Skip to content
This repository was archived by the owner on Oct 27, 2026. It is now read-only.

v0.3.0 — Trajectory, Feedback, Companion Dashboard

Choose a tag to compare

View all tags
@m3data m3data released this 10 Mar 03:23
· 12 commits to main since this release
v0.3.0
7c78aa1

What's new

Contextual query construction

The hook no longer treats each prompt in isolation. build_contextual_query() blends the current message embedding with up to 5 recent query embeddings using exponential decay weighting. Context weight is capped (default 0.4, structural ≥60% current dominance) and adjusts based on trajectory signal — converging conversations narrow context, diverging ones widen it. Graceful fallback when insufficient history.

Semantic trajectory tracking

Frenet-Serret local curvature (delta-kappa) computed over accumulated query embeddings. Outputs a trajectory signal: converging, diverging, stable, or insufficient_data. Anti-entrainment override: when converging + narrow results, the diversity pool widens automatically. Trajectory persists across hook invocations via file-locked numpy buffer.

Relevance feedback and auto-labelling

Every result surfaced by the hook is auto-labelled (useful or noise) based on similarity threshold. Human corrections via the dashboard write corrected:<user> entries. Labels feed into Bayesian relevance weights (prior=2.0) with latest-wins semantics per (file, query) pair. Append-only storage preserves full audit trail.

Companion dashboard (SPEC-003)

Local web app (dashboard/serve.py + dashboard/index.html) on port 8111. No external dependencies beyond stdlib. Renders:

  • Hit rate hero metric (session-scoped, colour-coded health thresholds)
  • Query timeline with expandable results (file path, section, score, source type, label)
  • Trajectory signal with delta-kappa values
  • Feedback stream (auto-labels + human corrections)

Click any result label to correct it. Auto-starts/stops via Claude Code session hooks.

Hook unification

Discovered and fixed: two separate hooks (sense-auto-query.py and hook.py) maintained separate state files, leaving the SPEC-001/SPEC-003 pipeline dead for all Claude Code invocations. Single entry point sense-mcp-hook now handles everything via shared session state at /tmp/sense-session-state.json.

Audit fixes

  • P0: ro_conn NameError in hook.py silently killed auto-labelling/trajectory/ambient search (swallowed by except Exception: pass)
  • COALESCE/GROUP BY mismatch in serve.py and feedback.py
  • Per-line JSONDecodeError handling in trajectory reader
  • build_contextual_query() return signature changed to (embedding, metadata) tuple

Stats

  • 188 tests, 0 failures
  • 6 commits since v0.2.0

Full Changelog: v0.2.0...v0.3.0

Assets 2
Loading