[timmy-capability] Ollama inference contention when loop and Timmy run simultaneously #70
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: Rockachopa/Timmy-time-dashboard#70
Reference in New Issue
Block a user
Delete Branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
DISCOVERED VIA INTERVIEW: When the autonomous development loop is running (using Ollama for inference), Timmy's chat responses timeout after 60-90s. They're competing for the same GPU.
Options:
This matters because the loop is supposed to TALK to Timmy each cycle. If Timmy can't respond while the loop runs, the interview pattern breaks.
Tags: [loop-generated] [timmy-capability]
Triage Assessment — Needs Investigation
Scoring: scope=1, acceptance=1, alignment=2 (total=4).
The four options listed need a decision. Recommended approach:
Option 4 first (cheapest): Check Ollama's
OLLAMA_NUM_PARALLELsetting. Default is 1. Setting it to 2 may solve contention without any code changes.If that doesn't work:
Option 2 (two-model split): Use a smaller model (e.g.,
qwen3:1.7b) for Timmy chat while the loop usesqwen3:30b. This is a config change, not code.Action: Someone needs to test
OLLAMA_NUM_PARALLEL=2and report back. Filing this as investigation, not implementation.[triage-generated]
Fixed in PR #238 (loop-cycle-60).
Changes:
BaseAgent.run()now retries on transient Ollama errors (ConnectError, ReadError, ReadTimeout, ConnectTimeout) with exponential backoff (2s, 4s, max 16s)TimmyWithMemory.chat()gets the same retry logicThis addresses the immediate symptom. For deeper fixes (queue-based scheduling, model-level separation), those can be separate issues.