.sh is supported as a single-skill distribution channel:

npx skills add https://github.com/Dimon94/cc-devflow --skill cc-next
npx skills add https://github.com/Dimon94/cc-devflow --skill cc-dev
npx skills add https://github.com/Dimon94/cc-devflow --skill cc-plan
npx skills add https://github.com/Dimon94/cc-devflow --skill cc-diagnose
npx skills add https://github.com/Dimon94/cc-devflow --skill cc-do
npx skills add https://github.com/Dimon94/cc-devflow --skill cc-review
npx skills add https://github.com/Dimon94/cc-devflow --skill cc-pr-review
npx skills add https://github.com/Dimon94/cc-devflow --skill cc-pr-land
npx skills add https://github.com/Dimon94/cc-devflow --skill cc-check
npx skills add https://github.com/Dimon94/cc-devflow --skill cc-act

Use cc-devflow init for the full pack, cc-devflow adapt for generated platform outputs, and skills add only when you want one skill at a time.

Configuration

CC-DevFlow reads layered YAML config before durable workflow documents are written:

~/.cc-devflow/config.yml
<repo>/.cc-devflow/config.yml
<repo>/.cc-devflow/config.local.yml

Precedence is deterministic: defaults < user < project < local < environment < CLI. output.document_language is machine-enforced and currently supports en and zh-CN. It controls durable planning, review, and handoff Markdown headings, prose, placeholders, evidence summaries, and PR/body drafts; code, commands, paths, schema keys, API names, and commit type/scope literals stay unchanged. Non-standard preferences belong under agent_preferences; they guide style but do not override workflow contracts.

version: 1
output:
  document_language: en
agent_preferences:
  general:
    - Start with the conclusion.
  documentation:
    - Keep headings short and avoid marketing language.

Useful commands:

npx cc-devflow config init --cwd /path/to/your/project --project
npx cc-devflow config set output.document_language zh-CN --cwd /path/to/your/project --project
npx cc-devflow config resolve --cwd /path/to/your/project --format policy
npx cc-devflow config doctor --cwd /path/to/your/project

See config/user-config.template.yml for the full sample.

Repository Format

Distributed skills live in .claude/skills/:

.claude/skills/<skill>/
├── SKILL.md
├── PLAYBOOK.md
├── assets/
├── references/
└── scripts/

Each shipped skill keeps its runtime contract local:

• SKILL.md has YAML frontmatter plus the Harness Contract
• PLAYBOOK.md has the Visible State Machine
• local resources stay beside the skill that owns them

The currently distributed skill folders are:

• .claude/skills/cc-next/
• .claude/skills/cc-dev/
• .claude/skills/cc-plan/
• .claude/skills/cc-diagnose/
• .claude/skills/cc-do/
• .claude/skills/cc-review/
• .claude/skills/cc-pr-review/
• .claude/skills/cc-pr-land/
• .claude/skills/cc-check/
• .claude/skills/cc-act/
• .claude/skills/cc-simplify/

Durable vs Ephemeral

• New change directories use REQ-<number>-<description> for requirements or FIX-<number>-<description> for bug fixes. REQ and FIX numbers advance independently, so the same number may exist in both prefixes. Parallel worktrees may also create repeated numbers; the full change key must use a specific description to distinguish the work.
• devflow/changes/<change>/ stores durable change truth in task.md, optional handoff/pr-brief.md, and Git commits. Real failures start in task.md#Failure Ledger; confirmed recurring lessons may also be compressed into incident postmortems under devflow/postmortems/.
• New planned changes default to one human-authored Markdown artifact: task.md, with frozen design in ## Contract Summary. Hotfix diagnosis does not require a task.md handoff unless the work is deliberately promoted into the PDCA tail. Legacy planning and review artifacts are readable fallback inputs only.
• Workflow state is Git-owned: keep task.md current, commit each completed stage/environment, and do not create extra process files.
• Use npm run verify:examples and npm run benchmark:skills to keep workflow truth and skill entrypoints small and measurable.
• devflow/workspaces/<change>/ stores ephemeral runtime scratch such as worker assignment, journals, prompts, and session logs.
• Regenerable files should not be persisted under devflow/changes/.

Artifact contract quick checks:

npm run verify:examples
npm run benchmark:skills

For complete artifact examples, start with docs/examples/START-HERE.md. Example version bindings live in docs/examples/example-bindings.json. Migration and authoring guidance for the minimized artifact contract lives in docs/guides/minimize-artifacts.md.

Development

git clone https://github.com/Dimon94/cc-devflow.git
cd cc-devflow
npm install
npm test
npm run verify

Publish validation:

npm run verify:publish

The main contributor guide is CONTRIBUTING.md. It explains the public surface rules, local CLI smoke tests, documentation rules, and PR expectations.

Discussion

Scan the QR code to join the cc-devflow WeChat group for feedback, usage notes, and feature requests.

If the QR code expires, please open an issue so the maintainers can refresh it.

Community

• Star the project if the workflow is useful: GitHub stars
• Open issues for reproducible bugs, stale docs, or missing platform adapters.
• Keep PRs focused: one skill, one CLI behavior, one compiler/adaptation fix, or one documentation cleanup.
• If a shipped skill changes, update its version, local CHANGELOG.md, examples, and affected public docs in the same PR.
• Read the Code of Conduct before participating in project discussions.
• Report vulnerabilities through the Security Policy, not public issues.

Star History

License

MIT