root/fusion
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(FN-882): add loop detection recovery with compact-and-resume

Browse Source 
- Add ContextLimitDetector to detect agent loops via repeated tool call patterns
- Implement compact-and-resume strategy: summarize conversation and restart agent from current step
- Add loop recovery to StuckTaskDetector with configurable attempt tracking and retry limits
- Extend executor with automatic loop recovery on context limit detection
- Add loop recovery support to pi executor with same compact-and-resume pattern
- Add comprehensive tests for context-limit-detector, stuck-task-detector loop detection, executor, and pi recovery
- Add changeset for patch bump to @gsxdsm/fusion
- Update README with loop detection and recovery documentation
This commit is contained in:
gsxdsm
2026-04-04 19:47:48 -07:00
parent 1348e78bc1
commit e28cbb1c1f
12 changed files with 846 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
README.md
View File

@@ -858,6 +858,31 @@ Automatically resolves:
```
Terminates and retries tasks with no agent activity for the specified duration (10 minutes in this example).
The detector distinguishes between two stuck reasons:
- **Inactivity** — no heartbeats at all (session appears dead). Immediately kills and re-queues.
- **Loop** — agent is active but not making step progress despite lots of activity (e.g., context growth causing repetitive behavior). Triggers compact-and-resume before falling back to kill/requeue.
**Loop Recovery (Compact-and-Resume):**
When a loop is detected, the system attempts one in-process recovery before killing the agent:
1. The stuck task detector calls the `onLoopDetected` callback before terminating the session
2. The executor compacts the conversation context using `session.compact()` (or falls back to deterministic compaction instructions)
3. After compaction, the agent receives a resume prompt asking it to review current state and take a different approach
4. If the agent loops again after compaction, the second detection falls through to the normal kill/requeue path
This one-attempt ceiling prevents recovery churn while giving the agent a chance to break out of context-growth loops. The recovery state is in-memory per `execute()` lifecycle — it does not persist across retries.
**Context-Limit Recovery:**
When an LLM provider returns a context-window overflow error (e.g., "prompt is too long", "exceeds the context window"), the executor catches it before the normal failure path:
1. If no prior compact-and-resume attempt exists, the executor compacts the session and resumes with a fresh prompt
2. If compaction fails or the ceiling has been reached, the error falls through to the normal failure/requeue path
This handles the common case where long-running agent conversations grow beyond the model's context window.
**Pause Behavior for In-Progress Tasks:**
Pausing a task that is currently executing will immediately terminate the agent session and move the task back to `todo`. When the task is later unpaused, the scheduler immediately picks it up (event-driven, no poll-cycle delay) and resumes execution from where it left off (step progress is preserved). The task is never left stranded in `in-progress` after a pause - both the error-throwing and graceful session exit paths move it to `todo`. Paused tasks are never marked as `failed`.