joogy06/agent-foundry/code-comprehension

code-comprehension (v1)

A cheap, standing, read-only generated comprehension front-door. Turns an arbitrary code repo into a generated, non-drifting PROJECT.md + docs/components/<id>/COMPONENT.md — built from a bounded synthetic component partition + structural code edges + per-component LLM intent — so the front-door stops being hand-maintained prose that drifts.

Status: PRODUCTION (D-A v1 ship). Design document: /path/to/project/docs/plans/2026-06-08-generated-front-door-design.md (read §12 → §13 → §11 → §1–§10). Classification: skill_text (Contract Map N/A). This skill is the consumer of synthetic contract-maps; it does not introduce a gated product surface.

What this skill is for

The forge→bob contract-driven pipeline (intent-extract, wiring-extract-static) only starts on a repo that already carries a signed progress/contract-map.yaml. Most real code repos (legacy apps, the user's other projects) have no durable full contract-map. This skill is the bounded standalone path that:

1. Synthesizes its own component partition (partition.py) — directory-primary + entry-point seeding + a hard cap, with bidirectional auto-resolution of over- and under-partition (no user pause). The synthetic map is UNSIGNED and written to .comprehension/ — never under `progress/` (collision hazard with a real signed map + poisoning a later gates.py G1).
2. Runs both extractors in a new `--standalone` mode where a bob claim is structurally impossible (no --claim-uuid, the heartbeat thread is never constructed, transition-request emission is unreachable). The run writes nothing under `.ledger/` → CB4-safe by absence.
3. Renders deterministic, zero-LLM docs (render_docs.py) — byte-identical on re-run given a fixed intent-map.

What this skill is NOT

• NOT an agent. It is a cheap standing pipeline invoked on demand. No design team, no challengers, no bob cycle.
• NOT a ledger writer. It touches nothing under .ledger/ or progress/. It cannot drive a real ledger transition (that is the CB4 kill-criterion, satisfied by absence).
• NOT SCIP-precise. v1 edges come from wiring-extract-static's Python AST + JS/TS regex generic fallback (+ FastAPI/Express framework-route plugins only where a repo uses them). Edges are labeled "structural (AST/regex)" with a confidence note. Real SCIP wiring is a v2 prerequisite.
• NOT a doc overwriter (by default). v1 generates to a shadow path for side-by-side review. --migrate (which overwrites real docs) is a separate, human-gated step and is off the v1 critical path.

CLI

code-comprehension build <repo>     # full pipeline → .comprehension/PROJECT.generated.md (shadow) + shadow COMPONENT.md
code-comprehension build <repo> --migrate   # HUMAN-GATED: archive + propose section-split + section-approval, then adopt

In practice the three scripts are invoked directly (this is a script skill, not an MCP tool):

python3 ~/.claude/skills/code-comprehension/scripts/comprehension_run.py --project-dir <repo> [--shadow]

The orchestrator (comprehension_run.py) owns the whole sequence; partition.py and render_docs.py are also independently runnable for testing.

Pipeline (v1)

