Pass --output-format json with --json-schema and unwrap CLI envelope

The Claude Code adapter previously passed --json-schema alone. On Claude
CLI 2.1.160 that combination still emits the model's conversational
preamble on stdout while the schema-enforced structured payload ships on
a sidecar channel the adapter cannot read. Result: callers requesting
structured output got prose that fails JSON parsing downstream — exactly
the failure mode the activity-core CUST-WP-0045 daily triage canary hit
on 2026-06-01 ("Triage report generated and returned via structured
output. Key signals:..." → json.loads error at column 1).

Fix: when --json-schema is set, also pass --output-format json. The CLI
then writes a JSON envelope on stdout. The adapter unwraps it by
probing a small allowlist of known text-bearing fields (result,
result_text, content, text, output). Unknown envelope shapes fall
through to raw stdout so the operator can introspect the structure and
extend the allowlist.

The unwrap path is only triggered when --json-schema was set, so non-
schema callers keep the existing raw-stdout behavior.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

llm_connect/claude_code.py

@@ -2,12 +2,12 @@
 Claude Code CLI adapter — runs the ``claude`` CLI as a subprocess.
 """
 
-import asyncio
-import json
-import os
-import subprocess
-from pathlib import Path
-from typing import Optional
+import asyncio
+import json
+import os
+import subprocess
+from pathlib import Path
+from typing import Optional
 
 from llm_connect.adapter import LLMAdapter
 from llm_connect.models import RunConfig, LLMResponse
@@ -26,23 +26,23 @@ class ClaudeCodeAdapter(LLMAdapter):
     length limits (compiled prompts can exceed 30 KB).
     """
 
-    def __init__(
-        self,
-        cli_path: Optional[str] = None,
-        model: Optional[str] = None,
-        config: Optional[LLMConfig] = None,
-    ):
-        self._config = config or LLMConfig(provider="claude-code")
-        self._cli_path = cli_path or self._resolve_cli_path()
-        self._model = model
+    def __init__(
+        self,
+        cli_path: Optional[str] = None,
+        model: Optional[str] = None,
+        config: Optional[LLMConfig] = None,
+    ):
+        self._config = config or LLMConfig(provider="claude-code")
+        self._cli_path = cli_path or self._resolve_cli_path()
+        self._model = model
 
     # ── LLMAdapter interface ────────────────────────────────────────
 
-    def execute_prompt(self, prompt: str, config: RunConfig) -> LLMResponse:
-        self._preflight_budget(config)
-        cmd = self._build_command(config)
-
-        timeout = config.timeout_seconds or self._config.timeout_seconds
+    def execute_prompt(self, prompt: str, config: RunConfig) -> LLMResponse:
+        self._preflight_budget(config)
+        cmd = self._build_command(config)
+
+        timeout = config.timeout_seconds or self._config.timeout_seconds
 
         try:
             result = subprocess.run(
@@ -65,7 +65,7 @@ class ClaudeCodeAdapter(LLMAdapter):
                 stderr=result.stderr,
             )
 
-        content = result.stdout
+        content = _unwrap_cli_json_envelope(result.stdout, config)
         prompt_tokens = estimate_tokens(prompt)
         completion_tokens = estimate_tokens(content)
 
@@ -86,10 +86,10 @@ class ClaudeCodeAdapter(LLMAdapter):
         self._consume_budget(config, response)
         return response
 
-    async def async_execute_prompt(self, prompt: str, config: RunConfig) -> LLMResponse:
-        """Native async implementation using asyncio.create_subprocess_exec."""
-        self._preflight_budget(config)
-        cmd = self._build_command(config)
+    async def async_execute_prompt(self, prompt: str, config: RunConfig) -> LLMResponse:
+        """Native async implementation using asyncio.create_subprocess_exec."""
+        self._preflight_budget(config)
+        cmd = self._build_command(config)
 
         timeout = config.timeout_seconds or self._config.timeout_seconds
 
@@ -117,7 +117,7 @@ class ClaudeCodeAdapter(LLMAdapter):
                 stderr=stderr_bytes.decode(),
             )
 
