Skip to content

06-model-stylesheet.md

Latest commit

 

History

History
96 lines (77 loc) · 4.87 KB

06-model-stylesheet.md

File metadata and controls

96 lines (77 loc) · 4.87 KB
Raw

06 - Model Stylesheet Pipeline

Engine-feature demo — this guide teaches a single mechanism: CSS-style model_stylesheet rules for per-node model selection. For the canonical attractor shape used in real work, start with Tutorial 00: The Convergence Loop.

Run it

Self-contained -- the goal is baked into the .dot, so no --param is needed. From the attractor repo root:

DOT="$PWD/examples/pipelines/06-model-stylesheet.dot"
mkdir -p /tmp/attractor-demo && cd /tmp/attractor-demo
attractor run "$DOT" --cwd .

See README.md in this folder for the run pattern and why the $DOT capture + cd + --cwd . are needed (box-node process-cwd alignment + dot-path resolution).

What This Exercises

  • Stylesheet parsing: The model_stylesheet graph attribute is parsed into StyleRule objects with selectors, specificity, and properties
  • CSS selectors:
    • * (universal, specificity=0): Applies to all nodes as a baseline
    • .class (class selector, specificity=2): Targets nodes with matching class attribute
    • #id (ID selector, specificity=3): Targets a specific node by its ID
  • Specificity resolution: Higher specificity wins. #critical_review (3) > .code (2) > * (0)
  • Explicit node attribute override: quick_fix has llm_model="gemini-2.5-flash-preview-05-20" directly on the node, which beats any stylesheet rule (highest precedence)
  • Recognized properties: llm_model, llm_provider, reasoning_effort

Pipeline Structure

start -> analyze -> refactor -> lint_check -> critical_review -> quick_fix -> done
         (.planning) (.code)    (.fast)       (#id + .code)     (.code + explicit)

Resolved Model Assignments

Node Class Matching Rules Winner (by specificity) Final Model
analyze planning *(0), .planning(2) .planning gpt-[5-9]* (openai)
refactor code *(0), .code(2) .code claude-sonnet-* (anthropic)
lint_check fast *(0), .fast(2) .fast gemini-*-flash (gemini, low)
critical_review code *(0), .code(2), #critical_review(3) #critical_review claude-opus-* (anthropic, high)

How these globs resolve — and how evergreen they are. Each glob is copied into node.attrs["llm_model"] verbatim, then the engine resolves it against the provider's LIVE model list at run time (newest stable match wins). Evergreen-ness depends on whether the provider keeps a stable TIER NAME across generations:

  • Anthropic (claude-sonnet-*, claude-opus-*) and Gemini (gemini-*-flash) — the tier persists, so the glob tracks new generations indefinitely.
  • OpenAI has no persistent tier (the generation is the name), so .planning uses a generation RANGE gpt-[5-9]* — tracks the newest through gpt-9, and needs a one-char bump at gpt-10.

A bare concrete id (e.g. claude-sonnet-4-6) is NOT resolved — it's passed to the provider as-is and 404s once retired. Pin a concrete id only for locked evals. | quick_fix | code | *(0), .code(2) | .code BUT node has explicit llm_model + llm_provider | gemini-*-flash on gemini (explicit override) |

Overriding a model glob on a node? Override the provider too. A glob is resolved against the node's provider, so llm_model="gemini-*-flash" needs llm_provider="gemini" alongside it — otherwise it inherits anthropic from the .code class and matches nothing. (Concrete ids bypass resolution and are provider-inferred; globs are not.)

Expected Behavior

  1. Stylesheet is parsed during the INITIALIZE phase (before execution)
  2. apply_stylesheet() walks all nodes:
    • For each node, finds all matching rules
    • For each property, keeps the highest-specificity match
    • Only sets properties the node doesn't already have explicitly
  3. During execution, each node's llm_model, llm_provider, and reasoning_effort are available in node.attrs for the backend to use
  4. The codergen handler passes these to the backend via the node parameter

Or run from a bundle / recipe

steps:
  - agent: attractor:pipeline-runner
    instruction: "Run the model stylesheet pipeline"
    context:
      pipeline_path: "examples/pipelines/06-model-stylesheet.dot"

What to Look For

  • After stylesheet application, inspect node attrs:
    • analyze.attrs["llm_model"] == "gpt-[5-9]*"
    • critical_review.attrs["llm_model"] == "claude-opus-*" (ID selector wins over .code class; resolved to a concrete opus id at run time)
    • quick_fix.attrs["llm_model"] == "gemini-*-flash" (explicit attribute wins)
  • Validation passes (stylesheet syntax is valid)
  • Each node's prompt.md is written with the correct model context
  • No stylesheet properties are applied to start/exit nodes (they have no LLM interaction)