Imported from Trampoline-AI/predict-rlm (.agents/skills/rlm/SKILL.md). Install upstream with npx skills add Trampoline-AI/predict-rlm --skill rlm. Copyright stays with the author.

Create a PredictRLM

Check for skill updates

Before planning or building, check whether the installed rlm skill has an update:

npx skills update rlm

Replace npx with pnpx, bunx, or the equivalent package runner in use. After an update, re-read the skill before continuing.

An RLM is a callable, pre-configured agent. It autonomously explores context, writes and executes code in a sandboxed REPL, calls tools, inspects results, and iterates until the task is done. Unlike a chat agent, an RLM is a function — you define its inputs, outputs, and tools, then call it from your code. It returns structured data, not chat messages.

Mental model

The caller supplies a DSPy signature. Its input and output fields are the function contract, and its docstring is the operating procedure. On each call:

1. PredictRLM prepares the typed inputs and acquires a sandboxed execution session. The default backend is a stateful Python REPL running through Deno and Pyodide/WASM.
2. The outer LM reads the signature and strategy, then writes Python code to inspect inputs, preserve intermediate state, branch, retry, and verify. Python variables persist across iterations of that call.
3. Sandbox code can use await predict(...) for focused sub-LM perception or extraction. Each predict() call has its own context window and can use a typed DSPy signature, Pydantic schemas, and dspy.Image.
4. Sandbox code can also use packages and modules supplied by Skill objects, call host-side tools for capabilities outside WASM, and access explicitly allowed network domains.
5. When the work is complete, the outer LM calls SUBMIT(...). PredictRLM validates the declared outputs, synchronizes file outputs to the host, and returns a dspy.Prediction with the output fields and a structured trace.

The separation is deliberate: the outer LM owns planning and code execution; the sub-LM owns narrow understanding tasks; deterministic libraries and tools do the work they are better suited for. Large context stays in files, sandbox variables, and focused subcalls instead of accumulating in one prompt.

application call
    -> DSPy signature: inputs + strategy + outputs
    -> outer LM <-> stateful sandbox REPL
                     |-> predict() -> sub-LM
                     |-> skill packages and modules
                     `-> host-side tools
    -> validated dspy.Prediction + trace

Core API

A minimal typed RLM has a signature, a configured PredictRLM, and a normal sync or async call:

The signature class docstring is the model instruction prompt passed to the outer LM. It is not merely developer documentation: put the RLM's workflow, decision rules, verification steps, and completion criteria there. Changing the docstring changes the model's runtime behavior.

import dspy
from pydantic import BaseModel, Field

from predict_rlm import CtxStr, File, PredictRLM
from predict_rlm.skills import pdf


class Analysis(BaseModel):
    summary: str = Field(description="Grounded summary of the documents")
    risks: list[str] = Field(description="Material risks supported by the documents")


class AnalyzeDocuments(dspy.Signature):
    """Inspect the documents, apply the criteria, and return a grounded analysis.

    Survey the files first. Extract only relevant evidence with focused
    predict() calls, verify important claims, then submit the typed result.
    """

    documents: list[File] = dspy.InputField(desc="Documents to inspect")
    criteria: CtxStr = dspy.InputField(desc="Criteria the outer LM must see in full")
    analysis: Analysis = dspy.OutputField(desc="Evidence-grounded analysis")


rlm = PredictRLM(
    AnalyzeDocuments,
    lm="openai/gpt-5.4",
    sub_lm="openai/gpt-5.1",
    skills=[pdf],
    max_iterations=30,
)

result = await rlm.acall(
    documents=[File(path="report.pdf")],
    criteria="Cover obligations, deadlines, and material risks.",
)
print(result.analysis.summary)

Use rlm(...) for a synchronous call and await rlm.acall(...) for an asynchronous call. Both return a dspy.Prediction; each declared output is available as an attribute such as result.analysis.

