Specfuse Loop

For engineers using AI coding agents — a local-first driver that structures features as verified, work-unit sequences so each agent session runs focused on one task instead of accumulating context drift.

A small, local-first executor for the Plan + Work Unit pattern in a single repository. You plan a feature as a sequence of gates, each a batch of self-contained work units with explicit acceptance criteria and verification. The loop dispatches each work unit to a fresh agent session, verifies the result itself, commits one squashed commit per unit, and stops at each gate so you can reflect — with the next gate already drafted and waiting for your review.

Specfuse Loop is one of three independently-adoptable projects under the Specfuse methodology suite:

• specfuse/codegen — deterministic source code from OpenAPI / AsyncAPI / Arazzo specifications.
• specfuse/loop — this project. Single-repo, spec-optional, lightweight. You author the task graph directly; no specification and no agent-coordination overhead are required.
• specfuse/orchestrator — multi-repo, spec-first, agent coordination across many component repositories.

Use any one without the others. The loop and the orchestrator share the same gate-cycle methodology (see docs/methodology.md); the loop is the lightweight surface for work that lives in one repo and may have no formal specification.

Why it exists

AI coding agents do well on narrow, well-scoped work and poorly on large, vague work. The loop's bet is that the leverage is in the planning: if you remove ambiguity up front — crisp work units with hard boundaries and machine-checkable verification — then execution can run with a fresh agent per unit, re-grounding from durable files each time rather than accumulating context drift. It is the Ralph loop idea applied at work-unit granularity, with the planning rigor Ralph's bare task list lacks.

How it works (in one minute)

• A feature lives in .specfuse/features/FEAT-YYYY-NNNN-slug/, with a PLAN.md (the task graph: gate order, work-unit membership, dependencies), one GATE-NN.md per gate, and one WU-*.md per work unit (frontmatter + the prompt body a fresh session receives). The loop also handles orchestrated features dispatched by the Specfuse Orchestrator, identified by INIT-YYYY-NNNN/FNN IDs — the loop treats both namespaces identically; only the ID root differs. Use .specfuse/scripts/gh_features.py to discover a target repo's open specfuse:feature issues as feature candidates; use .specfuse/scripts/adopt_feature.py <repo> <issue-number> (or the interactive /adopt-feature skill) to scaffold a dispatchable feature folder from a picked issue.
• The driver (specfuse run, from the specfuse-loop pip package) walks the current gate's ready work units, dispatches each as a fresh claude -p session, runs the unit's verification itself as the exit oracle, and commits one squashed, trailer-carrying commit per unit. A failed gate is retried with a fresh session carrying the failure evidence, up to three attempts, then escalated.
• Each gate ends with a closing sequence so reflection, a durable cross-feature LEARNINGS.md, documentation, and drafting the next gate all happen systematically. Non-terminal gates use a two-WU form (close-intermediate + plan-next); the terminal gate uses a single close WU. A legacy four-WU form (retrospective → lessons → docs → plan-next) is accepted but emits a lint warning.
• The gate is the human boundary. The driver runs unattended within a gate and stops at it; you review the next gate's draft and arm it. (Under automatic mode, safe gates can self-arm; the dangerous edges always pull you back in — see the methodology doc.)

Quickstart

In a target single-repo project:

Contributing to this repo? Run ./scripts/install-hooks.sh once after cloning to enable the pre-push hook (runs scripts/smoke-test.sh — same checks CI runs — before each git push). Bypass with git push --no-verify.

The driver installs from PyPI and the skills ship as a Claude Code plugin:

pipx install specfuse                           # or: uv tool install specfuse
#   one command for the whole suite: specfuse run / lint / monitor / stats / …
#   (or: python3 -m pip install specfuse, in a venv)

# in Claude Code, enable the skills plugin (one-time):
#   /plugin marketplace add specfuse/specfuse
#   /plugin install specfuse@specfuse

specfuse init /path/to/your-project             # scaffold .specfuse/ + wire .claude/ (--dry-run previews)

cd /path/to/your-project
$EDITOR .specfuse/verification.yml              # match the `code` gates to your stack
# author your first feature (in Claude Code: /draft-feature)
specfuse run --dry-run                           # show the gate walk, no dispatch
specfuse run                                     # the real run

