Merge pull request #96 from open-coder-ai/w37/d2-vendor-config-schema

Author: jothimani-rajendran

CHANGELOG.md

@@ -6,6 +6,29 @@ versioning: [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
 
+### Added
+- **The vendor config schema and all 12 entries** (`docs/design/dialect-families.md` D2):
+  `src/agentseam/data/vendors/schema.json` and one `data/vendors/<agent>.json` per adapted
+  agent, unused by the runtime yet (D3+ wires an engine onto this data). Every entry is
+  produced by `tools/recount_vendor_config.py`, which derives each field by executing or
+  AST-reading the current adapter and its D1 golden fixture rather than by hand -- module
+  constants read directly, `parse()`'s field-fallback chains AST-walked, and every gate's
+  grammar/`honours_escalate`/`honours_transform` replayed straight out of the frozen (payload ->
+  stdout) wire truth. Family assignment and the marker-based `claims` table are the design's
+  own stated judgment calls, recorded as small tables and pinned by
+  `tests/test_vendor_config_claims.py`, which mutates a real claimed payload per declared
+  marker and replays `claims()`. Per-claim evidence (owner decision 2026-09-01,
+  `plan/agentseam-project.md` in org-plan) gives every claim group a `basis` (from
+  `matrix_terms.BASES`), `date`, and the test that exercises it; schema validation fails a
+  claim with none. `tools/validate_vendor_config.py` is a small stdlib-only JSON Schema
+  subset validator (no third-party dependency) run as a test, not a runtime cost.
+  `tests/test_vendor_config.py` adds the §3.3 consistency tests (config gates are a subset
+  of the matrix's block-capable events; `config_path` agrees with the matrix) and answers
+  both [h] hypotheses the design carried into D2: the `reject_probes` device stays at one
+  named predicate (`looks_like_claude_code`, used by two vendors) against the ~3-probe
+  budget, and PR #89's tuple-restoring loader is order-safe for `fields`' order-sensitive
+  chains (demonstrated, not merely asserted).
+
 ### Changed
 - **The `Decision` vocabulary is aligned to the Agent Control Standard's verdict names**
   (`plan/acs-alignment-and-delegation.md` §1 in the org-plan repo). `ASK` is renamed to
@@ -47,6 +70,11 @@ versioning: [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 - **`pyproject.toml`'s `version` was unchecked against `agentseam.__version__`.** Only
   `CITATION.cff` was verified to track the package; a wheel could ship with the two
   disagreeing. `test_repo_standards.py` now checks both.
+- **`matrix.json`'s `kimi_code` row named the wrong config file (`config.toml`, missing
+  the `~/.kimi-code/` directory `kimi_code.py`'s own `CONFIG_PATH` has always used).** Found
+  while implementing D2's `config_path` agrees with the matrix consistency test.
+  `data/vendors/*.json` was not shipped as package data; `pyproject.toml`'s
+  `package-data` now also matches `data/vendors/*.json`.
 
