agent-orchestration-process 0.3.0__tar.gz → 0.4.0__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (31) hide show
  1. agent_orchestration_process-0.4.0/PKG-INFO +142 -0
  2. agent_orchestration_process-0.4.0/README.md +120 -0
  3. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/pyproject.toml +1 -1
  4. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/pyproject.toml.orig +1 -1
  5. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/cli.py +59 -11
  6. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/egress_bridge.py +4 -9
  7. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/models.py +8 -0
  8. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/prompt_transport.py +6 -2
  9. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/runner.py +277 -129
  10. agent_orchestration_process-0.4.0/src/agent_orchestration_process/sandbox_launch.py +137 -0
  11. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/zcode_protocol.py +4 -8
  12. agent_orchestration_process-0.3.0/PKG-INFO +0 -144
  13. agent_orchestration_process-0.3.0/README.md +0 -122
  14. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/LICENSE +0 -0
  15. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/__init__.py +0 -0
  16. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/__main__.py +0 -0
  17. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/batch.py +0 -0
  18. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/codex_auth.py +0 -0
  19. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/codex_routes.py +0 -0
  20. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/dsh_runner.mjs +0 -0
  21. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/egress.py +0 -0
  22. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/integration.py +0 -0
  23. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/isolation.py +0 -0
  24. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/locks.py +0 -0
  25. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/migration.py +0 -0
  26. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/model_catalog.py +0 -0
  27. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/model_listing.py +0 -0
  28. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/pricing.py +0 -0
  29. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/provider_versions.py +0 -0
  30. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/worktrees.py +0 -0
  31. {agent_orchestration_process-0.3.0 → agent_orchestration_process-0.4.0}/src/agent_orchestration_process/zcode.py +0 -0
@@ -0,0 +1,142 @@
1
+ Metadata-Version: 2.4
2
+ Name: agent-orchestration-process
3
+ Version: 0.4.0
4
+ Summary: Run bounded, non-interactive jobs across agent harnesses with isolation, exact resume, and provenance
5
+ Keywords: agent-orchestration,ai-agents,autoresearch,coding-agents,git-worktrees
6
+ Author: AOP contributors
7
+ License-Expression: MIT
8
+ License-File: LICENSE
9
+ Classifier: Development Status :: 3 - Alpha
10
+ Classifier: Environment :: Console
11
+ Classifier: Operating System :: POSIX :: Linux
12
+ Classifier: Programming Language :: Python :: 3
13
+ Classifier: Programming Language :: Python :: 3 :: Only
14
+ Classifier: Topic :: Software Development
15
+ Requires-Dist: model-prices>=0.0.3
16
+ Requires-Dist: pyyaml>=6.0.2
17
+ Requires-Python: >=3.11
18
+ Project-URL: Homepage, https://github.com/wakamex/agent-orchestration-process
19
+ Project-URL: Repository, https://github.com/wakamex/agent-orchestration-process.git
20
+ Project-URL: Issues, https://github.com/wakamex/agent-orchestration-process/issues
21
+ Description-Content-Type: text/markdown
22
+
23
+ # Agent Orchestration Process (AOP)
24
+
25
+ AOP is a command-line tool for running coding agents such as Claude Code and Codex as unattended jobs. It is for automation: CI jobs, batch experiments, benchmarks, and programs that hand work to an agent and need a dependable result back. One `aop run` command drives ten different agent CLIs, called harnesses here, and gives every run the same sandbox, deadline, and result record, so a caller does not need a separate adapter for each CLI's flags, sessions, and output.
26
+
27
+ - Run [Codex](https://github.com/openai/codex), [Claude Code](https://code.claude.com/docs/en/overview), [Cursor Agent](https://docs.cursor.com/en/cli/overview), [Devin CLI](https://docs.devin.ai/cli/index), [OpenCode](https://github.com/anomalyco/opencode), [Antigravity (`agy`)](https://github.com/google-antigravity/antigravity-cli), [Grok Build](https://github.com/xai-org/grok-build), [Hermes](https://github.com/NousResearch/hermes-agent), [DeepSeek Harness (`dsh`)](https://github.com/deepseek-ai/deepseek-harness), and Zcode through the same `run` and `resume` commands. Each keeps its own login, models, and tools.
28
+ - Isolate each run in a Git worktree or an empty sealed workspace, inside a [bubblewrap](https://github.com/containers/bubblewrap) sandbox, with optional network-level blocking of web access.
29
+ - Get one JSON result per run: the answer, timing, token usage, an API-equivalent cost, the harness session ID for exact resume, and whether the provider refused.
30
+ - Keep cleanup, checkpointing, and Git integration outside the agent's control.
31
+
32
+ [Agent Client Protocol (ACP)](https://agentclientprotocol.com/) connects agents to editors for interactive sessions, where a person chats, approves actions, and steers. AOP serves the opposite case: software running bounded jobs with nobody watching. The two are complementary.
33
+
34
+ ## Requirements
35
+
36
+ - Linux. AOP's sandbox is built on Linux namespaces.
37
+ - Python 3.11 or newer and [uv](https://docs.astral.sh/uv/).
38
+ - `bwrap` (bubblewrap) for the sandboxed profiles, with unprivileged user namespaces allowed. On Fedora and Debian, install the `bubblewrap` package. Ubuntu 24.04 and later also need `sudo sysctl kernel.apparmor_restrict_unprivileged_userns=0`.
39
+ - Git, for runs in a repository.
40
+ - At least one harness, installed and signed in. Codex needs version 0.150.1 or newer and Antigravity 1.1.16 or newer; [Harness adapters](docs/harnesses.md#supported-harnesses) lists the versions each adapter was last tested with.
41
+
42
+ ## Quickstart
43
+
