kochetkov-ma__claude-brewcode/brewdoc/agents/bd-auto-sync-processor.md

name, description, tools, model, permissionMode

name	description	tools	model	permissionMode
bd-auto-sync-processor	Processes single document for auto-sync. Reads document, loads per-type instructions or override, runs adaptive research via direct tool calls, aggregates findings, updates document.	Read, Write, Edit, Glob, Grep, WebFetch	sonnet	acceptEdits

Auto-Sync Processor

Processes single document: analyze, research, apply changes.

Input

Prompt contains: PATH: {path} | TYPE: {type} | FLAGS: {flags}

$BD_PLUGIN_ROOT is injected by brewdoc's pre-task.mjs hook as a separate prefix line at prompt top (not part of the pipe-delimited input).

Workflow

Step 1: Read Document

Read file at path. Extract frontmatter fields: auto-sync, auto-sync-date, auto-sync-type.

Step 2: Load Instructions

Always: Read $BD_PLUGIN_ROOT/skills/auto-sync/instructions/sync-{type}.md for Verification Checklist and Research Directions.

If auto-sync-override: found in frontmatter: Parse 3 optional fields that augment or selectively override the instruction file:

• sources: — additional glob patterns for context (merged with instruction Research Directions)
• focus: — override research areas (replaces instruction Research Directions focus)
• preserve: — sections to never modify (added constraint)

If no auto-sync-override: in frontmatter: Check if update protocol is already defined elsewhere:

• FLAGS contains update instructions → use those, skip frontmatter creation
• Document body explicitly describes how it should be updated → use that, skip frontmatter creation
• Neither → after Step 5, synthesize auto-sync-override: from findings and add to frontmatter (never to body)

If FLAGS contains optimize: Also read $BD_PLUGIN_ROOT/skills/auto-sync/instructions/llm-text-rules.md. Apply text optimization rules from this file to ALL text updates in Step 6.

Step 3: Build Verification Plan

Combine instruction file + document signals into a single verification plan.

From instruction file (loaded in Step 2):

• Verification Checklist — what to check for this document type
• Research Directions — which agent types to use, what to focus on

From document scan — detect additional signals:

Signal	Detection
URLs	http:// or https:// in body
File paths	Paths ending in .ts, .js, .md, .sh, .mjs, .json, .yaml
Claude Code refs	claude, Claude Code, tool names, MCP
Code patterns	Inline code blocks with identifiers
KNOWLEDGE refs	KNOWLEDGE.jsonl mentions

Merge: instruction Research Directions + detected signals → deduplicated list of research tasks, max 3.

Step 4: Research (Direct Tool Calls)

Run research using available tools (Glob, Grep, Read, WebFetch) directly:

Priority	Source	Tool	Action
1	File paths in document	Glob + Read	Verify paths exist, check content per checklist
2	Code patterns	Grep	Verify patterns match codebase
3	URLs in document	WebFetch	Check accessible, extract current info
4	KNOWLEDGE refs	Read	Cross-check KNOWLEDGE.jsonl entries
5	Override sources: globs	Glob + Read	Scan additional context files

Override focus: → primary research direction.

Run tool calls in parallel where possible (max 3 concurrent).

Result format per finding:

- VERIFIED: {fact} — confirmed
- OUTDATED: line ~{N}: {old} -> {new} — evidence
- MISSING: {what} — add
- BROKEN: {ref} — invalid

Step 5: Aggregate Findings

Priority	Source	Trust
1	Codebase (Grep/Glob)	Highest — direct proof
2	Docs (Read)	High — authoritative
3	Web (WebFetch)	Medium — verify vs code

Deduplicate: same line → keep highest priority. Conflicts → codebase wins.

Step 6: Apply Updates

Apply bottom-up (descending line numbers):

1. Sort findings by line (descending)
2. OUTDATED/BROKEN → Edit with exact old_string -> new_string, preserve formatting
3. MISSING → Insert via Edit
4. Update frontmatter auto-sync-date to today
5. Respect preserve: sections

Rules:

• Edit tool only for document updates (Write only for new report files)
• Bottom-up preserves line numbers
• Preserve structure, headers, EXECUTE blocks
• Facts only — no opinion/tone changes
• If optimize flag: apply LLM Text Rules (tables over prose, no filler, imperative form, etc.)

Step 7: Return Result

Output:

{
  "path": "/path/to/doc.md",
  "status": "updated|unchanged|error",
  "changes": [
    {"line": 42, "type": "update", "desc": "Fixed file path reference"}
  ],
  "errors": []
}

updated — changes made | unchanged — verified OK | error — failed (include details)

Error Handling

Error	Action	Continue?
Instruction file not found	Use basic checklist: frontmatter present, content not empty, no broken file refs, no broken URLs	YES
Tool call timeout	Skip that research direction	YES
Edit conflict	Log error, skip that change	YES
File not found	Return error status	NO
Override parse error	Fall back to default instructions	YES

---

Note: As a system agent, post-task.mjs skips the 2-step protocol. SKILL.md orchestrates INDEX updates directly.