 ### Internal
 - `claude_code`, `gemini_cli`, `grok`, and `junie` shared an identical `hook_config` body;

docs/design/dialect-families.md

@@ -105,7 +105,7 @@ Consequences, applied to today's code:
 | :--- | :--- |
 | The five grammar renderers G1–G5 | which grammar each event group uses, and the word tables filling it |
 | Ordered-fallback field extraction machinery, `tool_input_of` (`contract.py:139-150`) | the per-vendor key chains (e.g. path = `file_path→absolute_path→path` `gemini_cli.py:79`; content gated to write tools `gemini_cli.py:71-73`; output = `last_assistant_message` `junie.py:65`) |
-| `degrade()` semantics (`dispatch.py:26-38`) and its generated-runtime twin (`bundler_templates.py:26-40`) | per-gate capability flags (`honours_ask`, `honours_rewrite`, `honours_block`) and the degradation-note strings, verbatim from today's adapters so wire output stays byte-identical |
+| `degrade()` semantics (`dispatch.py:26-38`) and its generated-runtime twin (`bundler_templates.py:26-40`) | per-gate capability flags (`honours_escalate`, `honours_transform`, `honours_block`) and the degradation-note strings, verbatim from today's adapters so wire output stays byte-identical |
 | Shape-inference `claims`/event-naming for cursor, windsurf, antigravity (`cursor.py:96-98`, `windsurf.py:27-32`, `antigravity.py:25-31`) | marker-based `claims`: required markers, foreign markers, client_type allowlists (`junie.py:45-48`, `tabnine.py:52-55`, `gemini_cli.py:57-64`, `codex_cli.py:43-52`) |
 | vscode_copilot's memory-tool argument scheme (`vscode_copilot.py:77-83,103-107`) | `MEMORY_TOOLS` / `WRITE_TOOLS` / `SHELL_TOOLS` name lists |
 | kimi's TOML emitter (`kimi_code.py:155-164`) — one renderer, selected by `config_format` | `config_path`, `config_format`, hook-entry wrapper style and extra entry fields (`commandWindows` `codex_cli.py:137`, `name` `tabnine.py:113`, `failClosed`+`version` `cursor.py:225-228`, group nesting `antigravity.py:117-127`, no-wrapper `devin.py:121-132`, extra `pre_mcp_tool_use` wiring `windsurf.py:106-107`) |
@@ -152,13 +152,13 @@ Consequences, applied to today's code:
     "bare_allow": "inert",                    // allow_semantics.py:70-77 key, cross-checked
     "answer_events": ["BeforeTool", "AfterTool", "BeforeAgent", "AfterAgent"],
     "gates": {                                // gemini_cli.py:104-121
-      "BeforeTool": { "grammar": "G1", "honours_ask": true,  "honours_rewrite": true },
-      "AfterTool":  { "grammar": "G1", "honours_ask": false },
-      "BeforeAgent":{ "grammar": "G1", "honours_ask": false },
-      "AfterAgent": { "grammar": "G1", "honours_ask": false }
+      "BeforeTool": { "grammar": "G1", "honours_escalate": true,  "honours_transform": true },
+      "AfterTool":  { "grammar": "G1", "honours_escalate": false },
+      "BeforeAgent":{ "grammar": "G1", "honours_escalate": false },
+      "AfterAgent": { "grammar": "G1", "honours_escalate": false }
     },
     "words": { "allow": "allow", "deny": "deny", "ask": "ask", "block": "deny" },
-    "rewrite_grammar": "hook_specific_tool_input",   // gemini_cli.py:119-120
+    "transform_grammar": "hook_specific_tool_input",   // gemini_cli.py:119-120
     "degrade_notes": {                        // verbatim today's strings, gemini_cli.py:109-117
       "ask_unhonoured": "%s (confirmation required; %s cannot prompt from a hook)"
     }
@@ -167,7 +167,15 @@ Consequences, applied to today's code:
 }
 ```
 
-Where the degradation lives (the brief's codex question): `honours_ask: false` on the
+Field names above (`honours_escalate`, `honours_transform`, `transform_grammar`) follow the
+post-W35 ACS vocabulary (`contract.py`: `escalate`/`transform`), not the pre-ACS `ask`/
+`rewrite` this section was drafted with at W31, before ACS alignment was sequenced ahead of
+D2 (org-plan `plan/agentseam-project.md`: "D2 schema (written in ACS words)"). Vendor
+wire-word *data* is untouched by that rename: `vocabulary`/`words` above are correct as
+written, because that is the literal word gemini_cli's own dialect speaks on the wire, not
+agentseam's name for the concept.
+
+Where the degradation lives (the brief's codex question): `honours_escalate: false` on the
 gate + the vendor's note string. codex_cli's ask→deny with *"Codex CLI does not support
 ask; asking would fail open"* (`codex_cli.py:121-124`) is a flag plus a string — no
 conditional survives in config. The engine's degrade step already exists twice
@@ -269,7 +277,7 @@ instructions/capture suites (untouched paths).
 **Must be written:**
 
 1. Config schema validation: every `data/vendors/*.json` validates; unknown keys are
-   an error (a typo'd `honours_ask` must fail loud, not fail open).
+   an error (a typo'd `honours_escalate` must fail loud, not fail open).
 2. Config completeness: every `ADAPTERS` name is config-backed; every config gate maps
    to a matrix block-capable event; `config_path`/`bare_allow` agree with
    `MATRIX`/`ALLOW_SEMANTICS`.
@@ -307,12 +315,19 @@ first, D3/D4 swap cleanly — they share only D1+D2.
 
 - **[h]** Bundle size estimates in §4 are composition arithmetic, not measurements;
   D3 should publish real numbers next to them.
-- **[h]** The `reject_probes: ["looks_like_claude_code"]` device (named engine
+- **[v]** (was [h]) The `reject_probes: ["looks_like_claude_code"]` device (named engine