44
+ Install AOP from [PyPI](https://pypi.org/project/agent-orchestration-process/):
45
+
46
+ ```sh
47
+ uv tool install agent-orchestration-process
48
+ ```
49
+
50
+ Run a sealed task. It sees no repository and no local files, only what you pass with `--input`:
51
+
52
+ ```sh
53
+ aop run --agent claude --profile sealed --prompt "Reply with only the word copper."
54
+ ```
55
+
56
+ The answer goes to stdout; the run's details go to stderr:
57
+
58
+ ```text
59
+ aop: run_id=74a91b9f-d64e-45cd-9c56-b308c8840d58 session_id=fba843cd-3050-4672-a39b-1deaa63ef5d9 time=2.13s tokens=16887 accounting=complete calculated_cost=$0.000171 provider_reported_cost=n/a billing=subscription artifacts=/home/me/.aop/runs/74a91b9f-d64e-45cd-9c56-b308c8840d58
60
+ copper
61
+ ```
62
+
63
+ AOP records the run in a `.aop` folder in the directory you ran it from. Add `--json` for the full result as one object, abridged here:
64
+
65
+ ```json
66
+ {
67
+ "run_id": "5f668bbc-cad7-4852-b1b3-00b021952dd2",
68
+ "provider": "claude",
69
+ "model": "claude-haiku-5-5",
70
+ "status": "succeeded",
71
+ "final_message": "copper",
72
+ "session_id": "24ec7778-cd70-450f-a1f1-9090ae51b889",
73
+ "duration_seconds": 2.18,
74
+ "error": null,
75
+ "usage": {"input_tokens": 16883, "cached_input_tokens": 11227, "output_tokens": 4},
76
+ "accounting_status": "complete",
77
+ "calculated_cost": {"amount_usd": 0.00124527, "currency": "USD", "priced_as": "claude-haiku-5-5"},
78
+ "billing": {"route": "subscription", "credential_source": "claude-oauth"},
79
+ "refusal": null
80
+ }
81
+ ```
82
+
83
+ Continue the same harness session from any directory with the run ID:
84
+
85
+ ```sh
86
+ aop resume 5f668bbc-cad7-4852-b1b3-00b021952dd2 --prompt "Now reply with silver."
87
+ ```
88
+
89
+ To work on a repository, run from inside it. AOP creates a worktree for the task and leaves your checkout alone:
90
+
91
+ ```sh
92
+ aop init
93
+ aop run task-a --agent codex --prompt "Implement the parser and its tests"
94
+ ```
95
+
96
+ `aop init` adds `/.aop/` to `.gitignore`; commit that before integrating tasks. See [Integration](docs/integration.md) for merging a task's changes back.
97
+
98
+ ## Safety defaults
99
+
100
+ AOP is built to run agents unattended, so know what a run can do before you start one:
101
+
102
+ > [!WARNING]
103
+ > - Harness permission prompts are always bypassed. The agent runs its tools without asking; the AOP sandbox and profile are the boundary.
104
+ > - The network is open unless you pass `--no-web`. Without it, an agent can reach any host, including to send data out.
105
+ > - The agent can read the credentials of its own harness, which AOP copies into the run so the harness can sign in.
106
+ > - The `host` profile has no sandbox: the agent has your full user account and filesystem.
107
+ > - AOP limits a run only by wall-clock time with `--timeout`. It has no spending, token, or resource cap.
108
+
109
+ [Running unattended](docs/unattended.md) covers what the sandbox does and does not stop, every result status and exit code, recovery after a crash, and cleanup.
110
+
111
+ ## Execution profiles
112
+
113
+ `--profile` sets what a run can see and change. The default is `edit`.
114
+
115
+ | Profile | Use it when |
116
+ | --- | --- |
117
+ | `edit` | The agent should modify an isolated task worktree. |
118
+ | `review` | The agent should inspect a task worktree without modifying it. |
119
+ | `sealed` | The agent should see only explicit input snapshots, with no repository or inherited instructions. |
120
+ | `host` | The command needs the native host environment and filesystem, with no sandbox. |
121
+
122
+ Preview a run's boundary without starting it:
123
+
124
+ ```sh
125
+ aop run --agent codex --profile sealed --prompt "Answer using only declared inputs" --dry-run
126
+ ```
127
+
128
+ See [Execution profiles](docs/profiles.md) for exact filesystem, instruction, environment, credential, input, and output boundaries, and [CLI guide](docs/cli.md#disable-model-controlled-web-access) for `--no-web`.
129
+
130
+ ## Documentation
131
+
132
+ - [CLI guide](docs/cli.md): running and resuming tasks, inputs and artifacts, batches, models, and cleanup.
133
+ - [Execution profiles](docs/profiles.md): what each profile exposes.
134
+ - [Running unattended](docs/unattended.md): safety, statuses, exit codes, and recovery.
135
+ - [Harness adapters](docs/harnesses.md): per-harness support, versions, and options.
136
+ - [Token usage and pricing](docs/token-usage.md): the usage and cost fields, and upgrading old run records.
137
+ - [Integration](docs/integration.md): checkpointing and merging task changes.
138
+ - [Glossary](docs/glossary.md): the terms these pages use.
139
+
140
+ ## Development
141
+
142
+ See [CONTRIBUTING.md](CONTRIBUTING.md) for setting up a checkout and running the tests, including the opt-in tests against installed harnesses.
@@ -0,0 +1,120 @@
1
+ # Agent Orchestration Process (AOP)
2
+
3
+ AOP is a command-line tool for running coding agents such as Claude Code and Codex as unattended jobs. It is for automation: CI jobs, batch experiments, benchmarks, and programs that hand work to an agent and need a dependable result back. One `aop run` command drives ten different agent CLIs, called harnesses here, and gives every run the same sandbox, deadline, and result record, so a caller does not need a separate adapter for each CLI's flags, sessions, and output.
4
+
5
+ - Run [Codex](https://github.com/openai/codex), [Claude Code](https://code.claude.com/docs/en/overview), [Cursor Agent](https://docs.cursor.com/en/cli/overview), [Devin CLI](https://docs.devin.ai/cli/index), [OpenCode](https://github.com/anomalyco/opencode), [Antigravity (`agy`)](https://github.com/google-antigravity/antigravity-cli), [Grok Build](https://github.com/xai-org/grok-build), [Hermes](https://github.com/NousResearch/hermes-agent), [DeepSeek Harness (`dsh`)](https://github.com/deepseek-ai/deepseek-harness), and Zcode through the same `run` and `resume` commands. Each keeps its own login, models, and tools.
6
+ - Isolate each run in a Git worktree or an empty sealed workspace, inside a [bubblewrap](https://github.com/containers/bubblewrap) sandbox, with optional network-level blocking of web access.
