Skip to content

CLAUDE.md

Latest commit

 

History

History
132 lines (92 loc) · 5.66 KB

CLAUDE.md

File metadata and controls

132 lines (92 loc) · 5.66 KB
Raw

Claude Code Instructions

@AGENTS.md

CRITICAL: Always Verify

NEVER say "probably", "maybe", "might be", "could be", or other conditional hedging.

Before making ANY claim about:

  • Whether a file exists
  • Whether code is present
  • Whether a feature works
  • Whether something was done

You MUST verify it first using the appropriate tool (Read, Glob, Grep, Bash, etc.).

Do NOT:

  • Assume something exists without checking
  • Claim something works without testing
  • Say "it should be there" without reading the file
  • Trust previous agents' claims without verification

DO:

  • Read files before claiming they contain something
  • Run tests before claiming they pass
  • Check directory contents before claiming files exist
  • Verify configurations before claiming they're correct

Commit Message Format (REQUIRED)

This project uses Conventional Commits

When creating commits, you MUST use this format:

<type>[optional scope]: <description>

[optional body]

Quick Reference

Types (use these exactly):

  • feat: New feature → CHANGELOG Added section
  • fix: Bug fix → CHANGELOG Fixed section
  • perf: Performance improvement → CHANGELOG Changed section
  • refactor: Code refactor → CHANGELOG Changed section
  • docs: Documentation only → Omitted from CHANGELOG
  • test: Tests only → Omitted from CHANGELOG
  • chore: Maintenance/tooling → Omitted from CHANGELOG
  • security: Security fix → CHANGELOG Security section
  • breaking or add !: Breaking change → MAJOR version bump

Examples

# New feature
feat(loops): Add OODA loops MCP integration

# Bug fix
fix(analytics): Resolve concurrent write issue in loop-usage.jsonl

# Breaking change (note the !)
feat!: Remove deprecated thoughtbox_v1 tool

# With scope
perf(catalog): Cache hot-loops.json for faster sorting

# Documentation (not in changelog)
docs: Update README with loops documentation

Why This Matters

  1. Automated changelog: Commits automatically populate CHANGELOG.md
  2. Semantic versioning: Commit types determine version bumps
  3. Clear history: Easy to see what changed and why
  4. PR quality: GitHub Action validates format

If you create a commit, use conventional format. The changelog automation depends on it.

Agent Teams — Thoughtbox Integration

When participating in an Agent Team, bootstrap Thoughtbox as your reasoning substrate through the Code Mode surface (the only registered MCP tools are thoughtbox_search, thoughtbox_execute, and thoughtbox_peer_notebook):

  1. Load the cipher: read the thoughtbox://cipher MCP resource (reasoning-pattern shorthand, includes the multi-agent extension).
  2. Begin work — record decisions as thoughts via tb.thought inside thoughtbox_execute.

Hub coordination (workspaces, problems, proposals, channels) is exposed over MCP as tb.hub.* inside thoughtbox_execute (SPEC-V1-INITIATIVE Phase 4.1). Call tb.hub.register({ name }) or tb.hub.quickJoin({ name, workspaceId }) once and record the returned agentId. The FIRST registration in an MCP session becomes the implicit default identity for agentId-less calls — if you are a sub-agent sharing the parent's MCP session, pass your own agentId explicitly in every tb.hub.* mutation or your work is attributed to the parent. The local-mode HTTP surface (POST /hub/api) remains for non-MCP clients. The spawn-prompt templates in .claude/team-prompts/ use this tb.hub.* surface (rewritten in Phase 4.5).

What to Record as Thoughts

  • Key decisions and their reasoning
  • Hypotheses before investigation
  • Conclusions after analysis
  • Disagreements with other agents' conclusions

What NOT to Record

  • Routine file reads or searches
  • Every intermediate step
  • Information that's already in the codebase

Decided Architecture (v1 deployment) — NON-NEGOTIABLE

  • Execution plane: Google Cloud Run
  • Control plane / persistence: Supabase — Postgres, Auth, Storage
  • Billing: Stripe
  • Session routing: single-instance Cloud Run (maxScale=1); transport sessions in-process
  • NO Cloud Storage FUSE — all persistence goes through Supabase, containers are stateless
  • DUAL BACKEND: FileSystemStorage (local/self-hosted) + SupabaseStorage (deployed). Both implement same interfaces. Neither replaces the other.
  • Initiative spec: .specs/deployment/v1-initiative.md

Sub-Agent Dispatch Rules

  1. Hard constraints first: When dispatching sub-agents, put decided constraints from the initiative spec at the TOP of the prompt as non-negotiable. Current codebase state goes SECOND, labeled as "what needs to change."
  2. Review output against spec: Before presenting sub-agent output to the user, check it against the initiative spec. If it introduces infrastructure or approaches not in the spec, catch it before it reaches the user.
  3. No invented constraints: Never treat "minimize code changes" as a goal unless the user explicitly says so. Deployment initiatives exist because code needs to change.

Improvement Loop Learnings

Auto-generated learnings from autonomous improvement cycles

What Works

What Doesn't Work

  • Sub-agents optimizing for "no code changes" when the initiative exists to make code changes
  • Presenting sub-agent output without verifying it against the initiative spec
  • Writing sub-agent prompts before checking existing infrastructure. The prompt IS the plan — if you assume a workflow (e.g., local docker build + docker push) without checking what's already in place (e.g., Cloud Build triggers), the sub-agent executes the wrong plan correctly. Always audit existing infrastructure before writing implementation prompts.

Current Capability Gaps