-        content = stdout_bytes.decode()
+        content = _unwrap_cli_json_envelope(stdout_bytes.decode(), config)
         prompt_tokens = estimate_tokens(prompt)
         completion_tokens = estimate_tokens(content)
 
@@ -139,49 +139,92 @@ class ClaudeCodeAdapter(LLMAdapter):
         self._consume_budget(config, response)
         return response
 
-    def validate_config(self, config: RunConfig) -> bool:
-        try:
-            result = subprocess.run(
-                [self._cli_path, "--version"],
+    def validate_config(self, config: RunConfig) -> bool:
+        try:
+            result = subprocess.run(
+                [self._cli_path, "--version"],
                 capture_output=True,
                 text=True,
                 timeout=10,
             )
             return result.returncode == 0
-        except (subprocess.TimeoutExpired, FileNotFoundError, OSError):
-            return False
-
-    def _build_command(self, config: RunConfig) -> list[str]:
-        cmd = [self._cli_path, "--print"]
-        if self._model:
-            cmd.extend(["--model", self._model])
-
-        json_schema = _json_schema_arg(config)
-        if json_schema:
-            cmd.extend(["--json-schema", json_schema])
-        return cmd
-
-    def _resolve_cli_path(self) -> str:
-        configured = (
-            os.environ.get("LLM_CONNECT_CLAUDE_CLI_PATH")
-            or os.environ.get("CLAUDE_CLI_PATH")
-            or self._config.claude_cli_path
-        )
-        if configured and configured != "claude":
-            return configured
-
-        local_cli = Path.home() / ".local" / "bin" / "claude"
-        if local_cli.exists():
-            return str(local_cli)
-        return configured or "claude"
-
-
-def _json_schema_arg(config: RunConfig) -> str | None:
-    schema = (config.model_params or {}).get("json_schema")
-    if not schema:
-        return None
-    if isinstance(schema, str):
-        return schema
-    if isinstance(schema, dict):
-        return json.dumps(schema, separators=(",", ":"))
-    return None
+        except (subprocess.TimeoutExpired, FileNotFoundError, OSError):
+            return False
+
+    def _build_command(self, config: RunConfig) -> list[str]:
+        cmd = [self._cli_path, "--print"]
+        if self._model:
+            cmd.extend(["--model", self._model])
+
+        json_schema = _json_schema_arg(config)
+        if json_schema:
+            cmd.extend(["--json-schema", json_schema])
+            # With --json-schema alone the CLI prints conversational text on
+            # stdout while the structured payload ships on a sidecar channel
+            # callers cannot reach. --output-format json forces the structured
+            # response (wrapped in an envelope) onto stdout.
+            cmd.extend(["--output-format", "json"])
+        return cmd
+
+    def _resolve_cli_path(self) -> str:
+        configured = (
+            os.environ.get("LLM_CONNECT_CLAUDE_CLI_PATH")
+            or os.environ.get("CLAUDE_CLI_PATH")
+            or self._config.claude_cli_path
+        )
+        if configured and configured != "claude":
+            return configured
+
+        local_cli = Path.home() / ".local" / "bin" / "claude"
+        if local_cli.exists():
+            return str(local_cli)
+        return configured or "claude"
+
+
+def _json_schema_arg(config: RunConfig) -> str | None:
+    schema = (config.model_params or {}).get("json_schema")
+    if not schema:
+        return None
+    if isinstance(schema, str):
+        return schema
+    if isinstance(schema, dict):
+        return json.dumps(schema, separators=(",", ":"))
+    return None
+
+
+# Field names Claude Code's `--output-format json` envelope is known to use
+# for the model's primary textual response. Probed in order; the first match
+# wins. If none match (because the envelope shape is something we haven't
+# seen), we return the raw envelope string so the caller still gets the data
+# and can introspect it.
+_ENVELOPE_TEXT_FIELDS = ("result", "result_text", "content", "text", "output")
+
+
+def _unwrap_cli_json_envelope(stdout: str, config: RunConfig) -> str:
+    """Extract the model's payload from Claude CLI's --output-format json envelope.
+
+    Only attempts unwrap when --json-schema was set, because that's the only
+    code path that adds --output-format json to the CLI invocation. Other
+    paths keep raw stdout (current behavior preserved).
+    """
+    if not _json_schema_arg(config):
+        return stdout
+    text = stdout.strip()
+    if not text:
+        return stdout
+    try:
+        envelope = json.loads(text)
+    except json.JSONDecodeError:
+        return stdout
+    if not isinstance(envelope, dict):
+        return stdout
+    for key in _ENVELOPE_TEXT_FIELDS:
+        if key in envelope:
+            value = envelope[key]
+            if isinstance(value, str):
+                return value
+            if isinstance(value, (dict, list)):
+                return json.dumps(value)
+    # Unknown envelope shape — surface it raw so the operator can see it
+    # in the validation error and we can update _ENVELOPE_TEXT_FIELDS.
+    return stdout

