LLM roadmap: mission metadata + Cmd+K semantic discovery

[Th0rgal]

Summary

Use the existing mission LLM capability to improve discovery, navigation, and context quality across clients, with Cmd+K as the primary surface.

This issue proposes a phased roadmap with backend-first architecture so behavior is consistent for web and iOS.

Why

Current LLM usage is mostly title generation. We can get much more value by improving:

• mission metadata quality (title + short description)
• duplicate avoidance
• search relevance in Cmd+K
• fast navigation in long mission histories

Recommended roadmap

Phase 1 — Metadata foundation (backend-first)

1. Backend-managed title + short_description

• Generate both when a mission first gets a successful assistant response.
• Persist in mission store and emit events for client updates.

2. Metadata lifecycle updates

• Re-evaluate metadata every 10 messages (or milestone events), only if drift is significant.
• Keep updates best-effort and non-blocking.

3. Anti-duplicate title generation

• Include recent mission titles as negative context in prompt.
• Apply server-side near-duplicate guard and append disambiguator if needed.

Phase 2 — Cmd+K relevance improvements

4. Hybrid search pipeline

• Keep lexical search for recall.
• Re-rank top-N results with LLM for semantic relevance.
• Cache rerank responses by query hash + mission update timestamp.

5. Query expansion

• Expand short/ambiguous queries into likely synonyms/domains before retrieval.

6. Richer result cards

• Show title + short_description + status + backend in Cmd+K.

Phase 3 — Navigation and action intelligence

7. Jump-to-moment retrieval

• Query examples: “where did we fix session id?”, “show timeout root cause”.
• Return mission + message/event anchor + short rationale.

8. Contextual quick actions in Cmd+K

• Suggest actions like Resume, Open failing tool call, Follow-up mission, etc.

Data model additions

• missions.short_description (nullable text)
• missions.metadata_updated_at
• optional provenance fields (model/version/source) for generated metadata

API considerations

• Include short_description in mission list/detail payloads.
• Add optional endpoint for semantic search/rerank in Cmd+K flow.
• Ensure generated metadata updates are evented for live UI sync.

Acceptance criteria

• New missions receive backend-generated title + short description.
• Metadata refresh after ~10 messages works without blocking mission execution.
• Duplicate/near-duplicate titles are measurably reduced.
• Cmd+K top results improve for natural-language queries.
• Web and iOS show identical mission metadata behavior.

Non-goals (initially)

• Full vector database rollout for all content.
• Generating long descriptions or full narrative reports by default.

Notes

This roadmap intentionally favors incremental value:

• Phase 1 improves metadata quality immediately.
• Phase 2 improves search UX without major infra changes.
• Phase 3 adds high-value navigation and action intelligence once metadata/retrieval are reliable.