Distribution. Code ships via pip — the specfuse umbrella hard-depends on specfuse-loop (the driver) and every other component, so pipx upgrade specfuse / uv tool upgrade specfuse re-resolves the whole suite. Claude assets ship via the specfuse/specfuse plugin marketplace. specfuse init lays down .specfuse/ and wires .claude/; specfuse upgrade is the same idempotent operation over an existing scaffold. Every specfuse run self-provisions (version-syncs .specfuse/ from the installed package), so an upgrade reaches existing projects on their next run. (./init.sh is a deprecated v1.0 shim that delegates to specfuse init/upgrade; slated for removal.)

Flat commands are deprecated. specfuse-loop, specfuse-lint, specfuse-monitor, specfuse-monitor-lint and specfuse-stats still work as aliases until 1.0.0, with identical arguments and behaviour. Use the subcommands. specfuse doctor reports which flat names are still on your PATH; SPECFUSE_NO_DEPRECATION_WARNING=1 silences the notice in noisy hooks or CI. pip install specfuse-loop on its own remains supported for library use.

One driver per working tree. The driver holds an exclusive advisory lock on .specfuse/.loop.lock for the duration of a run; a second driver targeting the same checkout exits immediately with a clear error message. To run two features in parallel, use separate git worktree checkouts — each gets its own lock. --dry-run is exempt and may run alongside a live driver.

This repository is also a self-demonstrating reference installation: its own .specfuse/ contains a worked example feature (features/FEAT-LOOP-2026-0001-health-endpoint/). From the repo root you can run:

python .specfuse/scripts/lint_plan.py .specfuse/features/FEAT-LOOP-2026-0001-health-endpoint
python .specfuse/scripts/loop.py --dry-run

Layout

specfuse-loop/
├── LICENSE  NOTICE  CONTRIBUTING.md  README.md  .gitignore
├── init.sh                      deprecated v1.0 shim → delegates to `specfuse init`/`upgrade`
├── docs/
│   ├── getting-started.md       narrated first-feature + operator walkthrough
│   ├── methodology.md           the gate-cycle contract (shared with the orchestrator)
│   ├── skills.md                the skills catalog, ordered by lifecycle phase
│   ├── concepts/                why it exists; orchestrator mapping
│   │   ├── ralph-lineage.md     the Ralph / Gas Town lineage
│   │   └── architecture-addendum-gates-and-iterative-planning.md
│   └── dev/                     internal working notes (not user-facing)
└── .specfuse/                   canonical scaffold + worked example
    ├── README.md
    ├── roadmap.template.md  verification.yml.example  LEARNINGS.md
    ├── rules/result-contract.md
    ├── skills/verification/SKILL.md
    ├── scripts/{loop.py, lint_plan.py, gh_features.py, adopt_feature.py, gh_backend.py}
    ├── templates/{PLAN,GATE,WU}.template.md
    └── features/FEAT-LOOP-2026-0001-health-endpoint/   (the worked example)

specfuse init also ships the durable docs — methodology.md, skills.md, and concepts/ — into a target's .specfuse/docs/, so an initialized repo is self-documenting without this checkout.

Status

Early but exercised. The driver, linter, parsing, dependency ordering, draft/arm gating, the deterministic auto-close predicate, and verification wiring are all tested, and the loop dogfoods itself — its own .specfuse/features/ holds 20+ features taken through the full gate cycle, including multi-gate features whose forward-design model (each gate's plan-next drafts the next) has held across four consecutive gates.

What works today: single-feature and orchestrator-dispatched features; adopting a GitHub specfuse:feature issue into a dispatchable folder; GitHub issue-label state transitions for adopted features; per-gate auto-close on clean runs with a full-ceremony fallback when a gate goes off-plan; and a single-driver working-tree lock so two drivers can't corrupt one checkout.

Expect rough edges. The interfaces (WU contract, RESULT block, correlation-ID scheme, verification.yml shape) are stable; tooling around them is still hardening.

License

Apache License 2.0. See LICENSE.