[sanchitrk]

What

The session drain took one state snapshot per exchange and reused it for every tool-use and steer turn. A settings edit, a session model binding or a tool declared mid-chain only took effect at the next exchange, and nothing could step in between stages.

• Snapshot every turn. The drain is one flat loop that takes a fresh snapshot before each turn. Ordering is unchanged: tool results and steers continue first, a follow-up only once both are exhausted.
• Hooks service (runner/hooks.ts) with three seams, composed in fixed order in its live layer (no registry, no self-registering handlers):
  • prepareNextTurn(turn, next) runs between turns with the fresh snapshot and may replace it.
  • prepareRequest(request, { retry, failed? }) runs before every provider attempt, retries included.
  • finishTurn(turn) runs after the turn commits and returns "end" | "continue" | undefined. "end" wins over "continue" and leaves queued input for the next run. "continue" only sends a request if something was appended after the reply.
• Completed.pending: what the next run would promote (queued steers, else the oldest follow-up), read without promoting it.
• Resume ordering. An explicit resume or a recovered run answers a reply that stopped mid-chain before a queued follow-up goes in, and still promotes steers that land while it loads. Prompts and wakes are unchanged.

The seams are no-ops until a consumer is wired in.

Behaviour changes

• Settings edits, session bindings and plugin changes made during a tool chain apply at the next turn instead of the next exchange. A turn's tool batch still runs on its own snapshot.
• Every turn now re-reads settings and re-runs plugin setup. Not measured; expected to be small next to a provider request.
• A snapshot failure can now happen between turns. The previous turn is already committed, so the run stops cleanly and a later run resumes from it.
• Steers still promote all at once.

Verification

• New cases in runner.loop.test.ts: tool declared mid-chain reaches the next request, finishTurn end and continue, prepareRequest on a retry (replacement model reaches the request), pending input shown without promotion, a steer landing while a resume loads. Artifact: test/__artifacts__/runner.hooks.json.
• settings.loop.test.ts expectations updated from per-exchange to per-turn.
• vp check, vp run typecheck and the full harness suite pass.

Not covered

• finishTurn "continue" with an appended entry: nothing can append session entries yet.
• The per-turn cost of MCP reconcile, until the MCP work lands on this branch.
• A failing hook currently fails the turn; revisit when the first real consumer lands.

🤖 Generated with Claude Code