fix: recover Claude Code sessions after truncation

[contsun]

Summary

• detect Claude Code stop_reason=max_tokens and automatically continue ordinary truncations in the same session
• detect poisoned Claude Code sessions caused by truncated tool-call JSON (__unparsedToolInput, InputValidationError) and zero-token empty results
• reset poisoned agent sessions, save a recovery handoff, and inject it once into the next fresh session
• add prompt guidance to reduce oversized replies and long Bash heredocs
• add regression tests for stop reason propagation, poisoned-session detection, reset behavior, and recovery handoff injection

Why

When Claude Code hits max_tokens while emitting a large tool call, the tool-use JSON can be truncated. That leaves the session in a poisoned state: cache reads drop to zero, large prompt prefixes are billed as fresh input, and subsequent turns may return zero-token empty results. This patch separates ordinary truncation (same-session auto-continue) from corrupted sessions (fresh session + recovery handoff).

Test plan

• git diff --check
• go test ./agent/claudecode -run 'TestHandleResultPropagatesMaxTokensStopReason|TestHandleAssistantUnparsedToolInputMarksNextResultUnhealthy|TestHandleUserInputValidationErrorMarksNextResultUnhealthy|TestHandleResultZeroTokenEmptyMarksUnhealthy' -v
• go test ./core -run 'TestRecoveryHandoffInjectedOnceAfterUnhealthyReset|TestProcessInteractiveEvents_AutoContinuesSameSessionOnMaxTokens|TestProcessInteractiveEvents_UnhealthyResultResetsSessionInsteadOfEmptyResponse|TestUnsolicitedReader_UnhealthyResultResetsSession|TestI18n_AllKeysHaveEnglish' -v

Note: a full go test ./core run hit an existing TempDir cleanup flaky in TestCUJ_A3_ImageReachesAgent; rerunning that test alone passed.

[contsun]

Update: added detection for Claude Code's no-visible-output meta prompt (). This path was not a normal terminal result event, so the previous recovery logic would not trigger. Added and force-pushed the branch.

[contsun]

Update: added detection for Claude Code's no-visible-output meta prompt:

[Your previous response had no visible output. Please continue and produce a user-visible response.]

This path is emitted as an isMeta=true user event, not as a normal terminal result event, so the previous recovery logic did not trigger. The branch now marks that event as an unhealthy session, which causes cc-connect to reset the agent session, save a recovery handoff, and notify the user.

Added regression test:

TestHandleUserNoVisibleOutputMetaMarksSessionUnhealthy

Force-pushed the branch with commit c5f1d336.