weakincentives 0.9.0

Tools for developing and optimizing side effect free background agents

Weak Incentives (Is All You Need)

Weak Incentives is a Python library for building "background agents" (automated AI systems). It provides lean, typed, and composable building blocks that keep determinism, testability, and safe execution front and center without relying on heavy dependencies or hosted services.

The core philosophy treats agent development as a structured engineering discipline rather than an exercise in ad-hoc scripting. The library ships a set of focused abstractions that reinforce that rigor:

• Prompts as code. Prompts are typed Prompt objects composed from sections such as MarkdownSection, versioned deterministically, and scoped with chapters so long-form directives can stay dormant until policies open them.
• Structured I/O. Declaring a dataclass output type for a Prompt[OutputT] automatically builds a JSON schema, instructs the provider to obey it, and parses the reply back into a typed Python object.
• Stateful, replayable sessions. Session acts as a Redux-like state store, letting pure reducers respond to events (for example, ToolInvoked) so every change is observable, replayable, and snapshot-friendly.
• Typed and sandboxed tools. Tools are typed callables (Tool[ParamsT, ResultT]) with explicit contracts for inputs and outputs, plus built-in suites for planning, a secure in-memory VFS, and sandboxed Python evaluation.
• Provider-agnostic adapters. Adapters connect the framework to providers like OpenAI or LiteLLM by handling API calls, tool negotiation, and response parsing while keeping the agent logic model-agnostic.
• Configuration and optimization hooks. Structured logging via structlog, enforced deadlines, and a powerful prompt overrides system let teams A/B test and iterate on prompts through JSON files without touching the application source.

What's novel?

While other agent frameworks provide a toolbox of loose components, Weak Incentives offers an opinionated chassis that emphasizes determinism, type contracts, and observable workflows:

1. Redux-like state management with reducers. Every state change is a traceable consequence of a published event processed by a pure reducer. This thread of causality delivers replayability and visibility far beyond free-form dictionaries or mutable object properties.
2. Composable prompt blueprints with typed contracts. Prompts are built from reusable sections and chapters backed by dataclasses, so composition and parameter binding feel like standard software engineering instead of string concatenation.
3. Integrated, hash-based prompt overrides. PromptDescriptor content hashes ensure overrides only apply to the intended section version, and LocalPromptOverridesStore keeps the JSON artifacts in version control so teams can collaborate without risking stale edits.
4. First-class in-memory virtual filesystem. The sandboxed VFS ships as a core tool, giving agents a secure workspace whose state is tracked like any other session slice and avoiding accidental host access.
5. Lean dependency surface. Avoiding heavyweight stacks such as Pydantic keeps the core lightweight. Custom serde modules provide the needed functionality without saddling users with sprawling dependency trees.

In short, Weak Incentives favors software-engineering discipline—determinism, type safety, testability, and clear state management—over maximizing the number of exposed knobs.

The specs below dive into each area when you need exact contracts and deeper context:

• Observable session state with reducer hooks. A Redux-like session ledger and in-process event bus keep every tool call and prompt render replayable. Built-in planning, virtual filesystem, and Python-evaluation sections ship with reducers that enforce domain rules while emitting structured telemetry. See Session State, Prompt Event Emission, Planning Tools, Virtual Filesystem Tools, and Asteval Integration.
• Composable prompt blueprints with strict contracts. Dataclass-backed sections compose into reusable blueprints that render validated Markdown and expose tool contracts automatically. Specs: Prompt Overview, Prompt Composition, and Structured Output.
• Chapter-driven visibility controls. Chapters gate when prompt regions enter the model context, defaulting to closed until runtime policies open them. Expansion strategies and lifecycle guidance live in Chapters Specification.
• Override-friendly workflows that scale into optimization. Prompt definitions ship with hash-based descriptors and on-disk overrides that stay in sync through schema validation and Git-root discovery, laying the groundwork for iterative optimization. Review Prompt Overrides for the full contract.
• Provider adapters standardize tool negotiation. Shared conversation loops negotiate tool calls, apply JSON-schema response formats, and normalize structured payloads so the runtime stays model-agnostic. See Adapter Specification and provider-specific docs such as LiteLLM Adapter.
• Local-first, deterministic execution. Everything runs locally without hosted dependencies, and prompt renders stay diff-friendly so version control captures intent instead of churn. The code-review example ties it together with override-aware prompts, session telemetry, and replayable tooling.

