v0.5.1 — autonomous loop mode for create-spec

Adds an opt-in autonomous loop mode to /thinkkit:create-spec for hierarchical (large-repo) specs. Default single-pick behavior is unchanged.

What's new

• Loop mode, opt-in only. Activate by passing --loop or saying "do them all" / "run the whole priority list." Claude then powers through every non-complete subsystem without per-subsystem prompts.
• Compaction-safe. Every iteration re-reads SPECIFICATION/manifest.json from disk, so automatic context compaction between (or mid-) iterations resumes cleanly.
• Bounded parallelism. Each subsystem is sharded into ≤3 parallel Explore subagents based on a >10 projects / >500 source files heuristic, keeping main-context pressure in check.
• priority-order manifest field. Optional for default flow, required for loop mode. If absent when the user opts in, the skill derives a candidate order and asks for confirmation.
• Stop conditions. Loop exits cleanly on all-done, user interrupt, or two consecutive non-converging pressure-test rounds (escalates with loop-blocked: true).
• Guardrails preserved. No git commits/pushes/tags inside the loop; writes restricted to SPECIFICATION/.

Files changed

• skills/create-spec/SKILL.md — Phase 0 hierarchical branch routes to Step 9.5 on explicit opt-in; argument-hint surfaces --loop.
• skills/create-spec/references/large-repo-process.md — new Step 9.5 section, manifest schema gains optional priority-order, Step 9 sub-steps 3 and 6 narrowed to permit loop mode without weakening the single-session default.

v0.5.0 — create-spec scales to large repos

Extends the create-spec skill to handle large repos and monorepos with explicit user consent on depth vs. breadth trade-offs.

What's new

• Phase 0.5 Scale Triage — measures repo size and gates on user choice before committing to a strategy. Threshold: ≥2,000 source files or >10 top-level modules.
• Five tier choices — architectural-only, architectural + targeted drilldowns, hierarchical single-session (parallel Explore agents), hierarchical multi-session, or full depth anyway.
• Hierarchical output — when tier 3 or 4 is chosen, produces a SPECIFICATION/ directory (kept separate from source) with index.md, manifest.json, and subsystems/<name>.md.
• Hybrid subsystem discovery — build-file heuristics propose candidates, Claude refines, user confirms before deep work.
• Parallel exploration — tier 3 delegates subsystem surveys to Agent(Explore) subagents in parallel to fit large repos in one session.
• Multi-session resume — manifest tracks per-subsystem status (pending/in-progress/complete/stale) with last-spec-commit. Re-invocation shows the list and asks which subsystem to work on next.
• Incremental hierarchical updates — Phase 0 detects an existing SPECIFICATION/, diffs each complete subsystem against its recorded commit, and marks drifted ones stale.

Why

Running create-spec on a large monorepo (e.g., 20k+ source files across hundreds of projects) previously forced the skill to improvise on-the-fly, silently downgrading depth without user input. This release makes the trade-off explicit and gives the user real agency.

v0.4.0 — Smarter create-spec

Reworks the create-spec skill with three improvements:

• Behavioral focus over structural mirroring — Directory Structure becomes Module Organization; Dependencies becomes Capabilities and External Integrations
• Implementation gap analysis — new Section 10 surfaces test coverage gaps, security risks, performance concerns, reliability issues, and dead code during the deep review
• Incremental updates — detects an existing spec, diffs against the recorded commit, and updates only changed sections. Integrates with CLAUDE.md to enforce ongoing maintenance.

v0.3.0 — Tutorial skill

Adds /thinkkit:tutorial — an interactive walkthrough for new users covering all skills, the four architecture patterns, and common compositions.

v0.2.1 — fix marketplace validation

Fix marketplace.json validation error that caused "Marketplace sync failed" when adding as a marketplace. Removed unsupported top-level description key.

v0.2.0 — add create-spec, take-notes, resolve-against-transcript skills

New Skills

• create-spec — Reverse-engineer any codebase into a reconstruction-grade SPECIFICATION.md, with iterative pressure-testing
• take-notes — Real-time meeting notes assistant: feed terse shorthand, get expanded prose with speaker attribution, action items, and open questions
• resolve-against-transcript — Post-meeting reconciliation: compare a transcript against notes, walk through discrepancies interactively

Fixes

• map-the-repo: Fixed duplicate step numbering
• Added allowed-tools frontmatter to all new skills
• create-spec: Disambiguated trigger description from map-the-repo
• Added CLAUDE.md for development guidance
• Updated README with all eight skills