rules-to-hook（挂钩规则）

Hook Context Rules

Author and maintain .claude/context-rules.json for the context-inject.mjs hook. The engine supports both Claude Code and VS Code Copilot hook payload formats.

VS Code Copilot support

The engine auto-detects payloads from VS Code Copilot (camelCase fields, toolArgs as JSON string). Key differences from Claude Code:

• Event name: VS Code doesn't include the event in the payload. Pass it as a CLI argument: node.claude/hooks/context-inject.mjs preToolUse. The engine normalizes camelCase (preToolUse) to PascalCase (PreToolUse) used in rules.
• Tool input: VS Code sends toolArgs (JSON string), not tool_input (object). The engine parses it automatically.
• Output: VS Code only supports permissionDecision on preToolUse — flat JSON, no hookSpecificOutput wrapper. Context injection (text, hint, learnings) has no effect on VS Code.
• Supported inject types on VS Code: only block and allow (PreToolUse).

VS Code hook configuration

{
  "hooks": {
    "preToolUse": [{
      "type": "command",
      "bash": "node .claude/hooks/context-inject.mjs preToolUse"
    }],
    "postToolUse": [{
      "type": "command",
      "bash": "node .claude/hooks/context-inject.mjs postToolUse"
    }]
  }
}

Config Format

Rules are a JSON array. Each rule has an on event, a when filter (all keys optional, implicit AND), and an inject payload (exactly one key).

[
  {
    "on": "PreToolUse",
    "when": {
      "tool": "Write|Edit|MultiEdit|replace_string_in_file|create_file|multi_replace_string_in_file",
      "path": "src/core/**"
    },
    "inject": {
      "text": "core/ is wiring only — no business logic here."
    }
  },
  {
    "on": "PostToolUse",
    "when": { "tool": "Read|read_file", "path": "src/stages/**" },
    "inject": { "hint": "docs/stages-overview.md" }
  }

]

Block and allow rules control tool execution (PreToolUse only):

[
  {
    "on": "PreToolUse",
    "when": { "tool": "Bash", "command": "rm -rf" },
    "inject": { "block": "Destructive rm -rf is not allowed." }
  },
  {
    "on": "PreToolUse",
    "when": { "tool": "Read", "path": "docs/**" },
    "inject": { "allow": "Safe: reading documentation." }
  }
]

Precedence: block > allow > context. If any matched rule is a block, the tool call is denied even if other rules allow or inject context.

on values

PreToolUse / PostToolUse / UserPromptSubmit / SessionStart / SubagentStart / PostToolUseFailure / Stop / PreCompact

The installer registers hooks for all 8 events automatically.

when keys (all optional, implicit AND)

Platform Tool Names

Claude Code and VS Code Copilot use different tool names. Always include both sets in tool patterns to ensure rules fire on both platforms.

For write rules: Write|Edit|MultiEdit|replace_string_in_file|create_file|multi_replace_string_in_file For read rules: Read|read_file

inject keys (exactly one)

Block behavior per event

Stop safety guard

When a Stop hook blocks (meaning "don't stop, continue"), the agent retries stopping with stop_hook_active: true in the payload. The engine detects this flag and exits silently (process.exit(0)) before matching rules, allowing the agent to stop and preventing infinite loops.

Learnings System

Agents can record file-bound learnings that are injected as context when other agents interact with those files.

Storage

Learnings are stored in .claude/learnings.json:

[
  {
    "files": ["src/stages/**"],
    "learning": "Stage transforms must use AST nodes, not string manipulation",
    "timestamp": "2026-03-04T10:30:00Z"
  }
]

Recording learnings (CLI)

# Add a learning
node .claude/hooks/learn.mjs add --files 'src/stages/**' --learning 'Use AST nodes only'

# Add for multiple file patterns
node .claude/hooks/learn.mjs add --files 'src/core/**,src/stages/**' --learning 'Cross-cutting concern'

# List all learnings
node .claude/hooks/learn.mjs list

# List learnings matching a file
node .claude/hooks/learn.mjs list --files 'src/stages/elision.ts'

# Remove by index
node .claude/hooks/learn.mjs remove --index 0

# Update by index
node .claude/hooks/learn.mjs update --index 0 --learning 'Updated insight'

# Update file globs after rename/move
node .claude/hooks/learn.mjs repath --index 0 --files 'src/new-path/**'

# Find orphaned learnings (globs that match no files)
node .claude/hooks/learn.mjs check

When adding, the CLI shows overlapping existing learnings so you can curate (merge, update, or remove redundant entries).

When renaming or deleting files, run check to find orphaned learnings, then repath or remove to fix them.

