Skip to content

07-fidelity-modes.md

Latest commit

 

History

History
84 lines (64 loc) · 4.26 KB

07-fidelity-modes.md

File metadata and controls

84 lines (64 loc) · 4.26 KB
Raw

07 - Fidelity Modes Pipeline

Engine-feature demo — this guide teaches a single mechanism: fidelity attributes (full, summary, none) controlling how much prior context each node carries. 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="$PWD/examples/pipelines/07-fidelity-modes.dot"
mkdir -p /tmp/attractor-demo && cd /tmp/attractor-demo
attractor run "$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

  • Fidelity attribute on nodes: Each node explicitly sets how much prior context to carry
  • Fidelity attribute on edges: The edge from integration_test -> final_review overrides the node's fidelity
  • thread_id for session reuse: implement_auth and implement_rate_limit share thread_id="api-impl", meaning they reuse the same LLM session under full fidelity
  • All fidelity modes: truncate, full, compact (graph default), summary:medium, summary:high
  • Fidelity resolution precedence: edge fidelity (highest) > node fidelity > graph default_fidelity > system default ("compact")

Pipeline Structure

start -> architect -> implement_auth -> implement_rate_limit -> implement_logging -> integration_test -> final_review -> done
         (truncate)   (full, thread    (full, thread            (summary:medium)     (compact, from      (edge overrides
                       ="api-impl")     ="api-impl")                                  graph default)      to summary:high)

Fidelity Resolution for Each Node

Node Node Fidelity Edge Fidelity Graph Default Resolved Session
architect truncate -- compact truncate Fresh (goal + run ID only)
implement_auth full -- compact full Reused (thread "api-impl")
implement_rate_limit full -- compact full Reused (thread "api-impl")
implement_logging summary:medium -- compact summary:medium Fresh (~1500 token summary)
integration_test (none) -- compact compact Fresh (bullet-point summary)
final_review compact summary:high compact summary:high Fresh (~3000 token summary)

Key: Edge fidelity on integration_test -> final_review overrides final_review's own fidelity="compact".

Context Preamble Content by Mode

Mode Content
truncate "Goal: Build a REST API...\nRun ID: ..."
full Full conversation history preserved in LLM session
compact Bullet-point list: completed stages, outcomes, context.* values
summary:medium Stage outcomes with notes, active context.* values (~1500 tokens)
summary:high Detailed outcomes with failures, context updates (~3000 tokens)

Thread Reuse Behavior

implement_auth and implement_rate_limit both use thread_id="api-impl" with fidelity="full":

  • implement_auth starts a new LLM session with key "api-impl"
  • implement_rate_limit reuses that same session -- it sees the full conversation from implement_auth
  • This is powerful for multi-step implementations where each step builds on the previous one's context

Or run from a bundle / recipe

steps:
  - agent: attractor:pipeline-runner
    instruction: "Run the fidelity modes pipeline"
    context:
      pipeline_path: "examples/pipelines/07-fidelity-modes.dot"

What to Look For

  • architect/prompt.md should have minimal preamble (truncate mode)
  • implement_auth and implement_rate_limit should share a session thread
  • implement_logging/prompt.md should have a medium-detail summary preamble
  • integration_test/prompt.md should have a compact bullet-point preamble
  • final_review/prompt.md should have a high-detail summary (edge override worked)
  • Checkpoint resume: if the pipeline resumes from checkpoint, full fidelity degrades to summary:high for the first resumed node (session state can't be serialized)