fix(agent-core-v2): invalidate streamed attempt state when the llm requester retries below the turn

[kimi-agent-bot]

Related Issue

No tracking issue. Reported via the Mangrove platform (task 201346, container 9123219, kimi 0.42.0, KIMI_CODE_INFINITE_RETRY=1): a streaming APIConnectionError: terminated mid-response led to process exit 1.

Problem

With KIMI_CODE_INFINITE_RETRY=1, AgentLLMRequesterService retries failed requests inside its own for (;;) loop (and inside the projection-policy resend path) without telling the turn machine. llm.sent is emitted only once, before the loop. When a stream dies after emitting half a tool call (function part at index=0), the retry reuses stream index 0 while the turn's HistoryAccumulator and ToolCallIdNormalizer still hold the interrupted attempt's state:

1. the retried stream's new tool call at index=0 hits the stale assignedByIndex[0] mapping and is silently rewritten to the previous attempt's id;
2. the finished assistant message ends up with duplicate tool call ids [A, A, B];
3. createMachineTools.beginBatch did not dedupe, flushIfReady drained the same pending entry twice (pending.get(A) → undefined on the second read), and runBatch dereferenced entry.input.toolCall.id outside its try block;
4. the void runBatch(...) call had no rejection handler → unhandled rejection → process exit 1.

A secondary leak: a duplicate pending.set(id, ...) overwrote the first entry, leaving its promise pending forever.

What changed

Root fix — the turn machine now learns about every below-the-turn retry:

• AgentLLMRequesterService accepts a new onAttemptRetry override and invokes it before each projection-policy resend and before each infinite-retry backoff sleep.
• createMachineRequester maps that callback to a new llm.request.retrying requester event (LlmRequestEvent / LlmEvent).
• The turn machine handles llm.request.retrying in thinking with the same action as llm.sent (extracted as discardAttemptStream): roll back and recreate the accumulator and its response normalizer, so each attempt accumulates from zero. Parts of the interrupted attempt already forwarded to the parent machine/UI are not reclaimed — noted in docs/{en,zh}/llm.md.
• model-requester-impl handles the new event in its event switch (same reset as a re-sent request).

Hardening in agent/loop/machine/tools.ts so a duplicate id can never kill the process again:

• beginBatch dedupes expected ids;
• flushIfReady skips missing pending entries instead of a non-null assertion;
• execute settles a superseded pending entry with an error result instead of leaking its promise;
• all batch launches go through startBatch, which catches rejections into onBatchError, and the entry-setup lines moved inside runBatch's try.

Tests:

• test/agent/loop/machineTools.test.ts (new): beginBatch with duplicate ids runs one batch per unique id and settles the superseded call; executor failures reach onBatchError and settle every pending call. Red before the fix (reproduces the exact unhandled TypeError: Cannot read properties of undefined (reading 'input')).
• test/agent/llmRequester/llmRequesterService.test.ts: unit coverage for onAttemptRetry on both retry paths and the final-error path; an integration test wiring the real service → machine requester → turn machine where attempt 1 streams function(index=0, id=call_a) then fails with APIConnectionError and attempt 2 streams id=call_b, asserting the finished message has tool call ids ['call_b'] (duplicate ['call_a', 'call_a'] before the fix). Fake timers throughout.

agent-core-v2: 368 test files / 6569 tests passed, tsc --noEmit clean, oxlint --type-aware 0 errors (warnings at baseline), check-no-comments and check-import-boundaries OK.

Checklist

• I have read the CONTRIBUTING document.
• I have linked a related issue (external PRs: the issue must have a maintainer's /approve).
• I have added tests that prove my feature works.
• Ran gen-changesets skill, or this PR needs no changeset.
• Ran gen-docs skill, or this PR needs no doc update.

[changeset-bot]

🦋 Changeset detected

Latest commit: 72147a6

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@moonshot-ai/kimi-code	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[kimi-agent-bot]

@codex review

[pkg-pr-new]

pnpm dlx https://pkg.pr.new/@moonshot-ai/kimi-code@72147a6

npx https://pkg.pr.new/@moonshot-ai/kimi-code@72147a6

commit: 72147a6

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Can't wait for the next one!

Reviewed commit: 72147a69aa

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".