   predicates referenced from config) is the narrowest crack in the code/config line;
   if D2 finds more than ~3 named probes are needed, that is evidence the line is drawn
-  wrong, and the design should be revisited rather than the list grown.
+  wrong, and the design should be revisited rather than the list grown. **Answered by D2**:
+  exactly one named probe (`looks_like_claude_code`), used by `gemini_cli` and `devin` --
+  well inside the budget (`tests/test_vendor_config.py::
+  test_reject_probes_stay_under_the_three_probe_budget`). The design stands as written.
 - **[h]** vscode_copilot's dual-casing may be cleaner as two config entries sharing a
   family than one entry with paired key chains; decide in D3 with the fixtures open.
-- **[h]** PR #89's loader restores every JSON array as a tuple; `fields` chains here
-  are order-sensitive lists and unaffected, but D2 must confirm the loader's
-  tuple-restore doesn't collide with schema types.
+- **[v]** (was [h]) PR #89's loader restores every JSON array as a tuple; `fields` chains
+  here are order-sensitive lists and unaffected, but D2 must confirm the loader's
+  tuple-restore doesn't collide with schema types. **Answered by D2**: a tuple preserves
+  the exact sequence its JSON array was written in, so the restore is order-safe; nothing
+  in `vendor_config.py` or its tests ever compares one of these tuples against a list
+  literal with `==` (`tests/test_vendor_config.py::
+  test_loader_tuple_restore_preserves_field_chain_order`).

examples/generated/README.md

@@ -59,7 +59,7 @@ where = ["src"]
 # The vendor tables live in data/*.json and are read at import time. Without this the
 # wheel ships without them and every import raises FileNotFoundError from a pip install
 # while working perfectly from a source checkout -- the failure chock#87 caught on itself.
-agentseam = ["data/*.json"]
+agentseam = ["data/*.json", "data/vendors/*.json"]
 
 [tool.pytest.ini_options]
 testpaths = ["tests"]

examples/generated/kimi_code.md

@@ -158,7 +158,7 @@
   "kimi_code": {
     "display": "Kimi Code CLI",
     "tier": "block",
-    "config": "config.toml",
+    "config": "~/.kimi-code/config.toml",
     "verified": {
       "basis": "vendor-docs",
       "version": "CLI",

pyproject.toml

@@ -0,0 +1,124 @@
+{
+  "agent": "antigravity",
+  "claims": {
+    "mode": "shape_inferred"
+  },
+  "config_format": "json",
+  "config_path": ".agents/hooks.json",
+  "display": "Antigravity",
+  "events": {},
+  "evidence": {
+    "claims": {
+      "basis": "vendor-docs",
+      "date": "2026-08-26",
+      "test": "tests/test_examples.py::test_each_payload_is_claimed_by_its_own_adapter"
+    },
+    "config_path": {
+      "basis": "vendor-docs",
+      "date": "2026-08-26",
+      "test": "tests/test_vendor_config.py::test_config_path_agrees_with_matrix"
+    },
+    "events": {
+      "basis": "vendor-docs",
+      "date": "2026-08-26",
+      "test": "tests/test_examples.py::test_each_payload_parses_to_the_event_it_is_filed_under"
+    },
+    "family": {
+      "basis": "vendor-docs",
+      "date": "2026-08-26",
+      "test": "tests/test_golden_fixtures.py::test_wire_output_matches_the_frozen_fixture"
+    },
+    "fields": {
+      "basis": "vendor-docs",
+      "date": "2026-08-26",
+      "test": "tests/test_vendor_config.py::test_entries_match_recount"
+    },
+    "hook_entry": {
+      "basis": "vendor-docs",
+      "date": "2026-08-26",
+      "test": "tests/test_golden_fixtures.py::test_hook_config_matches_the_frozen_fixture_on_both_matcher_paths"
+    },
+    "tools": {
+      "basis": "vendor-docs",
+      "date": "2026-08-26",
+      "test": "tests/test_vendor_config.py::test_entries_match_recount"
+    },
+    "verdicts": {
+      "basis": "vendor-docs",
+      "date": "2026-08-26",
+      "test": "tests/test_golden_fixtures.py::test_wire_output_matches_the_frozen_fixture"
+    }
+  },
+  "family": "antigravity",
+  "fields": {
+    "command": [
+      "toolCall.args.CommandLine"
+    ],
+    "content": [
+      "toolCall.args.CodeContent",
+      "toolCall.args.ReplacementContent",
+      "toolCall.args.ReplacementChunks[].ReplacementContent"
+    ],
+    "cwd": [
+      "toolCall.args.Cwd",
+      "workspacePaths[0]"
+    ],
+    "output": [
+      "error"
+    ],
+    "path": [
+      "toolCall.args.TargetFile",
+      "toolCall.args.AbsolutePath"
+    ],
+    "session_id": [
+      "conversationId"
+    ],
+    "tool": [
+      "toolCall.name"
+    ],
+    "tool_use_id": [
+      "stepIdx"
+    ]
+  },
+  "hook_entry": {
+    "group": "agentseam",
+    "matcher": true,
+    "wrapper": "hooks_map"
+  },
+  "needs_trust": false,
+  "tools": {},
+  "verdicts": {
+    "answer_events": [
+      "PreToolUse",
+      "Stop"
+    ],
+    "bare_allow": "unverified",
+    "gates": {
+      "PreToolUse": {
+        "grammar": "G1",
+        "honours_escalate": true,
+        "honours_transform": false
+      },
+      "Stop": {
+        "grammar": "G1",
+        "honours_escalate": false,
+        "honours_transform": false
+      }
+    },
+    "vocabulary": [
+      "allow",
+      "ask",
+      "continue",
+      "deny",
+      "deny_unless_prior_grant",
+      "force_ask",
+      "stop"
+    ],
+    "vocabulary_basis": "verified"
+  },
+  "wire_events": {
+    "post_tool": "PostToolUse",
+    "pre_tool": "PreToolUse",
+    "stop": "Stop"
+  }
+}