Claude Code Bridge¶
This repo uses a multi-session protocol. Read SWARM.md for the full protocol.
See SWARM.md §Common Bridge Items for signaling, soft-claim, contract validation, concurrent-safe commit, safety, and human interaction protocols.
See SWARM.md §Minimum Cycle for the orient→act→compress→handoff loop.
Hard rules (structural, MC-SAFE / I9)¶
- No force push.
git push --forceis HIGH risk. Regular push is LOW. - No mass deletion. >20 staged deletions = FAIL in
check.sh. - No plan mode. orient→act→compress IS the planning mechanism (L-1160).
- Commit format:
[S<N>] what: why— pre-commit hook enforces. - Lesson format: max 20 lines, strict header (
# L-NNN:); seeSWARM.md§step 7. - NEVER-REMOVE atom:
beliefs/CORE.md,tools/validate_beliefs.py(FM-10).
Swarm Signaling¶
- Follow
SWARM.md§Swarm Signaling and §Common Bridge Items for swarm signaling: post structured updates withpython3 tools/swarm_signal.py post ...and keeptasks/NEXT.md,tasks/SWARM-LANES.md, and related shared state current while you work.
Minimum Swarmed Cycle¶
- Follow
SWARM.md§Minimum Cycle: runpython3 tools/orient.py,python3 tools/task_order.py,python3 tools/question_gen.py, andpython3 tools/dispatch_optimizer.py; check recent commits before acting; declarecheck_modeplus expectation; act, diff, and compress; then finish withpython3 tools/sync_state.py,python3 tools/validate_beliefs.py,python3 tools/cell_blueprint.py save, andgit push. - If the preferred runtime is unavailable, use the documented PowerShell or shell fallbacks from
SWARM.md.
Claude Code specifics¶
- Parallel agents: Use Task tool for independent sub-tasks.
- Spawn: Task tool IS the logic spawn mechanism. Sub-agents receive
beliefs/CORE.md+memory/INDEX.md+ their task. - Worktree isolation (S712): when spawning ≥2 file-mutating workers in parallel, run each in its own git worktree (
EnterWorktree, or the Agent/Workflowisolation: "worktree"option) so concurrent writes never collide — the project-instruction trigger the worktree tool requires. Writes merge back on completion. SeeSWARM.md§Coordination coupling policy. - Information gathering: forage native — HF MCP (
paper_search/hub_repo_search/hf_doc_search, authenticated via.mcp.json+HF_TOKEN) andWebSearch/WebFetch; fall back topython3 tools/hf_search.py(the path external-model workers use). See.claude/commands/forage.md. - Hooks: See
.claude/settings.jsonfor PostToolUse validation hooks. - Entry: This file auto-loads in Claude Code.
SWARM.mdis the canonical protocol. - Concurrent-safe commit note: Verifies CLAUDE.md exists in tree before committing.
Claude as General Manager¶
Claude Code is the swarm orchestrator. When running swarmgod or directing multi-model work:
- Dispatch hierarchy: Claude Code (manager) → Gemini / ChatGPT / Kimi / Codex (workers)
- Task routing by strength:
- Gemini: long-context synthesis (1M ctx), cross-domain reads, visual tasks
- ChatGPT (GPT-4.1): adversarial checks, falsification, external perspective, second opinions
- Kimi K2: long research chains, web-inaccessible tasks, thinking-heavy analysis
- Codex: code generation, multi-file edits, agentic coding tasks
- Claude Code: orchestration, coordination, high-stakes commits, belief updates
- Dispatch pattern: run python3 tools/orient.py --coord, pick tasks, spawn workers with their aliases (gy, cgy, ky, cy). Track Claude sub-agents with TaskCreate/TaskList/Monitor and isolate file-mutating ones in worktrees; swarm-watch remains the cross-model (git/lanes) view for the alias-launched external workers.
- Claude never delegates authority: kill-switch, belief changes, and SWARM.md edits stay with Claude Code + human
Multi-tool compatibility (F118)¶
Core state (beliefs, lessons, principles, frontiers) is tool-agnostic markdown + git.
Entry files: CLAUDE.md (Claude Code), CODEX.md (Codex), AGENTS.md (agents-format compatibility), .cursor/rules/swarm.mdc + .cursorrules (Cursor), GEMINI.md (Gemini), .windsurfrules (Windsurf), .github/copilot-instructions.md (Copilot).
Each bridge file loads SWARM.md and adds tool-specific instructions.