world-model-optimizer 0.2.0__py3-none-any.whl

wmo/harness/tools.py

@@ -0,0 +1,105 @@
+"""The rollout agent's tool surface: a deliberately minimal set, rendered compactly.
+
+The agent gets four tools: `bash` + file read/write against the environment, and `submit` to end
+the run. Everything else the agent needs it composes with bash — a small always-loaded tool schema
+keeps the prompt cheap, and capability comes from composition rather than tool count.
+
+Env tools become `Action`s the environment answers (the world model, in closed-loop eval); `submit`
+is handled by the runtime itself.
+"""
+
+from __future__ import annotations
+
+from pydantic import BaseModel, Field, ValidationError
+
+from wmo.core.parsing import extract_json_object
+from wmo.core.types import Action, ActionKind, JsonObject
+
+
+class ToolSpec(BaseModel):
+    """One tool the agent may call: a name, what it does, and its named arguments."""
+
+    name: str
+    description: str
+    arguments: dict[str, str] = Field(default_factory=dict)  # arg name -> one-line description
+
+
+BASH = ToolSpec(
+    name="bash",
+    description="Run a shell command in the environment; returns stdout+stderr and exit status.",
+    arguments={"command": "the shell command to run"},
+)
+READ_FILE = ToolSpec(
+    name="read_file",
+    description="Read a file from the environment.",
+    arguments={"path": "absolute path of the file to read"},
+)
+WRITE_FILE = ToolSpec(
+    name="write_file",
+    description="Write a file in the environment (parent dirs are created).",
+    arguments={"path": "absolute path to write", "content": "full file content"},
+)
+SUBMIT = ToolSpec(
+    name="submit",
+    description="Finish the task and submit your answer/result summary. This ends the run.",
+    arguments={"answer": "your final answer or a summary of what you did"},
+)
+READ_SKILL = ToolSpec(
+    name="read_skill",
+    description="Read the full body of a skill from your library (the prompt lists only names).",
+    arguments={"name": "the skill name to read"},
+)
+
+TOOL_REGISTRY: dict[str, ToolSpec] = {
+    t.name: t for t in (BASH, READ_FILE, WRITE_FILE, SUBMIT, READ_SKILL)
+}
+
+# `read_skill` is only useful when a harness ships skills, so it is not a default tool; the runtime
+# adds it automatically when constructed with a non-empty skill library.
+DEFAULT_TOOLS = [t.name for t in (BASH, READ_FILE, WRITE_FILE, SUBMIT)]
+
+
+def resolve_tools(names: list[str]) -> list[ToolSpec]:
+    """Map tool names to specs, raising on unknown names. `submit` is mandatory (ends the run)."""
+    unknown = [n for n in names if n not in TOOL_REGISTRY]
+    if unknown:
+        raise ValueError(f"unknown tools {unknown}; registry has {sorted(TOOL_REGISTRY)}")
+    if SUBMIT.name not in names:
+        raise ValueError(f"the {SUBMIT.name!r} tool is required (without it a run cannot end)")
+    return [TOOL_REGISTRY[n] for n in names]
+
+
+def render_tools(tools: list[ToolSpec]) -> str:
+    """Render the tool list for the system prompt, one compact block per tool."""
+    lines: list[str] = []
+    for tool in tools:
+        args = ", ".join(f'"{k}": <{v}>' for k, v in tool.arguments.items())
+        lines.append(f"- {tool.name}: {tool.description}\n  arguments: {{{args}}}")
+    return "\n".join(lines)
+
+
+class ToolCall(BaseModel):
+    """One parsed tool call from the agent's reply."""
+
+    tool: str
+    arguments: JsonObject = Field(default_factory=dict)
+
+
+def parse_tool_call(text: str) -> ToolCall | None:
+    """Parse the agent's reply into a ToolCall (`{"tool": ..., "arguments": {...}}`), or None.
+
+    Lenient about surrounding prose/fences (the JSON is extracted, not matched), strict about
+    shape: a reply whose JSON has no string `tool` field is not a call.
+    """
+    raw = extract_json_object(text)
+    if raw is None:
+        return None
+    try:
+        return ToolCall.model_validate_json(raw)
+    except ValidationError:
+        return None
+
+
+def to_action(call: ToolCall) -> Action:
+    """An env tool call as the normalized Action the environment answers."""