The main constructor surface is:

• signature — a DSPy signature class or compact string signature;
• lm — the outer LM that writes code;
• sub_lm — the LM behind the built-in predict() tool;
• skills — reusable instructions, PyPI packages, sandbox modules, and tools;
• tools — sync or async host callables exposed inside the sandbox;
• allowed_domains — the sandbox network allowlist, empty by default;
• max_iterations, max_llm_calls, and max_output_chars — execution budgets;
• output_dir — host collection root for declared File outputs;
• verbose and debug — human-readable run output and lifecycle diagnostics.

Use File or list[File] for large file inputs and generated artifacts. Use CtxStr for a direct string input, such as a rubric, whose full runtime value must be added to the outer LM prompt as well as exposed as a Python variable. Use a Skill when the sandbox needs reusable instructions, packages, modules, or bundled tools. Use tools= directly for specific host-side actions such as authenticated APIs, databases, native libraries, or host filesystem access.

Runtime kernel and extension model

Most RLMs should use only signatures, File, CtxStr, Workspace, skills, and host tools. Extend the kernel only when the workflow needs a new typed boundary, resource lifecycle, execution substrate, or correctness-critical event stream. Do not reach for a custom adapter merely because a signature uses a Pydantic model.

At construction, PredictRLM resolves direct options and RuntimeContribution module factories into one immutable RuntimeSpec: instructions, adapters, tools, packages, exactly one execution backend, event sinks, validators, and tool operations. Module factories run once at construction. Every invocation then creates a fresh RunContext for mutable state, prepared inputs, bindings, output reservations, cleanup callbacks, and evidence status.

Kernel lifecycle and ownership

construction
  direct options + RuntimeContribution modules
      -> immutable RuntimeSpec

invocation
  prepare typed inputs and output requirements
      -> compile path and artifact claims
      -> open adapter-owned resources
      -> validate requirements and acquire one ExecutionSession
      -> bind inputs and reserve outputs
      -> apply invocation-local prompt contributions
      -> run generated-code attempts
           -> after_execution durability hooks
      -> materialize submitted outputs while the session is active
      -> finalize input adapters in reverse order
      -> finalize and release the kernel-owned session
      -> run remaining LIFO RunContext cleanup

The ownership boundaries are strict:

• Adapters declare requirements and destinations before session acquisition.
• The ExecutionBackend creates the invocation-scoped ExecutionSession.
• Adapters may use supported backend or session capabilities, but never acquire, finalize, or release the session.
• Mutable provider clients, leases, baselines, and retry state belong in the invocation's PreparedInput or RunContext, never on a shared adapter.
• Sandbox destination overlaps fail before generated code runs.
• Adapter resources finalize on success, failure, setup failure, and cancellation. Framework-owned session finalization still runs after adapter errors.

Choose the smallest extension point

Use the first boundary that can express the requirement:

1. Signature docstring or CtxStr — task strategy or invocation-specific criteria the outer LM must read.
2. Skill — reusable outer-LM instructions plus sandbox packages, importable Python modules, or bundled host tools.
3. tools= — a narrow sync or async host callable. Use this for native libraries, credentials, APIs, databases, and host filesystem access. Return plain JSON-like data or Pydantic values that can be normalized at the transport boundary.
4. File, Workspace, or an existing adapter — ordinary copied files, generated files, or a mutable directory. Prefer these built-ins over a new type.
5. InputAdapter with PreparedInput.path(), .paths(), or .glob() — a custom typed input can first be materialized as host paths. The kernel owns destination normalization, overlap checks, copying or explicit mounting, backend requirements, and the model-visible sandbox paths.
6. Full InputAdapter lifecycle — an external resource needs a provider lease, live handle, per-attempt synchronization, or a custom session capability that cannot be represented as host paths.
7. OutputAdapter — a custom output needs a reserved destination and provider-specific materialization while the session is still active.
8. RuntimeContribution through modules= — several host-side extensions belong together and should compose as one reusable construction-time unit.
9. EventSink through events= — a consumer needs ordered lifecycle events during the run. Use prediction.trace when only the completed trace is needed.
10. ExecutionBackend through execution= — only for a genuinely new sandbox or execution substrate. Use sandbox_backend="jspi" or "sbx" for maintained backends.

interpreter= is a compatibility bridge for older CodeInterpreter-shaped integrations, not the root extension contract. New session-native integrations must implement ExecutionBackend and use execution=.

