Code Module Summaries

Overview

Build and maintain evidence-backed Markdown summaries for an arbitrary Git repository. Treat each module's last successful commit SHA as its incremental checkpoint. Timestamps and Markdown mtimes are display metadata only.

The tracked control directory is the source of truth:

• registry.json: stable module boundaries, layout, and summary paths.
• state/<module-id>.json: one independent checkpoint per module.
• modules/*.md: summaries when using the centralized layout.
• INDEX.md: links and module ownership; update only when mappings change.

Invocation

Interpret user arguments with this contract:

/code-module-summaries [repo]
  [--output <control-dir>]
  [--layout centralized|colocated|custom]
  [--update]
  [--module <id-or-exact-name>]...
  [--target <git-ref>]
  [--refresh-modules]
  [--summary-name <filename.md>]
  [--head-only]
  [--accept-summary-drift]
  [--adopt-existing-summary]
  [--rebaseline --reason <text>]

Defaults:

• repo: current Git root.
• --output: code-summaries/.
• --layout: centralized.
• --target: the registry's tracked_ref, normally HEAD.
• --summary-name: MODULE_SUMMARY.md for colocated summaries.

--layout custom places each module's Markdown at an arbitrary in-repository .md path declared per module in registry.json (summary_path), with no fixed directory shape. This is the layout for a shared conclusions/-style tree, for example enterprise/conclusions/apps/gateway_conclusion.md. A custom registry may also declare a top-level index_path (e.g. enterprise/conclusions/README.md): a managed overview file the agent maintains, excluded from module source diffs and unassigned-path checks. See REFERENCE.md §9 for exact rules. centralized and colocated are unchanged.

--module is repeatable and case-sensitive. It selects exact IDs first, then an exact unique name. Never use fuzzy matching for a write operation.

Required Reference and Helper

Before the first scan, mapping refresh, deletion, or history recovery, read REFERENCE.md completely.

Use the deterministic helper for every update:

python <skill-dir>/scripts/scan_changes.py --help

Do not replace it with git log --since, file mtimes, or an improvised diff.

Workflow

1. Validate scope

1. Resolve the Git root and freeze the full target commit OID.
2. Confirm the control directory is inside the repository and intended to be tracked by Git. It must not be the repository root or an ancestor of any module root.
3. Default to committed Git objects. If selected-module source files are dirty, stop. --head-only may explicitly ignore those worktree changes.
4. Never run fetch, pull, stash, reset, checkout, clean, or rebase implicitly.

2. First scan

When registry.json does not exist:

1. Discover real module boundaries from workspace/package manifests, runtime entry points, deploy units, ownership, tests, and import cohesion.
2. Aim for a useful map, not exactly ten modules. Do not split or merge strong package boundaries merely to hit a count.
3. Ensure every tracked source path has an owner. When needed, use one repository-root module with root .; deeper module roots take precedence. A new package first appears there and triggers a mapping refresh.
4. Present the candidate map and output paths before writing unless the user already supplied explicit module roots.
5. Create registry.json using the schema in REFERENCE.md.
6. Run plan before creating summaries and retain each returned checkpoint_token. If a summary already exists without state, read and preserve it before explicitly using --adopt-existing-summary.
7. Generate each summary from code at the frozen target. Cite real paths and symbols; do not turn plans or README claims into implemented behavior.
8. Run checkpoint separately for each successfully verified module. A failed module must not advance its state.

3. Incremental update

Run a read-only plan first:

python <skill-dir>/scripts/scan_changes.py plan \
  --repo <repo> \
  --control-dir <output> \
  [--module <selector>]... \
  [--target <ref>] \
  [--head-only] \
  [--accept-summary-drift] \
  [--adopt-existing-summary]

Exit code 2 or has_blockers: true means zero summary writes until the reported blocker is resolved. For a multi-module plan with one blocked module, either resolve all blockers and rerun, or run a new plan selecting only the unblocked module; never partially execute the blocker-containing plan.

A full plan also blocks on tracked paths not owned by any module or excluded by the registry. This catches newly added top-level packages instead of silently ignoring them.

Use --accept-summary-drift only after an earlier plan actually reported SUMMARY_DRIFT and the existing edits were reviewed. It cannot pre-authorize a future summary change.

Handle each selected module by plan status:

• new: read the whole declared module scope and create its summary.
• changed: inspect every reported A/M/D/R path, including both sides of a rename, then read only the surrounding code and dependencies needed to explain the behavioral change. Apply the minimum accurate summary edit.
• unchanged: do not reread source and do not rewrite the Markdown; only checkpoint the target so the same range is not scanned again.
• deleted: do not delete or archive automatically. Confirm whether to retire, remap, or replace the module.
• blocked: stop for that module and follow the recovery table in REFERENCE.md.

After verifying one module's Markdown:

python <skill-dir>/scripts/scan_changes.py checkpoint \
  --repo <repo> \
  --control-dir <output> \
  --module <module-id> \
  --target <full-target-oid> \
  --target-ref <same-ref-used-by-plan> \
  --plan-token <token-returned-for-this-module> \
  --summary-sha256-at-plan <hash-returned-for-this-module> \
  [--head-only] \
  [--accept-summary-drift] \
  [--adopt-existing-summary] \
  [--summary-unchanged]

Pass the same plan options to checkpoint. Use --summary-unchanged only after inspecting every reported change and confirming that none changes the maintainer-facing summary.

Checkpoint successful modules independently. This is what lets one long-stale module retain its own baseline while another is updated frequently.

4. Strict single-module mode

With --module:

• Do not rediscover all modules.
• Do not modify another module's summary or state.
• Do not update INDEX.md or registry.json.
• Treat a missing root, new package boundary, split, merge, or ambiguous rename as mapping drift; require --refresh-modules.
• Before finishing, verify the write set contains only the selected summaries and state/<selected-id>.json.

