Deep refactoring v2: type safety, error consistency, further decomposition

[bug-ops]

Overview

Second-pass architectural refactoring after epic #268. Focuses on remaining type safety, error consistency, and further decomposition.

Plan: .local/plan/refactoring-plan-v2.md

Issues

Phase 1 — Quick wins (parallel) ✅

• R2-8.2: replace tokio full with granular features in zeph-core #283 — Deduplicate EmbedFuture, fix anyhow leak in embed closures
• R2-2.1: deduplicate EmbedFuture, fix anyhow leak in embed closures #284 — Replace tokio "full" with granular features
• R2-9.2: replace is_summary bool with MessageKind enum #285 — Replace is_summary: bool with MessageKind enum

Phase 2 — Core improvements ✅

• R2-1.1: typed errors for Channel trait #286 — Channel trait typed errors
• R2-10.1: further decompose agent/mod.rs (streaming, persistence) #287 — Further agent/mod.rs decomposition (streaming, persistence)

Phase 3 — Cleanup

• R2-3.1: group Agent struct fields into inner structs #288 — Remove anyhow from zeph-channels and zeph-mcp
• R2-8.1: gate qdrant feature in zeph-skills compilation #289 — Group Agent struct fields into inner structs
• R2-5.1: remove anyhow from zeph-channels and zeph-mcp #290 — Gate qdrant feature in zeph-skills

Phase 4 — Typed agent errors

• Replace anyhow with typed AgentError in zeph-core agent module #293 — Replace anyhow with typed AgentError in zeph-core agent module

Dependency Graph

#283 (EmbedFuture) ──┐
                     ├──► #288 (remove anyhow)
#286 (Channel err) ──┘

#287 (mod.rs decomp) ──► #289 (field grouping)

#288, #289 ──► #293 (AgentError)

#284, #285, #290 — independent

Deferred (confirmed)

Item	Reason
File watcher dedup (R3.2)	~40 LOC each, stable
blake3 shared (R3.3)	Only one crate now
Test mock consolidation (R3.4)	Mocks legitimately differ
Embedder trait (R5.2)	Superseded by EmbedFn
Embedding newtype (R6.2)	No confusion risk

Metrics

Priority	Count	Est.
P1	6	~20h
P2	3	~5h
Total	9	~25h