Input adapters

An InputAdapter[T] selects signature fields through value_type and turns the caller's value into model-visible data and runtime requirements. Adapter instances are construction-time objects and may serve concurrent calls, so keep them immutable or concurrency-safe.

Use the current lifecycle names and responsibilities:

• prepare(field, value, ctx) -> PreparedInput runs before backend acquisition. It receives no backend or session. Return declarative paths, artifacts, sandbox-root reservations, metadata, requirements, and input instructions.
• open(field, prepared, ctx, backend) may acquire an adapter-owned client or lease after all inputs are prepared but before a session exists. Store the handle in ctx.state under an adapter-owned key.
• bind(field, prepared, ctx, session) -> BoundInput runs after acquisition. Use it only when the default artifact mounting is insufficient; return the final value and bindings visible to generated code.
• append_prompt(prompt, field, prepared, ctx) -> str may add invocation-local context to the outer action and forced-extraction prompts after binding. It must not mutate the canonical PredictRLM.signature.
• after_execution(field, prepared, ctx, session, result, error) runs after each completed generated-code attempt. Use it for durability or sync work, including changes made before that attempt raised an error.
• finalize(field, prepared, ctx, session, error) performs the final save and releases adapter-owned resources. Opened adapters finalize once in reverse order; session is None if acquisition failed.

For the common path case, implement only prepare():

class S3FileAdapter(InputAdapter[S3File]):
    name = "s3-file"
    value_type = S3File

    async def prepare(self, field, value, ctx) -> PreparedInput:
        local_path = await materialize_s3(value.uri, ctx=ctx)
        return PreparedInput.path(
            local_path,
            instructions=(f"{field.name} is available at its sandbox path.",),
        )

Register one-off adapters directly with PredictRLM(MySignature, adapters=[S3FileAdapter()]), or contribute them from a runtime module when they belong with tools, packages, events, or an execution backend.

Copy is the portable default. Set mode="mount" only for a required live directory view; unsupported backends fail rather than silently copying. .paths() exposes an explicit list, and .glob() performs deterministic, host-side selection with traversal and symlink-escape checks.

For a resource described completely by an ID, URI, or host path, return an Artifact in PreparedInput.artifacts and let the default bind() call session.mount(artifact). For a provider-managed live handle, acquire it in open(), keep it in ctx.state, and use a capability protocol implemented by the selected session from bind().

Do not use removed lifecycle names: input prepare_session() became open(), input mount() became bind(), and MountedInput became BoundInput. prepare_session() remains an output-adapter method. The reserved ctx_str adapter name may only be replaced by a CtxStrInputAdapter subclass.

Output adapters

An OutputAdapter[T] owns a custom typed output boundary:

• prepare_session() contributes pre-acquisition policy and requirements;
• reserve() claims a destination after session acquisition but before code;
• materialize() turns the submitted value into the caller-facing value while the session is active.

Output adapters have no input-style finalize() hook. Clean partial failures locally, and register provider cleanup that can outlive the session with ctx.add_cleanup(). If an output owns a sandbox path, place it in the reservation artifact's metadata["sandbox_path"]; the kernel rejects overlaps with inputs and other outputs before execution.

