Repository navigation
fix(aiops): remove implicit 15-second tool execution deadline #523
Description
Problem
On the v3 test build (v0.18.2-v3-test.8b1c1db8), disk diagnosis reports that du scanning timed out and falls back to host_bash.
Code inspection confirms that the shared tool decorator defaults to 15 seconds and the chat composition root explicitly applies that budget to the startup tool catalog, including du_summary. Workflow tools repeat the same default. This silently overrides longer tool-specific execution budgets. The v3 Graph Config also advertises a 15-second default although its ToolTimeout field is not wired to execution.
Reproduction
- Ask the agent to diagnose disk usage on a sufficiently large directory tree.
- Invoke
du_summarywhere scanning takes more than 15 seconds. - The generic wrapper expires before the filesystem tool's own budget; the agent may retry through host_bash.
The reported user symptom is confirmed; the exact affected live invocation has not been traced. This is a code-confirmed timeout defect, not a claim that every scan failure has the same cause.
Expected behavior / fix
- No implicit global 15-second timeout around tool execution.
- Zero/unset decorator timeout adds no deadline; explicitly configured positive timeouts remain supported.
- Preserve parent cancellation/deadlines, approvals, auditing, rate limits and metrics.
- Remove the chat/workflow hardcoded 15-second wrapper and the misleading v3 Graph default. Keep archived v1/v2 source unchanged.
- Cover default behavior, explicit timeout, caller cancellation/deadline, and a tool execution lasting over 15 seconds with race tests.
Scope and remaining limits
This fix removes the generic 15-second restriction, not all operational deadlines. The filesystem tool currently has a 60-second Manager round-trip limit, 30-second Edge per-path limit and 60-second Edge batch limit. host_bash has independent Manager/Edge/sandbox budgets. These must not be represented as unlimited execution after this fix.
Tracked on the current runtime requirements branch refactor/unified-agent-runtime, related to #503. No merge or production release is requested.
Activity
singchia commentedon Oct 16, 2026
Follow-up on the same branch: extend built-in execution budgets before introducing configurable/asynchronous long-task execution.
du_summary/find_large_files: 5 minutes per path; 20 minutes for four waves of 16 paths; Manager waits 21 minutes including the separate 30-second filesystem snapshot and response headroom. Shared Manager/Edge constants prevent mismatches.host_bash: default sandbox execution 5 minutes (existing maximum stays 300 seconds); Manager per-device deadline 5m30s and batch deadline accounts for all four waves. Shorter requested deadlines still apply. Operator policy overrides remain authoritative.cloud_bash: explicitly pass a 10-minute budget to the runner instead of inheriting its 2-minute default.Validation includes a real shell command completing after 31 seconds, a requested 1-second deadline stopping a command, and Manager deadline propagation/parent deadline preservation. Edge changes require an Edge rollout; updating Manager alone does not activate the new scan/sandbox limits. Rollback is restoring the previous Manager/Edge binaries; no database or wire schema changes.
RyoKusnadi commentedon Oct 17, 2026
@singchia could you please review these two PRs when you have a moment? Thank you!
RyoKusnadi commentedon Oct 17, 2026
Status update:
ToolInvokerand the fast coordinator deps. It is best merged after fix(aiops): make a zero tool timeout mean no deadline #524, and it is the part markedFixes #523.This issue stays open until #525 lands.
fix(aiops): remove implicit tool execution timeout
fix(aiops): extend built-in tool execution budgets
fix(skills): allow five minutes for journal reads