7
+ - Get one JSON result per run: the answer, timing, token usage, an API-equivalent cost, the harness session ID for exact resume, and whether the provider refused.
8
+ - Keep cleanup, checkpointing, and Git integration outside the agent's control.
9
+
10
+ [Agent Client Protocol (ACP)](https://agentclientprotocol.com/) connects agents to editors for interactive sessions, where a person chats, approves actions, and steers. AOP serves the opposite case: software running bounded jobs with nobody watching. The two are complementary.
11
+
12
+ ## Requirements
13
+
14
+ - Linux. AOP's sandbox is built on Linux namespaces.
15
+ - Python 3.11 or newer and [uv](https://docs.astral.sh/uv/).
16
+ - `bwrap` (bubblewrap) for the sandboxed profiles, with unprivileged user namespaces allowed. On Fedora and Debian, install the `bubblewrap` package. Ubuntu 24.04 and later also need `sudo sysctl kernel.apparmor_restrict_unprivileged_userns=0`.
17
+ - Git, for runs in a repository.
18
+ - At least one harness, installed and signed in. Codex needs version 0.150.1 or newer and Antigravity 1.1.16 or newer; [Harness adapters](docs/harnesses.md#supported-harnesses) lists the versions each adapter was last tested with.
19
+
20
+ ## Quickstart
21
+
22
+ Install AOP from [PyPI](https://pypi.org/project/agent-orchestration-process/):
23
+
24
+ ```sh
25
+ uv tool install agent-orchestration-process
26
+ ```
27
+
28
+ Run a sealed task. It sees no repository and no local files, only what you pass with `--input`:
29
+
30
+ ```sh
31
+ aop run --agent claude --profile sealed --prompt "Reply with only the word copper."
32
+ ```
33
+
34
+ The answer goes to stdout; the run's details go to stderr:
35
+
36
+ ```text
37
+ aop: run_id=74a91b9f-d64e-45cd-9c56-b308c8840d58 session_id=fba843cd-3050-4672-a39b-1deaa63ef5d9 time=2.13s tokens=16887 accounting=complete calculated_cost=$0.000171 provider_reported_cost=n/a billing=subscription artifacts=/home/me/.aop/runs/74a91b9f-d64e-45cd-9c56-b308c8840d58
38
+ copper
39
+ ```
40
+
41
+ AOP records the run in a `.aop` folder in the directory you ran it from. Add `--json` for the full result as one object, abridged here:
42
+
43
+ ```json
44
+ {
45
+ "run_id": "5f668bbc-cad7-4852-b1b3-00b021952dd2",
46
+ "provider": "claude",
47
+ "model": "claude-haiku-5-5",
48
+ "status": "succeeded",
49
+ "final_message": "copper",
50
+ "session_id": "24ec7778-cd70-450f-a1f1-9090ae51b889",
51
+ "duration_seconds": 2.18,
52
+ "error": null,
53
+ "usage": {"input_tokens": 16883, "cached_input_tokens": 11227, "output_tokens": 4},
54
+ "accounting_status": "complete",
55
+ "calculated_cost": {"amount_usd": 0.00124527, "currency": "USD", "priced_as": "claude-haiku-5-5"},
56
+ "billing": {"route": "subscription", "credential_source": "claude-oauth"},
57
+ "refusal": null
58
+ }
59
+ ```
60
+
61
+ Continue the same harness session from any directory with the run ID:
62
+
63
+ ```sh
64
+ aop resume 5f668bbc-cad7-4852-b1b3-00b021952dd2 --prompt "Now reply with silver."
65
+ ```
66
+
67
+ To work on a repository, run from inside it. AOP creates a worktree for the task and leaves your checkout alone:
68
+
69
+ ```sh
70
+ aop init
71
+ aop run task-a --agent codex --prompt "Implement the parser and its tests"
72
+ ```
73
+
74
+ `aop init` adds `/.aop/` to `.gitignore`; commit that before integrating tasks. See [Integration](docs/integration.md) for merging a task's changes back.
75
+
76
+ ## Safety defaults
77
+
78
+ AOP is built to run agents unattended, so know what a run can do before you start one:
79
+
80
+ > [!WARNING]
81
+ > - Harness permission prompts are always bypassed. The agent runs its tools without asking; the AOP sandbox and profile are the boundary.
82
+ > - The network is open unless you pass `--no-web`. Without it, an agent can reach any host, including to send data out.
83
+ > - The agent can read the credentials of its own harness, which AOP copies into the run so the harness can sign in.
84
+ > - The `host` profile has no sandbox: the agent has your full user account and filesystem.
85
+ > - AOP limits a run only by wall-clock time with `--timeout`. It has no spending, token, or resource cap.
86
+
87
+ [Running unattended](docs/unattended.md) covers what the sandbox does and does not stop, every result status and exit code, recovery after a crash, and cleanup.
88
+
89
+ ## Execution profiles
90
+
91
+ `--profile` sets what a run can see and change. The default is `edit`.
92
+
93
+ | Profile | Use it when |
94
+ | --- | --- |
95
+ | `edit` | The agent should modify an isolated task worktree. |
96
+ | `review` | The agent should inspect a task worktree without modifying it. |
97
+ | `sealed` | The agent should see only explicit input snapshots, with no repository or inherited instructions. |
98
+ | `host` | The command needs the native host environment and filesystem, with no sandbox. |
99
+
100
+ Preview a run's boundary without starting it:
101
+
102
+ ```sh
103
+ aop run --agent codex --profile sealed --prompt "Answer using only declared inputs" --dry-run
104
+ ```
105
+
106
+ See [Execution profiles](docs/profiles.md) for exact filesystem, instruction, environment, credential, input, and output boundaries, and [CLI guide](docs/cli.md#disable-model-controlled-web-access) for `--no-web`.
107
+
108
+ ## Documentation
109
+
110
+ - [CLI guide](docs/cli.md): running and resuming tasks, inputs and artifacts, batches, models, and cleanup.
111
+ - [Execution profiles](docs/profiles.md): what each profile exposes.
112
+ - [Running unattended](docs/unattended.md): safety, statuses, exit codes, and recovery.
113
+ - [Harness adapters](docs/harnesses.md): per-harness support, versions, and options.
114
+ - [Token usage and pricing](docs/token-usage.md): the usage and cost fields, and upgrading old run records.
115
+ - [Integration](docs/integration.md): checkpointing and merging task changes.
116
+ - [Glossary](docs/glossary.md): the terms these pages use.
117
+
118
+ ## Development
119
+
120
+ See [CONTRIBUTING.md](CONTRIBUTING.md) for setting up a checkout and running the tests, including the opt-in tests against installed harnesses.
@@ -4,7 +4,7 @@ build-backend = "uv_build"
4
4
 