Requirements

• Python 3.12+ (the repository pins 3.12 in .python-version for development)
• uv CLI

Install

uv add weakincentives
# optional tool extras
uv add "weakincentives[asteval]"
# optional provider adapters
uv add "weakincentives[openai]"
uv add "weakincentives[litellm]"
# cloning the repo? use: uv sync --extra asteval --extra openai --extra litellm

Tutorial: Build a Stateful Code-Reviewing Agent

Use Weak Incentives to assemble a reproducible reviewer that tracks every decision, stages edits safely, and answers quick calculations inline. The runtime already ships with a session ledger and override-aware prompts, so you avoid custom state stores or ad-hoc optimizers.

1. Model review data and expected outputs

Typed dataclasses keep inputs and outputs honest so adapters emit consistent telemetry and structured responses stay predictable. See Dataclass Serde Utilities and Structured Output via Prompt[OutputT] for the validation and JSON-contract details behind this snippet.

from dataclasses import dataclass


@dataclass
class PullRequestContext:
    repository: str
    title: str
    body: str
    files_summary: str


@dataclass
class ReviewComment:
    file_path: str
    line: int
    severity: str
    summary: str
    rationale: str


@dataclass
class ReviewBundle:
    comments: tuple[ReviewComment, ...]
    overall_assessment: str

2. Create a session, surface built-in tool suites, and mount diffs

Planning, virtual filesystem, and Python-evaluation sections register reducers on the session so every run supports plans, staged edits, and quick calculations. Mount diffs ahead of time so the agent can read them through the virtual filesystem without extra callbacks. Install the asteval extra (uv add "weakincentives[asteval]") before instantiating AstevalSection so the sandbox is available at runtime. Specs worth skimming: Session State, Prompt Event Emission, Virtual Filesystem Tools, Planning Tools, and Asteval Integration.

from pathlib import Path

from weakincentives.runtime.events import InProcessEventBus, PromptExecuted
from weakincentives.runtime.session import Session
from weakincentives.tools import (
    AstevalSection,
    HostMount,
    PlanningToolsSection,
    VfsPath,
    VfsToolsSection,
)


bus = InProcessEventBus()
session = Session(bus=bus)


diff_root = Path("/srv/agent-mounts")
diff_root.mkdir(parents=True, exist_ok=True)
vfs_section = VfsToolsSection(
    session=session,
    allowed_host_roots=(diff_root,),
    mounts=(
        HostMount(
            host_path="octo_widgets/cache-layer.diff",
            mount_path=VfsPath(("diffs", "cache-layer.diff")),
        ),
    ),
)
planning_section = PlanningToolsSection(session=session)
asteval_section = AstevalSection(session=session)


def log_prompt(event: PromptExecuted) -> None:
    print(
        f"Prompt {event.prompt_name} completed with "
        f"{len(event.result.tool_results)} tool calls"
    )


bus.subscribe(PromptExecuted, log_prompt)

Copy unified diff files into /srv/agent-mounts before launching the run. The host mount resolves octo_widgets/cache-layer.diff relative to that directory and exposes it to the agent as diffs/cache-layer.diff inside the virtual filesystem snapshot. PlanningToolsSection, AstevalSection, and VfsToolsSection all register reducers when constructed, so wire them up with the Session you'll pass through ToolContext when dispatching tools.

3. Define a symbol search helper tool

Tools are typed callables that return structured results. Add lightweight helpers alongside the built-in suites—in this case, a symbol searcher that reads from a repo mounted at /srv/agent-repo. Review the Tool Runtime Specification to match the handler, ToolContext, and ToolResult contracts.

from dataclasses import dataclass
from pathlib import Path

from weakincentives.prompt import Tool, ToolResult


@dataclass
class SymbolSearchRequest:
    query: str
    file_glob: str = "*.py"
    max_results: int = 5


@dataclass
class SymbolMatch:
    file_path: str
    line: int
    snippet: str


@dataclass
class SymbolSearchResult:
    matches: tuple[SymbolMatch, ...]


repo_root = Path("/srv/agent-repo")


def find_symbol(params: SymbolSearchRequest) -> ToolResult[SymbolSearchResult]:
    if not repo_root.exists():
        raise FileNotFoundError(
            "Mount a repository checkout at /srv/agent-repo before running the agent."
        )

    matches: list[SymbolMatch] = []
    for file_path in repo_root.rglob(params.file_glob):
        if not file_path.is_file():
            continue
        with file_path.open("r", encoding="utf-8") as handle:
            for line_number, line in enumerate(handle, start=1):
                if params.query in line:
                    matches.append(
                        SymbolMatch(
                            file_path=str(file_path.relative_to(repo_root)),
                            line=line_number,
                            snippet=line.strip(),
                        )
                    )
                    if len(matches) >= params.max_results:
                        break
        if len(matches) >= params.max_results:
            break

    return ToolResult(
        message=f"Found {len(matches)} matching snippets.",
        value=SymbolSearchResult(matches=tuple(matches)),
    )


symbol_search_tool = Tool[SymbolSearchRequest, SymbolSearchResult](
    name="symbol_search",
    description=(
        "Search the repository checkout for a symbol and return file snippets."
    ),
    handler=find_symbol,
)