Runtime modules

Use a zero-argument module factory to package related host-side contributions:

from predict_rlm import CallableTool, RuntimeContribution


def document_runtime() -> RuntimeContribution:
    return RuntimeContribution(
        instructions=("Treat document IDs as opaque provider references.",),
        adapters=(DocumentReferenceAdapter(),),
        tools=(
            CallableTool(
                name="fetch_metadata",
                function=fetch_metadata,
                description="Fetch provider metadata for one document ID.",
                schema={
                    "type": "object",
                    "properties": {"document_id": {"type": "string"}},
                    "required": ["document_id"],
                },
            ),
        ),
        packages=("pure-python-package",),
        events=(DocumentAuditSink(),),
    )


rlm = PredictRLM(MySignature, modules=[document_runtime])

PredictRLM(modules=...) composes host-side runtime behavior. Skill.modules is different: it copies Python files into the sandbox so generated code can import them. Contribution packages are deduplicated; duplicate adapter, tool, or tool-operation names are errors; and the resolved configuration must select exactly one execution backend.

Execution, events, hooks, and gates

• An ExecutionBackend.start(spec, ctx) returns an async context manager that yields one invocation-scoped ExecutionSession. The session implements code execution, package installation, artifact mounting and collection, cancellation, and finalization. Optional capability protocols add operations such as host-directory mounts or mutable-directory collection.
• EventSink implements async emit, flush, and close plus a strict flag. Use strict=False for monitoring whose failure must not fail the run. Use strict=True only when durable, complete evidence is part of correctness; a strict sink failure can prevent successful publication.
• DSPy callbacks report high-level RLM iteration progress. EventSink reports ordered kernel lifecycle evidence. They solve different problems.
• RuntimeHook(target=..., phases={"before", "after", "error"}) instruments a dotted function inside the sandbox and emits sanitized RuntimeHookEvent values to on_runtime_hook_event. Runtime hooks require the SBX backend in PredictRLM v1; do not design a portable extension around them.
• submit_confirmation is a gate after a valid SUBMIT: return feedback to continue the RLM loop, or None/"" to accept the submission.

Before implementing a kernel extension, read docs/custom-path-inputs.md for path-backed inputs, docs/custom-adapters.md for lifecycle patterns, docs/api.md for exact method signatures, and docs/observability.md for event sink guarantees. Import public contracts from predict_rlm; do not use removed shim modules such as predict_rlm.adapters, predict_rlm.artifacts, predict_rlm.events, predict_rlm.execution, or predict_rlm.kernel.

Test extensions across success, generated-code failure, setup failure, cancellation, and concurrent calls. If they own mutable external state, verify LIFO cleanup and that no state leaks between invocations.

What this skill builds

This skill creates or extends the reusable package around PredictRLM: schemas, the DSPy signature and strategy, capability definitions, a service wrapper, and smoke tests. Do not add RLM-GEPA optimization wiring here; use the separate rlm-gepa skill once the base RLM works and an optimization objective exists.

Work in two phases:

1. Plan — define the RLM with the user, research feasibility, and produce a concrete implementation plan.
2. Build — implement the approved plan, then smoke-test the generated package.

Phase 1: Plan

Keep the following sequence. It prevents a plausible-looking RLM from having an unusable boundary, missing runtime capability, or an unfalsifiable output.

Step 1: Goal definition

Understand what the user wants to build.

Ask:

• What is the desired outcome and what does success look like?
• What is the input material: documents, code, data, APIs, or stateful systems?
• What should the output be: structured data, modified files, a spreadsheet, or another artifact?

Then validate RLM fit. An RLM is a good fit when it needs one or more of:

• selective exploration of large inputs such as documents, datasets, or codebases;
• multi-step work with tools, for example extract → transform → validate;
• actions that modify a file or another controlled state boundary;
• parallel sub-LM calls across many items;
• file-to-file transformation such as PDFs to spreadsheets or documents to reports.

If a single LM call or a deterministic script is a better fit, say so and propose it instead.

Step 2: Input design

Define every input before defining implementation files.

For each input, decide its name, type, source, description, and whether the outer LM needs the complete value immediately.