5
5
  [project]
6
6
  name = "agent-orchestration-process"
7
- version = "0.3.0"
7
+ version = "0.4.0"
8
8
  description = "Run bounded, non-interactive jobs across agent harnesses with isolation, exact resume, and provenance"
9
9
  readme = "README.md"
10
10
  requires-python = ">=3.11"
@@ -4,7 +4,7 @@ build-backend = "uv_build"
4
4
 
5
5
  [project]
6
6
  name = "agent-orchestration-process"
7
- version = "0.3.0"
7
+ version = "0.4.0"
8
8
  description = "Run bounded, non-interactive jobs across agent harnesses with isolation, exact resume, and provenance"
9
9
  readme = "README.md"
10
10
  requires-python = ">=3.11"
@@ -27,6 +27,7 @@ from .runner import (
27
27
  AgentRunner,
28
28
  RunStore,
29
29
  adapter_for,
30
+ NATIVE_STATE_PROVIDERS,
30
31
  plan_inputs,
31
32
  provider_runtime_record,
32
33
  )
@@ -200,6 +201,20 @@ def build_parser() -> argparse.ArgumentParser:
200
201
  help="how Claude returns the model's thinking, such as official summaries "
201
202
  "(claude only)",
202
203
  )
204
+ run.add_argument(
205
+ "--workdir",
206
+ type=Path,
207
+ metavar="DIR",
208
+ help="with --profile host, run the harness in DIR itself, Git or not, "
209
+ "instead of a task worktree; TASK becomes optional and AOP keeps its "
210
+ "records in the current directory",
211
+ )
212
+ run.add_argument(
213
+ "--native-state",
214
+ action="store_true",
215
+ help="with --profile host, let the harness use its own home, settings, "
216
+ "memory and sessions instead of AOP's private copy (claude only)",
217
+ )
203
218
  run.add_argument(
204
219
  "--writable-workspace",
205
220
  action="store_true",
@@ -328,7 +343,7 @@ def main(argv: Sequence[str] | None = None) -> int:
328
343
  _report_policy(policy, json_output=args.json)
329
344
  return 0
330
345
 
