fix(eval): pass the deep-swe prompt after -- so a leading dash is not a flag

.amplifier/evaluation/deep-swe/src/deepswe_agents/amplifier_agent.py

@@ -106,8 +106,12 @@ def run_command(self, instruction_path: str) -> str:
         # SESSION_ID). `--output json` appends the final envelope -- including
         # metadata.durationMs -- to agent.log; it does NOT suppress the
         # human-readable `[usage]` lines, so nothing is lost by adding it.
+        #
+        # The `--` before the prompt is load-bearing: without it, any instruction
+        # whose text begins with `-` is parsed as a flag and the CLI exits with a
+        # usage error before the agent ever runs.
         return (
             f"amplifier-agent run -y --config {HOST_CONFIG_PATH} "
             f"--session-id {shlex.quote(SESSION_ID)} --output json "
-            f'"$(cat {instruction_path})"'
+            f'-- "$(cat {instruction_path})"'
         )

.amplifier/evaluation/deep-swe/src/deepswe_agents/amplifier_foundation.py

@@ -172,11 +172,15 @@ def run_command(self, instruction_path: str) -> str:
 
         No `--model` flag: it requires `--provider` alongside it, and the model is
         already pinned by `default_model` in settings.yaml.
+
+        The `--` before the prompt is load-bearing: without it, any instruction
+        whose text begins with `-` is parsed as a flag and the CLI exits with a
+        usage error before the agent ever runs.
         """
         return (
             f"amplifier run --bundle '{self._anchors_ref}' "
             f"--mode single --output-format json "
-            f'"$(cat {instruction_path})"'
+            f'-- "$(cat {instruction_path})"'
         )
 
     def populate_context_post_run(self, context) -> None:  # type: ignore[no-untyped-def]

.amplifier/evaluation/deep-swe/src/deepswe_agents/opencode_amplifier.py

@@ -71,7 +71,12 @@ def agent_install_steps(self) -> list[InstallStep]:
         ]
 
     def run_command(self, instruction_path: str) -> str:
+        # Two separators, doing two different jobs. The first `--` hands the rest
+        # of the line to opencode (click consumes only that one). The second is
+        # load-bearing for the same reason as the other adapters: without it, any
+        # instruction whose text begins with `-` is parsed as a flag and the agent
+        # never runs. Click forwards the second `--` verbatim via os.execvp.
         return (
             f"amplifier-opencode launch -- run --auto --model amplifier/{self.model} "
-            f'"$(cat {instruction_path})"'
+            f'-- "$(cat {instruction_path})"'
         )

.amplifier/evaluation/deep-swe/src/deepswe_agents/opencode_vanilla.py

@@ -144,4 +144,9 @@ async def _dump_opencode_log(self, environment) -> None:
             self.logger.warning(f"could not read opencode log: {exc}")
 
     def run_command(self, instruction_path: str) -> str:
-        return f'opencode run --model anthropic/{self.model} --auto "$(cat {instruction_path})"'
+        # The `--` before the prompt is load-bearing: without it, any instruction
+        # whose text begins with `-` is parsed as a flag, opencode prints its usage
+        # banner, and the agent never runs. yargs is configured with
+        # `populate--: true`, and opencode's run command merges `argv["--"]` back
+        # into the message, so the prompt still arrives intact.
+        return f'opencode run --model anthropic/{self.model} --auto -- "$(cat {instruction_path})"'