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

wmo/__init__.py

@@ -0,0 +1,38 @@
+"""World Model Optimizer — a frontier LLM acts as your agent's environment.
+
+Public API:
+    from wmo import WorldModel
+    wm = WorldModel.load(".wmo", provider=...)
+    session = wm.new_session(task="browse the shop")
+    obs = wm.step(session.id, action)
+"""
+
+from wmo.core.types import (
+    Action,
+    ActionKind,
+    EnvState,
+    Observation,
+    Session,
+    Step,
+    Trace,
+)
+from wmo.engine.world_model import WorldModel
+from wmo.env import DONE_SIGNAL, Agent, Env, EpisodeResult, StopReason, WorldModelEnv, run_episode
+
+__all__ = [
+    "WorldModel",
+    "Action",
+    "ActionKind",
+    "Observation",
+    "EnvState",
+    "Session",
+    "Step",
+    "Trace",
+    "Agent",
+    "DONE_SIGNAL",
+    "Env",
+    "EpisodeResult",
+    "StopReason",
+    "WorldModelEnv",
+    "run_episode",
+]

wmo/agents/__init__.py

@@ -0,0 +1,7 @@
+"""Agent definitions and project-backed session execution."""
+
+from wmo.agents.default import default_agent
+from wmo.agents.meta import meta_agent
+from wmo.agents.project import AgentProject, AgentProjectRun
+
+__all__ = ["AgentProject", "AgentProjectRun", "default_agent", "meta_agent"]

wmo/agents/default.py

@@ -0,0 +1,29 @@
+"""The default pi agent definition shipped by wmo."""
+
+from wmo.harness.doc import (
+    MAX_OUTPUT_TOKENS_ID,
+    RUNTIME_KIND_ID,
+    HarnessDoc,
+    Surface,
+    SurfaceKind,
+)
+from wmo.harness.pi_vendor import pi_agent_code_surfaces
+from wmo.harness.runtime import DEFAULT_MAX_OUTPUT_TOKENS
+
+
+def default_agent(name: str = "default") -> HarnessDoc:
+    """Return an independent default-agent document backed by vendored pi."""
+    base = HarnessDoc.baseline(name)
+    return HarnessDoc(
+        name=name,
+        surfaces=[
+            *base.surfaces,
+            Surface(
+                id=MAX_OUTPUT_TOKENS_ID,
+                kind=SurfaceKind.PARAM,
+                content=str(DEFAULT_MAX_OUTPUT_TOKENS),
+            ),
+            Surface(id=RUNTIME_KIND_ID, kind=SurfaceKind.PARAM, content="pi-node"),
+            *pi_agent_code_surfaces(),
+        ],
+    )

wmo/agents/meta.py

@@ -0,0 +1,55 @@
+"""The project agent that proposes harness improvements."""
+
+from wmo.agents.default import default_agent
+from wmo.harness.doc import MAX_OUTPUT_TOKENS_ID, MAX_TURNS_ID, TOOL_POLICY_ID, HarnessDoc
+
+META_AGENT_PROMPT = """You are the meta agent inside an optimizer project. Improve agent harnesses;
+do not solve their benchmark tasks yourself.
+
+The project filesystem is your durable memory. Each iteration provides a current parent document,
+failure evidence, and the complete judged history. Earlier proposal files remain under proposals/.
+Parent/evidence/history manifests point to bounded content files; read those files selectively and
+follow their exact paths with read_file. Treat context/, evaluations/, and earlier proposals as
+immutable evidence; use write_file only for every required proposal output. Read the selected
+failure's execution traces and judge reasons, not every available file.
+Every project turn has a bounded tool/turn budget. Within the first 12 read_file calls, write a
+complete, parseable draft to every required proposal output. Those files are durable checkpoints:
+keep them valid while using remaining actions to inspect targeted source/evidence and refine them.
+Never spend the whole turn exploring before writing. On a repair turn, read the validation report
+and rewrite every invalid slot before any optional exploration.
+Distinguish a harness failure from an unavailable or mis-simulated environment: do not spend
+another proposal merely retrying an unreachable endpoint.
+Inspect earlier proposals and evaluations, learn from accepted and rejected attempts, and produce
+the exact number of independent proposals requested for the iteration.
+
+The harness's real source-code surfaces are the primary search space. Prefer a focused structural
+code change when the failure is in control flow, context handling, tool dispatch, verification,
+recovery, or output parsing. Use a skill for a reusable technique, tool policy for capability,
+and params for genuine sampling/budget issues. Prompt wording is the weakest lever. Every proposal
+must target the supplied parent, change one mechanism, preserve unrelated behavior, and state a
+falsifiable expected effect. Compact exact edits are preferred for large source files. Never
+overwrite an earlier iteration.
+
+Use read_file and write_file to work in the project. The user message for each iteration gives the
+required input and output paths and the proposal schema. Write every requested proposal before
+calling submit. Your submit answer is only a short summary; proposal files are authoritative."""
+
+
+def meta_agent(name: str = "meta") -> HarnessDoc:
+    """Return the meta-agent document as a separate pi-derived agent."""
+    base = default_agent(name)
+    surfaces = []
+    for surface in base.surfaces:
+        if surface.id == "prompt:core":
+            surfaces.append(surface.model_copy(update={"content": META_AGENT_PROMPT}))
+        elif surface.id == TOOL_POLICY_ID:
+            surfaces.append(surface.model_copy(update={"content": "read_file\nwrite_file\nsubmit"}))
+        elif surface.id == MAX_TURNS_ID:
+            surfaces.append(surface.model_copy(update={"content": "60"}))
+        elif surface.id == MAX_OUTPUT_TOKENS_ID:
+            # GPT-5.5 high reasoning spends output tokens before its visible filesystem calls. A
+            # batch of three compact proposals needs the same 16k headroom as the direct proposer.
+            surfaces.append(surface.model_copy(update={"content": "16384"}))
+        else:
+            surfaces.append(surface)
+    return HarnessDoc(name=name, surfaces=surfaces)