331
- if args.command == "run" and args.profile == "sealed":
346
+ if args.command == "run" and (args.profile == "sealed" or args.workdir):
332
347
  manager = WorktreeManager.standalone(Path.cwd())
333
348
  elif args.command in {"resume", "cleanup"}:
334
349
  manager = _manager_holding_run(args.run_id)
@@ -344,6 +359,8 @@ def main(argv: Sequence[str] | None = None) -> int:
344
359
  request = RunStore(manager.state_dir / "runs").load_request(args.run_id)
345
360
  if request.profile == "sealed":
346
361
  _cleanup_sealed(manager, request)
362
+ elif request.workdir is not None:
363
+ pass # The caller's directory is not AOP's to remove.
347
364
  elif any(item.task == request.task for item in manager.list()):
348
365
  manager.remove(request.task, force=True)
349
366
  print(request.task)
@@ -396,8 +413,10 @@ def main(argv: Sequence[str] | None = None) -> int:
396
413
 
397
414
  if args.command == "run":
398
415
  task = args.task or str(uuid.uuid4())
399
- if args.task is None and args.profile != "sealed":
400
- raise AOPError("TASK is required unless --profile sealed is selected")
416
+ if args.task is None and args.profile != "sealed" and not args.workdir:
417
+ raise AOPError(
418
+ "TASK is required unless --profile sealed or --workdir is selected"
419
+ )
401
420
  preview = resolve_policy(
402
421
  args.profile,
403
422
  provider=args.agent,
@@ -405,6 +424,17 @@ def main(argv: Sequence[str] | None = None) -> int:
405
424
  input_names=tuple(Path(path).name for path in args.input_paths),
406
425
  ).to_dict()
407
426
  preview["provider"] = args.agent
427
+ workdir = AgentRunner._validated_workdir(args.workdir, args.profile)
428
+ if workdir is not None:
429
+ preview["workspace"]["content"] = "caller-directory"
430
+ preview["workspace"]["host_path"] = workdir
431
+ preview["working_directory"] = workdir
432
+ if args.native_state:
433
+ if args.profile != "host":
434
+ raise AOPError("--native-state applies only to the host profile")
435
+ if args.agent not in NATIVE_STATE_PROVIDERS:
436
+ raise AOPError(f"{args.agent} does not support --native-state")
437
+ preview["state"] = "harness-native"
408
438
  planned = plan_inputs(
409
439
  args.input_paths,
410
440
  excludes=args.input_excludes,
@@ -454,6 +484,8 @@ def main(argv: Sequence[str] | None = None) -> int:
454
484
  input_paths=args.input_paths,
455
485
  input_excludes=args.input_excludes,
456
486
  skip_external_links=args.skip_external_links,
487
+ workdir=args.workdir,
488
+ native_state=args.native_state,
457
489
  system_prompt=_read_system_prompt(args),
458
490
  tools=args.tools,
459
491
  provider_base_url=args.provider_base_url,
@@ -717,20 +749,36 @@ def _report_policy(
717
749
 
