🤖 Change interrupt sentinel from [INTERRUPTED] to [CONTINUE] #170
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
The [CONTINUE] sentinel is clearer to models, indicating they should continue their previous work rather than viewing it as an error state. This is a quick UX improvement while we work on more comprehensive handling of interrupted messages with thinking.
ammario
approved these changes
Oct 10, 2025
This was referenced Oct 10, 2025
ammario
pushed a commit
that referenced
this pull request
Oct 11, 2025
## Problem
Currently we always add the `[CONTINUE]` sentinel after partial
messages:
```
[user1] → [partial assistant] → [CONTINUE] → [user2]
```
But when a user message already follows the partial, the sentinel is
**redundant** - the user message itself provides the continuation
signal.
## Solution
**Conditionally skip the sentinel when unnecessary:**
```typescript
// Before: Always add sentinel
[user] → [partial] → [CONTINUE] → [user] → ...
// After: Only add when needed
[user] → [partial] → [user] → ... // Skip - user follows
[user] → [partial] → [CONTINUE] // Keep - last message
[user] → [partial] → [CONTINUE] → [assistant] // Keep - assistant follows
```
**Implementation:**
```typescript
if (msg.role === "assistant" && msg.metadata?.partial) {
const nextMsg = messages[i + 1];
// Only add sentinel if NO user message follows
if (!nextMsg || nextMsg.role !== "user") {
result.push(createSentinel());
}
}
```
## Benefits
- ✅ **Saves tokens** - One fewer message per resume-with-user-message
(common case!)
- ✅ **More natural** - Models handle "user interrupts then continues"
well
- ✅ **Cleaner history** - No redundant synthetic messages
- ✅ **Stateless logic** - Just looks at message array structure
## When Sentinel Is Added
- Partial is **last message** (empty resume)
- Partial followed by **assistant message** (rare but possible)
## When Sentinel Is Skipped
- Partial followed by **user message** (most common!)
- User interrupts stream and adds follow-up
- This is the typical resume pattern
## Testing
**Updated tests:**
- ✅ Multi-partial test now expects 5 messages (not 6)
- ✅ New test specifically for skip-sentinel behavior
- ✅ All 298 unit tests pass
**Test coverage:**
1. Sentinel added when partial is last message ✅
2. Sentinel skipped when user follows partial ✅
3. No synthetic messages when sentinel skipped ✅
## Risk Assessment
**Very low risk:**
- Only affects case where sentinel was already redundant
- Models naturally understand "user continues conversation"
- Existing behavior unchanged for empty resumes
- Stateless transform, no history dependencies
## Related
Builds on #170 which changed sentinel text from `[INTERRUPTED]` to
`[CONTINUE]`.
_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
When users interrupt a stream and then resume, we inject a synthetic
[INTERRUPTED]message to give the model context. However, this wording can confuse models into thinking there was an error or that they should stop.Solution
Change the sentinel text from
[INTERRUPTED]to[CONTINUE]to make it clearer that the model should continue its previous work.This is a quick UX improvement while we work on more comprehensive handling of interrupted messages with thinking (see #investigate-thinking-error branch).
Changes
[INTERRUPTED]→[CONTINUE]Testing
make testGenerated with
cmux