Session reducers accumulate structured state across prompt and tool events. When the symbol_search tool returns results, register a reducer that records the queries the reviewer explored along with the snippets that satisfied each one. Downstream sections can inspect this slice with session.select_all(ReviewedSymbol) to summarize the investigation history.

from dataclasses import dataclass

from weakincentives.runtime.events import ToolInvoked


@dataclass
class ReviewedSymbol:
    query: str
    matches: tuple[SymbolMatch, ...]


def track_reviewed_symbols(
    reviewed: tuple[ReviewedSymbol, ...],
    event: ToolInvoked,
    *,
    context: object,
) -> tuple[ReviewedSymbol, ...]:
    del context
    if event.value is None or not isinstance(event.value, SymbolSearchResult):
        return reviewed

    params = event.params
    reviewed_symbol = ReviewedSymbol(
        query=params.query,
        matches=event.value.matches,
    )
    return (*reviewed, reviewed_symbol)


session.register_reducer(
    SymbolSearchResult,
    track_reviewed_symbols,
    slice_type=ReviewedSymbol,
)

Attach custom tools to sections (next step) so the adapter can call them and record their outputs on the session alongside built-in reducers. The prompt can now chase suspicious references without delegating work back to the orchestrator.

4. Compose the prompt with deterministic sections and chapters

Sections render through string.Template, so keep placeholders readable and combine guidance with the tool suites into one auditable prompt tree. Long-form checklists or escalation playbooks often span many pages and only matter for specialized reviews; wrap them in a chapter so adapters can toggle visibility based on the user prompt. See the Prompt Class, Prompt Versioning & Persistence, and Chapters Specification for the rendering, hashing, and visibility rules that stabilize this structure.

from dataclasses import dataclass

from weakincentives import MarkdownSection, Prompt
from weakincentives.prompt import Chapter, ChaptersExpansionPolicy


@dataclass
class ReviewGuidance:
    severity_scale: str = "minor | major | critical"
    output_schema: str = "ReviewBundle with comments[] and overall_assessment"
    focus_areas: str = (
        "Security regressions, concurrency bugs, test coverage gaps, and"
        " ambiguous logic should be escalated."
    )


@dataclass
class ComplianceChapterParams:
    required: bool = False
    primary_jurisdictions: str = ""
    regulation_matrix_summary: str = ""
    escalation_contact: str = "[e-mail]"
    evidence_workspace: str = "gs://audit-artifacts"


overview_section = MarkdownSection[PullRequestContext](
    title="Repository Overview",
    key="review.overview",
    template="""
    You are a principal engineer reviewing a pull request.
    Repository: ${repository}
    Title: ${title}

    Pull request summary:
    ${body}

    Files touched: ${files_summary}
    """,
)


analysis_section = MarkdownSection[ReviewGuidance](
    title="Review Directives",
    key="review.directives",
    template="""
    - Classify findings using this severity scale: ${severity_scale}.
    - Emit output that matches ${output_schema}; missing fields fail the run.
    - Investigation focus:
      ${focus_areas}
    - Inspect mounted diffs under `diffs/` with `vfs_read_file` before
      commenting on unfamiliar hunks.
    - Reach for `symbol_search` when you need surrounding context from the
      repository checkout.
    """,
    tools=(symbol_search_tool,),
    default_params=ReviewGuidance(),
)


review_prompt = Prompt[ReviewBundle](
    ns="tutorial/code_review",
    key="review.generate",
    name="code_review_agent",
    sections=(
        overview_section,
        planning_section,
        vfs_section,
        asteval_section,
        analysis_section,
    ),
    chapters=(
        Chapter[ComplianceChapterParams](
            key="review.compliance",
            title="Compliance Deep Dive",
            description=(
                "Multi-page regulations guidance that only opens when the "
                "request demands a compliance audit."
            ),
            sections=(
                MarkdownSection[ComplianceChapterParams](
                    title="Regulatory Background",
                    key="review.compliance.background",
                    template="""
                    Compliance review requested.
                    Focus jurisdictions: ${primary_jurisdictions}

