feat: Claude Code patterns transfer (6 features + pipeline integration)

[VictorGjn]

Claude Code Patterns → modular-patchbay

Transfer de 6 features + 5 adapters + pipeline wiring extraits du source code Claude Code.

Features (3 phases)

Phase 1 — Standalone modules (6 source + 6 test files, 58 tests)

Feature	File
System Prompt Builder (static/dynamic boundary)	src/prompt/SystemPromptBuilder.ts
Reactive Compaction (5 signal types)	src/context/ReactiveCompaction.ts
Memory Pipeline (FS-backed + extraction)	src/memory/MemoryStore.ts
Context Collapse (4 strategies)	src/context/ContextCollapse.ts
Tool Use Summary (grouping + heuristics)	src/context/ToolUseSummary.ts
Agent/Skill Search (TF-IDF)	src/search/AgentSearch.ts

Phase 2 — Adapters + barrel export

Adapter	Bridges
systemFrameBuilderAdapter.ts	SystemPromptBuilder → buildSystemFrame()
reactivePackerWrapper.ts	ReactiveCompaction → packContext()
memoryStoreIntegration.ts	MemoryStore → context assembly
contextMiddleware.ts	ContextCollapse + ToolUseSummary → pipeline
agentSearchIntegration.ts	AgentSearch → agent registry

Phase 3 — Pipeline wiring + E2E

• systemFrameBuilder.ts → buildSystemFrameOptimized()
• packer.ts → packContextReactive
• contextAssembler.ts → assemblePipelineContextWithMemory()
• pipelineChat.ts → context middleware singleton
• server/routes/agents.ts → GET /agents/search
• server/services/teamRunner.ts → worktree auto-setup for agents with repoUrl
• Barrel export: src/claude-code-patterns/index.ts
• 2 E2E test files (8 test cases)

Architecture

systemFrameBuilderAdapter  → systemFrameBuilder.ts (buildSystemFrameOptimized)
reactivePackerWrapper      → packer.ts (packContextReactive)
memoryStoreIntegration     → contextAssembler.ts (assemblePipelineContextWithMemory)
contextMiddleware          → pipelineChat.ts (singleton)
agentSearchIntegration     → server/routes/agents.ts (GET /search)
worktreeManager            → server/services/teamRunner.ts (prepareAgentWorktree)

Testing

• ~69 unit tests + 2 E2E test files
• npx vitest run to verify locally

Key design choice

All integrations use adapter/wrapper pattern — zero modifications to existing function signatures, zero regression risk. Each adapter can be adopted incrementally.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 255ed2ab3d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Prevent truncated marker from increasing content size

The new truncation path appends a fixed "[... truncated by reactive compaction]" suffix without accounting for its length, so short files can become longer after “compaction” (for example, truncating a 5-character string to 1 character still yields a ~40-character result). In that case newTokens is reduced while the actual emitted content grows, which can silently violate token-budget assumptions and overflow downstream context windows.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Serialize agent hash fields without delimiter collisions

Building the cache key with [...].join('|') is ambiguous for free-form strings, so different agent metadata can collapse to the same hash whenever a field contains | (e.g., descriptions or tags), and the service will incorrectly reuse a stale _searchInstance instead of rebuilding the index. Using structured serialization per agent (rather than delimiter-joined text) avoids these false cache hits.

Useful? React with 👍 / 👎.