Metadata-Version: 2.5
Name: skilled-proposer
Version: 0.1.2
Summary: A GEPA instruction proposer for DSPy that writes generalizable, skill-informed instructions
Project-URL: Homepage, https://github.com/cmpnd-ai/skilled-proposer
Project-URL: Repository, https://github.com/cmpnd-ai/skilled-proposer
Author-email: Drew Breunig <dbreunig@gmail.com>
License-Expression: MIT
License-File: LICENSE
Keywords: dspy,gepa,llm,prompt-optimization
Classifier: Development Status :: 4 - Beta
Classifier: Intended Audience :: Developers
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Programming Language :: Python :: 3.13
Requires-Python: >=3.10
Requires-Dist: dspy>=3.0
Description-Content-Type: text/markdown

# skilled-proposer

A custom instruction proposer for [GEPA](https://dspy.ai/api/optimizers/GEPA/), the reflective prompt optimizer in [DSPy](https://dspy.ai). Drop it into `dspy.GEPA(instruction_proposer=...)` to get instructions that generalize instead of memorizing your training set, informed by reference skills you provide.

## Why

GEPA improves a program by asking a reflection model to rewrite each component's instruction based on execution traces and evaluator feedback. The stock proposer tells the reflection model to include "niche and domain specific factual information" from those traces in the new instruction. This helps with some tasks, but it can copy entities, numbers, and answers from your training examples into the prompt, and the prompt might perform worse on inputs it has never seen.

`SkilledProposer` uses a different meta-prompt. It gives the reflection model a three-step procedure. First, infer the task from the examples, because the assistant will only ever see the instruction. Second, diagnose why each failure happened and find the general rule that would have prevented it. Third, write the replacement instruction from those rules. The prompt then states one principle against overfitting. The proposer also adds three practical controls:

- Skills. Pass SKILL.md files, skill directories, or inline strings. The reflection model gets them as reference material, e.g., a prompting guide for your student model.
- Extra guidance. A plain string applied to every proposal.
- Length budgets. Cap the proposed instruction by words or tokens. The cap is enforced by a prompt constraint, then a compression pass, then truncation.

## Install

```bash
pip install skilled-proposer
```

Requires Python 3.10 or newer and dspy 3.0 or newer.

## Quickstart

```python
import dspy
from skilled_proposer import SkilledProposer

proposer = SkilledProposer(
    skills=[
        "./skills/prompt-engineering",                   # reads SKILL.md
        "./skills/prompt-engineering/models/openai.md",  # guidance for the student model
    ],
    additional_instructions="Write instructions in imperative voice.",
    max_words=300,
)

optimizer = dspy.GEPA(
    metric=metric,
    reflection_lm=dspy.LM("openai/gpt-5", temperature=1.0, max_tokens=32000),
    instruction_proposer=proposer,
    auto="medium",
)

optimized = optimizer.compile(program, trainset=train, valset=val)
```

## Skills

A skill is reference material for the reflection model. Each entry in `skills` can be:

- a path to a directory that contains a `SKILL.md` (the [Agent Skills](https://agentskills.io) layout)
- a path to a markdown file
- an inline string
- a `Skill(name=..., content=..., description=...)` object

If the file starts with YAML frontmatter, `name` and `description` are read from it and the block is stripped from the content. This repo ships an example at [`skills/prompt-engineering`](skills/prompt-engineering), a prompt optimization guide the reflection model can apply when rewriting instructions.

### Skills with subfolders

Loading a skill directory reads only its `SKILL.md`. Files in subfolders such as `models/` or `references/` are not loaded. This is deliberate. An agent browsing a skill can open those files when it needs them, but GEPA calls the proposer in a plain LM call with no filesystem, many times per run. What the reflection model should see is also known before the run starts, e.g., you know which student model you are optimizing. So you, the developer, pick the extra files and pass them alongside the parent skill:

```python
proposer = SkilledProposer(
    skills=[
        "./skills/prompt-engineering",                    # reads SKILL.md
        "./skills/prompt-engineering/models/openai.md",   # guidance for the student model
    ],
)
```

Each entry becomes its own `<skill>` block in the reflection prompt. Pass only the files that apply to your run. Inlining a whole skill folder would grow every proposal call for no benefit.

