microsoft / microsoft/amplifier
Recipe engine records a failed step as completed with empty output, corrupting resume
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 3.1k
- Forks
- 261
- Avg merge
- 3h 28m
- Merged PRs (30d)
- 13
Description
Environment
- amplifier 2026.07.13-51194ef
- core 1.6.0
- Linux (WSL2)
Component
amplifier-bundle-recipes / recipe engine
Problem
During a long multi-step recipe run, a step's agent sub-session died mid-turn. In one instance, the provider raised NotFoundError; in another, the model emitted a thinking-only response and the session ended with empty output. The engine recorded the step in the session's state.json as COMPLETED with empty output rather than failed.
Impact
operation=resumeskips the step- Downstream steps receive an empty string for its declared output
- The run continues silently corrupted
- Workaround: hand-editing the recipe session's state.json (remove the phantom completed step, rewind current_step_index), then resuming
- This is required after every such failure
Expected Behavior
A step whose sub-session errors — or returns empty output when the step declares an output — should be recorded as failed/retryable and re-run on resume, never completed.
Related Observation
Recipe session state was stored under a literal unexpanded placeholder directory with a doubled segment: ~/.amplifier/projects/{project}/recipe-sessions/<project-slug>/recipe-sessions/<session-id>/ — worth checking while in this area.
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start in the amplifier-bundle-recipes recipe engine's step-state handling and reproduce a failed or empty-output sub-session in a multi-step recipe. Inspect the session state.json and resume behavior, including the reported session-path placeholder. Done means failed or empty-output steps are retryable and resume does not skip them or pass empty output downstream.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- backend
- Issue type
- Bug
- Difficulty
- 3/5
- Estimated time
- 1-2 days
- Activity status
- Quiet
- Clarity
- Mostly clear
- Newbie friendliness
- 55/100