Agent architect kit
ThumbGate Pre-Action Checks self-improve from ranked lessons and repeated failures, hard-block detected secret leaks, and block matches in strict mode.From the repository description
npx -y skills add IgorGanapolsky/ThumbGate --skill agent-architect-kitAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
SKILL.md
5.1 KB, as published. Nobody here has run it
Agent Architect Kit ↔ ThumbGate Bridge
Status: shipped 2026-04-21. Test-backed. CLI-automatable. Safe to recommend publicly.
What this is
agent-architect-kit by @ultrathink-art is a CLAUDE.md + agent-role template kit that persists learnings as per-role markdown files (agents/state/memory/<role>.md). They solve the setup layer. ThumbGate solves the runtime-enforcement layer — PreToolUse hooks that actually block known-bad tool calls instead of hoping the agent reads its memory file.
The two stacks are complementary. A team running architect-kit graduates to ThumbGate when their markdown memory grows past the ~80-line ceiling their own memory-directive.md enforces, or when they want their operations meta-agent to do more than edit instructions — actually prevent the failing action.
This bridge turns that graduation into a one-shot command.
The bridge
What it does
scripts/integrations/architect-kit-memory-bridge.js walks an architect-kit memory directory, parses each per-role .md file, and emits one ThumbGate feedback entry per qualifying line:
| architect-kit section | → ThumbGate signal | notes |
|---|---|---|
## Mistakes | down with whatWentWrong | every entry becomes a thumbs-down lesson |
## Learnings | up with whatWorked | every entry becomes a thumbs-up memory |
## Stakeholder Feedback | up/down depending on keywords | "rejected", "broken", "wrong", "hate", etc. flip negative |
## Session Log | skipped | too granular to be useful in a searchable lesson DB |
Every ingested entry is tagged architect-kit, role:<name>, and the source section, so imports are auditable and rollbackable.
How to run it
Dry-run first (no writes, prints classification):
npm run integrations:architect-kit:import -- \
--dir=/path/to/agents/state/memory \
--dry-run --json
Real import (writes to ThumbGate feedback log):
npm run integrations:architect-kit:import -- \
--dir=/path/to/agents/state/memory
Single role (e.g. import only coder.md):
npm run integrations:architect-kit:import -- \
--dir=/path/to/agents/state/memory \
--role=coder
What you get after import
npm run feedback:stats # see new entries grouped by tag
npm run feedback:rules # regenerate prevention rules from the imported mistakes
The imported mistakes now feed the same pipeline as native ThumbGate feedback: lesson DB indexing, Thompson Sampling rollups, prevention-rule generation, and PreToolUse hook injection. The architect-kit operations agent's "edit the instructions" loop is now backed by hooks that can actually refuse a tool call.
Test proof
npm run test:architect-kit-memory-bridge
# 16 tests, 0 fail
Fixtures live in tests/fixtures/architect-kit-memory/ and mirror the exact format from architect-kit's memory-directive.md. The test-suite parity guard (tests/test-suite-parity.test.js) pins this test into the npm test chain — dropping the test accidentally now fails CI.
Positioning talking points (for outreach + docs)
- CLAUDE.md is the control plane. That's architect-kit's thesis and it's right. ThumbGate doesn't replace CLAUDE.md — it adds a runtime gate that enforces whatever CLAUDE.md says.
- Template memory → queryable memory. Per-role markdown files stop scaling past ~80 lines (architect-kit documents this limit themselves). ThumbGate's SQLite+FTS5+LanceDB store is what you migrate to when your memory deserves search.
- Operations becomes teeth, not talk. Their
operationsmeta-agent edits instructions. ThumbGate lets operations also register prevention rules that PreToolUse hooks honor — the same loop, but the agent literally cannot skip it.
What NOT to do
- Do not re-import the same memory twice. Architect-kit memory has no stable entry IDs; repeated imports duplicate lessons. Prune the source file (their own directive says to) before re-running, or run dry-run + diff first.
- Do not import the
Session Logsection. The bridge skips it deliberately. Session log entries are ephemeral task receipts, not lessons. - Do not assume operations-role edits propagate automatically. The architect-kit operations agent edits markdown instruction files; ThumbGate's prevention rules are regenerated by
npm run feedback:rules. Wire that into your Ralph Loop (or theirs) if you want continuous re-derivation.
Files added/touched by this integration
scripts/integrations/architect-kit-memory-bridge.js— parser + importer + CLItests/architect-kit-memory-bridge.test.js— 16 unit teststests/fixtures/architect-kit-memory/{coder,qa}.md— fixtures matching their formatpackage.json— newtest:architect-kit-memory-bridgeandintegrations:architect-kit:importscripts, chained intonpm testskills/agent-architect-kit/SKILL.md— this doc
When to reach back out to @ultrathink-art
If they ship stable entry IDs in their memory format, upgrade the bridge to do incremental imports (skip already-seen entries). Until then, treat this as a one-shot migration tool, not a sync daemon.