Injection

The learnings inject type in a context rule triggers lookup:

{
  "on": "PostToolUse",
  "when": { "tool": "Read|read_file", "path": "**" },
  "inject": { "learnings": true }
}

When matched, the engine reads .claude/learnings.json, finds entries whose files globs match the current file path, and injects them as additionalContext:

[Learnings for src/stages/elision.ts]
- Stage transforms must use AST nodes, not string manipulation
- Elision uses line-count thresholds

Prompt reminders

A UserPromptSubmit context rule reminds agents to record learnings on every prompt. It has no when filter, so it fires unconditionally. This rule is installed automatically by the installer.

Phase 0 — Bootstrap

Runs before any other phase, every time the skill is invoked.

Detection

Check whether the hook engine is installed:

1. .claude/hooks/context-inject.mjs exists
2. .claude/hooks/platform.mjs exists
3. .claude/hooks/learn.mjs exists
4. .claude/hooks/harness-eval.mjs exists
5. .claude/hooks/harness-format.mjs exists
6. .claude/settings.json has context-inject.mjs registered in all 8 events (PreToolUse, PostToolUse, UserPromptSubmit, SessionStart, SubagentStart, PostToolUseFailure, Stop, PreCompact)
7. .claude/hooks/node_modules/minimatch exists (dependency installed)
8. .claude/learnings.json exists

If all eight pass → skip to Phase 1.

If any is missing → run the installer:

node skills/rules-to-hook/install.mjs

The installer is idempotent. It:

• Creates .claude/hooks/ and copies the engine from skills/rules-to-hook/engine.mjs
• Copies the platform module from skills/rules-to-hook/platform.mjs
• Copies the learn CLI from skills/rules-to-hook/learn.mjs
• Copies harness modules from skills/rules-to-hook/harness-eval.mjs and skills/rules-to-hook/harness-format.mjs
• Ensures package.json has minimatch and installs dependencies
• Merges hook registrations into .claude/settings.json (all 8 events, .* matcher)
• Seeds an empty .claude/context-rules.json if absent
• Seeds an empty .claude/learnings.json if absent
• Seeds learnings context rules (injection on Read + prompt reminder) into context-rules.json

After the installer completes, confirm: "Hook engine installed. Proceeding to auto-discovery."

If the installer fails, stop and report the error — do not continue to later phases.

Phase 1 — Auto-discovery

Runs by default (and always first, even when invoked with add).

Discovery proceeds in five layers dispatched to subagents to keep the main context clean. Only structured summaries flow back. Do not propose rules until all layers complete.

Dispatch plan

Batch 1 (parallel):   Layer 1 · Layer 2 · Layer 3
Batch 2 (parallel):   Layer 3.5 · Layer 4  — both need Layer 3 output
Batch 3 (sequential): Layer 5  — needs Layer 3.5 + Layer 4 output

Create a tracked task for each batch (not each layer).

Batch 1 — Gather project data (3 parallel subagents)

Dispatch three general-purpose Task subagents in a single message. Each prompt must include the project root path and end with: "Return a structured summary only — not raw file contents."

Subagent → Layer 1: Project identity & toolchain

Prompt must instruct the subagent to:

1. Glob *.md at project root and read each file — extract purpose, guidelines, invariants.
2. Read the package manifest (package.json, pyproject.toml, Cargo.toml, or equivalent) — extract scripts, runtime, key deps.
3. Read config files (biome.json, .eslintrc*, tsconfig.json, prettier.config.*, etc.) — note codified conventions.

Expected return:

• Purpose (one sentence)
• Runtime & package manager
• Key dependencies (name → role)
• Codified conventions (list)

Subagent → Layer 2: Architecture map

Prompt must instruct the subagent to:

1. Run npx @anduril-code/ctx tree --depth 3 — identify source, doc, config, and test dirs.
2. Glob src/**/index.{ts,js} (or language equivalent) — read barrel files for module boundaries and public surfaces.
3. Check for types/, interfaces/, or similar dirs — confirm whether pure declarations (no runtime code, no cross-imports).

Expected return:

• Directory roles (table: dir → purpose)
• Module boundaries (what each barrel exports)
• Type-only directories (if any)

Subagent → Layer 3: Documentation deep-dive

Prompt must instruct the subagent to:

1. Run npx @anduril-code/ctx rank "architecture constraints invariants style" --maxResults 10. Read every result — do not guess from filenames.
2. Glob docs/**/*.md and read all remaining docs. For each doc: a. Use npx @anduril-code/ctx sections <file> to list headings. b. Use npx @anduril-code/ctx extract <file> --onlySections '<heading>' to read each section individually. c. For each section that contains actionable rules or constraints, produce a section record:

- section — the heading name - file — source doc path - content — condensed actionable rules (max 500 chars, drop examples/rationale) - keywords — 3-5 code-relevant terms (identifiers, directory names, patterns) that would appear in files this section governs

1. Skip sections that are purely informational (overview, changelog, credits).

Expected return:

• Array of section records (JSON), each with section, file, content, keywords

Wait for all three subagents. Collect their outputs.

