convez.ai
Agente
Chris-Graffagnino

creance

Chris-Graffagnino/creance

Ver no GitHub
InicialPluginScore · de 0 a 100

Creance — a tethered-autonomy harness for coding agents: one task → one issue → one branch → one PR, maker≠checker review gates, deterministic guards. MIT template repo.

Instalação

$npx convez-ai Chris-Graffagnino/creance

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 3 meses
Versão
—
Issues abertas
29
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

Creance

A tethered-autonomy harness for coding agents.

A creance is the light line a falconer uses to fly a hawk that has not yet earned free flight. That is this harness's whole philosophy in one word: treat an autonomous coding agent like a CI pipeline you don't trust — give it real autonomy, on a tether. Every run starts identically, the dangerous actions are deterministically impossible rather than discouraged, a different context grades the work, and every "done" claim is verified against an artifact rather than the model's say-so.

What you get

A complete, probe-tested workflow engine for running engineering work as one task → one issue → one branch → one PR, with an adversarial maker ≠ checker review gate. Three layers, each independently swappable:

LayerLives inYou change it when…
Methodology — runtime-neutral engine logic, written against abstract [roles].claude/workflow/**never (see the do-not-break list)
Adapter — maps each role to a concrete mechanism (shipped: Claude Code; spec'd: Codex CLI).claude/ skills, agents, hooks, settings.json, MODELS.mdyou port to a new runtime
Profile — every project-specific fact.claude/PROJECT.md, memory/constitution.md, specs/every new project (that's the template part)

Concretely, that includes:

  • /next-task — the full-ceremony per-task loop: pick the task, open the issue, cut the branch, implement with tests, verify, run an adversarial multi-reviewer gate, stop at PR.
  • Three read-only auditor subagents (spec / constitution / contract) dispatched in separate contexts, so the agent that wrote the code is never the one grading it.
  • A deterministic PreToolUse guard (hooks/guard.sh) that vetoes editing on the base branch, bulk staging, commits/pushes to base, and under-tier dispatch of the constitution reviewer — before execution, with regression tests in CI.
  • /triage — a read-only daily heartbeat with a dead-man-switch launcher contract (templates in docs/launchers/).
  • A model table (.claude/MODELS.md) — the only file naming models; capability tiers everywhere else, so a model swap is a one-line edit.
  • Conformance probes — one verifiable probe per role, because a binding that reads correctly can still be dead on a live driver (that exact failure was caught twice in this harness's source project).

Quickstart

Prerequisites

  • git — the engine's branch/issue/PR discipline assumes it.
  • gh (GitHub CLI), authenticated — issues, PRs, and review status all go through it.
  • bash on PATH — including on Windows (Git Bash). The guard hook is wired as bash "$CLAUDE_PROJECT_DIR/.claude/hooks/guard.sh"; on a machine without bash the guard is silently dead (the exact failure class in DESIGN-NOTES.md §"the guard was silently dead"). CI's wiring check runs on Linux and cannot catch this — run the guard probes (P-GD) on each workstation.
  • rg (ripgrep) — bundled with Claude Code; not guaranteed on other runtimes, where the verification greps need it installed separately.

Steps

  1. Use this template (GitHub) or clone it.
  2. cp .claude/PROJECT.template.md .claude/PROJECT.md and fill every <...> — it is the single source of project facts the engine reads.
  3. Write your memory/constitution.md (start from memory/constitution.template.md) — the project's non-negotiable principles. The constitution auditor enforces it as law.
  4. Fill the specs/ tree: copy specs/000-template/ to specs/001-<feature>/, rename spec.template.md → spec.md and tasks.template.md → tasks.md, then fill in a spec with US# acceptance criteria, a task backlog, and provider contracts if you have swappable seams. Keep the .template.md suffix on the originals — it keeps the skeleton out of the engine's specs/*/tasks.md fallback glob, so its placeholder tasks are never selectable by /next-task.
  5. Adapt the placeholders in AGENTS.md and add your toolchain's commands to .claude/settings.json → permissions.allow (e.g. Bash(npm test:*)) and to .github/workflows/ci.yml.
  6. Rewrite the [environment block] in .claude/skills/next-task/SKILL.md for your machine (the shipped one is the Windows PowerShell 5.1 worked example).
  7. Probe before you trust: run the conformance probes (.claude/adapters/claude-code-probes.md) and the verification greps in .claude/EXTRACTION.md §5. Do not run the harness unattended until they pass.

Reading order

Provenance

Extracted from a production project (a privacy-first iOS app built owner-absent under this workflow), where the harness ran real tasks, caught real constitution violations in review, and had its probe suite catch two live adapter failures before they could bite. The scar tissue is documented, not sanded off.

License

MIT

Ver o original no GitHub

Bom para

Relacionados