5. Mapping refresh

--refresh-modules is the only normal operation allowed to change module roots, IDs, summary paths, or layout. Produce a before/after mapping and ask before applying ambiguous splits, merges, or relocations. Retire old IDs; never silently reuse them for unrelated code.

After an approved mapping edit, run plan --mapping-refresh, rebuild the affected summary, then checkpoint with --mapping-refresh --reason <text> and the returned token/hash. Mapping-refresh plans also validate whole-repository ownership. Never delete state to bypass a revision mismatch.

After verified history rewriting, run plan --rebaseline, fully review the module at the new target, then checkpoint with --rebaseline --reason <text> and the returned token/hash. Rebaseline is not a date-based diff.

--refresh-modules, layout selection, and summary generation are Skill-level operations. The helper deliberately implements only deterministic plan and checkpoint; it does not guess repository architecture.

6. Batch subagent orchestration

For a whole-repository or large-scope initial generation (many new/changed modules), generate summaries in parallel. The helper never dispatches agents; you orchestrate them using the dispatching-parallel-agents skill. The per-module summary_path isolation is what makes this safe — each subagent writes exactly one file and shares no state.

6.1 Subagent model selection (mandatory, cost-first)

Never silently inherit the parent chat's model for batch subagents. Conclusion writing is mostly structured code reading + Markdown drafting; a top-tier parent model (e.g. Opus) is usually wasteful when multiplied across dozens of modules.

Before the first Task dispatch in a batch run:

1. Ask the user which model slug to use for the subagents (the model parameter on Task). Present a short recommendation table; do not invent slugs outside the currently available Task model list.
2. Wait for an explicit choice (or an explicit "use your recommended default"). Do not start parallel generation until confirmed.
3. Pass that slug as model on every summary-writing subagent in the run. Parent-side plan / checkpoint / registry edits stay on the parent model; only the per-module Markdown writers use the chosen cheap model.
4. Optional: allow a two-tier split if the user wants it — cheap default for stubs/small packages, a mid-tier model only for a few heavy modules the user names. Do not up-tier the whole batch without asking.

Recommended defaults (prefer the cheapest tier that still produces accurate maintainer docs; update names when the Task model list changes):

If the user declines to pick, use composer-2.5-fast for the whole batch and state that choice in the final report. Re-ask before switching tiers mid-run.

6.2 Dispatch steps

1. As the parent, run plan (full or a selected set) at a single frozen target_commit. Confirm zero blockers, then record each module's checkpoint_token, summary_sha256_at_plan, and summary_path.
2. Complete §6.1 (user-chosen or confirmed-default subagent model) before any parallel Task calls.
3. For every new/changed module, dispatch one subagent with that model and a self-contained prompt that includes: the module roots, the frozen target_commit (instruct it to read exact versions via git show <TARGET>:path), the exact output summary_path, the summary template (REFERENCE §7), the repository's existing conclusion style, and a hard boundary: write only its own summary file — never touch other summaries, registry.json, state/, or the index_path overview.
4. When all subagents return, the parent runs checkpoint for each module with that module's own token and hash. A failed or unverified module is left without a checkpoint so the next run still reports it as new/changed.
5. unchanged modules get no subagent; the parent checkpoints them directly (with --summary-unchanged only after review) or skips them.
6. Keep the overview/index file (INDEX.md, or the custom index_path README) as a final parent-authored step after the module summaries settle, never inside a module subagent.

Accuracy Invariants

• Persist full commit OIDs, never abbreviated SHAs.
• Compare Git trees from BASE to frozen TARGET; commit dates do not define the range.
• Require BASE to exist and be an ancestor of TARGET.
• Record one baseline per module; a repository-global baseline is insufficient for selective updates.
• Keep centralized and colocated layouts on the same tracked control plane.
• Exclude summaries, state, generated files, dependencies, build output, and caches from module source inputs.
• Preserve manual summary edits. --accept-summary-drift permits a merge only after those edits have been read and retained.
• A normal checkpoint requires the exact token from a blocker-free plan. Never invent, reuse, or omit it.
• Tokens bind the module state generation, target ref, target OID, mapping, source fingerprint, plan-time summary hash, and safety options; they are single-use.
• Never claim an update is complete before the summary and its checkpoint both succeed.

Summary Quality

Preserve each existing document's structure and tone. A summary should help a maintainer use, change, or extend the module:

• responsibility and explicit non-responsibilities;
• entry points, public interfaces, and core execution path;
• important classes/functions and data/config contracts;
• upstream/downstream dependencies;
• tests and operational boundaries;
• unresolved facts clearly marked as unverified.

Remove descriptions of deleted or renamed implementation. Do not add changelog noise unless the repository explicitly uses summaries as changelogs.

Final Response

Report:

1. frozen target commit;
2. selected modules and each module's previous checkpoint;
3. A/M/D/R evidence grouped by module;
4. summaries and state files written;
5. skipped unchanged modules;
6. blockers or mapping decisions still requiring confirmation.

Common Mistakes

• Using “ten days ago” or summary mtime as the baseline.
• Advancing one global checkpoint after updating only one module.
• Reading every module before checking the Git plan.
• Losing the old side of a rename or treating deletion as an addition.
• Rebuilding all summaries because one module changed.
• Overwriting manually edited Markdown without explicit acceptance.
• Guessing after force-push, shallow history, module split, or missing state.
• Dispatching batch summary subagents on the parent’s expensive model (e.g. Opus) without asking — always complete §6.1 first; default writers to composer-2.5-fast unless the user picks another slug.