agent-orchestration-process 0.2.0__tar.gz → 0.2.1__tar.gz

{agent_orchestration_process-0.2.0 → agent_orchestration_process-0.2.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agent-orchestration-process
-Version: 0.2.0
+Version: 0.2.1
 Summary: Run bounded, non-interactive jobs across agent harnesses with isolation, exact resume, and provenance
 Keywords: agent-orchestration,ai-agents,autoresearch,coding-agents,git-worktrees
 Author: AOP contributors

{agent_orchestration_process-0.2.0 → agent_orchestration_process-0.2.1}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "uv_build"
 
 [project]
 name = "agent-orchestration-process"
-version = "0.2.0"
+version = "0.2.1"
 description = "Run bounded, non-interactive jobs across agent harnesses with isolation, exact resume, and provenance"
 readme = "README.md"
 requires-python = ">=3.11"

{agent_orchestration_process-0.2.0 → agent_orchestration_process-0.2.1}/pyproject.toml.orig

@@ -4,7 +4,7 @@ build-backend = "uv_build"
 
 [project]
 name = "agent-orchestration-process"
-version = "0.2.0"
+version = "0.2.1"
 description = "Run bounded, non-interactive jobs across agent harnesses with isolation, exact resume, and provenance"
 readme = "README.md"
 requires-python = ">=3.11"

{agent_orchestration_process-0.2.0 → agent_orchestration_process-0.2.1}/src/agent_orchestration_process/cli.py

@@ -22,7 +22,7 @@ from .isolation import PROFILES, explain_profile, resolve_policy
 from .locks import exclusive_lock, task_lock_path
 from .model_catalog import ModelCatalog, ensure_catalog_fresh
 from .model_listing import AGENTS, AvailableModel, list_models
-from .models import RunRequest, RunResult
+from .models import PromptPart, RunRequest, RunResult
 from .runner import AgentRunner, RunStore, adapter_for, provider_runtime_record
 from .migration import migrate_runs, migration_roots, print_counts
 from .worktrees import (
@@ -409,9 +409,11 @@ def main(argv: Sequence[str] | None = None) -> int:
                     prefix="aop: ",
                     stream=sys.stderr,
                 )
+            prompt, prompt_parts = _read_prompt(args)
             result = AgentRunner(manager, adapter_for(args.agent)).run(
                 task=task,
-                prompt=_read_prompt(args),
+                prompt=prompt,
+                prompt_parts=prompt_parts,
                 base=args.base,
                 model=args.model,
                 inference_provider=args.inference_provider,
@@ -432,9 +434,11 @@ def main(argv: Sequence[str] | None = None) -> int:
             return _report_run(result, manager, json_output=args.json)
 
         if args.command == "resume":
+            prompt, prompt_parts = _read_prompt(args)
             result = AgentRunner(manager).resume(
                 run_id=args.run_id,
-                prompt=_read_prompt(args),
+                prompt=prompt,
+                prompt_parts=prompt_parts,
                 timeout_seconds=args.timeout,
                 artifacts=args.artifact,
                 input_paths=args.input_paths,
@@ -472,6 +476,13 @@ def _add_prompt_arguments(parser: argparse.ArgumentParser) -> None:
     prompts = parser.add_mutually_exclusive_group(required=True)
     prompts.add_argument("--prompt")
     prompts.add_argument("--prompt-file", type=Path)
+    prompts.add_argument(
+        "--prompt-parts",
+        type=Path,
+        metavar="PATH",
+        help='JSON list of prompt blocks, each {"text": ..., "cache": true|false}; '
+        "Claude caches the prompt through the one cached block (claude only)",
+    )
 
 
 def _add_artifact_arguments(parser: argparse.ArgumentParser) -> None:
@@ -520,15 +531,36 @@ def _read_system_prompt(args: argparse.Namespace) -> str | None:
         ) from error
 
 
-def _read_prompt(args: argparse.Namespace) -> str:
+def _read_prompt(args: argparse.Namespace) -> tuple[str, tuple[PromptPart, ...] | None]:
     if args.prompt is not None:
-        return args.prompt
+        return args.prompt, None
+    path = args.prompt_file or args.prompt_parts
     try:
-        return args.prompt_file.read_text()
+        text = path.read_text()
     except OSError as error:
+        raise AOPError(f"could not read prompt file {path}: {error}") from error
+    if args.prompt_file is not None:
+        return text, None
+    try:
+        blocks = json.loads(text)
+    except json.JSONDecodeError as error:
+        raise AOPError(f"prompt parts file {path} is not JSON: {error}") from error
+    if not isinstance(blocks, list) or not all(
+        isinstance(block, dict)
+        and set(block) <= {"text", "cache"}
+        and isinstance(block.get("text"), str)
+        and isinstance(block.get("cache", False), bool)
+        for block in blocks
+    ):
         raise AOPError(
-            f"could not read prompt file {args.prompt_file}: {error}"
-        ) from error
+            f"prompt parts file {path} must be a list of "
+            '{"text": string, "cache": boolean} objects'
+        )
+    parts = tuple(
+        PromptPart(chars=len(block["text"]), cache=block.get("cache", False))
+        for block in blocks
+    )
+    return "".join(block["text"] for block in blocks), parts
 
 
 def _report_policy(

{agent_orchestration_process-0.2.0 → agent_orchestration_process-0.2.1}/src/agent_orchestration_process/isolation.py