comprehension_run.py --project-dir <repo>          (claimless, read-only — NOT an agent)
  │  owns .wiring/ + .comprehension/ creation (the non-bob "single creator")
  │
  1. wiring-extract-static --standalone --contract-map-path none
  │     → .comprehension/.wiring/runs/<id>/static.jsonl   (first pass: src_component = unmapped_path:*)
  │
  2. partition.py  ── the load-bearing piece (§4) ──
  │     read static.jsonl + file tree + entry-points (reused from static.jsonl, NOT re-detected)
  │     → bounded component partition (directory-primary + entry-point seeds, ≤CAP)
  │     → AUTO-RESOLVE (§13): cap→collapse-tail into misc; giant/over-budget→auto-split
  │       if clean sub-boundary else auto-degrade to structural-only; ratify→auto-write lock
  │     → .comprehension/synthetic-contract-map.yaml   (UNSIGNED; never under progress/)
  │     → .comprehension/partition.lock (recomputed + diffed each run, per C9)
  │     → .comprehension/partition-report.json (partition + per-component cost; post-hoc)
  │
  3. wiring-extract-static --standalone RE-RUN with the synthetic map → real component ids on edges
  │
  4. intent-extract --standalone --contract-map-path .comprehension/synthetic-contract-map.yaml
  │     (Python/TS: real intent; other langs / degraded: structural-only, no LLM, confidence badge)
  │     → .comprehension/.wiring/intent-cache/<sha>.yaml   (content-hash cached)
  │
  5. render-bundle (in orchestrator):
  │     merge per-component intent caches → one {"components":[...]} intent-map (in memory)
  │     reconciler.reconcile(static_edges, [], manifest, comp_ids, ...) → component-edge view (in memory)
  │
  6. render_docs.py  (deterministic, ZERO-LLM)
  │     intent-map + component-edge view + manifests
  │     → .comprehension/PROJECT.generated.md (SHADOW) + .comprehension/components/<id>/COMPONENT.md
  │     → FRESHNESS:v1 + generated_from[] stamp ; edges labeled "structural (AST/regex)"
  │     → strip sampled_at / model_id / run-id for determinism

CB4 boundary (the central guarantee)

code-comprehension writes only under the repo's .comprehension/ scratch directory (and, on --migrate, the real docs after human approval). It writes nothing under .ledger/ or progress/. The two extractors run in --standalone mode where:

• No claim is required (--claim-uuid is not accepted in standalone mode).
• The heartbeat thread is never constructed (not merely --no-heartbeat — the object is never created, so there is no path that could touch a claim file).
• Transition-request emission is unreachable (guarded at the call site by the standalone flag).
• The output root is configurable under .comprehension/.

HARD acceptance test: snapshot .ledger/ and progress/ before and after a full code-comprehension build --standalone run → byte-identical (zero new/changed files). See tests/test_cb4_boundary.py.

This makes the run CB4-safe by absence: it cannot drive a real ledger transition because it never touches .ledger/ at all. On an ecosystem repo that has a live .ledger/, the run still writes nothing there.

Determinism (honest scoping)

• The render is byte-deterministic GIVEN a fixed intent-map (sorted keys, sampled_at/model_id/run-id stripped, FRESHNESS keyed to content hashes). tests/test_render_determinism.py deletes the render and rebuilds it from a fixed intent-map and asserts byte-identity.
• The intent itself is NOT byte-stable across cold LLM regens (temp=0 ≠ byte-identical; the cache evicts after 30 days). So cache.py never evicts content-addressed entries referenced by a current partition.lock, and model_id is pinned into generated_from[]. We do not claim cold-LLM-regen determinism.

Multi-language

Python + TS/TSX get real intent. Other languages fall through wiring-extract-static's generic fallback → structural-only components (no LLM intent; confidence: structural-only badge). Prose is never faked for unsupported languages.

Generated-ownership (adoption prerequisite — WP-9)

Generated docs carry a banner:

<!-- GENERATED by code-comprehension — do not hand-edit; edits are erased on regen; narrative → ARCHITECTURE.md -->

plus a generated_from[] hash. A freshness check fails on manual edits (hash mismatch). Before any repo adopts a generated PROJECT.md, the writer contracts (exit-with-docs, project-documentation, forge, agent-teams) must be amended to detect the banner and redirect human updates to ARCHITECTURE.md/history.md. v1 ships the banner + guard; the writer-contract migration is a deferred adoption prerequisite, NOT on the v1 build/dogfood critical path (the dogfood is shadow-only and never adopts).

Files

Reuse (unchanged)

• wiring-extract-static core (AST/regex generic fallback + framework plugins).
• intent-extract core (per-component LLM pass + content-hash cache).
• wiring-reconcile's pure `reconciler.py` merge core only (NOT its stubbed file/promotion lifecycle).
• intent-map-render's d1_sequence.render / d4_heatmap.render as pure functions (NOT its CLI, which expects evo .ledger/evo/... paths + carries HARD-RULE-5 gating that does not apply here).
• The FRESHNESS:v1 machinery + the 4th SessionStart freshness hook (cheap hash-check + nudge; never eager-blocking).