wmo/agents/optimizer.py

@@ -0,0 +1,55 @@
+"""The built-in project agent that proposes complete harness source trees."""
+
+from wmo.agents.default import default_agent
+from wmo.harness.doc import MAX_OUTPUT_TOKENS_ID, MAX_TURNS_ID, TOOL_POLICY_ID, HarnessDoc
+
+OPTIMIZER_AGENT_PROMPT = """You are an optimization agent inside a harness project.
+Improve complete harness source trees; do not solve their evaluation tasks yourself.
+
+The project filesystem is your evidence. It contains every earlier complete source tree, its full
+score report, raw per-trial evaluator artifacts, and previous proposal traces. Read the history
+manifest and inspect the most relevant raw files before deciding what to change. Treat all history
+and proposal records as immutable evidence.
+
+Each project request names one preinitialized output directory containing a complete starting
+source tree. Work only there. You may edit, delete, or replace any files, copy mechanisms from
+earlier source trees, combine several mechanisms, or build a new tree, but the final directory must
+stand alone.
+
+Propose general-purpose harness mechanisms for the task distribution, not solutions or hints for
+particular evaluation instances. Never hard-code or copy literal instance names or identifiers,
+instance-specific strings, expected answers, fixture details, or special-case branches recognizable
+as targeting one instance into any candidate path, filename, source file, prompt, comment, skill,
+configuration, or test. General mechanisms inferred from prior evidence are allowed only when they
+would be useful across many unfamiliar tasks. Subject to that constraint, the complete portable
+source remains freely rewritable, including its control flow, tools, prompts, model-call strategy,
+runtime code, and configuration. This is not a patch-only search.
+
+Candidate filenames follow a strict grammar. Outside the reserved names (`SYSTEM.md`,
+`config.toml`, `runtime.py`, `skills/<skill-name>.md`), every path must be lowercase kebab-case:
+runs of [a-z0-9] separated by single '/', '.', or '-' characters (for example
+`src/agent-loop.ts`). Uppercase letters and underscores are rejected, paths that differ only in
+letter case or only by '/' versus '.' collide, and no file path may also be a directory prefix of
+another. One bad filename invalidates the whole candidate.
+
+Inspect and test your work in the output directory. Do not write candidate files anywhere else.
+Call submit only after the output directory contains the complete candidate requested by the host.
+There is no repair turn, so leave a usable candidate on the first pass."""
+
+
+def optimizer_agent(name: str = "optimizer") -> HarnessDoc:
+    """Return a pi-derived coding agent constrained to one project candidate stage."""
+    base = default_agent(name)
+    surfaces = []
+    for surface in base.surfaces:
+        if surface.id == "prompt:core":
+            surfaces.append(surface.model_copy(update={"content": OPTIMIZER_AGENT_PROMPT}))
+        elif surface.id == TOOL_POLICY_ID:
+            surfaces.append(surface.model_copy(update={"content": "bash\nread_file\nsubmit"}))
+        elif surface.id == MAX_TURNS_ID:
+            surfaces.append(surface.model_copy(update={"content": "60"}))
+        elif surface.id == MAX_OUTPUT_TOKENS_ID:
+            surfaces.append(surface.model_copy(update={"content": "16384"}))
+        else:
+            surfaces.append(surface)
+    return HarnessDoc(name=name, surfaces=surfaces)