[Bug]: Merge Push fails if Codex isn't set up #31005

Activity

1. maxkloker

   added

   bugSomething is broken or behaving incorrectly.Something is broken or behaving incorrectly.

   needs-triageIssue needs maintainer review and initial categorization.Issue needs maintainer review and initial categorization.

   on Nov 4, 2026
2. Exotic209093 commented on Nov 5, 2026

   

   Exotic209093

   on Nov 5, 2026

   Contributor

   More actions

   I looked into this on main at 337c585 (reading the code, not reproducing it at runtime).

   The PR title/body, commit message and branch name all use the "Text generation model" setting, which defaults to Codex gpt-6-luna (packages/contracts/src/settings.ts:1320). Codex is enabled by default whether or not the CLI is installed (settings.ts:1400-1424), and the settings fallback (apps/server/src/serverSettings.ts:509-539) only moves away from a disabled provider, not one whose binary is missing. TextGeneration.generate has a fallback that checks installed and auth (apps/server/src/textGeneration/TextGeneration.ts:88-123), but it only fires when the usage window is exhausted.

   Workaround: Settings → Threads → Text generation, and pick a model from a provider you have installed. Settings → Source control → "Use a separate source control writer model" does the same for commit/PR/branch text only.

   It isn't Codex-specific: Claude fails the same way ("Failed to spawn Claude CLI process", ClaudeTextGeneration.ts:248). Codex is just the default.

   Related: #26962 (same cause, thread titles), #5182, #28906. #29607 (open) falls back when the selected provider isn't installed or signed in, and since it changes the shared generate path it should cover generatePrContent too. #31255 (open) makes the toast name the model and link to its setting.
3. Muzammil-Bit commented on Nov 6, 2026

   

   Muzammil-Bit

   on Nov 6, 2026

   Contributor

   More actions

   Note

   🤖 Claude Opus 5.5 responding on behalf of Muzammil Hussain

   This is fixed on main by #29607 (merged Nov 6). #26979 fixes it as well, but that PR has not merged and now conflicts with main.

   Both PRs make text generation move to another usable provider when the selected one can't start. Before them, it only moved when the selected model's usage window was exhausted. Commit messages, PR content and branch names all go through that same function.

   To check, I ran two dev servers, each with a fresh database. In both I set Settings → Providers → Codex → Binary path to a missing file (Codex still enabled, shown as "Not found"), left the text generation model at the default gpt-6-luna, picked Claude Haiku 5.5, and committed a new file with an empty commit message:

   	main dd059b4	#26979 1e238fd
   Result	Committed 9224977, "Add greet helper in greet.ts"	Committed 802d564, "Add greet helper"
   Provider that wrote the message (server trace)	Claude	Claude
   Failed to spawn Codex CLI process	none	none

   I tested the commit path, not the "create PR" button, because that button would open a real PR. Both go through the same fallback. Until a build with #29607 ships, the workaround is to pick an installed provider's model in Settings → Threads → Text generation model.

   👍React with 👍1Reacted by Yash Singh
4. Yash-Singh1 commented on Nov 6, 2026

   

   Yash-Singh1

   on Nov 6, 2026

   Collaborator

   More actions

   Should be fixed now.
5. Yash-Singh1

   closed this as completedon Nov 6, 2026