## Options

```python
SkilledProposer(
    skills=None,                   # skills, paths, or inline strings
    additional_instructions=None,  # guidance applied to every proposal
    base_instructions=None,        # replace the built-in meta-prompt
    max_words=None,                # word cap on proposed instructions
    max_tokens=None,               # token cap on proposed instructions
    prompt_model=None,             # (standalone GEPA only)
    max_examples=None,             # cap reflective examples per component
    on_error="keep",               # "keep" or "raise"
)
```

- `base_instructions` replaces the whole meta-prompt, including the anti-overfitting rules. If you still want those rules, include equivalent text in your replacement.
- `on_error="keep"` logs a failed proposal and keeps the current instruction, so a long GEPA run survives a flaky call. Use `on_error="raise"` during development so failures surface. Either way, LM/provider errors (`LMError`) always propagate, so a dead API key fails the run instead of silently keeping unchanged text for the whole run.
- `max_tokens` counts tokens with litellm's tokenizer when it can resolve your model name, and falls back to about 4 characters per token.

## Using the standalone gepa package

`dspy.GEPA` runs the proposer inside the reflection model's context, so you do not pass a model. The standalone [gepa](https://github.com/gepa-ai/gepa) package does not set a DSPy context, so pass the model yourself:

```python
proposer = SkilledProposer(
    skills=[
        "./skills/prompt-engineering",
        "./skills/prompt-engineering/models/openai.md",
    ],
    prompt_model=dspy.LM("openai/gpt-5", temperature=1.0, max_tokens=32000),
)
```

Then pass `proposer` wherever gepa accepts a `ProposalFn`.

## Using with Flex

dspy 3.3 added [`dspy.Flex`](https://dspy.ai/diving-deeper/flex/), a module that holds its whole implementation as Python source, which GEPA rewrites during optimization. GEPA sends Flex components to a built-in code proposer, and a custom `instruction_proposer` never sees them. The built-in prompt does not warn the reflection model against memorizing the training set, and with code the risk is worse than with instructions. The model can write a branch that matches one training input and returns its answer.

`SkilledCodeProposer` applies this package's approach to Flex source. The reflection model gets the same three step procedure, a rule against overfitting written for code, your reference skills, and your extra guidance. Every proposal is checked before it is used. It must parse and define a class with a `forward` method, or the current source is kept.

dspy has no `code_proposer` hook yet, so this package patches the built-in proposer for the duration of a `compile` call:

```python
import dspy
from skilled_proposer import SkilledCodeProposer, SkilledProposer, use_code_proposer

optimizer = dspy.GEPA(
    metric=metric,
    reflection_lm=dspy.LM("openai/gpt-5", temperature=1.0, max_tokens=32000),
    instruction_proposer=SkilledProposer(skills=["./skills/prompt-engineering"]),
    auto="medium",
)

code_proposer = SkilledCodeProposer(
    skills=["./skills/prompt-engineering"],
    additional_instructions="Prefer few predictors and plain Python.",
)

with use_code_proposer(code_proposer):
    optimized = optimizer.compile(program, trainset=train, valset=val)
```

Notes:

- This feature requires dspy 3.3 or newer. The rest of the package still works with dspy 3.0.
- dspy logs a warning that a custom `instruction_proposer` skips code components. Under the patch the warning is expected and harmless, because the patched proposer handles them.
- The patch is a bridge. We are proposing a `code_proposer` parameter for `dspy.GEPA`; once it lands, pass `SkilledCodeProposer` there and drop the patch.
- `SkilledCodeProposer` takes `skills`, `additional_instructions`, `base_instructions`, `prompt_model`, `max_examples`, and `on_error`, with the same meanings as `SkilledProposer`. There is no length budget for code.

## Limits

- v0.1 is text only. Rich values such as `dspy.Image` are stringified in the reflective examples, so the reflection model cannot see them. Multimodal support is planned for v0.2.

## License

MIT