Batch 2 — Path affinity + enforcement audit (2 parallel subagents)

Dispatch two general-purpose Task subagents in a single message. Paste the section records from Layer 3 into both prompts.

Subagent → Layer 3.5: Path affinity via codebase grep

Prompt must instruct the subagent to:

1. For each section record, Grep the codebase for each keyword.
2. For each keyword hit, note which top-level source directory contains it (e.g. src/api, src/components, tests).
3. Rank directories by hit density: hits for this section's keywords divided by total files in the directory.
4. Assign the top 1-3 directories as path globs (e.g. src/api/**).
5. If a section's keywords match across 5+ top-level directories with no clear winner (no directory has >30% of hits), assign src/** and flag as broad: true.
6. If a section's keywords produce zero hits, flag as noMatch: true.

Expected return:

• Enriched section records: original fields plus paths (string array of globs) and optional broad / noMatch flags

Subagent → Layer 4: Enforcement audit

Prompt must instruct the subagent to:

1. Read every file in .claude/hooks/ — source code, not just filenames. Understand what each hook enforces (blocks, rewrites, denies, injects).
2. Read .claude/context-rules.json if it exists — list every rule.
3. Cross-reference each section record from Layer 3 against existing hooks and rules. Mark each as:

- Enforced — already covered → skip - Unenforced — no coverage → candidate - Partial — partially covered → may need complementary rule

Expected return:

• Existing hooks summary (what each enforces)
• Existing rules list
• Section status table (section → Enforced / Unenforced / Partial)

Wait for both subagents. Collect their outputs.

Batch 3 — Verify & derive candidates (1 subagent)

Dispatch one general-purpose Task subagent. Paste the enriched section records from Layer 3.5 and the section status table from Layer 4 into the prompt.

Subagent → Layer 5: Verification & per-section rule generation

Prompt must instruct the subagent to:

1. Filter to unenforced / partial sections only (from Layer 4 status).
2. For every section that asserts a codebase property (e.g. "types/ is pure", "no imports from cli/"), use Grep to confirm it actually holds. Drop any section whose assertions do not hold.
3. For each verified section, emit one rule per path in its paths array: {"on": "PreToolUse", "when": {"tool": "Write|Edit|MultiEdit|replace_string_in_file|create_file|multi_replace_string_in_file", "path": "<glob>"}, "inject": {"text": "<Section Name>: <condensed content>"}}

- Prefix text with the section name for self-describing context. - Cap text at 500 characters. If longer, condense to actionable rules only (drop examples and rationale).

1. For each verified section, also emit one PostToolUse Read rule per path in its paths array: {"on": "PostToolUse", "when": {"tool": "Read|read_file", "path": "<glob>"}, "inject": {"hint": "<most relevant source doc path>"}}

- Use hint (not text) to keep Read injections lightweight. hint is prefixed with "Related: " — the agent decides whether to follow up. - The hint value should be the relative path to the most relevant source doc from the section's _source field. - Apply the same deduplication as step 5: if multiple sections target the same path glob, pick the single most relevant doc to hint rather than emitting multiple Read rules for the same path.

1. Deduplication: if two sections produce rules with the same path and overlapping content, merge them into one rule with combined text (separated by newline).
2. For sections flagged noMatch: true, include the candidate but annotate with (no matching code found) — the user will assign a path or skip.
3. For sections flagged broad: true, include the candidate but annotate with (broad — keywords matched everywhere).
4. Format each candidate as a JSON rule object.

Expected return:

• Verified sections (with Grep evidence)
• Candidate rules (JSON array, each annotated with source section, file, path, and any flags)

Wait for completion.

Presentation

Use the Layer 5 output to present a numbered table of candidate rules. Each row must include:

• Scope — the path glob this rule targets
• The rule (compact JSON or summary)
• Source — which doc/file and section established the constraint
• Verification — what Grep/check confirmed it
• Flags — broad or no match if applicable

#  on           scope              inject                          source
───────────────────────────────────────────────────────────────────────────
0  PreToolUse   src/api/**         text: "API Design: Use camel…"  docs/style.md §API Design
1  PostToolUse  src/api/**         hint: docs/style.md             docs/style.md §API Design
2  PreToolUse   src/components/**  text: "Components: Functional…" docs/style.md §Components
3  PostToolUse  src/components/**  hint: docs/style.md             docs/style.md §Components
4  PreToolUse   **/*.test.*        text: "Test Patterns: descri…"  docs/style.md §Test Patterns
5  PostToolUse  **/*.test.*        hint: docs/style.md             docs/style.md §Test Patterns
6  PreToolUse   src/**  (broad)    text: "Naming: Use snake_cas…"  docs/style.md §Naming
7  PostToolUse  src/**  (broad)    hint: docs/style.md             docs/style.md §Naming

Ask: "Accept all (a), pick by number (e.g. 1,3), or skip (s)?"