718
750
 
719
751
  def _manager_holding_run(run_id: str) -> WorktreeManager:
720
- """The state of the run: the repository's, or the current directory's.
752
+ """The state that holds the run, found from any directory.
721
753
 
722
- Sealed runs record themselves under the current directory even inside a
723
- repository, so a resume or cleanup looks in both places.
754
+ The repository or current directory comes first, since sealed runs
755
+ record themselves under the current directory even inside a repository.
756
+ Otherwise the run's archived result names its state root, and failing
757
+ that the registered state roots are searched.
724
758
  """
725
759
  try:
726
760
  repository = WorktreeManager.discover(Path.cwd())
727
761
  except AOPError:
728
762
  repository = None
729
- if repository is not None and (repository.state_dir / "runs" / run_id).is_dir():
730
- return repository
731
- if repository is None or (Path.cwd() / ".aop" / "runs" / run_id).is_dir():
732
- return WorktreeManager.standalone(Path.cwd())
733
- return repository
763
+ for state_dir in (
764
+ repository.state_dir if repository is not None else None,
765
+ Path.cwd() / ".aop",
766
+ _archived_state_root(run_id),
767
+ *registered_state_roots(),
768
+ ):
769
+ if state_dir is not None and (state_dir / "runs" / run_id).is_dir():
770
+ return WorktreeManager(state_dir.parent)
771
+ return repository or WorktreeManager.standalone(Path.cwd())
772
+
773
+
774
+ def _archived_state_root(run_id: str) -> Path | None:
775
+ try:
776
+ recorded = json.loads(
777
+ (results_archive_dir() / f"{run_id}.json").read_text()
778
+ )["state_root"]
779
+ except (OSError, ValueError, KeyError, TypeError):
780
+ return None
781
+ return Path(recorded) if isinstance(recorded, str) else None
734
782
 