• Use File or list[File] for large files. Inputs are copied into the sandbox under /sandbox/input/<field>/; the RLM accesses them on demand.
• Use ordinary str, primitives, or lean Pydantic models for metadata and configuration.
• Use CtxStr only for a small-to-moderate string that the outer LM must see in full, such as a rubric or task instruction. CtxStr is input-only and only valid as a direct field annotation in a class-based DSPy signature.
• Do not put raw document contents or large tables directly in a prompt when a File reference and focused extraction can keep the context small.

Confirm the input boundary before moving on.

Step 3: Output design

Define the structured output before selecting packages or writing prompts.

For each output field, decide its name, type, description, and whether it is a primitive, a Pydantic model, File, or list[File].

Push for specific, observable fields. Use Pydantic Field(description=...) for non-obvious model fields. Model only data the caller needs; do not expose internal IDs or intermediate reasoning.

Ask what users check first, which computed values matter, and whether they need output files. Confirm the schema before moving on.

Step 4: Research feasibility

Research autonomously, then report a clear feasibility assessment.

1. Find domain libraries and existing project patterns.
2. Check sandbox compatibility. The default execution environment is Pyodide in WASM: pure-Python wheels and Pyodide-built packages work; native binaries and ordinary C extensions do not.
3. Identify network needs and list exact domains for allowed_domains.
4. Identify host-side needs. Put native libraries, authenticated APIs, database access, heavy host filesystem work, and unsupported packages behind typed host-side tools.
5. Check whether pdf, spreadsheet, or docx built-in skills already cover the task.
6. Check whether the boundary already fits File, Workspace, CtxStr, an existing adapter, or a PreparedInput path declaration before designing a full runtime extension.

Treat a package that cannot run in Pyodide as a design decision, not a surprise at implementation time: use a host-side tool or change the approach.

Step 5: Capability design

Choose the smallest capability surface that makes the workflow work.

Built-in skills

State which built-in skills are needed and why:

• pdf for PDF rendering, text extraction, modification, and redaction;
• spreadsheet for Excel workbooks, formulas, and formatting;
• docx for reading and writing Word documents.

Custom skills

Create a Skill only for a capability that can be reused across different RLMs. A skill teaches the model how to use one thing well—for example, how to inspect and modify PDFs, work with spreadsheets, or query a docket system. Its instructions should remain useful when attached unchanged to another RLM that needs the same capability.

Put workflow-specific instructions in the DSPy signature docstring instead. The signature docstring is the outer LM's instruction prompt, so it owns the current RLM's objective, input-specific procedure, business rules, output requirements, verification steps, and completion criteria.

Use this test:

• If the guidance explains how to use a reusable capability, put it in a Skill.
• If the guidance explains what this RLM must do, put it in the signature docstring.
• If the text names this RLM's fields, output schema, one-off policy, or workflow stages, it is almost certainly signature-specific.

Do not create a skill merely to move instructions out of a long signature, and do not duplicate the same guidance in both places. A one-off host callable belongs directly in tools=; bundle it into a skill only when the callable and its operating guidance form a capability other RLMs can reuse.

For each genuinely reusable skill, specify:

• name — concise capability identifier;
• instructions — general operating guidance independent of the current RLM;
• packages — Pyodide-compatible sandbox packages required by the capability;
• modules — host paths mapped to reusable importable sandbox modules;
• tools — host callables that belong with the reusable capability.

instructions should be an operating manual, not a short package description or list of functions. A useful skill normally includes:

1. when to use the capability and how to choose among its approaches;
2. an ordered workflow for inspecting, modifying, or producing artifacts;
3. runnable code patterns for the operations models commonly get wrong;
4. correctness invariants and checks to perform before SUBMIT;
5. known failure modes, sandbox limitations, and recovery guidance.

Built-in skills are intentionally substantive: the PDF skill explains visual rendering versus text extraction and parallel predict() calls; the DOCX skill covers document structure, editing patterns, formatting traps, and limitations; the spreadsheet skill covers formula policy, recalculation, rendering, style preservation, and output verification.