tests/test_claude_code.py

@@ -39,15 +39,65 @@ def test_execute_prompt_passes_json_schema_to_claude_cli(monkeypatch):
             "--print",
             "--json-schema",
             '{"type":"object"}',
+            "--output-format",
+            "json",
         ],
         "input": "Produce a report.",
         "capture_output": True,
         "text": True,
         "timeout": 42,
     }
+    # Stdout shape that does not match any known envelope field is returned
+    # verbatim so the caller can introspect and we can extend the field list.
     assert response.content == '{"summary":"ok","recommendations":[]}'
 
 
+def test_execute_prompt_unwraps_cli_json_envelope_result_field(monkeypatch):
+    """With --output-format json the CLI wraps the model payload in an
+    envelope. The adapter unwraps the textual result so the caller still
+    sees the model's structured-output JSON, not the envelope."""
+    def fake_run(cmd, input, capture_output, text, timeout):  # noqa: ANN001
+        envelope = {
+            "type": "result",
+            "result": '{"summary":"ok","recommendations":[]}',
+            "total_cost_usd": 0.001,
+        }
+        import json as _json
+        return SimpleNamespace(
+            returncode=0,
+            stdout=_json.dumps(envelope),
+            stderr="",
+        )
+
+    monkeypatch.setattr("llm_connect.claude_code.subprocess.run", fake_run)
+    adapter = ClaudeCodeAdapter(cli_path="/custom/claude")
+
+    response = adapter.execute_prompt(
+        "Produce a report.",
+        RunConfig(model_params={"json_schema": {"type": "object"}}),
+    )
+
+    assert response.content == '{"summary":"ok","recommendations":[]}'
+
+
+def test_execute_prompt_no_unwrap_without_json_schema(monkeypatch):
+    """Without --json-schema we do not pass --output-format json, so the
+    envelope unwrap path stays inert and raw stdout passes through."""
+    def fake_run(cmd, input, capture_output, text, timeout):  # noqa: ANN001
+        return SimpleNamespace(
+            returncode=0,
+            stdout='{"result":"this is just stdout, not an envelope"}',
+            stderr="",
+        )
+
+    monkeypatch.setattr("llm_connect.claude_code.subprocess.run", fake_run)
+    adapter = ClaudeCodeAdapter(cli_path="/custom/claude")
+
+    response = adapter.execute_prompt("Plain prompt.", RunConfig())
+
+    assert response.content == '{"result":"this is just stdout, not an envelope"}'
+
+
 def test_claude_code_adapter_prefers_env_cli_path(monkeypatch):
     monkeypatch.setenv("LLM_CONNECT_CLAUDE_CLI_PATH", "/home/me/bin/claude")