mirror of
https://github.com/openprose/prose.git
synced 2026-09-19 05:55:05 +08:00
dan
624e75ad89
* Refresh OpenProse skill conventions * Align OpenProse skill with service system conventions * Fix OpenProse PR review inconsistencies * Clarify OpenProse state backend loading * Enforce OpenProse smoke run state gate
101 lines
3.0 KiB
Markdown
101 lines
3.0 KiB
Markdown
---
|
|
name: refine
|
|
kind: pattern
|
|
---
|
|
|
|
# Refine
|
|
|
|
Iteratively improve a result through delegation rounds until a quality threshold is met.
|
|
|
|
### Metadata
|
|
|
|
- `version`: 0.2.0
|
|
- `role`: coordinator
|
|
|
|
### Slots
|
|
|
|
- `refiner`
|
|
- `evaluator`
|
|
|
|
### Config
|
|
|
|
- `max_rounds` (integer, default: 3): Maximum number of refinement rounds
|
|
- `threshold` (number, default: 0.8): Score at which the result is accepted
|
|
|
|
### Invariants
|
|
|
|
- The loop is bounded by `max_rounds`
|
|
- The evaluator scores only the current result against the original task
|
|
- The refiner receives its prior output and evaluator feedback on retries
|
|
- The final output is the first result meeting `threshold`, or the last attempted result when the budget is exhausted
|
|
|
|
### Shape
|
|
|
|
- `self`: manage refinement rounds, pass evaluator feedback to refiner
|
|
- `delegates`:
|
|
- `refiner`: produce or improve a result
|
|
- `evaluator`: score the result 0..1 and suggest improvements
|
|
- `prohibited`: none
|
|
|
|
### Requires
|
|
|
|
- Pattern instance receives:
|
|
refiner: string -- service or system name for the refiner
|
|
evaluator: string -- service or system name for the evaluator
|
|
task_brief: string -- the task
|
|
max_rounds: number -- (optional, default 3)
|
|
threshold: number -- (optional, default 0.8) score at which to stop
|
|
|
|
### Ensures
|
|
|
|
- Round 1: refiner produces initial result from the task brief
|
|
- Evaluator scores the result (0..1) and provides specific improvement suggestions
|
|
- If score >= threshold: return immediately
|
|
- If score < threshold: refiner receives the result, score, and suggestions
|
|
- Each round accumulates improvement — refiner sees its own prior output
|
|
- Returns when threshold met or max_rounds exhausted
|
|
- `result`: the final output
|
|
- `score`: the final score
|
|
- `rounds_used`: number of rounds
|
|
|
|
### Delegation
|
|
|
|
```prose
|
|
let current_result = null
|
|
let current_score = 0
|
|
let improvement_suggestions = null
|
|
|
|
repeat max_rounds as round:
|
|
let current_result = call refiner
|
|
task_brief: task_brief
|
|
current_result: current_result
|
|
current_score: current_score
|
|
improvement_suggestions: improvement_suggestions
|
|
|
|
let evaluation = call evaluator
|
|
task_brief: task_brief
|
|
result: current_result
|
|
|
|
current_score = evaluation.score
|
|
improvement_suggestions = evaluation.suggestions
|
|
|
|
if current_score meets threshold:
|
|
return {
|
|
result: current_result,
|
|
score: current_score,
|
|
rounds_used: round
|
|
}
|
|
|
|
return {
|
|
result: current_result,
|
|
score: current_score,
|
|
rounds_used: max_rounds
|
|
}
|
|
```
|
|
|
|
### Notes
|
|
|
|
The refiner does not know it is in a refinement loop. The evaluator does not know its score drives iteration.
|
|
|
|
Different from `retry-with-learning`: refinement improves work that is mediocre — the result exists but is not good enough. Retry-with-learning recovers from failure — the result is broken or absent. Refinement uses a continuous quality score (0..1) and improvement suggestions. Retry uses binary failure detection and failure analysis. A result that scores 0.4 needs refinement. A result that throws an error or returns nothing needs retry.
|