                    The attached regulation matrix may span many pages. Only
                    cite sections that apply to this pull request.
                    """,
                    default_params=ComplianceChapterParams(),
                ),
                MarkdownSection[ComplianceChapterParams](
                    title="Compliance Checklist",
                    key="review.compliance.checklist",
                    template="""
                    - Summarize gaps against: ${regulation_matrix_summary}
                    - Escalate urgent findings to: ${escalation_contact}
                    - Link all evidence in: ${evidence_workspace}
                    """,
                    default_params=ComplianceChapterParams(),
                ),
            ),
            default_params=ComplianceChapterParams(),
            enabled=lambda params: params.required,
        ),
    ),
)


requires_compliance_review = True  # derived from user metadata
compliance_params = ComplianceChapterParams(
    required=requires_compliance_review,
    primary_jurisdictions="SOX §404, PCI-DSS",
    regulation_matrix_summary="See 12-page compliance dossier in appendix.",
    evidence_workspace="gs://audit-artifacts/octo-widgets",
)

expanded_prompt = review_prompt.expand_chapters(
    ChaptersExpansionPolicy.ALL_INCLUDED,
    chapter_params={
        "review.compliance": compliance_params,
    },
)


rendered = expanded_prompt.render(
    PullRequestContext(
        repository="octo/widgets",
        title="Add caching layer",
        body="Introduces memoization to reduce redundant IO while preserving correctness.",
        files_summary="loader.py, cache.py",
    ),
    ReviewGuidance(),
    compliance_params,
)


print(rendered.text)
print([tool.name for tool in rendered.tools])

Set requires_compliance_review = False (and skip the chapter parameters) when the user prompt does not request a regulated-industry audit—the compliance chapter stays closed and the oversized guidance never reaches the model.

5. Evaluate the prompt with an adapter

Adapters send the rendered prompt to a provider and publish telemetry to the event bus; the session wiring above captures PromptExecuted and ToolInvoked events automatically. Pass the chapter-expanded prompt plus the same parameter dataclasses you used for rendering so the adapter sees the specialized compliance guidance. For payload formats and parsing guarantees see Adapter Evaluation and Native OpenAI Structured Outputs.

from weakincentives.adapters.openai import OpenAIAdapter


adapter = OpenAIAdapter(
    model="gpt-4o-mini",
    client_kwargs={"api_key": "sk-..."},
)


response = adapter.evaluate(
    expanded_prompt,
    PullRequestContext(
        repository="octo/widgets",
        title="Add caching layer",
        body="Introduces memoization to reduce redundant IO while preserving correctness.",
        files_summary="loader.py, cache.py",
    ),
    ReviewGuidance(),
    compliance_params,
    bus=bus,
    session=session,
)


bundle = response.output
if bundle is None:
    raise RuntimeError("Structured parsing failed")


for comment in bundle.comments:
    print(f"{comment.file_path}:{comment.line} → {comment.summary}")

If the model omits a required field, OpenAIAdapter raises PromptEvaluationError with provider context rather than silently degrading.

6. Mine session state for downstream automation

Selectors expose reducer output so you can ship audit logs without extra plumbing. Planning reducers keep only the latest Plan; register a custom reducer before PlanningToolsSection if you need history. See Session State and Snapshot Capture and Rollback for selector and rollback rules.

from weakincentives.runtime.session import select_latest
from weakincentives.tools import Plan, VirtualFileSystem


latest_plan = select_latest(session, Plan)
vfs_snapshot = select_latest(session, VirtualFileSystem)


if latest_plan:
    print(f"Plan objective: {latest_plan.objective}")
    for step in latest_plan.steps:
        print(f"- [{step.status}] {step.title}")
else:
    print("No plan recorded yet.")


if vfs_snapshot:
    for file in vfs_snapshot.files:
        print(f"Staged file {file.path.segments} (version {file.version})")

7. Override sections with an overrides store

Persist optimizer output so the runtime can swap in tuned sections without a redeploy. LocalPromptOverridesStore is the default choice: it discovers the workspace root, enforces descriptors, and reads JSON overrides from .weakincentives/prompts/overrides/. Refer to the Prompt Overrides specification to keep namespace, key, and tag hashes aligned.

