This release adds 2 notable features for engineering teams evaluating rollout.
✓ No known CVEs patched in this version
Summary
AI summaryAdded PromptKind classification and multi-step conflict resolution with evidence gathering in Librarian.
Full changelog
Bundle 2 — Staff infrastructure cleanup
S3: Mock provider classifier refactor
Extracted the mock provider's substring dispatch into a pure classifyPromptKind(message) -> PromptKind function plus a per-kind counter. Tests can now assert exactly which classifier branches were exercised by a run, instead of guessing from fallthrough counts.
- New
PromptKindunion (8 kinds:memory_need,entity_extraction,task_inference,relevance_filter,conflict_resolution,fact_extraction,compression,unknown). kindCountstracker onMockProvider, exposed viagetKindCounts()/resetKindCountsExternal()and module-levelgetMockKindCounts()/resetMockKindCounts().complete()classifies once at the top, increments the counter, then dispatches viaswitch (kind).- Behavior preserved bit-for-bit;
unknownfalls through to the existing canned researcher JSON path.
S4: Multi-step Librarian conflict resolution
Before resolveWithReasoning calls the LLM, gather up to 5 sibling facts on the same entity (excluding the conflicting key) ranked by confidence and feed them to the model as evidence. The prompt now asks "Which value is more consistent with the related facts above?" as a new consideration. Best-effort: if the tx query fails, the helper returns [] and the original single-call behavior runs unchanged.
- New helper
gatherConflictEvidence(existing, tx)withMAX_CONFLICT_EVIDENCE_FACTS=5and 200-char summary truncation. - New helper
formatConflictEvidence(evidence)renders as markdown bullet list. resolveWithReasoningrewired to await the gather step before the LLM call.- New
librarian.conflict_multi_stepcounter in metrics for dashboard observability.
Tests
tests/mock-provider/run_b2_classifier_tests.ts(21 tests, all green)- New script:
npm run test:b2-mock-classifier - Backward compat verified on 9 suites: mock-provider, mock-failure-modes, router, user-operating-rules, compliance-task-relevance, file-change-recall, mid-turn-attend, tool-call-retrieval, m2-tool-result-autowrite.
Diffstat
src/lib/providers/mock.ts+212 / -46src/librarian/index.ts+59 / -1src/lib/metrics.ts+1- 5-file version bump to 0.3.25
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About nfemmanuel/iranti
Persistent shared memory for AI coding agents. Stores facts as `entity/key/value` triples with hybrid semantic search, task checkpoints, and conflict resolution — shared across Claude Code, Codex CLI, and GitHub Copilot.
Related context
Related tools
Beta — feedback welcome: [email protected]