---
purpose: Open issues and known limits of the lfcs-orchestrator + lfcs-pricing system. Snapshot at a point in time — decays fast. Update as items resolve or new ones surface. Referenced from orchestrator CLAUDE.md §10 (pointer-only).
authored_by: claude-code-orchestrator
created: 2026-05-09
last_reviewed: 2026-05-09
---

# Known limits — orchestrator + pricer

This is a status snapshot, not a contract. Items here move (resolved → deleted; new gaps → appended). Do not treat absence of an item here as a guarantee that the system handles it.

## Open

- **Voice path on intake gate not yet built.** Q-1..Q-6 are typed for now. Voice deferred until `lfcs-pricing` Task 0 voice-spike completes.
- **Auto-relay v2 (file bus + daemon) is the ~85% solution.** True full autonomy (auto-prompt-injection across Claude Code sessions) requires CLI hooks Claude Code doesn't expose. v3 gap.
- **Gap-pattern library bootstrap thin.** Library starts with G1–G5 derived from 2630. Expect rapid expansion in first 3 jobs of orchestrator-run; review `standards/gaps-pattern-library.md` after each session and add patterns Rocky-flagged.
- **Standard LFCS Exclusions Library is a v1 reconstruction** from 2630 + 2629 patterns. Expect tightening as more jobs run through. Review monthly or after every 3 bids.
- **No methodology auto-write.** All proposals stay in `_Internal/lessons-learned/proposals-pending.md`. If proposals queue gets long (>20 unresolved), orchestrator will surface and ask Rocky to triage.
- **Autonomous-answering system live but unbattle-tested (2026-05-08 build).** Priority chain (orchestrator §5.0) + precedent-finder skill + auto-decision rules (`standards/standing-context.md` §13) + assumptions log (`standards/assumptions-log-template.md`) shipped together. First real-bid run pending. Self-eval against 2630 gaps: 4/4 caught (`_Internal/audit-trail/autonomous-answering-self-eval-2026-05-08.md`). Calibration of similarity weights + confidence thresholds expected over the next 3 bids.
- **Daemon autonomous mode (2026-05-08 evening build).** Daemon spawns `claude -p` headless on file change — both agents run end-to-end without Rocky terminal input. State machine: phase-1-intake → phase-2 → phase-3 → phase-3-sealed → phase-3.5-envelope → phase-4-sanity → phase-5-blind (skip if not blind) → phase-6-submit → phase-7-close. Safety caps: MAX_ITERATIONS_PER_BID=30, MAX_WALL_HOURS_PER_BID=4, PER_CALL_TIMEOUT_SECONDS=600. Rocky-input gate writes `rocky-question-{bid_id}.md`; Rocky drops `rocky-answer-{bid_id}.md` to resume. Dry-run simulator confirmed end-to-end (7 spawns, 1 touchpoint, clean phase-7 termination). Live trial: budget API quota carefully — each spawn is a real `claude -p` invocation. Run `python lfo_handoff_daemon.py --interactive` for legacy notify-only mode.
- **2627 BCS-leak lesson candidate not yet written.** Queued at session 80b1f893 close. Pending `lfo-lessons-write` invocation at Phase 7 of bid 2627 (currently mid-Phase-3-sealed). Distillation: BCS-quarantine metadata MUST NOT appear in worker-readable bus; whitelist > blacklist for path control; file rotation > block-edit when leak redaction needed; author identity (no $-figure) breaks BCS independence via anchoring bias.

## Resolved (recent)

(append items here as they close, with date + brief note. Trim entries older than 30 days during review.)