v0.5.0 — retrieval keywords

Minor release focused on retrieval quality and the feedback loop.

Retrieval keywords

Decisions now carry a curated keywords field — the synonyms and code identifiers an engineer might use in a task description that the decision's own wording doesn't contain.

• Matched at serve time alongside pattern/rationale; keywords are ordinary tokens, so over-used ones self-deprecate via idf and admission still requires the same evidence floor (#61)
• Populated everywhere decisions are born: bootstrap extraction, feedback refinement, and agent submission over MCP (#62)
• metatron enrich-keywords backfills existing corpora in batched LLM calls with per-batch progress and a cost estimate (#63)
• Visible and editable in the curation UI: chips in the decision drawer, a comma-separated field in the editor (#64)
• The global synonym/alias table is retired — vocabulary gaps are now closed per decision, validated by replaying recorded production queries (#65)

Serving and feedback fixes

• Scope matching handles glob scopes (src/services/**), the literal global scope, and architectural-area names ("billing") via segment containment (#58)
• Helpfulness scoring counts binary helpful/unhelpful feedback (not just graded ratings) and centers against a leave-one-out corpus baseline, so positively-skewed model ratings still produce a usable ordering signal (#59)
• Near-duplicate candidates are skipped at submission and refinement — including resubmissions of human-rejected decisions — protecting the curation queue (#60)

Install / upgrade

pip install -U getmetatron
# or
uv tool upgrade getmetatron

Docker: kerbelp/getmetatron:0.5.0 (also :0.5, :latest).

After upgrading, run metatron enrich-keywords once per repo to backfill keywords on your existing canonical decisions.

Full changelog: v0.4.0...v0.5.0

v0.3.0

Added

• Per-repo database catalog — each repository gets its own store under ~/.metatron, with one-time auto-migration from the legacy single-file metatron.db.
• Employee attribution — local identity with metatron whoami, served/feedback events stamped with the actor, and metatron import to merge contributors' databases into a curator's catalog.
• New React web UI wired to the live API — Agent Impact, candidate review, and drill-down side drawers.
• metatron export <repo> for handing a repository's decisions to another store.

Changed

• Renamed "priors" to "decisions" across the project.
• SQLite runs in autocommit for real-time reads, with WAL and a busy timeout for concurrent multi-process access.
• Terminology and UX polish — "canonical" glossed as "approved & served to agents"; timestamps surfaced on candidates and decisions.

Fixed

• Boot errors surfaced instead of an endless splash.
• Agent-activity status vocabulary and last-active timestamps corrected.
• Assorted UI fixes (helpful-rate NaN, feedback-loop crash, contrast, ingest counts).

Hero demo

Two-minute product demo: extracting decisions from a codebase, human curation with an AI judge, an agent consuming decisions over MCP, feedback refined into a new decision, and that decision reused by another engineer. Assets for the README and site.

v0.2.1

Release v0.2.1 with MIT license, glama.json and CI integration.