nfemmanuel/iranti

Bundle 6: A3 iterative exploration + A5 tool-using Attendant

This release lands two Attendant-agency features from the A-series backlog:

A3 — Iterative exploration (bounded refinement)

• New planRefinementPass() pure helper decides whether to run a bounded retry observe() when the initial retrieval call returns zero facts.
• Widens entity hints to 2-segment prefixes (e.g. project/iranti/custom_scope → project/iranti) and unions them with the original hints so host-supplied specific hints (including pendingToolCall-derived entities) always survive the retry.
• Extracts stopword-filtered fallback terms from latestMessage for future multi-step work.
• Gated off on post-response phase. Max 1 extra observe() call per attend, by design.
• Five refinement outcomes surfaced on AttendResult.refinementPass: not_needed, attempted_added, attempted_empty, declined_no_hints, declined_post_response.

A5 — Tool-using Attendant (planned, not executed)

• New planAttendantToolCalls() pure helper produces up to 3 planned (not executed) follow-up tool calls from a small palette: search_related, observe_entity, query.
• Follows the A4 plan-proposal pattern — the host sees the plan and can act on it, but the Attendant does not itself call these tools yet.
• Four proposal bases: brief_entities, drift_tokens, objective_signals, empty.
• New AttendResult.attendantToolPlan field with proposed[], basis, note.

Tests

• 28 new tests covering both pure-unit (helpers) and integration (attend() wiring) layers:
  • 10 A3 pure unit + 8 A5 pure unit
  • 5 A3 integration + 5 A5 integration
• All 28 green on first run.
• Backward-compat regression gate green across all 10 prior suites after the union-preservation fix to tool-call-retrieval.

Package versions

• iranti → 0.3.29
• @iranti/sdk → 0.3.29
• iranti (Python) → 0.3.29