Repository navigation
feat(server): start Codex threads on GPT-6.1-Sol at medium effort - #543
Merged
Merged
Conversation
incognitojam
added a commit
that referenced
this pull request
Sep 29, 2026
… none (#544) A Codex turn where the user never picked an effort ran at medium, whatever the effort picker showed as default. The composer only sends options the user chose explicitly, so the adapter sent no `effort` on `turn/start`. The turn's collaboration mode then filled in a hardcoded `medium` (`input.effort ?? "medium"` in `CodexSessionRuntime`). The picker showed Codex's catalog default, which is Low for GPT-5.6-Sol and GPT-6.1-Sol. Upstream has the same behaviour. ## Fix When a turn's selection names no reasoning effort, the Codex adapter now sends the effort the provider snapshot marks as that model's default. Clients show the same value, so the label and the turn agree. The driver builds the snapshot before the adapter and gives the adapter that lookup. An explicit choice still wins. A model the snapshot has no effort options for, such as a custom model, keeps the old fallback. This is independent of #543, which changes the default shown for Codex models to medium where supported. With both, a new GPT-6.1-Sol thread shows Medium and sends Medium explicitly. Added the `codex-default-effort` fork ledger entry. ## Validation - **Live run** on a dev server with isolated state and Codex CLI 0.158.0. I created a GPT-5.6-Sol thread with no effort option; the effort picker shows Low for that model. Codex's thread settings for the turn recorded `effort: "low"` and `reasoning_effort: "low"`, and the agent replied. Before the change, an untouched GPT-6.1-Sol turn whose picker showed Low recorded `medium`. - **Tests:** a new `CodexAdapter.test.ts` case sends turns with no effort, an explicit `high`, and a model with no default. The turns send the default, `high`, and nothing, respectively. The `CodexAdapter`, `CodexProvider` and `CodexSessionRuntime` suites pass. Three npm-update tests in `CodexDriver.test.ts` fail locally on this machine and fail the same way on `main`. - **Checks:** server typecheck and ledger check pass. --- Written by an agent (Claude Code, claude-opus-5-5).
GPT-6-Sol moves to legacy models. Manifest catalog entries can now badge models that Codex discovers, so GPT-6.1-Sol is marked new.
GPT-6-Sol stays the first fallback for Codex CLIs that do not offer GPT-6.1-Sol. A turn without an explicit effort runs at the collaboration mode's medium fallback, so the effort picker now marks medium as the default instead of the catalog's own default.
incognitojam
force-pushed
the
styal/gpt-6-1-sol
branch
from
September 29, 2026 20:53
06f9e8e to
71b3fab
Compare
incognitojam
pushed a commit
that referenced
this pull request
Sep 29, 2026
…ingdotgg#9921) Fork adaptation: Fork #544 builds the Codex adapter after the provider snapshot so it can read each model's default effort, so the adapter and the text generation that now reads the snapshot's models are both created after it. The manifest's legacy check keeps fork #543's catalog lookup helper and adds the Codex family fallback to it. (cherry picked from commit a92161a) Upstream-PR: 9921
incognitojam
pushed a commit
that referenced
this pull request
Sep 29, 2026
…11347) Co-authored-by: maria-rcks <254055478+maria-rcks@users.noreply.github.com> Fork adaptation: Already present. The fork chooses its own Codex and Claude defaults (codex-default-model: GPT-6.1-Sol, then GPT-6-Sol and GPT-6-Astra; Claude Opus 5.5 from fork #405), fork #405 already gave Fable 5.1 a medium default effort, and fork #543 makes medium the default effort for every Codex model that offers it, which covers GPT-6-Astra. This commit records provenance only. (cherry picked from commit 1f73a89) Upstream-PR: 11347
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Note
New Codex threads start on GPT-6.1-Sol at medium effort. GPT-6.1-Sol is listed as a current model with a "New" badge, and GPT-6-Sol moves to legacy models. The effort picker now marks the effort that a turn without an explicit choice actually runs at.
Problem
OpenAI released GPT-6.1-Sol today. Codex CLI 0.159.0 lists it, but styal filed it under legacy models, because the model manifest's current Codex list only knew GPT-6-Sol. New Codex threads still defaulted to GPT-6-Sol.
The effort picker also misreported the effort a new thread runs at. It marked the effort Codex's catalog calls default, which is Low for GPT-6.1-Sol and GPT-5.6-Sol. But a turn without an explicit effort sends the collaboration mode's
mediumfallback, which overrides the catalog default. An untouched GPT-6.1-Sol thread showed Low and ran at Medium.Medium is also the effort to start at. In the independent FrontierCode 1.1 run through Codex, GPT-6.1-Sol scores 45.5 at low and 50.2 at medium, with no gain above medium, for $0.36 per task at medium. OpenAI's DeepSWE numbers show the same step from low to medium (64.4% to 73.0%).
Change
gpt-6.1-solreplacesgpt-6-solin the current Codex models, and aproviders.codexcatalog entry gives it anewbadge.classifyModelsnow copies a catalog entry's badge onto a model that Codex discovers. The entry can only badge or classify a model; it never adds one that Codex doesn't list.DEFAULT_MODELisgpt-6.1-sol. The preference order is now GPT-6.1-Sol, GPT-6-Sol, GPT-6-Astra, then older models, so Codex CLIs before 0.159.0 keep GPT-6-Sol.mapCodexModelCapabilitiesmarksmediumas styal's default whenever the model supports it, instead of Codex's catalog default. Today that matches what untouched turns run at, because the collaboration mode falls back to the same constant. fix(server): send Codex the effort shown as default when a turn names none #544 makes the server send the shown default explicitly. Only GPT-6.1-Sol and GPT-5.6-Sol show a different default than before; the other current models already defaulted to Medium in Codex's catalog.codex-default-modelledger entry and the styal differences page. Upstream still defaults to GPT-6-Astra at2cbc24f.This is written for the fork rather than imported. Upstream has an open PR that adds the same manifest entry and badge support: pingdotgg#14294. Upstream intake will meet that change in
ModelManifest.tsandmodel-manifest.json.Installed styal servers fetch the model manifest from
main, so after merge they list GPT-6.1-Sol as current and GPT-6-Sol as legacy without an update. The badge, the new default and the effort default need the new server.Validation
CODEX_HOMEso the global Codex install and its cache were untouched:gpt-6.1-soland no effort option answered a real prompt. The server trace showssendTurnwithprovider.model: gpt-6.1-sol, and Codex's thread settings recordedeffort: "medium". That is the run where the composer, before this change, showed Low.mainopens new threads on GPT-6-Sol and files GPT-6.1-Sol under legacy models.ModelManifest.test.tscovers badging discovered models.CodexProvider.test.tscovers the medium effort default and the GPT-6.1-Sol → GPT-6-Sol → GPT-6-Astra preference. The Codex provider, adapter, session runtime and manifest suites pass. Three npm-update tests inCodexDriver.test.tsfail locally on this machine; they fail the same way onmain.main)Written by an agent (Claude Code, claude-opus-5-5).