Update Source Map

Overview

Build (or refresh) a structured index of a project directory. The skill always produces a Markdown SOURCE_MAP.md (primary, human-readable) and a JSON source_map.json (programmatic). A hand-curated curated_summaries.json is preserved across regenerations so curated per-file knowledge compounds over time.

When this skill runs

Trigger on requests like: - "build a source map for this project" - "update the source map" - "create a project index" - any task where Codex is about to do work in a workspace it doesn't already know well

Do not trigger for: code refactoring, single-file edits, paper reading, or tasks unrelated to project navigation.

Workflow

Step 1 — Decide mode (auto-detect)

Run scripts/detect_existing.py <workspace_root> to decide between create and update mode. Detection priority:

1. <workspace>/x_temp/SOURCE_MAP.md + x_temp/source_map.json → update
2. <workspace>/SOURCE_MAP.md + source_map.json → update
3. <workspace>/.codex/source_map.json → update
4. Otherwise → create

The detection JSON has fields: mode, md_path, json_path, found_at. Pass --force-mode create|update to override.

If detection returns partial: true (only one of the two files exists), warn the user before proceeding.

Step 2 — Confirm the Run Plan

Before any execution, report the plan:

• Which mode (create / update) and why
• Which workspace to scan
• Where the outputs will go (default: <workspace>/x_temp/)
• Any custom exclusions or extensions
• Estimated file count from a quick find (to set expectations)

If the user explicitly asked you to build or update the source map, proceed after stating the plan. If the user only asked what would happen, stop after the plan.

Step 3 — Setup

Create the output directory if it doesn't exist:

mkdir -p <workspace>/x_temp

For update mode with diff support, also keep the previous inventory:

mv <workspace>/x_temp/file_inventory.jsonl <workspace>/x_temp/file_inventory.prev.jsonl

(only if regeneration would otherwise lose it)

Step 4 — Run the pipeline

Two equivalent options:

Option A — one-shot CLI (recommended for simple cases):

python3 scripts/regenerate.py <workspace_root> \
    [--output-dir x_temp] \
    [--force-mode create|update] \
    [--exclude "extra_dir1,extra_dir2"] \
    [--keep-prev-inventory]

Option B — modular (use when intermediate inspection is useful):

# 1. Scan
python3 scripts/scan_workspace.py <workspace_root> \
    --output <workspace>/x_temp/file_inventory.jsonl \
    [--exclude "..."]

# 2. Build
python3 scripts/build_source_map.py \
    --inventory <workspace>/x_temp/file_inventory.jsonl \
    --md-out <workspace>/x_temp/SOURCE_MAP.md \
    --json-out <workspace>/x_temp/source_map.json \
    --curated-out <workspace>/x_temp/curated_summaries.json \
    --mode create|update \
    [--existing-curated <workspace>/x_temp/curated_summaries.json] \
    [--prev-inventory <workspace>/x_temp/file_inventory.prev.jsonl]

Step 5 — Verify (do not skip)

After building, always check the following. Each check has a known failure mode from real past runs:

1. JSON validity: python3 -c "import json; json.load(open('<json_path>'))" — must succeed
2. File count consistency: MD + PDF count in source_map.json["totals"] must equal len(files) — catches partial writes
3. No type-mismatch in tables: spot-check 1 row of the Markdown — r["type"] should appear in the type column (known bug: hardcoded "PDF" leaked into MD rows in earlier versions)
4. No root-file-as-folder bug: top-level rollup table should only contain real folders, not files at workspace root
5. Curated summaries not lost: if update mode, curated_summaries.json count should be ≥ previous count minus deleted files
6. No notes.assets/ leakage: find <workspace>/x_temp -name "*.png" -o -name "*.jpg" should return nothing
7. Original files untouched: find <workspace> -newer <workspace>/x_temp/SOURCE_MAP.md -type f should not return anything (or only items inside x_temp/)

If any check fails, fix the source (build_source_map.py) and re-run. Do not patch the output Markdown by hand.

Step 6 — Hand off

Tell the user: - Mode used (create / update) - Number of files indexed - Where the artifacts live - Any diff highlights (in update mode) - Whether the curated_summaries.json was preserved / extended

Append the standard completion audit block (see "Completion audit" below).

Quick reference — script purposes

Reference docs (load on demand)

• references/format-spec.md — exact schema of SOURCE_MAP.md, source_map.json, curated_summaries.json. Read when you need to know the output structure precisely.
• references/create-vs-update.md — the auto-detection rules, when to force-mode, and how curated summaries survive updates. Read when mode decision is unclear.
• references/exclusions-default.md — what's filtered out by default, how to add custom exclusions, and the "no modify" file boundary. Read before scanning an unfamiliar project.

Boundaries (non-negotiable)

• ❌ Never modify files outside the output directory (default: <workspace>/x_temp/)
• ❌ Never read PDF contents — metadata only
• ❌ Never index image assets (*.assets/, *.images/, etc.)
• ❌ Never touch files with names containing no modify, no-edit, readonly, do not touch
• ❌ Never overwrite curated_summaries.json entries that were curated by hand
• ❌ Never reformat a regeneration silently — if file count changes, surface the diff to the user

Completion audit

At the end of every run, append:

Completion audit:
- Mode: create | update
- Files indexed: <N> (MD: <M>, PDF: <P>)
- Outputs: <md_path>, <json_path>, <curated_path>
- Curated summaries preserved: yes (N entries) | no (create mode)
- Diff: <N> added, <M> removed (update mode only)
- Original files touched: no
- All 7 verify checks passed: yes | <failed checks>

Lessons embedded in this skill

The skill encodes mistakes from a prior real-world run on a 43-file / 66MB personal knowledge base:

1. Type-mismatch bug: hardcoding "PDF" in a loop over a folder that contains a README.md — fixed by using r["type"] everywhere
2. Root-file-as-folder bug: top-level rollup loop treated root-level files as if they were subdirectories — fixed by skipping len(parts) == 1 in compute_top_rollups
3. Double-slash in tree root: render_tree passed both name and a trailing slash — fixed by single source of truth in tree_root
4. Lost curated knowledge: if curated summaries were stored in the build script, they were lost on every regeneration — fixed by extracting to curated_summaries.json
5. PDF binary read attempts: avoid read_text() on PDFs; only inspect metadata

When you add new scripts, check the verify checklist in Step 5 for additional bugs to add.