ai-dp

claude-ai-dev-pipeline-plugin

An autonomous feature-development pipeline for Claude Code (ai-dp namespace). It runs end-to-end without interactive questions — the agent makes every hard call itself, records each one, and finishes with a readable report.

investigation → decisions → plan → plan-improve → plan-improve-apply
              → plan-wrapup → implementation → codereview → report

State lives in your project at .claude/ai-feature-pipeline/<feature>/. Artifacts (context.md, decisions.md, plan.md, review.md, report.md, state.json) are plain files you can read, diff, and commit.

This is the autonomous sibling of dp. dp pauses at proposal and wrap-up for your approval; ai-dp never stops to ask. The two are fully namespaced and can be installed side by side (see Coexistence with dp).

What makes it autonomous

• No interactive questions. No step ever calls AskUserQuestion. Every ambiguity is resolved by the agent.
• A decisions document. Before planning, the decisions step writes decisions.md — a list of every hard call, each with the options considered and the one SELECTED (plus a one-line rationale). It's your audit trail, and you override a run by editing it.
• No proposal gate. There is no separate proposal step; plan.md simply opens with a Proposal section (what/why, root cause for bugs, brief approach).
• A final report. The last step writes report.md: how it was implemented, problems encountered, any tech debt added, and what tests were added — in plain, readable prose.

Quick start — Claude Code

In a running session:

/plugin marketplace add kostia-official/claude-ai-dev-pipeline-plugin
/plugin install ai-dp@claude-ai-dev-pipeline-plugin

From the terminal:

claude plugin marketplace add kostia-official/claude-ai-dev-pipeline-plugin
claude plugin install ai-dp@claude-ai-dev-pipeline-plugin

Then in any project:

/ai-dp:dev-pipeline rewrite auth to use refresh tokens

To resume:

/ai-dp:dev-pipeline continue auth-rewrite
# or pass the path directly:
/ai-dp:dev-pipeline .claude/ai-feature-pipeline/auth-rewrite/

Upgrade later:

/plugin marketplace update claude-ai-dev-pipeline-plugin
/reload-plugins

Prerequisites

• Bun on $PATH — every hook and CLI script runs on Bun. brew install bun or curl -fsSL https://bun.sh/install | bash.
• git — the codereview and report steps read the run diff.

No bun install needed: the scripts use only Bun/Node builtins.

What the pipeline does

#	Step	Output	Notes
1	ai-dp:investigation	context.md	Wide context; records any ambiguity instead of asking.
2	ai-dp:decisions	decisions.md	Resolves every hard call itself; options + SELECTED + rationale.
3	ai-dp:plan	plan.md	Opens with the Proposal section, then the detailed plan.
4	ai-dp:plan-improve	review.md	Self-review of the plan.
5	ai-dp:plan-improve-apply	patched plan.md	Applies fixes; decides unclear ones itself and logs them.
6	ai-dp:plan-wrapup	finalized plan.md	Consistency check, then auto-advances (no sign-off).
7	ai-dp:implementation	code changes	No : any / no dirty as; Stop hook blocks completion until typecheck + lint pass.
8	ai-dp:codereview	review + fixes	Embedded reuse/quality/efficiency review; Promise.all is never auto-applied.
9	ai-dp:report	report.md	Readable summary: implementation, problems, tech debt, tests.

Once started, the run chains through all nine steps on its own.

Hook enforcement

• Stop hook (enforce-pipeline-progress.ts) keeps the run moving: it hard-blocks a turn from ending while the current step is still pending. It also blocks finishing implementation until typecheck + lint are recorded as passing.
• The strict-typing rules (no : any, no dirty as) are self-policed by the implementation skill and backed by the project's own lint — not a separate static-analysis hook.

Coexistence with dp

ai-dp is fully namespaced so it never collides with dp:

• Env vars: AIDP_SESSION_ID, AIDP_PLUGIN_ROOT, AIDP_STATE_DIR (vs dp's DP_*).
• Runs: .claude/ai-feature-pipeline/<feature>/ (vs dp's .claude/feature-pipeline/). Each plugin's Stop hook only sees its own runs.
• Logs: /tmp/ai-dp-logs/ (vs dp's /tmp/dp-logs/).

Install both; they won't interfere.

Session scoping

Each session sees only the runs it created or explicitly resumed, so two concurrent sessions in one project don't block each other. A SessionStart hook captures the session id and exports it (AIDP_SESSION_ID) via additionalContext + $CLAUDE_ENV_FILE; the orchestrator and skills pass it to advance.ts. Explicit resume transfers ownership to the current session.