fix(integrations): dispatch Amp via execute mode

src/specify_cli/integrations/amp/__init__.py

@@ -1,5 +1,8 @@
 """Amp CLI integration."""
 
+from collections.abc import Mapping, Sequence
+from typing import Any
+
 from ..base import MarkdownIntegration
 
 
@@ -18,3 +21,32 @@ class AmpIntegration(MarkdownIntegration):
         "args": "$ARGUMENTS",
         "extension": ".md",
     }
+
+    def build_exec_args(
+        self,
+        prompt: str,
+        *,
+        model: str | None = None,
+        output_json: bool = True,
+        integration_args: Sequence[str] | None = None,
+        integration_options: Mapping[str, Any] | None = None,
+    ) -> list[str] | None:
+        self.validate_runtime_config(integration_args, integration_options)
+        args = [self._resolve_executable()]
+        # Operator-injected extra args go before --execute: the flag takes the
+        # prompt as an optional inline value, so anything appended between the
+        # two would be consumed as the message instead.
+        self._apply_extra_args_env_var(args)
+
+        args.extend(["--execute", prompt])
+
+        if output_json:
+            # Amp's structured output is --stream-json (Claude Code-compatible
+            # stream JSON), valid only alongside --execute.
+            args.append("--stream-json")
+
+        # `model` is deliberately dropped: Amp has no model-selection flag.
+        # `-m/--mode` takes an agent mode (low/medium/high/ultra or a plugin
+        # mode), not a model identifier, so forwarding the caller's model onto
+        # it would silently select the wrong thing.
+        return args

tests/integrations/test_integration_amp.py

@@ -1,5 +1,7 @@
 """Tests for AmpIntegration."""
 
+from specify_cli.integrations import get_integration
+
 from .test_integration_base_markdown import MarkdownIntegrationTests
 
 
@@ -8,3 +10,75 @@ class TestAmpIntegration(MarkdownIntegrationTests):
     FOLDER = ".agents/"
     COMMANDS_SUBDIR = "commands"
     REGISTRAR_DIR = ".agents/commands"
+
+    def test_build_exec_args_uses_execute_mode(self):
+        """Amp dispatches through execute mode, not the inherited `-p`.
+
+        The Amp CLI has no `-p`/`--prompt` flag; passing one aborts with
+        `error: unknown option '-p'` before the agent runs (#4580).
+        """
+        integration = get_integration(self.KEY)
+
+        args = integration.build_exec_args(
+            "/speckit.specify build a login page",
+            output_json=False,
+        )
+
+        assert args == [
+            "amp",
+            "--execute",
+            "/speckit.specify build a login page",
+        ]
+        assert "-p" not in args
+
+    def test_build_exec_args_requests_stream_json(self):
+        """`--stream-json` is Amp's structured-output flag, used with --execute."""
+        integration = get_integration(self.KEY)
+
+        args = integration.build_exec_args("/speckit.plan add OAuth", output_json=True)
+
+        assert args == [
+            "amp",
+            "--execute",
+            "/speckit.plan add OAuth",
+            "--stream-json",
+        ]
+        assert "--output-format" not in args
+
+    def test_build_exec_args_omits_model_flag(self):
+        """Amp exposes no model-selection flag, so `model` is not forwarded.
+
+        `-m/--mode` takes an agent mode (low/medium/high/ultra), not a model
+        identifier, so remapping the caller's model onto it would be wrong.
+        """
+        integration = get_integration(self.KEY)
+
+        args = integration.build_exec_args(
+            "explain this repository",
+            model="gpt-5",
+            output_json=False,
+        )
+
+        assert args == ["amp", "--execute", "explain this repository"]
+        assert "--model" not in args
+        assert "-m" not in args
+        assert "gpt-5" not in args
+
+    def test_build_exec_args_applies_extra_args_before_execute(self, monkeypatch):
+        """Operator-injected flags precede `--execute` so they stay global.
+
+        `--execute [message]` takes the prompt as an optional inline value, so
+        injecting between the flag and the prompt would consume the prompt.
+        """
+        monkeypatch.setenv("SPECKIT_INTEGRATION_AMP_EXTRA_ARGS", "--no-notifications")
+        integration = get_integration(self.KEY)
+
+        args = integration.build_exec_args("check the build", output_json=True)
+
+        assert args == [
+            "amp",
+            "--no-notifications",
+            "--execute",
+            "check the build",
+            "--stream-json",
+        ]