+    return Action(kind=ActionKind.TOOL_CALL, name=call.tool, arguments=call.arguments)

wmo/harness/vendor/manifest.sha256

@@ -0,0 +1,58 @@
+SHA256 (pi-agent/CHANGELOG.md) = fa69af54b48c7f9a841b78a62a7d948e518f15504a9c7c8ea1f6d37abd96984e
+SHA256 (pi-agent/LICENSE) = 0457f5bcec3b3b211605dfb5d1a49042fd638f3686a410fe099c24a25af13c48
+SHA256 (pi-agent/README.md) = 966f93b5ba3ddc1c9b472ab3a63878244dc89f860826bc7b29a8d35fe7fd931a
+SHA256 (pi-agent/VENDOR.md) = a8ba9ec63411f27769c4d885344bd1c1e9f38be9b285a9f44c5853cf8551abcc
+SHA256 (pi-agent/docs/agent-harness.md) = 5a455054ad5a7ad377787289c70ceeaaaf0c0cbe363c766392e7d9e742df31e8
+SHA256 (pi-agent/docs/durable-harness.md) = 52895b5c89e6927ba0a4ee44ef5fa5fd3c09273b00c6ea202cedc66692b10dfd
+SHA256 (pi-agent/docs/hooks.md) = 5977b25b4e7dc2316bb7cd917955aeb1882fd1d77e7fbcbcb990e4fba5e26fc6
+SHA256 (pi-agent/docs/models.md) = b108da745f4afaf64a65b4bd2b50388efa0964556e9e9907eea2cdade3f6a323
+SHA256 (pi-agent/docs/observability.md) = 2749c2dbe8e1856e18b5ba0e61a27ffaffbcab3df4e24c3fd109cab84cc18fe5
+SHA256 (pi-agent/package.json) = e04b0a69effd94098f80258d5fb1b5a6807f8c39baba8d328a8936adc16cded3
+SHA256 (pi-agent/src/agent-loop.ts) = 64b9ba4aca90a6ff6a12b7b9392bd1ae4a95e3c463bf8d2ffe66750e69e483ea
+SHA256 (pi-agent/src/agent.ts) = 28bd9032a9ec7f1b6b3a55c6c876c0332879d51030483cb191cde8b53be9646f
+SHA256 (pi-agent/src/harness/agent-harness.ts) = 4f08ed1c3697540d2668746b8db94f18b3215c30a0d42b75f2162be51b0d72c6
+SHA256 (pi-agent/src/harness/compaction/branch-summarization.ts) = a3073bc85c273e6a6fda866a4da80b0115273cba9dc94e15ec357db049dd7e86
+SHA256 (pi-agent/src/harness/compaction/compaction.ts) = 1789101d10d4c54bc10e11977ef643f601647046a28a22f988d0178d3770888d
+SHA256 (pi-agent/src/harness/compaction/utils.ts) = 559c5915660a53a24aeca8ba9018be41f67aacb4ca1c40089e820c5881545827
+SHA256 (pi-agent/src/harness/env/nodejs.ts) = 07f510286e412254ed12bcb7ca3e3e24d757b6d53fc76bb1f40ccec7d8c302af
+SHA256 (pi-agent/src/harness/messages.ts) = 8630aa25733cdebf61ccc048dae3c93e6eab2b1cc414af2a00e37276478355ed
+SHA256 (pi-agent/src/harness/prompt-templates.ts) = 81aee96ee83455d767b9388be2f51e9c271b6b1730cabb09c307a470c5e94fe0
+SHA256 (pi-agent/src/harness/session/jsonl-repo.ts) = 6b2e679a44b70aab82639aade743d16c599704e0beed591bfd3b99f7b37f4598
+SHA256 (pi-agent/src/harness/session/jsonl-storage.ts) = 2225e272c0ad2813de834e7faecc09fa8e065ee18a9c38842e83da98246b38a8
+SHA256 (pi-agent/src/harness/session/memory-repo.ts) = ca40f7a3d51b77fe44dfe0520c20846255e08028bf99fd3afb836239a532b322
+SHA256 (pi-agent/src/harness/session/memory-storage.ts) = cf9f78e88b7bdda862434bd5485f13b037c90061123b6d7a0774a9093504b3fc
+SHA256 (pi-agent/src/harness/session/repo-utils.ts) = daf8b870a89399c259f850879cb0ebd93463adebfb79ccdacfe9050f8256b840
+SHA256 (pi-agent/src/harness/session/session.ts) = 14f9623b300fe09b98d03bf940dc79f9f6b55217135fd2d0c9c9868bd39434af
+SHA256 (pi-agent/src/harness/session/uuid.ts) = bf5ab460b3843dc497a6632b6a91e21af9fb1b2595ca9d45d93d9c232476f6e2
+SHA256 (pi-agent/src/harness/skills.ts) = 96041f33a64660b2efcf32d23b5607daa66df44ffef33b0b1c9ddac1c4d56693
+SHA256 (pi-agent/src/harness/system-prompt.ts) = a3a8f4e2e343562277d348e99762447131632aea963c9ee4633369e5d1a9e9f4
+SHA256 (pi-agent/src/harness/types.ts) = 4c1bae2a31545246b74cac6efd7463f8c8e584083c9e1bb23a23d1ff39df779f
