🤖 Fix stream error amnesia: commit partial progress to history #331
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
When stream errors occur, accumulated parts (text, tool calls, reasoning) are now committed to chat.jsonl before writing error state to partial.json. This ensures resumptions/retries don't lose progress from failed streams. Changes: - Modified PartialService.commitToHistory to strip error metadata and commit accumulated parts instead of deleting errored partials - Error metadata remains transient (UI-only), while progress is persisted - Added tests verifying errored partials with parts are committed correctly Net LoC: -5 (removed duplication, single source of truth for commit logic)
ammar-agent
force-pushed
the
stream-error
branch
from
a7a3ca8 to
57b469b
Compare
Prevents pollution of chat history with empty assistant messages when streams error before any content is generated (e.g., immediate network failures, auth errors, rate limits). Changes: - Added empty parts guard to commitToHistory shouldCommit logic - Added test verifying empty errored partials are skipped - Cleanup still happens (partial.json deleted) This completes the error recovery fix - preserves progress when present, skips empty placeholders when not.
ammar-agent
changed the title
ammario
approved these changes
Oct 19, 2025
github-merge-queue bot
pushed a commit
that referenced
this pull request
Oct 19, 2025
## Summary
Adds integration test to verify that stream error recovery preserves
context (no amnesia bug).
## Changes
- **Debug IPC for testing**: Added `DEBUG_TRIGGER_STREAM_ERROR` IPC
channel
- **StreamManager debug method**: `debugTriggerStreamError()` triggers
artificial stream errors that follow the same code path as real errors
- **Integration test**: Single error + resume scenario verifies context
preservation via **structured markers**
## Test Design
**Structured-marker approach** for precise validation:
**Test Flow:**
1. Generate unique nonce for test run (random 10-char identifier)
2. Model counts 1-100 using structured format: `${nonce}-<n>: <word>`
(e.g. `ai7qcnc20g-1: one`)
3. Collect stream deltas until ≥10 complete markers detected
4. Trigger artificial network error mid-stream
5. Resume stream and wait for completion
6. Verify final message has **both** properties:
- **(a) Prefix preservation**: Starts with exact pre-error streamed text
- **(b) Exact continuation**: Contains next sequential marker
(${nonce}-11) shortly after prefix
**Validation:**
- Pre-error content captured from stream-delta events (user-visible data
path)
- Stable prefix truncated to last complete marker line (no partial
markers)
- Assertions directly prove both amnesia-prevention properties
- No coupling to internal storage formats or metadata
**Why this approach:**
- **Precise**: Detects exact continuation (not just "some work done")
- **Unambiguous**: Random nonce makes false positives virtually
impossible
- **Robust**: Structured format less likely to confuse model than
natural language
- **Fast**: Haiku 4.5 completes in ~18-21 seconds
## Bug Fix
Also fixed event collection bug in `collectStreamUntil`: properly track
consumed deltas to avoid returning the same event multiple times.
Previous logic returned first matching event on every poll, causing
duplicate processing.
## Related
Follow-up to #331 which fixed the amnesia bug by preserving accumulated
parts on error.
## Test Results
✅ Test passes reliably in ~18-21 seconds
✅ Validates **exact** prefix preservation and continuation
✅ No flaky failures from timing issues
✅ Integration tests pass: 1 passed, 1 total
_Generated with `cmux`_
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Problem: Stream Error Amnesia
After stream errors, the Agent would "forget" minutes of accumulated work and continue from an old state. This caused frustrating loss of progress during retries.
Root Cause: Errored partials were deleted without committing accumulated parts to history. The original logic conflated two separate concerns:
Solution
Modified
PartialService.commitToHistoryto:Error state remains in
partial.jsonfor UI display, but doesn't prevent progress persistence.Changes
commitToHistoryto striperror/errorTypebefore commitImpact
✅ Amnesia fixed - Agent retains full context across error → retry cycles
✅ No empty messages - Immediate errors don't pollute history
✅ Net -4 LoC - Cleaner logic, single source of truth for commits
Verification
Before fix:
After fix:
Generated with
cmux