v0.3.33

Added

• Added write-time reconciliation ("reconcile-on-write"): after trace ingestion
  persists a durable record, Lerim runs a scoped context-curation pass over just that
  record plus its active semantic neighbors, so a genuine update supersedes the record
  it replaces at write time instead of waiting for the next periodic curate pass. It
  reuses the existing curator, the supersede change-kind, and the bi-temporal
  valid_until / superseded_by_record_id columns, so a stale record drops out of
  default retrieval immediately — no new DSPy signature and no schema migration.
• Added a scoped context-curator mode: run_context_curator(..., seed_record_ids=...)
  and ContextCuratorPipeline(seed_record_ids=...) review a seed record plus its active
  semantic neighbors (load_seed_and_neighbors) and skip single-record health review. A
  new protected_record_ids guard in apply_context_curation_plans keeps a just-written
  record from being retired by the same trace that created it; the periodic curator
  remains the safety net for anything the scoped pass leaves untouched.

v0.3.32

Added

• Added a per-record feedback signal and earned confidence: a new record_feedback
  table, a confidence column on records, and ContextStore.record_feedback /
  list_feedback. Confidence is updated in place (not versioned) and never bumps the
  index generation, so feedback stays cheap and does not churn FTS/embeddings.
• Added the lerim_context_feedback MCP tool, a lerim feedback <record_id> <used|correct|wrong|confirm> CLI command, and an api_feedback HTTP endpoint
  (POST/GET /api/records/{record_id}/feedback). Feedback references existing records
  only — it is not a memory_save primitive.
• Earned confidence and recency now factor into hybrid retrieval ranking, so confirmed
  records rise and repeatedly-wrong records fade. Added a relevance floor so retrieval
  returns fewer or no results instead of padding out to limit with irrelevant matches.
• Added automatic content-based secret/PII redaction at the ingestion edge
  (src/lerim/redaction.py), applied in write_session_cache (all native adapters) and
  write_compact_trace (generic-trace import) so secrets never reach records or skills.
• Added dashboard feedback controls, real earned-confidence display, and a
  provenance / proof-of-value view showing a record's source session and evidence.
• Added docs/contracts/feedback-and-confidence.md documenting the feedback + confidence
  JSON contracts across store, API, HTTP, CLI, and MCP.

Changed

• Corrected the Reciprocal Rank Fusion constant from RRF_K = 2 to the standard 60,
  so fused order reflects agreement between the semantic and lexical retrievers rather
  than either retriever's noisy rank-1 pick.
• Redaction is now automatic at the ingestion edge, building on the earlier manual
  scripts/redact_trace.py helper.
• Skill-stewardship validation now verifies that a proposal's cited evidence records
  actually exist and that the proposed text is supported by them, and treats AGENTS.md
  as a first-class instruction target when co-located with SKILL.md.

---

Install: pip install lerim==0.3.32 · Docker ghcr.io/lerim-dev/lerim:0.3.32

Full changelog: v0.3.31...v0.3.32

v0.3.31

Added

• Added bundled research and compliance signal profiles, registered as
  first-class source profiles alongside coding, support, ops, and generic.
• Added realistic multi-turn example traces for support, incident, research,
  and compliance workflows, replacing the earlier two-line toy traces.
• Added before/after demo guides for support ops, incident ops, research, and
  compliance, each showing real extracted records captured through extraction.
• Added a custom and non-coding agents front-door guide covering the profile,
  clean-trace, import, query, and improvement-loop path.
• Added a defense-in-depth redaction helper (scripts/redact_trace.py) that
  strips common secrets and PII from a trace before import.
• Added an agent improvement loop diagram to the README and custom-agents guide.

Changed

• Updated the README with a Custom and Non-Coding Agents section, the bundled
  profile table, and links to the vertical demos.
• Updated the docs navigation, examples index, and custom-trace submission guide
  to cover the new profiles, demos, and redaction helper.
• Updated the custom-trace-folder and context-retrieval diagrams to use neutral
  support and research agent examples instead of removed third-party logos.

---

Install: pip install lerim==0.3.31 · Docker ghcr.io/lerim-dev/lerim:0.3.31

Full changelog: v0.3.30...v0.3.31

v0.3.30

Fixed

• Fixed dashboard operation detail logs so they honor the selected project instead of falling back to shared logs.
• Fixed project-scoped operations activity by paging through the session catalog until matching project runs are found.
• Hardened source-session catalog migrations and FTS rebuilds so older or malformed catalogs recover into a healthy indexed state for graph and source views.

v0.3.29

Fixed

• Fixed dashboard graph project switching by enforcing selected-project scope on graph responses and clearing stale graph data during project reloads.
• Made dashboard record, graph, artifact, operations, and source views fail closed instead of preserving stale rows after scoped load failures.
• Fixed the Records tab All statuses filter so it returns active and archived project records instead of falling back to the active-only default.
• Hardened the dashboard launcher against mismatched backend versions so an old local API cannot silently serve stale project data to a newer dashboard.
• Made invalid project selections return clear errors for status, memory, and Clinic endpoints, and kept refine reports degraded instead of crashing when the session catalog is unavailable.

Published artifacts:

• PyPI: lerim 0.3.29
• Docker: ghcr.io/lerim-dev/lerim:0.3.29

v0.3.28

Fixed

• Updated the offline release integration expectation for project-list count fields so the release workflow validates the new dashboard project counts.

Includes

• Dashboard project selection fixes and active/archived/total record count clarity from v0.3.27.

v0.3.27

Changed

• Made dashboard project selection visibly affect graph, records, brief, Clinic, source, operations, overview, and insights views, while keeping shared runtime surfaces clearly labeled.
• Separated current active, archived, and total persisted record counts across project lists, status payloads, and Run Clinic.

Fixed

• Fixed plain missing project names being interpreted as relative paths, so invalid dashboard project selections now return a clear error instead of falling back to the current workspace.
• Scoped graph queries, record filters, run detail reads, and artifact history through the selected project, including child workspace paths.
• Kept lerim serve and dashboard session-backed endpoints available in a degraded state when the source-session catalog is unhealthy.

v0.3.26

Added

• Added project selection to the dashboard surfaces that read project-scoped data, including overview, analytics, context, graph, traces, memory artifacts, operations, and Run Clinic.
• Added the Run Clinic dashboard artifact surface and project-scoped generated artifact history views.

Changed

• Kept selected projects in dashboard navigation URLs while leaving shared Skills and Settings views unscoped.
• Labeled shared runtime and multi-project operation rows clearly when they appear inside a project-scoped dashboard view.

Fixed

• Scoped dashboard status, queue, source-session, graph, artifact-history, and record-open reads to the selected project, including child workspace paths.
• Prevented cross-project artifact history and mixed-project operation totals from leaking into project-scoped dashboard views.

v0.3.25

Added

• Added README and docs guidance for registering skill targets, reviewing skill update proposals, and using opt-in auto-apply.

Changed

• Refined the Skills dashboard proposal review into a desktop queue/detail layout with explicit risk labels and cleaner applied-proposal state.
• Reduced normal dashboard tab gutters so pages use the available workspace with compact responsive padding.

v0.3.24

[0.3.24] - 2026-05-31

Added

• Added skill stewardship registration, refresh, proposal review, diff preview, and apply/reject flows across the CLI, API, and dashboard.
• Added scoped record loading, artifact-surface validation, and opt-in auto-apply policy controls for registered instruction targets.

Fixed

• Hardened skill proposal application so stale concurrent apply attempts cannot revert files after a proposal has already been marked applied.