Skip to content

CLAUDE.md

Latest commit

 

History

History
80 lines (61 loc) · 3.78 KB

CLAUDE.md

File metadata and controls

80 lines (61 loc) · 3.78 KB
Raw

Claude Code Context

This file is the entrypoint for Claude Code working in this repository.

Default Skill Policy

Claude Code must use:

  • .agent/ for project-neutral process skills;
  • context/ for langgraph-cpp conventions and foundation / core / langgraph library skills.

Minimal Load Policy (cost control)

Prefer constraining how much you read over deleting skills from the pack.

  1. Smallest skill set only — one routing row; do not preload all of context/skills/.
  2. Authority first — skim context/AUTHORITY.md and relevant context/CONVENTIONS.md sections before specialty skills (stream / subgraph / HITL / provider-http / performance).
  3. Simple-change budget — focused edits: ≤2 context skills and ≤2 .agent skills (plus definition-of-done at handoff). Do not open five-plus context skills by default.
  4. Progressive disclosure — open a specialty skill only when the change actually hits that topic.
  5. Never full-pack reads — inventory tables are for lookup, not mandatory reading.

At the start of a task

  1. Read .agent/README.md only as needed to pick process skills.
  2. Read context/AGENTS.md (or context/README.md) for routing — not both end-to-end.
  3. Read context/AUTHORITY.md when contracts/schemas/conflicts apply; otherwise defer.
  4. Classify the task; pick the single best routing row.
  5. Read that minimal .agent/skills/*.md + context/skills/*.md set before editing.
  6. Use .agent/skills/definition-of-done.md before final handoff.
  7. When changing context/ or layering, run scripts/check-context-skills.sh and/or scripts/check-dependency-policy.sh.

Project Context

Before editing code, read:

  1. README.md for user-facing scope, build commands, and examples.
  2. docs/AI_INDEX.md for the current design evidence chain.
  3. docs/PRD.md for product and MVP requirements.
  4. docs/ROADMAP.md and docs/internal/WBS.md for staged implementation intent.
  5. docs/LIMITATIONS.md for current boundaries and deferred work.
  6. AGENTS.md for the full repository map, build/test commands, and non-negotiable rules.
  7. For foundation/core/langgraph edits: relevant sections of context/CONVENTIONS.md, plus context/AUTHORITY.md / context/STACK.md only when contracts or build options are in scope.

Skill Routing

Use .agent process skills by task type (see AGENTS.md for the full table). Additionally load these context library skills when the task touches the corresponding layer:

Task Context skills
src/foundation/** context/skills/foundation.md, context/skills/coding-standards.md
src/core/** / RuntimeServices / RuntimeContainer context/skills/core.md, context/skills/coding-standards.md
src/langgraph/** context/skills/langgraph.md, context/skills/coding-standards.md
Stream / projection context/skills/stream-projection.md
Subgraph / namespace context/skills/subgraph.md
Interrupt / HITL context/skills/hitl-interrupt.md
Provider / HTTP context/skills/provider-http.md
Graph concurrency / executors context/skills/concurrency.md
Checkpoint / store / storage context/skills/persistence.md
Tools / auth / redaction / edge context/skills/security.md
Tests / examples context/skills/testing-examples.md
Performance / hot paths context/skills/performance.md

Repository Rules

The rules in AGENTS.md are authoritative for Claude Code:

  • project identity and scope;
  • repository map;
  • build and test commands;
  • non-negotiable C++ runtime rules;
  • definition of done;
  • dual-layer skill loading with context/.

Conflict resolution for design facts follows context/AUTHORITY.md. If this file conflicts with AGENTS.md on process/routing, follow AGENTS.md.