v0.3.0

v0.3.0 — Durable everything

Every conversation turn now runs inside a durable Vercel Workflow run, and the
agent loop itself is AI SDK 7's WorkflowAgent — every LLM call and every tool
call is an individually durable, auto-retried step. GitHub files remain the only
store; Workflow is the execution container, never the database.

Highlights

🔁 Durable chat turns (Vercel Workflow)

• POST /api/chat starts a durable Workflow run and streams run.readable
  back; the turn keeps executing even if the connection drops
• Same-browser resume: refresh mid-response and the turn replays, then keeps
  streaming (run id persisted locally). Cross-device resume is deliberately out
  of scope — it would require a database this architecture refuses to have

🧠 One agent brain (AI SDK 7 WorkflowAgent)

• The streamText-in-one-step loop is gone: chat turns and background loops now
  share a single agent layer (src/app/api/agent/)
• Each LLM call and each tool call is its own durable workflow step, visible
  and retryable in the Workflow dashboard
• Tool context crosses step boundaries via serializable toolsContext — no
  closures

🌐 Live web search

• New webSearch tool backed by DeepSeek V4 Flash's native server-side search,
  reached through DeepSeek's Anthropic-compatible endpoint (the OpenAI-style
  protocol cannot express provider-executed tools)
• Search runs as one durable step; the card shows the cited answer with source
  links. The neutral tool contract keeps the backend swappable — your main
  model choice is untouched

🛠 Robustness

• Deterministic tool errors ("file not found", "access denied") return to the
  model as results instead of burning the runtime's automatic retries
• Fixed client-side validation crashes on raw model stream parts that share
  type names with UI chunks (reasoning-delta / text-delta / tool-input-*)

Infrastructure preview: background loops

The durable loop engine ships in this release as infrastructure: a goal goes
in, an autonomous agent works it step by step, and every increment is
checkpointed to memory/loops/ the moment it happens — surviving deploys and
restarts. The agent-facing startLoop door is intentionally not enabled yet;
in-conversation progress and the no-repo demo story land together in a coming
release.

Known trade-offs

• Server-measured reasoning duration has no WorkflowAgent equivalent yet; the
  thinking timer is client-side and live-session only

Full Changelog: v0.2.1...v0.3.0