fix(session): discard a failed attempt's parts before retrying and stop retrying past an executed tool

[filipeforattini]

Problem

Effect.retry(SessionRetry.policy) in packages/redcode/src/session/processor.ts wraps the whole stream consumption. The parts a failed attempt had already persisted (text, reasoning, step-start, tool parts) stayed in the assistant message, so the retried stream appended duplicates next to them. Worse, the retry replayed the unchanged streamInput, so a tool the model had already executed in the failed attempt (a gateway ending with finish_reason: network_error after a completed edit, see llm/ai-sdk.ts) was requested and executed a second time. The existing retry test only checked in-memory reasoning state, not what was persisted.

Fix

• The processor mints part ids through nextPartID(), which records them in ctx.attemptParts for the attempt in flight (reasoning, tool, step-start, step-finish, patch, text).
• When the retry policy decides to retry (its set hook, before the wait starts), discardAttempt settles the pending tool calls and removes the recorded parts through Session.removePart, which the core projector turns into a DB delete. The message never shows the failed attempt's output next to the retry's.
• updateToolCall flags ctx.attemptExecuted once a tool part leaves pending. A failure after that skips the retry layer and goes straight to halt, so the turn ends with the error visible and no tool executes twice. The check happens before the schedule so no misleading retry status is published.
• Terminal (non-retryable) failures keep their partial parts as before.

Test

Two regression tests in packages/redcode/test/session/processor-effect.test.ts, both failing on main and passing here:

• discard the failed attempt's parts before retrying: a text chunk then a midstream server_error, followed by a clean reply. Before: ["", "two"] text parts and two step-starts. After: exactly ["two"], one step-start, two provider calls, no error.
• do not retry after a tool call already ran: a completed tool call then finish_reason: network_error. Before: continue, a second provider request. After: stop, one provider call, tool executed once and still completed, message.error set.

mise x -- bun test --timeout 30000 test/session: 514 pass, 0 fail. mise x -- bun run typecheck: clean.

https://claude.ai/code/session_01KAcoL6wpgVEs2ebUrUGpHS

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}