For rules flagged (no match), prompt: "No matching code found for ''. Assign a path glob or skip? (glob / s)"

Write accepted rules to .claude/context-rules.json (create if absent, merge if exists).

Phase 2 — User Additions

After auto-discovery (or when invoked with add), ask: "Anything else to add? (y/n)"

If yes, collect one answer at a time:

1. Event — PreToolUse / PostToolUse / UserPromptSubmit / SessionStart / SubagentStart / PostToolUseFailure / Stop
2. Trigger keys — show only valid keys for the chosen event:

- PreToolUse / PostToolUse: tool, path, command - UserPromptSubmit: prompt

1. Inject type — text / hint / block / allow

- For block: available on PreToolUse, PostToolUse, UserPromptSubmit, Stop - For allow: only available when event is PreToolUse

1. Inject content:

- For hint: list existing .md files nearby and suggest one

1. Preview the JSON rule and ask: "Add this rule? (y/n)"
2. On confirm, append to .claude/context-rules.json.

Repeat until user says no.

Phase 3 — List

When invoked with list, display current rules as an ASCII table:

#  on               when                          inject
─────────────────────────────────────────────────────────────────
0  PreToolUse       tool=Edit|Write path=src/**   text: "..."
1  PostToolUse      tool=Read path=src/stages/**  hint: docs/...
2  UserPromptSubmit prompt=test                   text: "..."

If .claude/context-rules.json does not exist, print: No rules found.

Phase 4 — Remove

When invoked with remove <index>:

1. Show the full rule at that index.
2. Ask: "Remove this rule? (y/n)"
3. On confirm, remove the entry and rewrite the file.

If the index is out of range, print: Index <n> is out of range (0–<max>).

Phase 5 — Smoke Test

Runs automatically after writing rules in Phase 1 or Phase 2. For each accepted rule, verify the hook engine actually fires by simulating a matching tool call.

Procedure

1. For each rule in .claude/context-rules.json, construct a minimal matching payload: {"hook_event_name": "<rule.on>", "tool_name": "<first alternative from rule.when.tool>", "tool_input": {"file_path": "<synthetic path matching rule.when.path>"}}

- For when.tool: use the first pipe-separated alternative (e.g. "Write" from "Write|Edit|MultiEdit|..."). - For when.path: expand the glob to a concrete path (e.g. "src/core/example.ts" for "src/core/**"). - For when.command: set tool_input.command to a string matching the regex. - For when.prompt (UserPromptSubmit): set "prompt" to a matching string instead of tool_name/tool_input.

1. Pipe the payload to the engine and capture stdout: echo '<payload>' | node.claude/hooks/context-inject.mjs
2. Parse the JSON output and check:

- context rules (text / hint): output must contain additionalContext with the expected value. - block rules: output must contain "permissionDecision": "deny". - allow rules: output must contain "permissionDecision": "allow". - no output (exit 0, empty stdout): rule did not fire — fail.

1. Also construct one non-matching payload per rule (wrong tool name or path outside the glob) and confirm the engine produces no output (empty stdout or exit 0). This guards against overly broad rules.

Reporting

Present results as:

Rule  on           scope          inject   result
──────────────────────────────────────────────────
0     PreToolUse   src/api/**     text     PASS ✓ (matched, context injected)
1     PostToolUse  src/api/**     hint     PASS ✓ (matched, hint injected)
2     PreToolUse   src/api/**     text     FAIL ✗ (no output — rule did not fire)

If any rule fails:

• Show the payload that was sent and the raw output received.
• Suggest likely causes: wrong tool name, path glob mismatch, or missing config file.
• Ask: "Fix the failing rules and re-test? (y/n)"

If all rules pass: "All rules verified — hooks are firing correctly."

Validation

Apply before writing any rule:

• when must have at least one key — reject rules with an empty when object. Exception: UserPromptSubmit rules may omit when entirely to fire on every prompt (the engine treats missing when as {}, matching unconditionally).
• path globs must be non-empty strings.
• hint values should exist on disk — warn if the file is not found, but do not block.
• inject must have exactly one key — reject if zero or more than one key is present.
• block is valid on PreToolUse, PostToolUse, UserPromptSubmit, and Stop — reject on other events.
• allow is only valid on PreToolUse — reject on other events with: "allow rules only work on PreToolUse events."
• source key is only meaningful on SessionStart; agent_type on SubagentStart; error on PostToolUseFailure.

The engine also enforces this validation at runtime. Invalid rules are ignored instead of being partially applied.

Notes

• Prefer hint over text for long or optional context — it costs fewer tokens.
• The hook fires on every matching tool call — keep when filters targeted.
• Phase 0 may create context-inject.mjs and modify settings.json — only during initial bootstrap.