This reusable HTML capability shows the expected shape at a smaller scale:

from predict_rlm import Skill


html_skill = Skill(
    name="html",
    instructions="""Use BeautifulSoup to inspect and extract static HTML documents.

## Workflow

1. Read the source without modifying it and record its path and byte size.
2. Parse with `html.parser`. Remove scripts, styles, and `noscript` blocks, but
   preserve headings, lists, links, tables, labels, and form controls.
3. Inventory the semantic structure before extracting: title, headings, main
   content container, tables, links, and repeated navigation/footer regions.
4. Use deterministic selectors and extraction first. Use `predict()` only for
   interpretation of one bounded, cleaned section with a typed result; never
   send an entire large page when a section or table is sufficient.

## Parse and inventory

    from bs4 import BeautifulSoup

    with open(path, "r", encoding="utf-8", errors="replace") as handle:
        soup = BeautifulSoup(handle.read(), "html.parser")

    for node in soup(["script", "style", "noscript"]):
        node.decompose()

    root = soup.find("main") or soup.find("article") or soup.body or soup
    title = soup.title.get_text(" ", strip=True) if soup.title else None
    headings = [
        {
            "level": node.name,
            "text": node.get_text(" ", strip=True),
            "id": node.get("id"),
        }
        for node in root.find_all(["h1", "h2", "h3", "h4", "h5", "h6"])
    ]

## Extract links and tables

    links = [
        {"text": a.get_text(" ", strip=True), "href": a.get("href")}
        for a in root.find_all("a", href=True)
    ]

    tables = []
    for table in root.find_all("table"):
        rows = [
            [cell.get_text(" ", strip=True) for cell in row.find_all(["th", "td"])]
            for row in table.find_all("tr")
        ]
        tables.append(rows)

## Correctness and verification

- Preserve document order, exact link targets, table row order, element IDs,
  and the nearest heading as source context. Do not invent text hidden behind
  missing JavaScript data.
- Check that the selected root is nonempty, expected headings and tables were
  found, and extracted row/link counts match the parsed elements.
- Keep raw source text for quoted evidence. Re-open and verify important quotes
  or values against the original element before submitting.

## Failure modes

- Static HTML may omit content rendered by JavaScript. Report that limitation
  and use a host-side browser tool when rendered DOM access is required.
- `errors="replace"` can hide encoding damage; flag replacement characters in
  relevant text and retry with an identified encoding when necessary.
- `rowspan` and `colspan` require explicit normalization when table geometry
  matters. Do not pretend a flattened row list preserves merged-cell layout.
""",
    packages=["beautifulsoup4"],
)

The example is reusable because it teaches HTML handling without mentioning any one RLM's fields, output schema, business rules, or completion policy.

Host-side tools

For each tool, define the typed signature, a useful docstring, return shape, and why it must run on the host. Return plain JSON-like data or Pydantic models; PredictRLM normalizes Pydantic values, including nested values, to mappings and lists before they cross the sandbox boundary.

Prefer narrow, typed tools. A tool is a system boundary, not an escape hatch for arbitrary host access.

Kernel extensions

If the workflow needs a custom runtime boundary, specify it in the plan rather than discovering it during implementation:

• the caller-facing signature type and model-visible value;
• why built-in files, workspaces, skills, or host tools are insufficient;
• the smallest extension point from the kernel decision order above;
• all sandbox paths, copy versus mount behavior, and backend requirements;
• adapter-owned resources, invocation-local state, persistence points, and cleanup on success, failure, cancellation, and partial setup;
• whether the extension is portable across JSPI and SBX;
• the RuntimeContribution composition boundary when several pieces belong together; and
• focused lifecycle and concurrency tests.

Step 6: Strategy and architecture

Write the strategy that belongs in the DSPy signature docstring. It is the RLM's playbook:

1. survey and understand the available inputs;
2. gather evidence with file skills, predict(), and host tools;
3. process, cross-check, and synthesize the evidence;
4. emit the declared data and write any output files in their output locations.

