Phase 1 target: deliver the medium architecture end-to-end

[pheidon]

Summary
Track the first end-to-end release of MailPlus Intelligence using the recommended medium architecture.

Phase 1 target

• full metadata/thread index
• incremental sync
• selective semantic extraction
• wiki/entity promotion
• raw email remains only in MailPlus

Prioritized implementation checklist

• Roadmap: MailPlus remains canonical raw archive; Pheidon becomes the intelligence layer #1 Establish and document the canonical-store boundary
• Decide metadata/thread storage backend and schema strategy #9 Choose the storage backend and schema direction
• Build metadata and thread index for all MailPlus mail #2 Build the metadata/thread index
• Define MailPlus locator and export contract #10 Define the MailPlus locator and export contract
• Implement incremental MailPlus exporter/sync job #3 Implement incremental exporter/sync behavior
• Add observability, checkpointing, and auditability for sync/extraction jobs #15 Add operational safety rails: checkpointing, logs, and auditability
• Add search and on-demand raw message fetch workflows #4 Deliver search and on-demand raw message fetch
• Define classification heuristics and evaluation set for important-vs-noise mail #12 Define measurable classification heuristics and noise suppression
• Classify mail into VIP, project, admin, financial, travel, legal, and ignore/noise lanes #5 Implement classification lanes
• Design selected-message text cache policy #11 Decide selected-message text cache boundaries
• Define semantic output schema for summaries, entities, obligations, and events #13 Define semantic output schema
• Generate thread summaries, entity updates, and obligation candidates from selected mail #6 Generate summaries, entities, obligations, and events
• Design promotion/review workflow for wiki, memory, and reminders #14 Define promotion/review workflow
• Promote distilled email intelligence into wiki and memory with review gates #7 Promote distilled intelligence into wiki/memory with review gates

Questions this phase should answer well

• What’s my history with this person?
• Did I already commit to this?
• What admin, travel, financial, or legal follow-ups are pending?

Out of scope for phase 1

• full semantic chunking on most mail
• broad entity-graph magic before the basics are reliable
• overbuilt proactivity before recall quality is proven

Definition of done

• End-to-end flow exists from MailPlus source → index → retrieval → selected extraction → approved promotion.
• Raw mail remains canonical in MailPlus.
• Memory surfaces stay distilled and high-signal.

[jmcte]

All checklist items are now closed:

• Roadmap: MailPlus remains canonical raw archive; Pheidon becomes the intelligence layer #1 Canonical-store boundary — documented in docs/architecture/
• Build metadata and thread index for all MailPlus mail #2 Metadata/thread index — index_writer.py, mapper.py (PR feat: attachment indexing, cache, queue, exporters, CLI, and ops runbooks #77)
• Implement incremental MailPlus exporter/sync job #3 Incremental sync — sync.py with checkpoint management (PR feat: phase 2 — sync, extraction, LLM extractor, scheduler, live adapter #78)
• Add search and on-demand raw message fetch workflows #4 Search and retrieval — mpi search / mpi thread CLI (PR feat: attachment indexing, cache, queue, exporters, CLI, and ops runbooks #77)
• Classify mail into VIP, project, admin, financial, travel, legal, and ignore/noise lanes #5 Classification lanes — classifier.py, suppression.py (PR feat: attachment indexing, cache, queue, exporters, CLI, and ops runbooks #77)
• Generate thread summaries, entity updates, and obligation candidates from selected mail #6 Summaries, entities, obligations, events — extractor.py + llm_extractor.py (PR feat: phase 2 — sync, extraction, LLM extractor, scheduler, live adapter #78)
• Promote distilled email intelligence into wiki and memory with review gates #7 Wiki/memory promotion with review gates — queue.py, exporters.py, mpi queue CLI (PR feat: attachment indexing, cache, queue, exporters, CLI, and ops runbooks #77)
• Decide metadata/thread storage backend and schema strategy #9 Storage backend — SQLite, documented in docs/storage-backend-decision.md
• Define MailPlus locator and export contract #10 MailPlus locator/export contract — docs/mailplus-locator-export-contract.md
• Design selected-message text cache policy #11 Selected-message text cache — cache.py, docs/selected-message-text-cache-policy.md
• Define classification heuristics and evaluation set for important-vs-noise mail #12 Classification heuristics and noise suppression — docs/classification-heuristics.md, docs/noise-suppression-policy.md
• Define semantic output schema for summaries, entities, obligations, and events #13 Semantic output schema — semantic_contract.py, docs/semantic-output-schema.md
• Design promotion/review workflow for wiki, memory, and reminders #14 Promotion/review workflow — docs/promotion-review-workflow.md
• Add observability, checkpointing, and auditability for sync/extraction jobs #15 Operational safety rails — docs/operational-safety.md, audit log, mpi doctor

Definition of done met: end-to-end flow exists from fixture corpus → index → search/retrieval → deterministic + LLM extraction → queue review → dry-run export. Raw mail remains canonical in MailPlus. Memory surfaces are distilled and human-gated.

---

Generated by Claude Code