[RAGE-5][P0] Synthesize native Prolog skill candidates plus positive/negative trigger corpora #43

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-4
Blocked on public contract where necessary: lost-rob0t/prolog-rlm#173

Goal

For patterns classified as skill, materialize an immutable native skill candidate ready for upstream selection/outcome evaluation.

Output

Produce:

concise discovery metadata;
lean operational SKILL.md;
canonical Prolog-owned routing/dependency/graph metadata through the #173 contract;
references/resources for large knowledge;
script/tool references when deterministic procedure should not live in prose;
positive trigger cases;
adversarial negative/near-miss trigger cases;
provenance explaining the observed workflows that motivated the candidate;
candidate scope: global/project/environment.

Invariants

emitted skill remains compatible with the canonical open Agent Skills form chosen upstream;
no executable arbitrary Prolog embedded in Markdown/YAML;
skill text cannot grant tools, capabilities, authority, or promotion;
synthesis may use a model, but the output is candidate data only;
deterministic fallback/fixture path must exist without a live provider.

Acceptance

one real normalized workflow produces a complete native candidate;
positive and adversarial negative trigger fixtures are generated/stored;
source provenance is inspectable from candidate identity;
rerunning unchanged evidence is deduplicated by content/evidence fingerprint;
non-skill classifications cannot enter this path accidentally.

Parent: #6 Depends on: RAGE-4 Blocked on public contract where necessary: lost-rob0t/prolog-rlm#173 ## Goal For patterns classified as `skill`, materialize an immutable native skill candidate ready for upstream selection/outcome evaluation. ## Output Produce: - concise discovery metadata; - lean operational `SKILL.md`; - canonical Prolog-owned routing/dependency/graph metadata through the #173 contract; - references/resources for large knowledge; - script/tool references when deterministic procedure should not live in prose; - positive trigger cases; - adversarial negative/near-miss trigger cases; - provenance explaining the observed workflows that motivated the candidate; - candidate scope: global/project/environment. ## Invariants - emitted skill remains compatible with the canonical open Agent Skills form chosen upstream; - no executable arbitrary Prolog embedded in Markdown/YAML; - skill text cannot grant tools, capabilities, authority, or promotion; - synthesis may use a model, but the output is candidate data only; - deterministic fallback/fixture path must exist without a live provider. ## Acceptance - [ ] one real normalized workflow produces a complete native candidate; - [ ] positive and adversarial negative trigger fixtures are generated/stored; - [ ] source provenance is inspectable from candidate identity; - [ ] rerunning unchanged evidence is deduplicated by content/evidence fingerprint; - [ ] non-skill classifications cannot enter this path accidentally.

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