Skip to content

fix(provider): make readiness probe timeouts configurable - #33275

Merged
juliusmarminge merged 10 commits into
pingdotgg:mainfrom
johanjohnthomas:fix/configurable-codex-probe-timeout
Nov 12, 2026
Merged

juliusmarminge merged 10 commits into
pingdotgg:mainfrom
johanjohnthomas:fix/configurable-codex-probe-timeout

Conversation

@johanjohnthomas

@johanjohnthomas johanjohnthomas commented Nov 9, 2026 •

Copy link
Copy Markdown
Contributor

Provider readiness checks have hardcoded deadlines. On a remote Linux cluster, Codex was authenticated and working, but T3 Code's check timed out. Raising the installed timeout to 60 seconds let a probe complete successfully in 22.2 seconds. Patching the executable has to be repeated after updates.

Add T3CODE_<PROVIDER>_PROBE_TIMEOUT_MS overrides for Codex, Claude, Cursor, Antigravity, Grok, Kiro, fx, Pi, Muse, OpenCode, and ACP Registry agents. Each override controls that provider's existing readiness deadlines, including nested discovery/startup limits that would otherwise cut the configured wait short. Without an override, each phase keeps its current default. Invalid, nonpositive, fractional, or overflowing values fall back to the phase default.

Codex's discovery deadline follows its overall status-check budget. Other providers can have several sequential probe phases, so this is a per-phase limit rather than a total refresh deadline. The guide lists the variables and explains how to keep them in a remote server's launcher, systemd service, or provider-instance environment. Turn execution, installation-command, update-check, and shutdown deadlines are unchanged; optional usage/config enrichment keeps its shorter limits.

Closes #33266. Related: #27230 and #31023. This configures an existing capability without changing defaults or adding UI, following CONTRIBUTING.md's focused-configuration exception.

Verification:

  • Before the change, the new Cursor tests failed: a successful 20-second catalog read timed out despite a 30-second override, and a stalled read ignored a 1-second override. The existing Codex regression also fails on unmodified upstream for a successful 22-second probe.
  • The changed-file check ran 270 tests across 21 files; all passed. This includes namespace/validation cases, slow-success and stalled-cancellation regressions, provider status checks, discovery, and OpenCode runtime/driver tests. Another 19 tests passed across the Grok, Kiro, ACP Registry, and Claude capability suites (289 tests across 25 files in total).
  • Typechecking passed for the server and all 11 provider packages: corepack pnpm --filter t3 --filter '@t3tools/provider-*' run typecheck.
  • Changed-file lint, formatting, and the server GraphQL generation check passed.
  • vp run check:changed c73b38b379542cde87ec8bc855b70e7fae78fb5c could not pass its broader workspace checks: Knip cannot load the uninstalled Expo/Astro configs, and the desktop typecheck cannot find Electron types. The filtered checkout contains server/provider dependencies. The separate typecheck above passed. The commit hook runs the same workspace Knip check, so it was bypassed after running these checks explicitly.

The Linux cluster timings are the user's reported measurements; I did not rerun that live setup. The issue does not attribute the latency to a particular plugin.

Model: GPT-6.1-Sol. Harness: Codex CLI in T3 Code.

Loading
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

needs-rebase Conflicts with its base branch. Closes after 7 days with no activity. size:L 100-499 changed lines (additions + deletions). vouch:unvouched Author is not yet trusted in the VOUCHED list.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Allow users to configure provider readiness timeouts

3 participants