Shadow v0.6.1 — Recall-first search descriptions
Behavior fix on top of v0.6.0. Single-commit patch.
Why
Real-use feedback: Shadow was not calling shadow_memory_search on its own. Root cause was wording — the three search-tool descriptions framed the call as reactive ("use when the user references the past"), and the consuming LLM honored that too strictly. Settled decisions and anti-patterns lived in memory but were not recalled before proposing.
What changed
The descriptions of shadow_memory_search, shadow_memory_similar, and shadow_search (cross-entity) now describe themselves with explicit proactive triggers:
- Before naming / refactor / library / approach proposals
- Before opining on conventions
- Before estimating effort
- Whenever the user mentions a decision, pattern, convention, or "lo que ya hicimos"
The recall cost (<50ms) is also stated, removing the implicit incentive to skip "just in case it's expensive."
Zero code changes — descriptions only. Clients that cache tools/list will pick up the new wording on the next session (re-initialize).
Companion (per-installation)
A convention memory titled "Recall-first: search memory antes de proponer" was teach'ed into the local Shadow DB so it surfaces in contextKnowledge on shadow_check_in. That memory lives per-install; only the description rewrite is in the repo.
Numbers
- 1 commit (
412e096) +chore(release): 0.6.1 - 3 lines of description text changed
- 0 new tests, 0 regressions
- 0 schema migrations, 0 new env vars
Upgrade
git pull && npm run build && shadow daemon restart. No further action.
Full diff: v0.6.0..v0.6.1.