Skip to content

feat: outcome-feedback learning loop + run-telemetry to R_eff/decay #365

Description

@explosivebit

Context

From a study of the external ECC project (github.com/affaan-m/ECC). ECC's continuous-learning-v2 observes sessions via hooks → atomic "instincts" (confidence 0.3–0.9, project-scoped) → /evolve clusters them into skills; its state-store also logs skillRuns[] {outcome, tokensUsed, durationMs, userFeedback} and skillVersions[] {contentHash, promoted, rolledBack}.

Proposed

  1. Outcome-feedback learning loop: a mechanism (forgeplan + Hindsight) that mines recurring session patterns into confidence-scored candidates and promotes high-confidence ones (e.g. seen in ≥2 projects, avg confidence ≥0.8) into durable artifacts (NOTE / claim / hypothesis). forgeplan already has the promotion primitives — forgeplan_claim / _hypothesis_promote; this wires the front of the loop. This is the Ruflo-style learning layer NOTE-004 said we'd adopt.
  2. Run telemetry → R_eff: log per-artifact/agent execution outcomes (outcome / tokens / duration / userFeedback) and let forgeplan_decay / _calibrate factor actual usage outcomes — not just graph structure — into R_eff.

Why

Closes the missing "learning layer": forgeplan + Hindsight capture sessions today but don't auto-synthesise durable, confidence-scored knowledge, and R_eff reflects structure, not observed outcomes.

Refs: ECC skills/continuous-learning-v2/SKILL.md, schemas/state-store.schema.json (skillRuns[]/skillVersions[]); our NOTE-004 (Ruflo) + Hindsight. Surfaced by ECC competitive-analysis; filed from ForgePlan/marketplace.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions