dogma: Skill for Claude Code

Corpus Sweep

Governing Axiom

This skill enacts the Endogenous-First axiom from MANIFESTO.md and is governed by AGENTS.md § Programmatic-First Principle. Before sweeping any corpus, check whether a prior sweep result is already cached in .cache/sources/ or recorded in the active session scratchpad — re-sweeping a cached corpus wastes tokens and violates Local Compute-First.

"Scaffold from existing system knowledge and external best practices." — MANIFESTO.md § 1 Endogenous-First

---

When to Use

• A synthesis or implementation phase requires patterns found across ≥20 source documents.
• The Orchestrator needs a cross-cutting summary before delegating to a Synthesizer.
• A research epic requires an initial landscape report before forming hypotheses.
• You are auditing agent files, research docs, or session scratchpads for a recurring pattern.

Do not use this skill for:

• Single-document reads (use read_file directly).
• Live web fetches (use the source-caching skill instead).
• Tasks where the required context fits within a single grep_search or semantic_search call.

---

Workflow

1. Pre-warm cache — before reading any source file, run:

   uv run python scripts/fetch_all_sources.py
   

   Skip if .cache/sources/manifest.json was updated within the current session.

2. Read source index — identify the target corpus:

   • For research sources: read .cache/sources/manifest.json to list all cached URLs.
   • For internal docs: list docs/research/, docs/guides/, or .github/agents/ as appropriate.
   • Record the total document count before batching.

3. Batch-read docs — divide the corpus into batches of 5–10 files. For each batch:

   • Use read_file (for long docs) or grep_search (for keyword scanning).
   • Record findings per document: filename, key claim, relevant quote (≤50 words), and cross-references noted.
   • Stop scanning a document once sufficient signal is found — avoid reading full content when a heading scan suffices.