Choose one RLM when the work is a coherent workflow with shared state and a reasonable iteration budget. Chain RLMs only when stages have genuinely different capabilities, output artifacts consumed by later stages, or separate model/iteration requirements. For every chain, define typed stage boundaries and a DAG before implementation.

Choose initial max_iterations, max_llm_calls, and max_output_chars from the workload. Record required allowed_domains, whether a capable sub_lm is needed for extraction, and whether the caller needs output_dir for generated files.

Feasibility checklist

Before finalizing the plan, verify:

• Every sandbox package is Pyodide-compatible or replaced by a host tool.
• Network domains are explicit and minimal.
• Host-side tools cover every unsupported host capability.
• The iteration and LM-call budgets are plausible for the workflow.
• Large inputs stay as File references or on-demand metadata.
• Output schemas are specific enough to validate.
• Any chained stages have typed, necessary boundaries.

Plan output

Deliver a plan containing:

1. Overview and RLM-fit decision.
2. Delivery scope and file manifest.
3. Complete input and output schemas.
4. Complete signature strategy.
5. Skill and host-tool contracts.
6. Service architecture, including a typed DAG when chained.
7. Feasibility constraints, alternatives, and initial budgets.
8. Smoke-test cases and exact verification commands.

Phase 2: Build

Implement the approved design with the smallest maintainable package layout. Do not scaffold benchmarks or RLM-GEPA optimization speculatively. Add an eval layer only when the user has examples and a scoring need; use rlm-gepa for optimization wiring.

Default package layout

my_rlm/
├── pyproject.toml
├── src/
│   └── my_rlm/
│       ├── __init__.py
│       ├── schema.py       # Pydantic inputs and outputs
│       ├── signature.py    # DSPy signature and strategy
│       ├── service.py      # PredictRLM wiring
│       ├── skills.py       # Only when custom skills are needed
│       └── tools.py        # Only when host-side tools are needed

Create schema.py, signature.py, service.py, package exports. Add files only when the chosen boundary needs them.

Dependencies

Install predict-rlm in the target repository without hardcoding a version:

uv add predict-rlm

Add the examples extra only when using built-in document or spreadsheet skills, and add any domain dependency only where it runs. Do not include GEPA or visualization extras in an agent-only project.

Schema

from pydantic import BaseModel, Field


class KeyDate(BaseModel):
    name: str = Field(description="What deadline or effective date this is")
    date: str = Field(description="ISO-8601 calendar date")


class DocumentAnalysis(BaseModel):
    summary: str = Field(description="Grounded Markdown summary")
    key_dates: list[KeyDate] = Field(
        default_factory=list,
        description="Dates supported by the input documents",
    )

Signature and strategy

Treat the signature docstring as the RLM's executable strategy, not a one-line summary. Strong instructions define source authority, discovery, bounded tool use, intermediate state, evidence rules, validation, recovery from missing or conflicting information, the output contract, and a stopping condition.

import dspy

from predict_rlm import CtxStr, File

from .schema import DocumentAnalysis


class AnalyzeDocuments(dspy.Signature):
    """Analyze a related document set into a grounded, structured report.

    Read `criteria` first and treat the supplied documents as the sole authority
    for document facts.

    ## Method

    1. Inventory every file before deep reading. Record its exact name, format,
       page or sheet count, likely role, and which criteria it may answer.
    2. Inspect each file with deterministic, format-appropriate tools. Search
       and extract structure to narrow the work; render pages or use focused
       `predict()` calls only on bounded sections that need interpretation.
       Never send the whole document corpus through one giant sub-LM call.
    3. Keep a compact evidence map in Python state with the source file, page,
       sheet, section, or other locator for each relevant fact. Retain evidence,
       not repeated full-document text.
    4. Reconcile overlapping sources. Prefer explicit and more current evidence,
       deduplicate repeated facts, and preserve meaningful disagreements.

