speccraft-cli 0.1.0

speccraft — trust-graded knowledge base CLI (init a .speccraft/ KB in a repo)

Speccraft

A trust-graded knowledge base your codebase carries with it — so agents (and you) build on what's ratified as true, not on guesses.

---

Speccraft turns a repo's scattered, half-lost knowledge — intent, invariants, the reasons behind decisions, the gotchas in your dependencies, the bugs nobody wrote down — into a cited, versioned, human-ratified knowledge base that lives inside the repo (.speccraft/) and stays honest as the code changes.

It's built for the hard case: an existing (brownfield) codebase whose design rationale has already evaporated, worked by AI coding sessions across Claude Code, Codex, and OpenCode.

Cite or it didn't happen. Nothing becomes truth except through a human.

Contents

• Why it exists
• Quickstart
• What it captures
• The steady-state loop
• Enforcement
• Layout
• Multi-tool
• License

---

Why it exists

Every codebase knows things nobody wrote down: why a constant is 30, which alternative was rejected, that lapsed users still get paid features, that this dependency version has a known CVE. Code records the outcome of a decision, never the decision. AI agents, lacking that context, confidently repeat known mistakes and drift from intent.

Speccraft's answer is a grounded ratchet: seed a knowledge base from human-adjudicated truth, validate every change against it, and let error be caught rather than compound.

1	Cite or it didn't happen Every claim is pinned to path:line @<commit-sha>.
2	Provenance is never blurred Each fact is tagged by how it's known: derived (machine-harvested), inferred (agent hypothesis), elicited (your own words), decision (an ADR), external (dependency knowledge, sub-graded by source).
3	Nothing becomes ratified except through a human Agents propose; you rule. A fact's trust state moves pending → ratified → challenged, never silently.
4	Deterministic before generative Anything a regex or git can harvest is harvested by a no-LLM tool that cannot hallucinate; agents only interpret, and their output enters as an unratified hypothesis.

---

Quickstart

pipx install speccraft-cli
speccraft init /path/to/your-repo

Want unreleased main? pipx install git+https://github.com/swshukla/speccraft.git instead.

Requirements: python3, git. No API keys, no services, no LLM calls in the tooling itself — the harvesters are pure stdlib.

speccraft init scaffolds your-repo/.speccraft/, runs every harvester (pinned to your last code commit), installs the four speccraft-* procedures for Claude Code / Codex / OpenCode, wires AGENTS.md/CLAUDE.md, arms the git hooks, and writes KB-STATUS.md.

From there:

Step	Action
1	Tune .speccraft/kbforge.yaml — set components and risk_paths (auth / money / truth-critical path patterns).
2	Run speccraft-interview in a Claude Code (or Codex/OpenCode) session — the one irreducibly human step; intent and invariants land in kb/normative/ and everything downstream grounds in it.
3	Run the extraction passes, then work through .speccraft/QUEUE.md with speccraft-ratify.
4	It's self-sustaining after that. Just build — each commit runs the ship loop, drift keeps citations honest, recall grounds each task.

Go deeper: speccraft/forge/README.md (full walkthrough) · speccraft/forge/SPEC.md (system reference) · speccraft/forge/workflow-execution.html (visual tour)

---

What it captures

Aspect	How	Provenance
Structure — routes, models, tests, churn, module map	deterministic harvest	derived
Product intent, stage, monetization	founder interview	elicited
Invariants (the rules that must always hold)	founder interview	elicited
Capability map — what's marketed vs what the code does	agent extraction	inferred
Data sources — what data already flows in	agent extraction	inferred
Integrations — why each 3rd-party exists, what's used, who depends on what	agent extraction	inferred
Assumptions & tradeoffs — recovered from code "scars"	residue harvest → hypothesis cards	derived + inferred
Consistency — contradictions, duplicates, proposed conventions	harvest + classify	derived + inferred
Dependencies + version-pinned best-practices / gotchas	inventory + sourced cards	derived + external
Decisions — captured at the moment a tradeoff is made	ADR-lite lane	decision
Findings — the consolidated bug / work list	flows from rulings	mixed

---

The steady-state loop

 session start
      │
      ▼
 speccraft-recall ─── ground in ratified truth
      │
      ▼
   build  ───────────  speccraft-decide (log tradeoffs)
      │                speccraft-diverge (on conflict)
      ▼
 git commit ─── pre-commit lane guard
      │
      ▼
 ship loop: drift → re-pin → re-harvest → kb: commit → KB-STATUS refresh
      │
      ▼
 speccraft-ratify ─── founder rules the queue, occasionally
      │
      ▼
 feeds the next session

---

Enforcement

A git pre-commit hook keeps machine-proposed and human-ratified knowledge distinguishable: commits touching the founder/machine lanes (kb/normative/, kb/derived/, ledger/) require an explicit flag — KB_RATIFY=1 (a human ruling) or KB_SHIPLOOP=1 (the automatic re-pin).

It's not a lock against an intruder — you own the repo — it's a signature: when a fact is ratified, the git history shows a human stood behind it, so an agent can run autonomously without quietly promoting its own guesses to truth. A post-commit hook runs the ship loop on every code commit.

Sessions write only QUEUE.md, kb/decisions/, kb/inferred/, and proposed rows in findings/. Everything else is founder/machine.

---

Layout

your-repo/.speccraft/
├── kbforge.yaml       product profile: repo, components, risk_paths
├── KB-STATUS.md       auto-refreshed briefing (pin, open queue, invariants)
├── QUEUE.md           the one adjudication queue — rulings are git commits
├── kb/
│   ├── derived/       machine-harvested, regenerated each commit (never hand-edited)
│   ├── inferred/      agent hypotheses (pending-ratification)
│   ├── normative/     ratified intent, invariants, conventions (human-only)
│   └── decisions/     ADR-lite, captured at decision time (append-only)
├── ledger/DIV-*.md    ruled divergences (fix-code / fix-model / accepted-deviation)
└── findings/FINDINGS.md   the consolidated bug / work list

---

Multi-tool

One source, three runtimes.

	Claude Code	Codex	OpenCode
Rules file	CLAUDE.md (imports AGENTS.md)	AGENTS.md	AGENTS.md
Procedures	.claude/skills/	Agent Skills standard	OpenCode commands
Git hooks	identical	identical	identical

The git hooks enforce identically regardless of which tool — or which human — makes the commit.

---

License

MIT — do what you like with it.