@@ -157,7 +157,7 @@ def resolve_policy(
         host={
             "access": selected.host_access,
             "runtime_paths": (
-                ["/usr", "/bin", "/lib", "/lib64", "/etc runtime files"]
+                ["/usr", "/bin", "/sbin", "/lib", "/lib64", "/etc runtime files"]
                 if selected.host_access == "runtime-only"
                 else ["/"]
             ),

{agent_orchestration_process-0.2.0 → agent_orchestration_process-0.2.1}/src/agent_orchestration_process/models.py

@@ -57,6 +57,15 @@ class InferenceRoute:
         return cls(**fields)
 
 
+@dataclass(frozen=True)
+class PromptPart:
+    """One content block of a prompt; `cache` asks Claude to cache the prompt
+    up to and including this block. The text itself stays in RunRequest.prompt."""
+
+    chars: int
+    cache: bool = False
+
+
 @dataclass(frozen=True)
 class RunRequest:
     run_id: str
@@ -90,6 +99,9 @@ class RunRequest:
     allowed_urls: tuple[str, ...] = ()
     # A sealed run's private, initially empty /workspace is writable.
     writable_workspace: bool = False
+    # The prompt split into content blocks, in order; their lengths add up to
+    # len(prompt). None sends the prompt as one block.
+    prompt_parts: tuple[PromptPart, ...] | None = None
     # How `aop migrate-results` converted an older record, if it did.
     migration: dict[str, Any] | None = None
 
@@ -106,6 +118,10 @@ class RunRequest:
         if fields.get("tools") is not None:
             fields["tools"] = tuple(fields["tools"])
         fields["allowed_urls"] = tuple(fields.get("allowed_urls", ()))
+        if fields.get("prompt_parts") is not None:
+            fields["prompt_parts"] = tuple(
+                PromptPart(**part) for part in fields["prompt_parts"]
+            )
         fields["inputs"] = tuple(
             InputSnapshot.from_dict(item) for item in fields["inputs"]
         )

{agent_orchestration_process-0.2.0 → agent_orchestration_process-0.2.1}/src/agent_orchestration_process/prompt_transport.py

@@ -1,11 +1,14 @@
 """Prompt transport independent of the OS argument and environment limits."""
 
+from collections.abc import Sequence
 from contextlib import contextmanager
 import json
 import os
 from pathlib import Path
 import tempfile
 
+from .models import PromptPart
+
 
 # Native protocols remain in their adapters. Runtime-file is only for CLIs
 # whose one-shot parser accepts text exclusively as an argument.
@@ -47,6 +50,28 @@ def stdin_prompt(provider: str, prompt: str) -> str | None:
     return None
 
 
+def claude_message(prompt: str, parts: Sequence[PromptPart]) -> str:
+    """A stream-json user message whose content blocks follow `parts`.
+
+    Claude Code marks its own system blocks and the context it appends after
+    the prompt with one-hour cache marks, and a five-minute mark must not
+    precede a one-hour one, so a caller's mark uses one hour too.
+    """
+    content = []
+    start = 0
+    for part in parts:
+        block: dict[str, object] = {
+            "type": "text",
+            "text": prompt[start : start + part.chars],
+        }
+        if part.cache:
+            block["cache_control"] = {"type": "ephemeral", "ttl": "1h"}
+        content.append(block)
+        start += part.chars
+    message = {"type": "user", "message": {"role": "user", "content": content}}
+    return json.dumps(message) + "\n"
+
+
 @contextmanager
 def prompt_stream(prompt: str | None):
     """A seekable stdin avoids blocking on pipe writes before the timeout starts."""

{agent_orchestration_process-0.2.0 → agent_orchestration_process-0.2.1}/src/agent_orchestration_process/runner.py

