Repository navigation
Releases
v1.2.0 — The entrance card learns to forget (a retention contract with teeth)
Closes a growth defect seen downstream: the framework tells the Archivist to update pipeline_status.md at every close, but no rule ever rolled anything off — so the entrance card, read on the hottest paths (boot, every statusline render, SessionStart, the PreToolUse gates), grows without bound. A production adopter's card had reached ~96,000 characters (~43k tokens) of mission history duplicated verbatim from its CHANGELOG lane; applying this contract cut it −94% with zero information loss.
What's new
📇 Entrance-card retention contract (ADR 103) — the Retention Policy in GOVERNANCE.md now manages pipeline_status.md itself. The card holds current state only: bounded cells behind a leading legal [STATE] token, at most 3 one-line pointers under ## 📜 Recent Missions, and a ## 🧭 Navigation routing table naming a home for every information class the card no longer carries. Budget: ≤ 12,000 chars hard, > 10,000 warn (counted after CRLF normalization). Mission history lives where it always belonged — the CHANGELOG lane the Archivist already writes.
✂️ Verify-then-trim (Archivist Hard Rule 8) — deletion gets a sanctioned, provable moment. At every mission close the Archivist first verifies the CHANGELOG already holds the closing mission's narrative (append-first refusal: trim only history the CHANGELOG lane already carries), then writes the new current-state cells, demotes the prior mission to one pointer line, and drops pointers beyond 3.
🛡 A validator with teeth — new scripts/check-entrance-card.mjs: 8 hard rules on the shared entrance-card parser (single parsing truth — the validator can never disagree with the hooks), wired as pre-commit gate 3 with an 18-test suite on the zero-dep node:test runner. Alongside it: a > 10k warn lane in the pipeline-status narrative check and a SessionStart over-budget banner, so drift is announced at boot long before the hard gate bites. The new gate validated its own landing commit.
🔧 Statusline fix at the source — the template card's Active Persona cell is now the bare persona name (the emoji is presentation, applied by the statusline). The decorated idiom had silently rendered the status chip "⚙️ Unknown" since it shipped.
Stays out of your way
The starter template already conforms — a fresh project sees no new burden, and the Day-0 welcome block stays. Existing adopters: run npm run check:entrance-card. If your card is over budget, follow the verify-then-trim procedure in lfe-archivist/SKILL.md — append any mission narrative your CHANGELOG lane is missing first, then trim; the pre-trim card stays one git show away.
Under the hood
- Test suite: 1179 passing / 0 failing (+29 for the validator and the two hook lanes).
- Budgets ship as framework defaults with the GOVERNANCE "project tunability" override idiom — projects re-bound them in their own
LLM_AGENT_GUIDE.md. - Recorded as ADR 103 in
.docs/architecture/framework-decisions.md— the contract was proven on a downstream production repo first, then ported here generic.
v1.1.0 — The Inspector gets eyes (+ a scripted path for human rejections)
Closes a reliability gap seen downstream: the Inspector could pass a change on green tests while the rendered UI was wrong, and a human rejection at the final step had no route back into the pipeline — so fixes degraded to unverified in-place patching.
What's new
👁 Visual verification gate (ADR 102) — a new opt-in Inspector sub-skill, lfe-visual-check, renders the changed UI surface (preview/browser MCP, with a manual fallback), reasons over it, and presents it for a human visual sign-off. It is artifact-free — it saves no image files. A new visual-gate enforcement hook makes that sign-off mandatory: it denies the Inspector→Archivist close of a UI-touching slice until the inspection record carries the human's visual_confirmed + visual_signoff. It is the framework's first unconditional-deny floor (it holds even under warn posture), with fail-safe ALLOW on every ambiguous path so it can never deadlock a legitimate close.
🔁 Human-rejection rework loop (ADR 101) — a rejection at the Inspector's finalization gate now re-enters the Builder through a re-entrant rework loop (typed rework_directive sentinel; up to 5 distinct rounds; orthogonal to the mechanical Cycle Guard). The fix always travels back through the Builder and is fully re-verified; patch-in-place at the final step is structurally closed — src/** is writable only under the Builder persona.
Stays out of your way
Both ship inert on a fresh project: the visual sub-skill is off by default and arms only when a change actually touches a UI file class; nothing persists images; the starter stays a clean [BLANK CANVAS] scaffold. Non-visual work gains no new approval burden.
Under the hood
- The gate has its own injected-I/O test seam; new typed coordination fields are validated; the enforcement chain + gate inventory are reconciled across every documentation representation.
- Test suite: 1150 passing / 0 failing.
- Framework decisions recorded as ADR 101 + ADR 102 in
.docs/architecture/framework-decisions.md. SeeUSER_MANUAL.mdfor the visual sign-off ritual.
Claude-LFE v1.0.0
Claude-LFE v1.0.0 — first public release
A ready-to-clone Claude Code project scaffold built on Library-First Engineering: documentation is the source of truth, the AI executes a disciplined assembly line, and the rules are mechanically enforced rather than left to vigilance.
What's wired and ready
- Persona assembly line — Architect → Builder → Inspector → Archivist, each tool-locked, handing off through files in
.plans/rather than conversation. - The skill set — grill → PRD → vertical slices → plan → plan-critique → build → TDD → inspect → archive, dispatched by the framework.
- Enforcement hooks — persona path-locking, a plan-critique gate, frontmatter validation, checkpoint-flip coordination, a boot-precondition handshake, the warn-first C1–C4 git/mission gate family, a continuous statusLine, and a SessionStart orientation reminder.
- Self-measurement — a skill-accuracy eval harness and a directive-voice census, both joined to
npm test.
Getting started
Clone (or Use this template), open in Claude Code, and run /lfe-boot → /lfe-extract-domain to seed your product's domain. See USER_MANUAL.md for day-to-day usage and LLM_AGENT_GUIDE.md for the agent protocol.
Further improvements land as updates.