agentic-swe

Agentic SWE

Claude codes your PRs. You review the receipt. Then merge.

An open-source autonomous SWE pipeline that runs in your editor or CI, writes every decision into your repo, and gives you a shareable audit trail of what the AI did and why.

What you get:

• Structured PRs — Claude works through a state machine (lean / standard / rigorous), not one mega-prompt
• Cost-attributed decisions — every phase has a dollar amount and an artifact in .worklogs/<id>/
• Audit trail — /receipt renders a shareable summary suitable for a PR description, Slack, or a compliance ticket

Quickstart

npm install -g @agentic-swe/agentic-swe
claude --plugin-dir "$(agentic-swe path)"

Then in Claude Code:

/work Add retry logic to the API client
# ...pipeline runs, opens a PR...
/receipt

→ See Install & first run for Cursor, Codex, OpenCode, Gemini CLI, and Claude Code plugin marketplace alternatives.

What /receipt looks like

/receipt reads .worklogs/<id>/ and renders the work item as markdown. Sample from test/fixtures/receipt/lean-happy/:

# /work add-retry-logic — Add retry logic to the API client

| Field | Value |
|---|---|
| Work ID | add-retry-logic |
| Track | lean |
| Status | completed |
| Duration | 47 min |
| Cost | $1.84 |
| PR | https://github.com/example/repo/pull/142 |

## Decisions made (6)

1. **feasibility → lean-track-check** ($0.08) — lean signal → feasibility.md#L1-L20
2. **lean-track-check → lean-track-implementation** ($0.04) — verdict: simple
3. **lean-track-implementation → validation** ($1.33) — implementation complete → implementation.md
4. **validation → pr-creation** ($0.21) — tests green
5. **pr-creation → approval-wait** ($0.18) — PR opened
6. **approval-wait → completed** ($0.00) — approved by suraj

## Human gates respected (1)

- `approval-wait` resolved by user at 2026-05-17T14:47:00Z — approved by suraj

## Loop counters

- `self_review_iter`: 0
- `doubt_cycles`: 1
- `code_review_iter`: 0

## Verifiable references

- All artifacts: `test/fixtures/receipt/lean-happy/`
- Audit log: `test/fixtures/receipt/lean-happy/audit.log` (9 entries)

Every line above is computed from .worklogs/<id>/ — no LLM summary, no hallucinated PRs. Reproduce locally:

node scripts/render-receipt.cjs --work-dir test/fixtures/receipt/lean-happy

Docs: agentic-swe.github.io/agentic-swe-site

Pipeline at a glance

After feasibility, lean-track-check sets pipeline.track in state.json. Tracks merge into PR creation → approval-wait → completed.

%%{init: {'theme': 'dark', 'fontFamily': 'ui-sans-serif, system-ui, -apple-system, Segoe UI, sans-serif'}}%%
flowchart TD
    start(["/work — start or resume"])
    feasibility["feasibility"]
    check{"lean-track-check<br/>sets pipeline.track"}
    lean["Lean track<br/>lean-track-implementation<br/>validation · pr-creation"]
    std["Standard track<br/>design → verification → test-strategy<br/>implementation → self-review<br/>validation · pr-creation"]
    rig["Rigorous track<br/>design → design-review<br/>verification → test-strategy<br/>implementation → self-review<br/>code-review → permissions-check<br/>validation · pr-creation"]
    gate{{"approval-wait<br/>human gate"}}
    done(["completed"])

    start --> feasibility --> check
    check -->|lean| lean
    check -->|standard| std
    check -->|rigorous| rig
    lean --> gate
    std --> gate
    rig --> gate
    gate --> done

    classDef accent fill:#1f6feb,stroke:#58a6ff,color:#ffffff,stroke-width:2px
    classDef step fill:#21262d,stroke:#30363d,color:#e6edf3,stroke-width:1px
    classDef branch fill:#21262d,stroke:#388bfd,color:#c9d1d9,stroke-width:2px
    classDef decide fill:#21262d,stroke:#d29922,color:#ffdfb8,stroke-width:2px
    classDef gateNode fill:#21262d,stroke:#a371f7,color:#e6edf3,stroke-width:2px