+SHA256 (pi-agent/src/harness/utils/shell-output.ts) = 7449d6f119933b1c0a22b336f303675adbb529a460465617e527eaadbfc80ba6
+SHA256 (pi-agent/src/harness/utils/truncate.ts) = e76c92247c1fc5c6d67c69c615bc509eecd703a6e326177a646ec0ab1d56decb
+SHA256 (pi-agent/src/index.ts) = 9462120578a090d58dca0f28ef8af4d056f2d43b6516cd0ee596c91a2f840aef
+SHA256 (pi-agent/src/node.ts) = 63d8dc9b22d93b3b1172ed1f1e83cc62bb370789f56f8dd31affb5c2a0a8dff3
+SHA256 (pi-agent/src/proxy.ts) = 504a0c801264c91d5346893e2e77570d611bf748cd1d0d900564370b4afd97fb
+SHA256 (pi-agent/src/types.ts) = 1247af46daf185be14846871c110259246008f1166982b672abad34f7b2c6022
+SHA256 (pi-agent/test/agent-loop.test.ts) = 567a12ff9a4fa34622b4830eda501ae22d1629a09f8a8f7dd5e9710bbeab794b
+SHA256 (pi-agent/test/agent.test.ts) = 501f535b30908ac8993873dfc77a9ffed52090f3419d8a758e35cf7c20533a01
+SHA256 (pi-agent/test/e2e.test.ts) = 73d60e6e90f0e88805c3953bfaa93e27efcb8f64186cced226ccea469d1b6dee
+SHA256 (pi-agent/test/harness/agent-harness-stream.test.ts) = cc4e7a9b744a809e9af8c5f77226358b5ae3a14e90483019a07581a5c76a6117
+SHA256 (pi-agent/test/harness/agent-harness.test.ts) = 3914ff96ae93ff1a5a28f9c020c293078dc131be84a952a6a8fc9e42ab7585c5
+SHA256 (pi-agent/test/harness/compaction.test.ts) = 6df83f072c9d7c5a9b5274062888bfa2f4c9165639f02b665795b398e5cfec81
+SHA256 (pi-agent/test/harness/nodejs-env.test.ts) = 8d3eb536d6965181669025e9730d011d9351620de683a35059c63adefa61cfe5
+SHA256 (pi-agent/test/harness/prompt-templates.test.ts) = a9b4834539edf0d9e89340c995f908490a6b2b9823af815460e0fadabd7b0e68
+SHA256 (pi-agent/test/harness/repo.test.ts) = c56dc9a448bda5612c1b645552e2f6581d9e827509fb615b94a26c9fc4c18a13
+SHA256 (pi-agent/test/harness/resource-formatting.test.ts) = 7bc3d164339f2642327a576a905389e12b7081bec77e4315712c9f7576115a15
+SHA256 (pi-agent/test/harness/session-test-utils.ts) = bdc012094e8dc5dfc53309885f616e8136fdf0f4cb9f1c0b027663ff0bb342fe
+SHA256 (pi-agent/test/harness/session-uuid.test.ts) = 6e7ac91166160d5a8643bca59f1b59f1de60a5a17424f6dacf39a11c8d9d08c8
+SHA256 (pi-agent/test/harness/session.test.ts) = cd6c9c51180080c5b4791a4ab9e28fd843cf2a7053ae79edafe721abd69db327
+SHA256 (pi-agent/test/harness/skills.test.ts) = 5cb717a1c0fdad7ce4a5ddea6c4d51aa91fc4a5863af4aaf543519f6a3f39af3
+SHA256 (pi-agent/test/harness/storage.test.ts) = e9dc317eb6159623427adae36c55b59c78b417e57a9cc6168afc791102612bf2
+SHA256 (pi-agent/test/harness/system-prompt.test.ts) = 40dafee3aedf176cd02c19c932fbc16442e016a5b29b254d7c76bd0197db3c5f
+SHA256 (pi-agent/test/harness/truncate.test.ts) = 45eaa130c61d02dbbd7a0861cfcac6ce592c136decbcc1a7c0e8be117f23a55e
+SHA256 (pi-agent/test/scratch/simple.ts) = 225766902e8c6e6e1829b59bf61dc793cfbaa8ac6df43494778fe6e7c0596db3
+SHA256 (pi-agent/test/utils/calculate.ts) = 125727abe1b8c6380650f5180fa561fe52dc6858d6b8bf1ee5f3dfa2234b2d41
+SHA256 (pi-agent/test/utils/get-current-time.ts) = 0192fa9720f3713c7ca6be54968c49f681eef6d01ea944cf418863b9c56d027b
+SHA256 (pi-agent/tsconfig.build.json) = a298e6122b4fc4d958c929033a78cb2eb2f95a0123ce516d246d0481aeaee4a0
+SHA256 (pi-agent/vitest.config.ts) = bd9d2fe6aeec2b28b13eb972c73d4e97feff28d4d183f49376eee8f351729b93
+SHA256 (pi-agent/vitest.harness.config.ts) = c031af7542703dfd655e85424c35fd6f980fdb36dc4ccc067efc87b5d700b99b