nfemmanuel/iranti

Bundle 8 — M6 sub-turn reasoning loop

Final bundle in the M-series backlog. The Attendant can now re-score its own partial assistant response on a mid-turn attend and fire one bounded extra observe() retry with broadened hints harvested from the partial text. This is A3's refinement pattern re-applied on response progress rather than initial-pass emptiness — the "most agentic" sub-turn loop from the memo.

New

• planSubTurnLoop (pure helper in src/staff/subTurnLoop.ts) — five gates (phase, partial presence, length, budget, novelty), extracts novel tokens + type/id hints from the partial, proposes them for a widening retry, bounded to one extra observe() call. Pure, deterministic, fully unit-testable.
• AttendInput.partialResponse — new field on mid-turn attend carrying the in-progress assistant response.
• AttendResult.subTurnLoopPlan — echoed on all three return paths (post-response closeout, memory_not_needed, main success) so hosts can safely destructure on every attend response.

How it works

1. Mid-turn attend arrives with partialResponse set.
2. Baseline retrieval runs as usual.
3. planSubTurnLoop inspects the partial text, diffs its tokens against the baseline (latest message + original hints), and extracts any novel type/id entity patterns.
4. If any novel signal is found, the Attendant fires ONE extra observe() call with the UNION of allObserveEntityHints and the proposed hints — same preservation fix as B6's refinement retry so specific host-supplied hints are not dropped.
5. Net-new facts are merged into the response; outcome flips from attempted_empty → attempted_added when the retry surfaces new facts.

Tests

• 26 new unit tests in tests/staff/run_b8_subturn_loop_tests.ts covering all gates, token/hint extraction, budget caps, dedup, stopword filtering, entity-shaped baseline decomposition, and plan shape invariants.
• Registered as npm run test:b8-subturn-loop.
• Regression gate green across b5-b8, mid-turn-attend, a4, b4, tool-call-retrieval, m2, user-operating-rules, compliance, file-change-recall, artifact-capture, b2, mock-failure-modes, mock.

Files

• src/staff/subTurnLoop.ts (new)
• tests/staff/run_b8_subturn_loop_tests.ts (new)
• src/attendant/AttendantInstance.ts (3 return paths + AttendInput/AttendResult fields)
• package.json (+ test:b8-subturn-loop script, version bump)
• clients/typescript/package.json, clients/python/pyproject.toml, clients/python/iranti.py, src/api/server.ts (version bump)

Backlog status

This closes the M-series backlog (B1-B8). Full sequence shipped: B1 attendant agency, B2 mock classifier, B3 host contracts, B4 skip/drift, B5 objective + auto-checkpoint, B6 refinement + tool plan, B7 reasoning budget + council mode, B8 sub-turn loop.