lsm / neokai / 27559611059 / 28
82%
dev: 82%

DEFAULT BRANCH: dev
Ran
Files 351
Run time 14s
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst
coverage: 18.735% (+0.03%) from 18.704%
27559611059.28

push

github

web-flow 
feat(daemon): rank scope lessons by task relevance before prompt injection (#2081)

* feat(daemon): rank scope lessons by task relevance before prompt injection

selectActiveLessonsForTask previously sliced the 3 newest active lessons
with no task-specific filtering. This wasted prompt budget on irrelevant
lessons and explained the low lessons_reused metric (1/9).

Add lightweight relevance scoring based on:
- tag overlap between lesson.appliesTo and task.labels (+10 per match)
- keyword overlap between task text and lesson rule/why (+2 per match)
- lesson confidence (+3 * confidence)
- relative recency within the active set (0-5 tiebreaker)

Lessons are now sorted by score desc, then updatedAt desc, before slicing
to the limit. Includes unit tests for tag ranking, keyword ranking,
confidence fallback, and tie-breaking.

* fix(daemon): cap recency tiebreaker below smallest relevance signal

Normalize recency to 0-0.99 so a single keyword match (+2) or tag match (+10) always outranks any recency delta. Prevents unrelated newly-updated lessons from displacing task-specific ones in the top-3 slot.

* fix(daemon): make confidence and recency true tiebreakers below keyword matches

- Scale confidence to 0-0.5 (was 0-3) so max confidence + max recency (0.49) = 0.99 < a single keyword match (+2).

- Scale recency to 0-0.49 (was 0-0.99).

- Trim appliesTo and labels before exact tag-overlap scoring so whitespace from judge output does not miss matches.

- Update fallback test to use confidence 1.0 vs 0.0 so the assertion remains stable against recency.

19547 of 104332 relevant lines covered (18.74%)

9.59 hits per line

Source Files on job daemon-online-rpc-1 - 27559611059.28