    ## Evidence rules

    - Every conclusion and key date must trace to a source locator.
    - Never turn missing, ambiguous, or contradictory evidence into a confident
      claim. Mark it unresolved and explain what is missing.
    - Use `criteria` to control emphasis and format, but never let it justify an
      unsupported fact or silently exclude a supplied file.

    ## Verification and stopping

    Before submitting, confirm that every file was inspected, every criterion is
    answered or explicitly unresolved, important facts were cross-checked, and
    the result satisfies the declared schema. Stop only after those checks pass.

    ## Output contract

    Return one `DocumentAnalysis`. Its summary must be a standalone grounded
    report that follows `criteria`; `key_dates` must contain only supported dates
    with their meaning and normalized value. Use an empty list when no key date
    is supported.
    """

    documents: list[File] = dspy.InputField(desc="Documents to inspect")
    criteria: CtxStr = dspy.InputField(desc="Rubric to apply in full")
    analysis: DocumentAnalysis = dspy.OutputField(desc="Grounded result")

Callers pass a plain string for criteria; CtxStr makes it both a sandbox variable and a full prompt appendix. Use an ordinary str when the RLM can inspect the value from the sandbox instead.

Service

import dspy

from predict_rlm import File, PredictRLM
from predict_rlm.skills import pdf

from .signature import AnalyzeDocuments


class DocumentAnalyzer(dspy.Module):
    def __init__(
        self,
        *,
        lm: dspy.LM | str | None = None,
        sub_lm: dspy.LM | str | None = None,
        max_iterations: int = 30,
        verbose: bool = False,
        debug: bool = False,
    ) -> None:
        self.predictor = PredictRLM(
            AnalyzeDocuments,
            lm=lm,
            sub_lm=sub_lm,
            max_iterations=max_iterations,
            verbose=verbose,
            debug=debug,
            skills=[pdf],
        )

    async def aforward(self, documents: list[File], criteria: str):
        return await self.predictor.acall(documents=documents, criteria=criteria)

Use tools=[callable] when the callable name is suitable, or tools={"name": callable} to expose a deliberate tool name. Use skills=[...] for reusable sandbox capabilities. Pass output_dir to PredictRLM when File or list[File] outputs must be collected below a caller-selected root.

Runtime reference

Core constructor

PredictRLM(
    signature,
    lm=None,
    sub_lm=None,
    max_iterations=30,
    max_llm_calls=50,
    max_output_chars=50_000,
    verbose=True,
    tools=None,
    skills=None,
    adapters=(),
    execution=None,
    modules=(),
    events=(),
    allowed_domains=None,
    output_dir=None,
    interpreter=None,
    runtime_hooks=None,
    submit_confirmation=None,
    debug=False,
)

lm drives outer code generation; sub_lm serves sandbox predict() calls. Both accept a DSPy LM or model string. predict() is always available.

The kernel extension contracts are described above. Use execution= for a new session-native backend and reserve interpreter= for CodeInterpreter-shaped compatibility integrations. Consult docs/api.md for exact signatures rather than guessing lifecycle parameters.

Files, skills, and tools

• File(path="report.pdf") represents an input file; File.from_dir(path) creates a sorted recursive list of file references.

• Input files are copied to /sandbox/input/<field>/. File outputs are written under /sandbox/output/<field>/ and synchronized back after execution.

• Skill(name, instructions, packages, modules, tools) bundles reusable capability. Skill tool-name collisions are errors.

• pdf, spreadsheet, and docx are the built-in skills:

  from predict_rlm.skills import docx, pdf, spreadsheet

• Host tools run outside WASM. Make their docstrings, arguments, and return data clear enough for generated code to call safely.

Sandbox constraints

• Use pure-Python or Pyodide-compatible packages in sandbox skills.
• Do not expect subprocesses, native binaries, or ordinary C extensions.
• Default network access is denied; set a minimal allowed_domains list when the sandbox must reach a network service.
• Use host-side tools for native code, credentials, databases, and host files.