refactor: remove self-learning and the optimizer package (P1.4)

Self-learning is excluded from the community edition (ADR-0004); the
project owner reworks it separately. Cost tracking stays.

- agent: delete behaviour evolution (evolve-schema, iterative GEPA,
  compose-evolution, evolution gates, orchestrator and deploy),
  improvement and correction detection, AgentLearning, the eval-dataset
  builder, optimization-capture and the agent_insights tool. Secret
  detection moves from eval-dataset to secret-patterns. The LLM wiring
  keeps only the Haiku adapter and judge call used by skill audit and
  hygiene, so @ax-llm/ax is now declared by agent instead of server.
- server: delete /api/evolution/*, the evolution and optimizer services,
  GEPA expansion of the first chat message, the prompt_optimization and
  monthly_assessment jobs, the idle-session reviewer and its
  session-reviewer persona, prompt injection of past corrections and
  thumbs-down recorded as correction signals. Existing job rows of the
  removed types are ignored, not deleted.
- core: delete OptimizationLogStore; stop re-exporting the improvement
  signal and evolution run stores.
- web: delete the Evolution screen and the two job types in the
  automation editor.
- delete packages/optimizer.

Also logs UF-027 (self-learning loops read data nothing collects) and
UF-028 (the Settings prompt-shape choice is ignored by chat).
This commit is contained in:
shumeiko.it
2026-10-20 00:54:58 +02:00
parent b734c1d3d2
commit 2e1999a3f6
112 changed files with 336 additions and 16165 deletions

View File

@@ -1,5 +1,6 @@
/**
* LLM-as-Judge — rubric-based scorer used by the evolution loop.
* LLM-as-Judge — rubric-based scorer used by skill audit, skill hygiene and
* loop jobs.
*
* Scores a candidate response against an expected ground truth along
* three weighted dimensions:
@@ -10,9 +11,8 @@
*
* Then applies a length penalty (soft cap) multiplied into the final score.
*
* The textual feedback produced is the input that drives GEPA's reflective
* mutations — it must explain *what* went wrong (or right) so the next
* prompt variation can address it.
* The textual feedback must explain *what* went wrong (or right) so the
* caller can act on it.
*
* Model-agnostic — caller provides the llmCall function. Defaults assume
* a Haiku-class judge via LiteLLM, but any OpenAI-compatible backend works.
@@ -42,7 +42,7 @@ export interface JudgeScore {
conciseness: number;
/** Multiplier applied to the weighted score (0..1). */
lengthPenalty: number;
/** Free-form feedback used as reflective-mutation input for GEPA. */
/** Free-form feedback explaining the score. */
feedback: string;
/** True when the judge successfully parsed a structured response. */
parsed: boolean;