Skip to content

13-task-segmentation.md

Latest commit

 

History

History
142 lines (101 loc) · 3.96 KB

13-task-segmentation.md

File metadata and controls

142 lines (101 loc) · 3.96 KB

13 — Task Segmentation

Segment lifecycle, resolution scoring, and segment analytics.

Purpose

A session can contain multiple tasks. MOA tracks each task as a segment so learning is based on discrete outcomes instead of whole-session guesses.

Segments answer:

  • What task was being attempted?
  • Which intent was assigned, if any?
  • Which tools and skills were used?
  • How many turns and tokens did it cost?
  • Did it resolve?
  • What learning should be recorded from the outcome?

Data Model

TaskSegment lives in crates/moa-core/src/types/segments.rs; persistent rows live in Postgres task_segments.

Important fields:

  • id
  • session_id
  • tenant_id
  • segment_index
  • intent_label
  • intent_confidence
  • task_summary
  • started_at
  • ended_at
  • resolution
  • resolution_signal
  • resolution_confidence
  • tools_used
  • skills_activated
  • turn_count
  • token_cost
  • previous_segment_id

ActiveSegment is the lighter projection stored in session VO state.

Segment Detection

Query rewriting produces QueryRewriteResult:

  • rewritten_query
  • high-level intent
  • is_new_task
  • task_summary
  • suggested tools and clarification metadata

When a turn is prepared, SegmentTracker uses the query rewrite metadata and session events to decide whether to:

  • keep the current active segment
  • create the first segment
  • close the previous segment and start a new one

The event log records SegmentStarted and SegmentCompleted events.

Intent Classification

New segments can be classified against active tenant intents:

  1. Build text from task summary and first user message.
  2. Embed the text with the configured embedding provider.
  3. Query active tenant intent centroids in Postgres.
  4. Accept the nearest match when cosine distance is below the configured threshold.
  5. Store the label and confidence on the segment.
  6. Append intent_classified to learning_log.

If a tenant has no active intents, classification returns no match. New tenants therefore start blank.

Segment Counters

During a turn, the orchestrator records:

  • tool names used
  • skill names activated
  • completed turn count
  • token cost

The active VO state and task_segments row stay in sync through session store calls.

Resolution Scoring

Resolution scoring combines five signal classes:

Signal Meaning
Tool outcome Whether tools completed, failed, or produced useful output
Verification Whether tests/checks/verification commands succeeded
Continuation Whether the next user message indicates success, rework, abandonment, or a new task
Self-assessment Whether the agent response claims completion or uncertainty
Structural Whether turns, cost, and duration are anomalous for the tenant/intent baseline

The scorer outputs:

  • resolved
  • partial
  • unknown
  • failed
  • abandoned

Scoring phases:

  • immediate: when a segment appears idle or completed
  • deferred: after a later user message gives continuation evidence
  • final: when cancellation or timeout closes the segment

Each score updates the segment row and appends resolution_scored to learning_log.

Materialized Views

Segment rows drive learning views:

View Use
skill_resolution_rates Ranks skills by tenant/intent resolution outcomes
intent_transitions Tracks common task-to-task flows
segment_baselines Provides structural baselines for resolution scoring

Refresh is handled through the session store's materialized-view refresh path.

Compaction Interaction

Segment events are durable boundaries. History compaction can summarize older events, but segment start/completion records remain part of replay and analytics.

Learning Flow

User messages
  -> query rewrite
  -> segment start/continue/complete
  -> tool and skill counters
  -> resolution score
  -> learning_log
  -> skill ranking and intent learning

Task segmentation is the measurement layer that makes the rest of MOA's learning pipeline reliable.