735
783
 
736
784
  def _migrate_results(runs_dirs: Sequence[Path], *, dry_run: bool) -> int:
@@ -5,11 +5,10 @@ Usage: egress_bridge.py SOCKET PROXY_PORT [HOST:PORT ...] -- COMMAND...
5
5
  Inside the namespace, 127.0.0.1:PROXY_PORT reaches the controller's CONNECT
6
6
  proxy through SOCKET, and each HOST:PORT (a loopback endpoint) is forwarded to
7
7
  the same endpoint outside. The controller decides what is allowed, so this
8
- script holds no policy. It runs as the sandbox's PID 1: it forwards signals to
9
- the command, reaps orphans, and exits with the command's status.
8
+ script holds no policy. It forwards stop signals to the command and exits
9
+ with the command's status; bubblewrap's own process 1 collects orphans.
10
10
  """
11
11
 
12
- import os
13
12
  import signal
14
13
  import socket
15
14
  import subprocess
@@ -88,12 +87,8 @@ def main(argv):
88
87
  child = subprocess.Popen(command)
89
88
  for number in (signal.SIGTERM, signal.SIGINT, signal.SIGHUP):
90
89
  signal.signal(number, lambda signum, frame: child.send_signal(signum))
91
- # As PID 1, reap every orphan until the command itself exits.
92
- while True:
93
- pid, status = os.wait()
94
- if pid == child.pid:
95
- code = os.waitstatus_to_exitcode(status)
96
- return code if code >= 0 else 128 - code
90
+ code = child.wait()
91
+ return code if code >= 0 else 128 - code
97
92
 
98
93
 
99
94
  if __name__ == "__main__":
@@ -114,6 +114,10 @@ class RunRequest:
114
114
  # that leave their input are skipped instead of failing the run.
115
115
  input_excludes: tuple[str, ...] = ()
116
116
  skip_external_links: bool = False
117
+ # A host run in this directory itself instead of a task worktree, and
118
+ # whether the harness keeps its own home and state there.
119
+ workdir: str | None = None
120
+ native_state: bool = False
117
121
  # How `aop migrate-results` converted an older record, if it did.
118
122
  migration: dict[str, Any] | None = None
119
123
 
@@ -203,6 +207,10 @@ class RunResult:
203
207
  provider_duration_seconds: float | None = None
204
208
  provider_status: str | None = None
205
209
  provider_error: str | None = None
210
+ # Set when the provider refused the input: its category and explanation
211
+ # when reported, the models that refused, the messages the harness
212
+ # withdrew, and the line of events.jsonl where the refusal begins.
213
+ refusal: dict[str, Any] | None = None
206
214
  accounting_status: str = "complete"
207
215
  usage_schema: str = TOKEN_USAGE_SCHEMA
208
216
  # The .aop state directory the run was recorded in, as listed in the
@@ -91,6 +91,10 @@ sys.argv = sys.argv[1:]
91
91
  flag = '--prompt' if '--prompt' in sys.argv else '-q'
92
92
  i = sys.argv.index(flag) + 1
93
93
  sys.argv[i] = pathlib.Path(sys.argv[i]).read_bytes().decode('utf-8')
94
- sys.path.insert(0, str(pathlib.Path(sys.argv[0]).resolve().parent))
95
- runpy.run_path(sys.argv[0], run_name='__main__')
94
+ if sys.argv[0] == '-c':
95
+ code = sys.argv.pop(1)
96
+ exec(compile(code, '<string>', 'exec'), {'__name__': '__main__'})
97
+ else:
98
+ sys.path.insert(0, str(pathlib.Path(sys.argv[0]).resolve().parent))
99
+ runpy.run_path(sys.argv[0], run_name='__main__')
96
100
  """