v2.3.1 — bugfixes #4 #5 #6

Three bugfixes against v2.3.0, surfaced from real-world build sessions.

Fixed

#4 — Engine state preserved across SessionEnd until pipeline complete (high)
session-end-handler.js previously deleted engine state files unconditionally, wiping mid-flight state when a subagent stop or inactivity timeout fired during a long /build. Now reads data.snapshot.project.state and only deletes when state === "complete". Mid-flight, missing, and malformed states are preserved.

#5 — Phase 7 test author/reviewer separation now programmatically gated (medium)
v2.3.0 prescribed test author/reviewer separation in prose only — the implementing agent could write tests inline and skip the author dispatch. Added required dep dep-test-author-spawned and gate gate-test-author-before-done blocking task-done. Build engine: 17 deps, 11 gates. commands/build.md Phase 7 Step A requires an inline update-engine proof string after the author subagent is dispatched.

#6 — FRD fast-track skip codified as first-class workflow (low)
Added optional audit-marker dep dep-fast-track-mode and a fourth decision-tree bullet to commands/build.md Phases 2/3/4. Users who explicitly say "go straight to build, FRD covers it" now follow a documented path with audit-trail clarity.

Notes

All three regression test suites were authored by sub-subagents (not the implementer) and reviewed by ucai:reviewer against the qa skill's 10 anti-gaming verdicts — the v2.3 discipline applied to itself.

5 files modified, 3 new test files, 18 regression tests total.

v2.2.0 — Enforcement Engine (never-forget)

What's New

Programmatic phase enforcement — /build and /ship pipelines now use a ContingencyEngine that mechanically blocks phase transitions until dependencies are met. This is the difference between "Claude usually follows the instructions" and "Claude literally cannot skip Phase 4."

Added

• Enforcement engine integration — never-forget library vendored into scripts/lib/never-forget/
• Build pipeline: 16 dependencies, 8 tasks, 10 logic gates, 128 shadow reactions
• Ship pipeline: 13 dependencies, 9 tasks, 7 logic gates
• 5 new scripts: engine-factory.js, engine-gates.js, update-engine.js, setup-build-engine.js, setup-ship-engine.js
• Gate checks before every phase in both /build and /ship
• Engine state updates after every phase with proof of work
• All hooks inject engine status — SessionStart, UserPromptSubmit, Stop, PreCompact, SessionEnd

How It Works

# Before Phase 4 (Design):
engine-gates.js --pipeline build --task task-design
→ {"allowed": false, "blockers": ["Complete codebase exploration before designing"]}

# After Phase 2 completes:
update-engine.js --pipeline build --dep dep-codebase-map --state complete --proof "12 files mapped"

# Now Phase 4 is unblocked:
engine-gates.js --pipeline build --task task-design
→ {"allowed": true, "blockers": []}

Backward Compatible

If the engine state file is missing or corrupted, everything falls back to instruction-based flow. No breaking changes.

Full changelog: CHANGELOG.md

v2.1.0 — Eliminate all ambiguity

Comprehensive audit of all 11 commands, 8 agents, 8 skills, and docs found 38+ issues: vague instructions, conflicting guidance, missing decision criteria, undefined terms, and phase gaps. This release fixes all of them.

Added

• /cancel-ship command — cancel an active ship pipeline with cleanup warnings for worktrees, branches, and PRs
• Decision Guide in workflow-guide.md — quick answers for which explorer, /build vs /ship, /init timing
• Terminology section — defines "build order step" vs "milestone" vs "phase"
• Failure Recovery table in /build — how to recover from interruption at each phase
• Build order sequencing algorithm in /plan — topological sort with concrete steps
• Acceptance criteria capture in /build Phase 1
• Agent synthesis steps in /debug, /plan, /docs
• Mood selection criteria in /plan Phase 4P
• Version bump arbitration in /release
• Completion promise definition in /iterate
• scripts/lint-plugin.js — automated consistency linter (24 checks)

Changed

• TodoWrite → tasks/todo.md across 6 commands, 7 agents, 3 skill files
• /ship phase count correctly documented as 9 phases (0-8)
• /build — fixed approval gate, added QA skill early, severity-ranked consolidation, phase transitions
• /plan Phase 5P — restructured into Prioritize/Sequence/Cross-cutting/Approve sub-steps
• Lessons consolidation — clarified: SessionStart warns, /ship auto-consolidates, others don't
• PreCompact scope — clarified it fires for ALL commands
• Decision criteria centralized in workflow-guide.md (single source of truth)

Peer-reviewed by GPT-5.4 via Copilot CLI. All 5 findings incorporated.