@@ -39,6 +39,7 @@ from .models import (
     InputFile,
     InputSnapshot,
     InferenceRoute,
+    PromptPart,
     RunArtifact,
     RunRequest,
     RunResult,
@@ -1221,7 +1222,11 @@ class ClaudeAdapter:
             command,
             cwd=worktree.path,
             environment=environment,
-            prompt=prompt_transport.stdin_prompt(self.provider, request.prompt),
+            prompt=(
+                prompt_transport.claude_message(request.prompt, request.prompt_parts)
+                if request.prompt_parts is not None
+                else prompt_transport.stdin_prompt(self.provider, request.prompt)
+            ),
             timeout_seconds=request.timeout_seconds,
             is_response=self._is_response,
         )
@@ -1396,6 +1401,8 @@ class ClaudeAdapter:
             command.extend(["--system-prompt", request.system_prompt])
         if request.thinking_display is not None:
             command.extend(["--thinking-display", request.thinking_display])
+        if request.prompt_parts is not None:
+            command.extend(["--input-format", "stream-json"])
         if request.session_id:
             command.extend(["--resume", request.session_id])
         else:
@@ -4205,7 +4212,7 @@ def _isolated_root_command(bwrap: str) -> list[str]:
         "--tmpfs",
         "/tmp",
     ]
-    for path in ("/usr", "/bin", "/lib", "/lib64"):
+    for path in ("/usr", "/bin", "/sbin", "/lib", "/lib64"):
         source = Path(path)
         if source.exists():
             command.extend(["--ro-bind", path, path])
@@ -4214,6 +4221,9 @@ def _isolated_root_command(bwrap: str) -> list[str]:
         "/etc/ca-certificates",
         "/etc/crypto-policies",
         "/etc/hosts",
+        # The dynamic linker finds libraries in folders registered under
+        # /etc/ld.so.conf.d only through this cache.
+        "/etc/ld.so.cache",
         "/etc/localtime",
         "/etc/nsswitch.conf",
         "/etc/pki",
@@ -5775,8 +5785,10 @@ class AgentRunner:
         allowed_urls: Sequence[str] = (),
         writable_workspace: bool = False,
         thinking_display: str | None = None,
+        prompt_parts: Sequence[PromptPart] | None = None,
     ) -> RunResult:
         self._validate_mode(mode)
+        prompt_parts = self._validated_prompt_parts(prompt, prompt_parts)
         if writable_workspace and profile != "sealed":
             raise AOPError("--writable-workspace applies only to the sealed profile")
         if thinking_display is not None:
@@ -5822,6 +5834,7 @@ class AgentRunner:
             allowed_urls=allowed_urls,
             writable_workspace=writable_workspace,
             thinking_display=thinking_display,
+            prompt_parts=prompt_parts,
         )
         worktree = (
             self._create_sealed_workspace(request.run_id)
@@ -5854,6 +5867,7 @@ class AgentRunner:
         timeout_seconds: float | None = None,
         artifacts: Sequence[str] = (),
         input_paths: Sequence[str | os.PathLike[str]] | None = None,
+        prompt_parts: Sequence[PromptPart] | None = None,
         _task_lock_held: bool = False,
     ) -> RunResult:
         parent_request = self.store.load_request(run_id)
@@ -5865,6 +5879,7 @@ class AgentRunner:
                 )
             self.adapter = adapter_for(parent_request.provider)
         self._validate_mode(parent_request.mode)
+        prompt_parts = self._validated_prompt_parts(prompt, prompt_parts)
         self._validate_no_web(
             parent_request.no_web, parent_request.mode, parent_request.profile
         )
@@ -5952,6 +5967,7 @@ class AgentRunner:
             allowed_urls=parent_request.allowed_urls,
             writable_workspace=parent_request.writable_workspace,
             thinking_display=parent_request.thinking_display,
+            prompt_parts=prompt_parts,
         )
         return self._execute(
             request,
@@ -6308,6 +6324,7 @@ class AgentRunner:
         allowed_urls: tuple[str, ...] = (),
         writable_workspace: bool = False,
         thinking_display: str | None = None,
+        prompt_parts: tuple[PromptPart, ...] | None = None,
     ) -> RunRequest:
         return RunRequest(
             run_id=str(uuid.uuid4()),
@@ -6335,6 +6352,7 @@ class AgentRunner:
             allowed_urls=allowed_urls,
             writable_workspace=writable_workspace,
             thinking_display=thinking_display,
+            prompt_parts=prompt_parts,
         )
 
     def _validate_provider_base_url(self, url: str | None, profile: str) -> None:
@@ -6356,6 +6374,27 @@ class AgentRunner:
         if parsed.scheme not in {"http", "https"} or not parsed.hostname:
             raise AOPError(f"--provider-base-url must be an http(s) URL: {url!r}")
 
+    def _validated_prompt_parts(
+        self, prompt: str, parts: Sequence[PromptPart] | None
+    ) -> tuple[PromptPart, ...] | None:
+        if parts is None:
+            return None
+        if self.adapter.provider != "claude":
+            raise AOPError(f"{self.adapter.provider} does not support --prompt-parts")
+        parts = tuple(parts)
+        if not parts or any(part.chars <= 0 for part in parts):
+            raise AOPError("--prompt-parts needs at least one part, none of them empty")
+        if sum(part.chars for part in parts) != len(prompt):
+            raise AOPError("prompt parts do not add up to the prompt")
+        # Anthropic accepts four cache marks per request, and Claude Code
+        # places three: two on its system prompt and one on the context it
+        # appends after the prompt.
+        if sum(part.cache for part in parts) > 1:
+            raise AOPError(
+                "Claude allows one cached prompt part; Claude Code uses the other marks"
+            )
+        return parts
+
     def _validated_framing(
         self, system_prompt: str | None, tools: Sequence[str] | None, no_web: bool
     ) -> tuple[str, ...] | None: