D3: the F1 hook_json family engine — claude_code, codex_cli, kimi_code, devin folded to engine + config

[open-coder-ai]

What this changes

Four of the five F1 hook_json adapters (claude_code, codex_cli, kimi_code, devin) are re-expressed as one shared engine driven by their data/vendors/<agent>.json entries (docs/design/dialect-families.md D3); their hand-written modules are deleted. vscode_copilot deliberately stays a dialect module (see Notes). The golden wire fixture suite passes byte-for-byte with zero changes under tests/fixtures/golden/ or to tests/test_golden_fixtures.py — that empty diff is the acceptance gate for the whole fold.

What lands:

• adapters/_hook_json.py — the family engine: marker claims(), ordered field-fallback parse(), G1/G2 verdict renderers plus the shared hookSpecificOutput context/transform bodies, every function taking the vendor entry as data. adapters/_hook_entry.py renders hook_config() (hooks-map / bare-map / flat-list wrappers, windows-wrapper extras, kimi's TOML emitter). adapters/_family.py binds one entry behind the unchanged adapters.get() surface (all module attributes tests rely on — EVENT_MAP, DECISION_VOCABULARY, BLOCKING_EVENTS, render_config, … — are derived from the entry). adapters/_probes.py holds looks_like_claude_code, now shared by the engine and gemini_cli.

• Schema + entries gain the §3.1 renderer data D2 deferred: verdict words, degrade_notes (verbatim strings from the deleted adapters, so the wire cannot move), reason_defaults, note_style, echo, context-event lists — plus three claims fields that turn notes-only caveats into checked data (accept_names, accept_when_all, reject_markers_unless_probe) and hook_entry.bare. All entries regenerated via tools/recount_vendor_config.py --write, never edited by hand; devin's PermissionRequest gate is now recounted explicitly by replaying the frozen pre_tool payload with the event name swapped (the generalization W37's report suggested).

• bundle() composes engine + inlined VENDOR literal for config-driven agents and still splices the other eight legacy-style. Bundle line counts, measured at this PR's base vs. this PR (the design's §4 [h] asks for real numbers; estimates were +20–30%):

  bundle	before	after	growth
  claude_code	494	650	+32%
  codex_cli	452	636	+41%
  kimi_code	458	624	+36%
  devin	451	650	+44%

  Slightly above the §4 estimate; the numbers and the payoff side (624 deleted adapter lines → ~390 shared engine lines + data) are recorded in the design doc's §7.

Claim check

• No capability claim is widened without a mechanism behind it — the matrix is untouched; config gates ⊆ matrix block-capable events is still asserted by test_config_gates_are_a_subset_of_matrix_block_capable_events, and devin's added PermissionRequest gate maps to canonical pre_tool, already block-capable in its row
• Any new/changed MATRIX row carries a verified record — no matrix rows changed
• Payload shapes come from a primary source — no new payload shapes; every new config string is replayed against tests/fixtures/golden/<agent>.json, the D1-frozen capture of the real adapters

Checks

• pytest -q passes — 1349 passed, 4 skipped (baseline was 1320/4; +29 tests, none removed)
• ruff check . and ruff format --check . pass (--select I a no-op)
• Runtime path is still stdlib-only (check_stdlib_only.py and the AST check both green)
• Commits are signed off (git commit -s)

Also run locally: examples/generate.py --check (13 pages unchanged), docs/assets/gen_brand_assets.py --check, CLI smoke (--version, agents, matrix, doctor), and the 12-agent × 4-outcome bundler subprocess replay.

Notes for the reviewer

• Sizing: 4 of 5 folded, per the brief's own sizing clause. vscode_copilot is not foldable inside the design's §3.1 config/code line: its claims() accepts by three disjoint paths (envelope marker; lowercase-first names with their own conditional reject list; a memory-tool call carrying no event name at all), and parse() branches on tool ∈ MEMORY_TOOLS ∧ command ∈ MEMORY_WRITE_COMMANDS to select a second field-chain set — branching on payload content, which §3.1 says is dialect code. Encoding that as config means either a rule interpreter (the "worse programming language" §3.1 refuses) or a named accept-probe, i.e. widening the reject_probes crack to the accept side — a design decision I'm surfacing rather than making. The §7 dual-casing [h] is answered accordingly in the design doc.
• Schema extensions are the design's own §3.2 word/note tables, not new expressiveness: strings, lists, and flat maps only, all validated with additionalProperties: false and covered by new mutation tests (a typo'd word/note key fails loud). Mutation checks were run beyond the schema: a mutated note string fails the golden replay; a flipped honours_transform visibly changes the wire.
• Recount for config-driven vendors: fields/wire_events are read off the bound entry (the entry is parse() there — the engine executes it, so a second derivation source no longer exists); gates are still replayed from the frozen fixtures, and behaviour is held by the golden suite plus the untouched per-adapter suites. The word/note tables live in tools/recount/tables.py with the claims tables, every string exercised by the fixture replay on every run.
• Two deliberate direct-call (never reachable through dispatch.handle, untested before and after) behaviour notes: respond() with a raw WARN at a G2 gate is now uniformly silent (the old claude_code emitted a bare hookEventName body; codex/kimi already returned ""), and test_function_local_imports_are_left_alone samples vscode_copilot now that the engine bundles legitimately contain no function-local imports.