Description
Coding Agent Plan Extraction preserves implementation plans that would otherwise remain buried inside long coding-agent transcripts. During normalization, it identifies supported plan signals in the session, converts substantive plan content into Artifact-shaped entities, and routes those Artifacts alongside the parent Conversation. The multi-target path keeps each destination’s identity and update behavior explicit.
The capability distinguishes durable work products from ordinary chat turns. A plan receives its own title, body, provenance, and source identity while the Conversation remains the full chronological record. Reprocessing the same or extended session is idempotent, so backfill and live tail can share extraction without multiplying Artifacts on every pass.
Selling Points
- Promotes reusable plans out of transient conversation scrollback.
- Preserves the original Conversation while giving plans their own Artifact identity.
- Runs during both historical import and live session updates.
- Uses deterministic source identities to prevent duplicate plan rows.
- Keeps destination-specific routing explicit rather than storing mixed entity shapes together.
User Story
As a developer revisiting an earlier project, I want the plan my coding agent produced to be discoverable as a standalone Artifact, not only as a fragment inside a thousand-message Conversation. I can open the plan directly while still following its provenance back to the session that produced it.
Extension Surface
The rating is fixed_core because plan detection and shaping are shipped normalization rules in packages/ingestion/src/coding-agent/normalize/plan-extract.ts, and multi-target routing is fixed in packages/ingestion/src/coding-agent/backfill/session-multitarget.ts. Composition can inject destination writers, but users cannot add arbitrary transcript-to-entity extractors through a public registration surface.
Capabilities & Limits
Capabilities include plan-signal detection, meaningful-content filtering, plan title and body assembly, Artifact routing, Conversation-plus-Artifact multi-target writes, and replay-safe identity. Extraction operates on normalized transcripts, so provider-specific message details are already reconciled. The limit is intentional: it extracts recognized in-transcript plans only; todos, memories, and arbitrary tool outputs are not automatically promoted, and production Memory Bank extraction remains a separate unshipped successor.
Implementation Map
packages/ingestion/src/coding-agent/normalize/plan-extract.tspackages/ingestion/src/coding-agent/backfill/plan-session.tspackages/ingestion/src/coding-agent/backfill/session-multitarget.tspackages/ingestion/src/coding-agent/route/route-entity.ts