[RAGE-4][P0] Mine repeated workflow patterns and classify the correct improvement mechanism #44

New issue

Open

opened 2026-09-20 01:40:01 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-09-20 01:40:01 +00:00

Owner

Parent: #6
Depends on: RAGE-1, RAGE-2

Goal

Turn normalized observations into evidence-backed candidate improvements without forcing every repeated behavior into SKILL.md.

Required classification

skill
project instruction
reference / knowledge
script / tool
hook / policy
agent role
AgentProlog config
memory / preference
ignore / insufficient evidence

Pattern detection

Aggregate repeated signals such as:

recurring corrections or clarifications;
repeated verifier failures and the repair that resolves them;
repeated ordered tool sequences;
repeated project instructions;
repeated routing/context mistakes;
stable preferences scoped to a project/environment.

One unusual session must not become a global candidate.

Requirements

evidence count + recency + scope + confidence remain inspectable;
deterministic heuristics form the baseline;
optional model assistance may propose classification but cannot erase evidence or force promotion;
distinguish global vs project vs environment scope;
ambiguous cases remain unresolved/insufficient-evidence;
output is closed candidate-classification data consumed by later slices.

Acceptance

repeated behavior is distinguished from one-off behavior;
fixture covers every classification outcome above;
conflicting evidence remains represented rather than flattened;
candidate rationale lists the concrete evidence records that caused it;
classification can run offline deterministically.

Parent: #6 Depends on: RAGE-1, RAGE-2 ## Goal Turn normalized observations into evidence-backed candidate improvements without forcing every repeated behavior into SKILL.md. ## Required classification ```text skill project instruction reference / knowledge script / tool hook / policy agent role AgentProlog config memory / preference ignore / insufficient evidence ``` ## Pattern detection Aggregate repeated signals such as: - recurring corrections or clarifications; - repeated verifier failures and the repair that resolves them; - repeated ordered tool sequences; - repeated project instructions; - repeated routing/context mistakes; - stable preferences scoped to a project/environment. One unusual session must not become a global candidate. ## Requirements - evidence count + recency + scope + confidence remain inspectable; - deterministic heuristics form the baseline; - optional model assistance may propose classification but cannot erase evidence or force promotion; - distinguish global vs project vs environment scope; - ambiguous cases remain unresolved/insufficient-evidence; - output is closed candidate-classification data consumed by later slices. ## Acceptance - [ ] repeated behavior is distinguished from one-off behavior; - [ ] fixture covers every classification outcome above; - [ ] conflicting evidence remains represented rather than flattened; - [ ] candidate rationale lists the concrete evidence records that caused it; - [ ] classification can run offline deterministically.

nsaspy referenced this issue

2026-09-20 01:40:38 +00:00

[EPIC] Personal skill refinery: observe workflows, synthesize candidates, RAGE-evaluate, and import external skills #6