sds-mode/code-to-docs-skill/code-to-docs-digest

Overview

Load structured context from an existing code-to-docs vault. Read-only, never writes files.

Related Skills

Invocation

Skill(skill: "code-to-docs:code-to-docs-digest", args: "<vault-path> [--scope <module,...>] [--focus issues|architecture|all]")

• <vault-path> — path to existing code-to-docs vault (required)
• --scope — comma-separated module names to load in full (default: overview only)
• --focus — architecture (default), issues, or all

Model

Haiku — read-and-present task. No analysis or generation required.

Execution

Step 1: Validate Vault

1. Check that <vault-path>/_state/analysis.json exists. If missing, stop and report:

> No code-to-docs vault found at <vault-path>. Run /code-to-docs:code-to-docs to generate one first.

1. Validate _state/analysis.json against the schema in ../code-to-docs-references/output-structure.md "State File Validation" section (required fields and types). If the file exists but is malformed or missing required fields, abort with a specific error naming the problem — do not partially load or guess. Suggest regenerating with /code-to-docs:code-to-docs or /code-to-docs:code-to-docs-update.

An absent schema_version or module_index is not a validation failure — it marks a v1 vault, which digest reads fine via the fallbacks noted below. Digest is read-only, so it never migrates a v1 vault; that is code-to-docs:code-to-docs-update's job.

Graceful degradation (Steps 2–4): if any other referenced file (System Overview, Dependency Map, System Map, _state/synthesis.md, or a Health/ file) is missing — as happens with partial or older vaults — note its absence in the summary and continue with the files that do exist. Only a missing or invalid _state/analysis.json is fatal.

Step 2: Load Baseline

Read:

• _state/analysis.json — module inventory, dependency graph, timestamps
• Architecture/System Overview.md — architecture narrative and diagrams

Step 3: Load Focus Content

Step 4: Load Scoped Modules

• For modules listed in --scope: load the full Modules/{Name}.md. This is the reader-facing doc, which is what a human catching up actually wants.
• For all other modules: load only a light overview, using the cheapest source available:

Prefer the v2 path whenever module_index is present: the whole non-scoped inventory comes free from a file Step 2 already read, replacing N partial document reads with zero. Do not load the whole ## Beginner section (prerequisites, key concepts, walkthrough, example) for non-scoped modules on either path — it is far larger than an overview and blows the token budget.

Step 5: Present Context Summary

Output a structured summary including:

• Architecture overview (from System Overview)
• Module inventory with status
• Focus-specific details (architecture maps or health issues)
• Scoped module deep-dives

Token Budget

These are soft, best-effort targets — the skill has no token counter, so treat them as guidance and use the line-count proxies below rather than exact measurement (rough rule: ~4 characters ≈ 1 token).

When in doubt, prefer the shorter summary — the user can always pull more in with --scope.

Red Flags

Do NOT do any of the following:

• Writing any files — this skill is strictly read-only. That includes migrating a v1 state file: read it as-is via the fallbacks and leave it alone
• Running generation or analysis — if the user wants updates, suggest /code-to-docs:code-to-docs-update instead
• Loading all modules without `--scope` or `--focus all` — only load overviews by default to stay within token budget
• Reading every `Modules/{Name}.md` to build the module inventory when module_index is present — it already carries every module's purpose and complexity, at no extra read
• Reading `_state/modules/*.md` reports — those are analysis internals for generate/update. Digest presents the reader-facing vault
• Using a non-Haiku model — this is a simple read-and-present task; Haiku is sufficient