atomr-agents-meetings-harness

Meetings harness

atomr-agents-meetings-harness turns a diarized SttConversation into a structured MeetingAnalysis: an attendee roster, a linear append-only ledger of notes and actions with owners, and a tiered summary stack that regenerates dynamically as the meeting unfolds.

atomr-agents-meetings-harness-web is the optional axum + React review UI on top of it (port 7100, side-by-side with the STT UI on 7000).

When to use it

• You have a finished diarized transcript (an SttConversation from stt-harness) and want attendees, notes, action items with owners, and a TL;DR.
• You have a live STT run and want incremental analysis that grows monotonically (notes/actions append; tail-segment summary revises; earlier segments freeze) so a UI can stream it.
• You want both records (transcript + analysis) joined under the same conversation_id in whichever Checkpointer backend is configured.
• Don't use it for raw transcription — that's stt-harness.
• Don't use it for general-purpose summarization of arbitrary text — it expects the diarized turn structure.

Build a MeetingsHarness

use std::sync::Arc;
use atomr_agents_meetings_harness::{
    BatchExtractionLoop, InMemoryMeetingsStore, IterationCapTermination,
    MeetingsHarness, MeetingsHarnessSpec, RuleBasedExtractor, RunMode,
};
use atomr_agents_stt_harness::{ConversationStore, InMemoryConversationStore};

let transcripts: Arc<dyn ConversationStore> =
    Arc::new(InMemoryConversationStore::new());
let analyses = Arc::new(InMemoryMeetingsStore::new());
let extractor = Arc::new(RuleBasedExtractor::new());

let spec = MeetingsHarnessSpec::new("meetings", "claude-opus-4-7")
    .with_mode(RunMode::Batch)
    .with_max_iterations(32);

let harness = MeetingsHarness::new(
    spec,
    transcripts.clone(),
    analyses.clone(),
    extractor,
    BatchExtractionLoop,
    IterationCapTermination::new(32),
);

let analysis = harness.run("call-7").await?;  // loads transcript by id

The model id is required on the spec — the caller picks it. The default RuleBasedExtractor is deterministic and LLM-free (useful for tests + the rest of the system end-to-end); production deployments swap in their own MeetingExtractor impl that drives an Agent.

Live mode

use atomr_agents_meetings_harness::{StreamingExtractionLoop, RunMode};

// Subscribe BEFORE the STT harness runs so no events are missed.
let stt_events = stt_harness.events();
let spec = MeetingsHarnessSpec::new("meetings", model_id)
    .with_mode(RunMode::Live { segment_turn_count: 8 });
let loop_strategy =
    StreamingExtractionLoop::new(stt_events, transcripts.clone(), "call-7".to_string());
let harness = MeetingsHarness::new(spec, transcripts, analyses, extractor, loop_strategy, term);

In live mode the loop:

• pulls new turns from the configured transcript store as the STT broadcast announces them,
• runs the extractor over the new window only,
• appends notes/actions to the tail (never reorders),
• revises the in-flight tail SegmentSummary until it exceeds segment_turn_count, then finalizes and opens a new one,
• recomposes summary_levels.running only when a segment finalizes,
• regenerates summary_levels.tldr only on finalize.

harness.cancel() (or POST /api/meetings/:id/stop) signals cooperative cancellation.

Data model

MeetingAnalysis holds attendees, notes, actions, and a SummaryLevels { segments, running, tldr }. See docs/meetings-harness.md for the full schema.

Foreign keys: Action.owner_attendee_id references Attendee.id. The tool layer validates the reference on every insert/update.

Tools

Tool	Purpose
list_turns	Page transcript turns.
get_turn	Full detail for one turn.
upsert_attendee	Idempotent attendee add/merge.
append_note	Append a note (linear, append-only).
append_action	Append an action (owner_id must resolve).
update_action	Patch a single action in place.
revise_tail_segment	Rewrite the in-flight tail segment summary.
finalize_segment	Freeze the tail; next revise opens a new one.
regenerate_running	Recompute summary_levels.running.
set_title	Set/replace the meeting title.
finalize	Set the TL;DR and mark state = Final.

There is no delete_note / delete_action by design — append-only is enforced at the tool surface.

Persistence

MeetingsStore trait + InMemoryMeetingsStore (default) + CheckpointerMeetingsStore (feature state). Analyses are filed under workflow_id = "meetings-harness", run_id = analysis.id — i.e. the same conversation_id as the source SttConversation, so both records join naturally.

CLI

# Analyze an existing transcript (requires --features meetings).
atomr-agents meetings analyze \
    --conversation-id call-7 \
    --model claude-opus-4-7 \
    --mode batch

# Serve the review UI (requires --features meetings-web).
atomr-agents meetings serve --bind 127.0.0.1:7100

See also

• atomr-agents-stt-harness — produces the transcripts this harness consumes.
• docs/meetings-harness.md — full architecture / data shapes / web API reference.