Skip to main content

Changelog

v0.8.1 — Positioning, Trust & Distribution (2026-07-09)​

Patch release focused on making the product story match the engineering quality (per competitive analysis).

  • README and all docs now lead with local-first, auto-injecting, self-aware context + memory layer + Trust Contract.
  • Hooks positioned as primary UX.
  • reporecall stats / get_stats prominently show freshness.
  • New npm run benchmark:tokens.
  • New competitive positioning doc.
  • Full sync across README, CLAUDE.md, website/docs, competitor matrix.

See root CHANGELOG for details.

v0.8.0 — Trust-Contract Remediation (2026-07-09)​

A breaking MCP surface change focused on freshness honesty, leaner injected context, smaller storage, and less agent tool-choice confusion.

Breaking — six-tool MCP surface​

The public MCP surface collapsed to exactly six tools:

  • search_context, search_code, explain_flow, memory, refresh_context, get_stats.

Standalone navigation and memory tools were folded into action-based verbs — search_code action=read_chunk; explain_flow action=callers|callees|stack_tree|imports|symbol|resolve_seed; memory action=recall|explain|list|store|forget. Destructive, wiki, business, and topology standalone tools were removed from public MCP registration; CLI Lens/explain JSON is now the structured export path for business and wiki data. See MCP Tools for the full mapping.

Added — freshness honesty​

  • Staleness metadata and warning banners across MCP responses and hook-injected context, with explicit empty-index repair guidance.
  • indexedCommit stamping on completed index/refresh passes (including no-change verification passes).
  • Auto-refresh on stale daemon indexes, with debouncing and a freshness-aware repair path.
  • Storage visibility in stats: metadata DB bytes, free bytes, target count, and target-alias count.
  • Design note for the tool collapse at docs/design/wp5-tool-collapse.md.

Changed — leaner context & compact storage​

  • Prompt-context injection is leaner: broad fallbacks reduced, vanished evidence disclosed, advisory text counted against budgets, and compression expansion points use search_code action=read_chunk.
  • Target aliases are deduped, capped per target, and pruned on startup for legacy databases (compact metadata storage).
  • Hook installation templates now surface curl/daemon failures instead of silently swallowing them.
  • Common test/spec fixture directories are ignored by default as code evidence.

Fixed​

  • Empty code indexes no longer block refresh_context or independent memory actions.
  • Memory write/read actions are no longer falsely coupled to code-index freshness.
  • Legacy indexes without commit stamps are treated honestly as stale in git repos.
  • The architecture lens no longer crashes with Cannot read properties of undefined (reading 'communities') (bound community-membership lookup in src/visualize/data-extractor.ts).

Earlier releases​

  • v0.7.1 — Self-evaluation patch: reporecall index generates deterministic wiki/business pages; intent-classifier routing fixes; single-transaction bulk delete; auth over-matching and confident-wrong seed fixes.
  • v0.7.0 — Capability evidence & business-context export: generic capability evidence for trace/architecture/change; productAreas[] / businessPages[] in lens --json; documented Codex support.
  • v0.6.x — Lens architecture dashboard and the generated wiki layer.
  • v0.5.0 — Topology-aware search (Louvain communities, hubs, surprises).
  • v0.4.0 — Intent-based retrieval modes (replacing the old R0/R1/R2 model).
  • v0.3.0 — Memory V1 (rules, facts, episodes, working context).
  • v0.1.0 – v0.2.0 — Initial release and core retrieval: Tree-sitter parsing, hybrid search, SQLite FTS, MCP integration, CLI, and Claude Code hooks.

For the complete, authoritative history, see the CHANGELOG.md shipped with the package.