v2.0.1 — Consistency fixes

Fixed

• Remove semicolons from stop-handler.js to match project-wide no-semicolon convention
• Rename receiving-code-review skill to review-responder for consistent agent-noun naming
• Fix workflow guide claiming skills "load automatically" — clarified that commands load them
• Update CLAUDE.md to remove stale "uses semicolons" note and bump version references

Changed

• Command descriptions improved for clarity: /build ("guided... with approval gates"), /init ("analyze codebase... with project conventions"), /iterate ("autonomous loop... until done or max iterations")
• Version bumped to 2.0.1 in plugin.json, marketplace.json, CLAUDE.md

v2.0.0 — Autonomous Execution Engine

What's New

Ucai v2.0 adds an autonomous execution engine on top of the existing methodology layer. Give a spec, get a PR.

/ship — Zero-gate spec-to-PR pipeline

The headline feature. 8 phases, zero approval gates:

• Spec → Explore → Detect Infrastructure → Implement → Verify Loop → Light Review → Create PR → Cleanup
• Runs in a worktree by default (user is never blocked)
• Deterministic test execution + auto-fix loop
• Auto-formatting via PostToolUse hook
• Optional CI watching (--ci-watch)
• Commit per milestone for clean PR history

/bootstrap — Infrastructure scaffolding

For projects without tests, linting, or CI:

• Detects tech stack (JS/TS, Python, Go, Rust)
• Recommends standard tools (Vitest, pytest, Prettier, ESLint, GitHub Actions, etc.)
• Scaffolds with one approval gate
• Creates a real example test (not a dummy)

PostToolUse auto-formatting

Every Write and Edit now runs through your project's formatter automatically. Detects Prettier, Black, Ruff, gofmt, rustfmt. Caches detection. Fails silently — never blocks.

Utility scripts

• detect-infra.js — Detect test/lint/format/CI commands from project files
• run-tests.js — Deterministic test runner with JSON output
• consolidate-lessons.js — Auto-consolidate lessons when >100 entries

Enhanced hooks

All lifecycle hooks (SessionStart, Stop, UserPromptSubmit, PreCompact, SessionEnd) are now ship-pipeline aware — context injection, state recovery, and cleanup.

Backward compatible

All 9 existing commands (/init, /plan, /build, /debug, /iterate, /cancel-iterate, /review, /docs, /release) are unchanged.

---

Full changelog: https://github.com/Joncik91/ucai/blob/main/CHANGELOG.md

v1.2.1 — Project scale assessment

Added

• Project scale assessment: /plan Phase 1P classifies projects as Mini/Small/Normal/Large based on scope signals, calibrating build order step count and milestone granularity throughout the planning flow

Changed

• /plan Phase 5P build order references project scale with explicit step count ranges; exceeding the upper bound requires justification
• /plan Phase 3F.5 milestone scoping replaces "more is better" bias with scale-calibrated counts: Mini/Small → 1-2, Normal → 2-3, Large → 3+

Fixed

• Sequential todo.md updates — mark each phase complete individually instead of batching

Full Changelog: v1.2.0...v1.2.1

v1.2.0 — Cherny Methodology Upgrade

What's New

Embeds Boris Cherny's Claude Code methodology into ucai's existing architecture. No restructuring — every command, agent, hook, and skill stays. The methodology principles get woven into the existing fabric.

Persistent Task Tracking

• /build and /debug write tasks/todo.md with checkable phase items
• Hooks inject active task into every prompt and surface progress during compaction

Self-Improvement Loop

• Corrections captured in tasks/lessons.md (append-only, persistent across sessions)
• Loaded by /build, /debug, /review, /docs at session start
• SessionStart announces count and warns at >100 entries

QA/TDD Integration

• /build Phase 7 now writes automated tests (unit/integration/E2E) before manual testing
• /debug Phase 5 writes regression tests before review agents
• Both load Skill(ucai:qa) for framework guidance

Elegance & Quality Gates

• /build Phase 5: elegance checkpoint for non-trivial changes (>50 lines or >3 files)
• /build Phase 6: staff engineer self-check (abstraction, shortcuts, integration, error paths)

Autonomous Debug

• /debug Phase 3: diagnosis + fix plan in single approval gate (was two gates)
• /debug Phase 4: autonomous execution after approval

Lessons-Aware Commands

• /review feeds known patterns to reviewer agents
• /docs extracts gotchas from lessons for documentation

Naming Cleanup

• Removed senior- prefix from all skill references (22 occurrences across 10 files)

---

Full changelog: https://github.com/Joncik91/ucai/blob/main/CHANGELOG.md