nfemmanuel/iranti

Bundle 7: S5 Archivist reasoning budget + S6 council mode

Lands the last two S-series Staff agency items as pure deterministic helpers, following the A4/A5 propose-but-don't-execute pattern.

S5 — Archivist reasoning budget

• planArchivistReasoningPass() in src/staff/archivistReasoning.ts is a pure helper over a candidate-entry snapshot. Never mutates, never calls an LLM, never hits I/O.
• Emits bounded proposals (default budget 5 per cycle) in priority order:
  1. compress — clusters of ≥3 duplicates sharing the same entityType/entityId/key
  2. flag_drift — clusters with confidence spread > 30
  3. demote — single entries below the review threshold AND unaccessed for ≥ staleAccessDays
  4. review_stale — any single entry unaccessed for ≥ 2× staleAccessDays
• Each proposal carries a snapshot metadata block (targetCount, minConfidence, maxConfidence, oldestLastAccessedAt, newestLastAccessedAt).
• runArchivist() now calls this pass after escalation processing. Proposals fire as staff events (reasoning_proposal_emitted) and ride along on ArchivistReport.reasoningProposals for downstream consumption by the Resolutionist.

S6 — Cross-Staff consultation (council mode)

• planCouncilConsultation() in src/staff/council.ts is a pure helper that proposes peer-Staff consultations before a caller finalises a decision.
• Rule set:
  • Librarian with conflict → consult Attendant for relevance; also Archivist if confidence spread ≥ 25
  • Attendant with a topic → consult Librarian for source-reliability; Archivist if ≥3 low-confidence injections
  • Archivist on reasoning proposals → consult Librarian for conflict history
  • Resolutionist on pending escalation → consult Archivist and Librarian
• Self-consultation dropped, budget capped (default 2), minimum-confidence floor enforced (50).
• AttendantInstance.attend() surfaces councilConsultationPlan on all three return paths (post-response shell, memory_not_needed, main success). Low-confidence injection count drives the Archivist consult rule on the main path.

Tests

• 29 new tests under tests/staff/run_b7_reasoning_council_tests.ts:
  • 14 archivist reasoning pure-unit tests
  • 15 council mode pure-unit tests
• All green on first run.
• Backward-compat regression gate green across 10 prior suites (b4-skip-drift, b5-objective-autocheckpoint, b6-refinement-tool-plan, a4-plan-proposal, tool-call-retrieval, mid-turn-attend, m2-tool-result-autowrite, file-change-recall, user-operating-rules, compliance-task-relevance, artifact-capture).

Package versions

• iranti → 0.3.30
• @iranti/sdk → 0.3.30
• iranti (Python) → 0.3.30