from pathlib import Path

from weakincentives.prompt.overrides import (
    LocalPromptOverridesStore,
    PromptDescriptor,
    PromptOverride,
    SectionOverride,
)


workspace_root = Path("/srv/agent-workspace")
overrides_store = LocalPromptOverridesStore(root_path=workspace_root)

descriptor = PromptDescriptor.from_prompt(review_prompt)
seed_override = overrides_store.seed_if_necessary(
    review_prompt, tag="assertive-feedback"
)

section_path = ("review", "directives")
section_descriptor = next(
    section
    for section in descriptor.sections
    if section.path == section_path
)

custom_override = PromptOverride(
    ns=descriptor.ns,
    prompt_key=descriptor.key,
    tag="assertive-feedback",
    sections={
        **seed_override.sections,
        section_path: SectionOverride(
            expected_hash=section_descriptor.content_hash,
            body="\n".join(
                (
                    "- Classify findings using this severity scale: minor | major | critical.",
                    "- Always cite the exact diff hunk when raising a major or critical issue.",
                    "- Respond with ReviewBundle JSON. Missing fields terminate the run.",
                )
            ),
        ),
    },
    tool_overrides=seed_override.tool_overrides,
)

persisted_override = overrides_store.upsert(descriptor, custom_override)

rendered_with_override = review_prompt.render(
    PullRequestContext(
        repository="octo/widgets",
        title="Add caching layer",
        body="Introduces memoization to reduce redundant IO while preserving correctness.",
        files_summary="loader.py, cache.py",
    ),
    overrides_store=overrides_store,
    tag=persisted_override.tag,
)


print(rendered_with_override.text)

The overrides store writes atomically to .weakincentives/prompts/overrides/{ns}/{prompt_key}/{tag}.json inside the workspace described in the Prompt Overrides specification. Optimizers and prompt engineers can still drop JSON overrides into that tree by hand—checked into source control or generated during evaluations—without subclassing PromptOverridesStore. Because sections expose stable (ns, key, path) identifiers, overrides stay scoped to the intended content so teams can iterate on directives without risking accidental drift elsewhere in the tree.

8. Ship it

You now have a deterministic reviewer that:

1. Enforces typed contracts for inputs, tools, and outputs.
2. Persists plans, VFS edits, and evaluation transcripts inside a session.
3. Supports optimizer-driven overrides that fit neatly into CI or evaluation harnesses.

Run it inside a worker, bot, or scheduler; the captured session state keeps each evaluation replayable. For long-lived deployments, follow the Prompt Overrides specification to keep overrides and tool descriptors in sync.

Logging

Weak Incentives ships a structured logging adapter so hosts can add contextual metadata to every record without manual dictionary plumbing. Call configure_logging() during startup to install the default handler and then bind logger instances wherever you need telemetry:

from weakincentives.runtime.logging import configure_logging, get_logger

configure_logging(json_mode=True)
logger = get_logger("demo").bind(component="cli")
logger.info("boot", event="demo.start", context={"attempt": 1})

The helper respects any existing root handlers—omit force=True if your application already configures logging and you only want Weak Incentives to honor the selected level. When you do want to take over the pipeline, call configure_logging(..., force=True) and then customize the root handler list with additional sinks (for example, forwarding records to Cloud Logging or a structured log shipper). Each emitted record contains an event field plus a context mapping, so downstream processors can make routing decisions without parsing raw message strings.

Development Setup

1. Install Python 3.12 (for example with pyenv install 3.12.0).

2. Install uv, then bootstrap the environment and hooks:

   uv sync
   ./install-hooks.sh
   

3. Run checks with uv run so everything shares the managed virtualenv:

   • make format / make format-check
   • make lint / make lint-fix
   • make typecheck (Ty + Pyright, warnings fail the build)
   • make test (pytest via build/run_pytest.py, 100% coverage enforced)
   • make check (aggregates the quiet checks above plus Bandit, Deptry, pip-audit, and markdown linting)

Integration tests

Provider integrations require live credentials, so the suite stays opt-in. Export the necessary OpenAI configuration and then run the dedicated make target, which disables coverage enforcement automatically:

export OPENAI_API_KEY="sk-your-key"
# Optionally override the default model (`gpt-4.1`).
export OPENAI_TEST_MODEL="gpt-4.1-mini"

make integration-tests

make integration-tests forwards --no-cov to pytest so you can exercise the adapter scenarios without tripping the 100% coverage gate configured for the unit test suite. The tests remain skipped when OPENAI_API_KEY is not present.

Documentation

• AGENTS.md — operational handbook and contributor workflow.
• specs/ — design docs for prompts, planning tools, and adapters.
• ROADMAP.md — upcoming feature sketches.
• docs/api/ — API reference material.

License

Apache 2.0 • Status: Alpha (APIs may change between releases)