[codex] Implement phase 2 write interpretation

[MaxLinCode]

What changed

This replaces the write-path slot-extractor seam with a unified interpretWriteTurn stage and rewires the router to use classifyTurn -> interpretWriteTurn -> writeCommit -> decideTurnPolicy.

Why

Phase 1 landed PendingWriteOperation, but the active router path still pre-gated schedule extraction and committed only flat extracted schedule fields. This change moves operation-aware interpretation into one LLM call, keeps commit responsible for required-field derivation from operationKind, and updates the code map/docs to match the landed architecture.

Impact

Write-capable turns now use one interpretation boundary instead of a separate router-owned slot-extraction pass. The commit layer now evaluates grouped field paths, preserves the existing confidence thresholds, and still keeps classifier-owned entity resolution in place for now.

Validation

• pnpm --filter @atlas/core test
• pnpm --filter @atlas/web test
• pnpm --filter @atlas/integrations test
• pnpm test
• pnpm typecheck

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
atlas-web	Error		Mar 28, 2026 9:46pm

[MaxLinCode]

Suggestion for follow-up architecture work:

PR 72 is a solid step forward, but the split between turnType and operationKind still feels more redundant than ideal on the write path. Right now we need to reason about turnType for gating (should this turn be interpreted for writes?) and operationKind for workflow contract (what fields are required, and do we merge or reset prior state?). That separation is defensible, especially for clarification_answer, but it also creates a mismatch surface where discourse classification and workflow interpretation can diverge in surprising ways.

For a future pass, I’d suggest tightening that boundary so only one layer owns write-workflow semantics. A few possible directions: keep turnType purely discourse-oriented and make operationKind exist only on write-capable turns; or reduce the classifier to a simpler write vs non-write decision and let the write interpreter own workflow kind end-to-end. Either way, the goal would be to reduce overlap, make invariants easier to state, and shrink the number of cases where a turn can be “write-like” in one representation but not the other.