Add producer mode to mc-prompter: rundown, replanning rail, and evidence-gated LLM coverage (Phase C)

[bmadcode]

What this is

Phase C plus the docs, PR 3 of 3 (PR 1: #4 classic teleprompter, PR 2: #5 voice-follow). This adds producer mode: a rundown file, a deterministic show-clock state machine with live replanning, rate-limited visual cues, and an opt-in local LLM (Ollama) that judges talking-point coverage under a hard evidence gate. This is the free-talk show feature: "I have 30 minutes, these 5 points in this order, an intro, and a wrap; keep me on track."

What you get

• The rundown artifact: a markdown file with frontmatter (duration, cue density, wrap reserve) and segments that are either scripted prose (prompted normally) or bullet talking points (tracked for coverage). Spec in references/rundown-spec.md, starter template shipped through mc-setup's assets, exact time-suffix parsing that rejects rather than guesses.
• The producer state machine (pure stdlib, fake-clock tested): GO LIVE / hold / resume / end show clock; sticky monotonic coverage where the human is always the authority; and real replanning: remaining time is redistributed across pending segments every tick with the wrap reserve protected, green/yellow/red always computed against the replan, and a "DROP: segment, or Ns each" suggestion when the math stops working. A producer that only turns red is a nag; this one re-plans.
• The cue engine: one active cue, per-density budgets (hands-off to chatty), release at VAD pauses, attention tier for time-critical calls, 15 s expiry, broadcast vocabulary (WRAP, STRETCH, 30 seconds, OVER). The binding contract lives in references/cueing.md.
• The Ollama tick (opt-in, [llm] lane, default qwen3:4b): structured outputs, no thinking tokens, prefix-cache-friendly prompt split, adaptive cadence that backs off under CPU pressure, and a deterministic evidence gate: every coverage claim must carry a verbatim transcript quote that (a) actually appears in the transcript and (b) relates to the point, so a hallucinated judgment cannot silently swallow a reminder. Without Ollama the deterministic rail and all time cues still work; the skill never pretends otherwise.
• UI: the ambient rail (show clock, current segment and its replanned time, NEXT point, timing colors) on the prompter and as a transparent OBS overlay; bullets segments get a large-type talking-points view with covered points struck through; the phone remote gains a producer tab (GO LIVE, hold, end, and per-point done/skip/make-current taps); scripted-to-bullets segment handoff is manual or anchor-driven with stale-anchor protection.
• Integration and docs: [prompter] and [llm] studio config tables with a surgical backfill for existing 1.0 studios (the 0.x classifier is untouched), an optional mc-setup interview step, an ollama row in the dependency check, Manny routing, PIPELINE.md and mc-script handoff notes, the user-guide walkthrough, README row, and TODO fast-follows.

Verification

• 460+ unit tests green under the CI quality gate command across the module; genericity lint clean; the shipped rundown template parses under the shipped parser
• Live end-to-end on this machine with the full real stack (nemotron ASR + qwen3:4b via Ollama): GO LIVE, voice-follow through a scripted intro, automatic segment handoff, two genuinely spoken talking points covered while a never-mentioned point was correctly left uncovered (the evidence gate blocked a persistent 0.9-confidence hallucinated claim from the real model), a DROP replan cue under overrun, and human skip/make-current overrides from the phone remote
• Adversarial 3-lens review found 20 issues (show-clock corruption at end-show, off-air speech permanently covering points, a stale-anchor handoff cascade, the template parse failure); all fixed with regression tests, then the live end-to-end was re-run green

🤖 Generated with Claude Code

https://claude.ai/code/session_01BbneC1PckyV8vkDzhGYkpN

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 3362b931-4d06-4c2c-a616-ede7f03431a1

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat-prompter-c

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}