All notable changes to this project will be documented in this file.
The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
- Review Enforcement in git-workflow skill and execute.md - reviewer agents (spec-reviewer, code-reviewer) MUST be spawned; inline reviews are now violations
- Pre-merge checklist requiring
.claude/reviews/spec/and.claude/reviews/code/files to exist - Test task detection and implementation source injection in execute.md - prevents test agents from hallucinating interfaces
- Defensive "For Test Tasks" section in babyclaude.md - blocks if implementation sources not provided
- Worktree path injection in babyclaude spawn with
cwd: worktreePathfor proper isolation
git-branch-guard.shswitched from blocklist to allowlist approach - only permits safe git commands (add, status, diff, log, show, ls-files, check-ignore, rev-parse, symbolic-ref, config --get/--list, commit)- Rationalizations table updated with "I'll do the review myself" violation
- Red flags updated with inline review warnings
- Branch collision bug - parallel babyclaude agents now operate in isolated worktrees instead of shared working directory
preserve-state.shPreCompact hook now always outputs valid JSON (even for no-op cases)- Phase-aware resume commands in preserve-state.sh (outline/execute/verify/ship)
batch-checkpoint-gate.sh: improved checkpoint message formatting with proper newline handling via jq
- babyclaude Stop prompt hook: use
{"ok": true/false}response format instead of{"decision": "allow/block"}(Claude Code expects the former)
sanitize-bash.sh: use"decision": "approve"instead of invalid"decision": "allow"verify-gate.sh: output to stderr instead of invalidhookSpecificOutputJSON (Stop hooks don't support it)
homepage,repository,license, andkeywordsfields to plugin.jsonpermissionModeto all 8 agents for proper permission handlingonce: trueto session-startup hook (prevents duplicate execution)pre-task-gate.shhook implementing constraint C-4 (review verdict gate)- PreToolUse/Task matcher in hooks.json for pre-task validation
- Permission deny rules in settings.local.json for dangerous git operations
- "Next Stage" sections to all 4 commands for workflow continuity
- Frontmatter
hooks.Stopto 5 agents (babyclaude, catastrophiser, cynic, git-perfectionist, taxonomy-extremist) - LLM-based Stop hook (
type: "prompt") for babyclaude completion evaluation sanitize-bash.shPreToolUse hook withupdatedInputpattern for command sanitizationoutput-schemato all 4 commands for structured JSON outputskill-rules.jsonfor skill auto-activation with intent/path pattern matchingskill-activation-hook.shUserPromptSubmit hook for auto-suggesting relevant skills- Execution tracing with
trace-start.shandtrace-end.sh(SubagentStart/SubagentStop) .claude/traces/directory for span-based execution timingallowed-toolsto all 4 skills with appropriate tool restrictions
- Commands restructured:
flags,merge_strategy,colormoved from frontmatter to body documentation - Plugin version bumped from 1.0.0 to 1.1.0
- Author name updated to full name in plugin.json
- Agent-specific SubagentStop hooks moved from hooks.json to agent frontmatter
- Invalid
context: forkfield from all agents (skill-only field) colorfield from command frontmatter (not a valid field)- SubagentStop section from hooks.json (replaced by frontmatter hooks + global tracing)
- Agent frontmatter now uses only valid fields per Claude Code spec
- Initial release
- 4 commands: outline, execute, verify, ship
- 8 agents: babyclaude, catastrophiser, cynic, spec-reviewer, code-reviewer, taxonomy-extremist, conflict-resolver, git-perfectionist
- 4 skills: brain-jam-plan, git-workflow, strict-enforcement, shipping-methodology
- Hook infrastructure with hooks.json and shell scripts