cc-warehouse - architecture review board

share.py · relocate.py

Cost, confirmed still real. Zero in-process private-helper calls across 8 test files now (5 for relocate, 3 for share, including the newer --EXPOSED path). Every exercise of this logic is a run_ccw subprocess.

Solution direction. Promote thin public transforms - redact_payload(text, patterns), scan_secrets(text), rewrite_bytes(path, text, patterns) - so tests assert against the transform directly.

Tier: VERIFIED at 4824098 · Why 3rd (was 1st): unchanged Strong reasoning, but C12/C6 both now have a PROVEN live consequence this review; C3's cost is real but has not (yet) caught a live defect.

render.py · build.py · archive.py

flowchart LR
  A[one payload] --> M[render_markdown 1185-1199]
  A --> H[render_html 2187-2204]
  A --> B[build_manifest 2238-2295]
  A --> W[write_session_folder 603/652/663]
  M --> P1[build_conversation + parse_session]
  H --> P2[build_conversation + parse_session]
  B --> P3[build_conversation]
  W --> P4[parse_session x2 + sha256 again]
  W -.iter_projection_files.-> M
            

Problem, grown since last review. render.py's own redundancy is unchanged (build_conversation x3, parse_session x2, sha256 x3 per write). NEW: archive.py's write_session_folder parses/hashes again on top of that, and build.build() can call BOTH write_projection and write_session_folder for the same payload when both trees are configured - up to build_conversation x6, parse_session x5, sha256 x7 per head. Multiplicative, not additive.

Solution direction. Hoist parsing to the orchestrator; build models once, pass into every emitter and into archive.py's own decision points.

Contract ties: R8, R9; DESIGN section 4 · Tier: VERIFIED at 4824098

reports.py · cli.py · sweep.py · build.py · migrate.py · relocate.py · import_tree.py · archive.py

Problem, grown from 4 to 6 copies. Two new modules since the last review each added their own lock-acquire shell instead of reusing one. cli.py's report/tally half also grew from 4 to 6 instances, plus a 7th partial in _run_reindex.

Solution direction. Unchanged - a with_lock(root, name) -> BatchReport runner plus a shared render_batch(report) -> exit_code in reports.py.

Contract ties: R9, R10; R14 untouched · Tier: AGENT-REPORTED at 4824098

parser.py · render.py

Fan-out is larger than described: beyond the ~7 originally-named sites, a census found kind comparisons at roughly 15+ sites, clustered in _phase_meta (render.py:694-761) alone.

Solution direction. A typed discriminant (Literal or enum, stdlib, R7-safe) driving dispatch through mappings, so a missing arm is a type error.

Contract ties: F6, R7 · Tier: AGENT-REPORTED at 4824098

archive.py

Problem. _current_manifest (archive.py:454-499) fuses a file read with three PURE dict comparisons (source_hash, config, renderer_version) that decide "is this folder current." No way to unit-test a renderer_version mismatch without writing 5 real files first - the same shape C11 already names, in a module every incremental build/migrate call goes through.

Solution direction. A pure _manifest_matches(manifest: dict, source_hash, options) -> bool taking an already-loaded dict.

Same shape as C11 · Tier: AGENT-REPORTED at 4824098 · archive.py's first appearance on this board

render.py

Problem. The turn walk is still written twice: markdown via _turn_body (:999-1015) + _user_md (:1018-1036); HTML via _claude_inner (:1984-1997) + _turn_html (:2000-2072). Leaf rendering confirmed still single-owner via _render_block. Correction: the old citation _render_turn no longer exists by that name.

Solution direction. One turn-walk yielding typed events, each emitter supplying only leaf rendering.

Contract ties: R9 · pairs with C1 · Tier: AGENT-REPORTED at 4824098

sweep.py · migrate.py · relocate.py

Problem, unchanged. sweep.py and migrate.py's walker scaffolds are still near-verbatim duplicates. New adjacent finding: relocate._scan_content (:298-413) has its own THIRD walker of the same general shape, serving a different job.

Solution direction. One walk_transcripts(root, *, classify) owning the scaffold.

Tier: AGENT-REPORTED at 4824098

archive.py · import_tree.py

Problem. Only two booleans exist (is_subagent, is_session); the one caller needing the 3-way classification, import_tree._kind (:84-101), reinvents it as a plain str compared by literal equality - the C7 shape, at lower stakes (one caller today).

Solution direction. A small classify(data) -> Literal[...] combinator any future second caller can share.

Same shape as C7 · Tier: AGENT-REPORTED at 4824098

catalog.py

Unchanged. _short_key (catalog.py:146-163) still welds the pure prefix-band math to the conn.execute collision query in the same loop iteration.

Solution direction. Split the pure "choose a length" function from the collision lookup. Same shape as the new C13.

Contract ties: F5 · Tier: VERIFIED at 4824098