fix: add stream retry cap and context recovery hint for provider errors

[roomote-v0]

Related GitHub Issue

Closes: #12087

Description

This PR attempts to address Issue #12087 where provider errors during API requests cause the agent to lose context and hallucinate about previously completed tasks.

Two key improvements:

1. Stream retry cap (MAX_STREAM_RETRIES = 5): Both the first-chunk error retry path (in attemptApiRequest) and the mid-stream error retry path (in recursivelyMakeClineRequests) now have a maximum retry limit. Previously, only context window errors had a cap (MAX_CONTEXT_WINDOW_RETRIES), while stream/provider errors could retry indefinitely with exponential backoff. After exceeding the limit, the error is presented to the user for manual retry, which resets the counter.

2. Context recovery hint on retry: When a retry occurs (retryAttempt > 0), a [CONTEXT RECOVERY NOTE] is prepended to the user content. This helps weaker models re-orient to the current task instead of latching onto earlier completed task context. The hint explicitly instructs the model to focus on the most recent request and not repeat previously completed tasks. Additionally, the last user message in apiConversationHistory is updated with the hint and refreshed environment details.

Feedback and guidance are welcome.

Test Procedure

• Unit tests added in src/core/task/__tests__/stream-retry-limit.spec.ts (17 tests, all passing)
• Existing grace-retry-errors.spec.ts tests still pass (11 tests)
• All lint checks pass
• All type checks pass across the monorepo

To run the new tests:

cd src && npx vitest run core/task/__tests__/stream-retry-limit.spec.ts

Pre-Submission Checklist

• Issue Linked: This PR is linked to an approved GitHub Issue (see "Related GitHub Issue" above).
• Scope: My changes are focused on the linked issue (one major feature/fix per PR).
• Self-Review: I have performed a thorough self-review of my code.
• Testing: New and/or updated tests have been added to cover my changes (if applicable).
• Documentation Impact: I have considered if my changes require documentation updates (see "Documentation Updates" section below).
• Contribution Guidelines: I have read and agree to the Contributor Guidelines.

Documentation Updates

• No documentation updates are required.

Additional Notes

The context recovery hint is designed to be model-agnostic and uses clear, directive language to help even weaker models (like glm-5 via OpenAI Compatible endpoints) maintain focus on the current task after error recovery. The retry cap of 5 was chosen to match a reasonable threshold -- with exponential backoff starting at the configured base delay, 5 retries provides sufficient recovery window while preventing runaway loops.

Interactively review PR in Roo Code Cloud