[Bug] New models not reachable on subscription auth: Claude Opus 5.5 (client version gate, recurrence of #1963) and GPT-6 Sol (missing from Codex discovery)

[chrispatil]

Prime Agent 0.9.5, macOS, Anthropic and OpenAI Codex subscriptions via OAuth. Both models were added via models.json since the bundled catalog stops at claude-opus-5 and gpt-6-astra.

Claude Opus 5.5 — spawning on anthropic/claude-opus-5-5 fails with:

400 invalid_request_error: Claude Code 2.1.261 does not support this model;
version 2.1.280 or newer is required.  error_code: claude_code_version_too_old

This is the same gate as #1963 (fixed in #1993 by bumping claudeCodeVersion); the string in the bundled @earendil-works/pi-ai Anthropic provider (dist/providers/anthropic.js) is now 2.1.261 and Anthropic wants ≥ 2.1.280. Since it has fallen behind twice, it may be worth reading the current Claude Code version at build time rather than pinning it.

GPT-6 Sol — openai-codex/gpt-6-sol never becomes executable: Codex model discovery (getExecutableModels) returns gpt-6-astra for my account but not gpt-6-sol. I can't tell from outside whether that's OpenAI's rollout or a client identifier gate like the Anthropic one.

Flagging since other subscription users will hit the same walls. Happy to test a build.

[felipecsl]

as a temporary workaround, Prime Agent lets models.json headers override the hard-coded header. Add this to ~/.prime/agent/models.json:

   {
     "providers": {
       "anthropic": {
         "headers": {
           "user-agent": "claude-cli/2.1.280"
         }
       }
     }
   }

[chrispatil]

Thanks Felipe. I can confirm this works for Opus 5.5. Seth mentioned that there’s a fix coming today that will cover Sol/Luna as well, https://bsky.app/profile/sethkarten.ai/post/3mw5uhxfzns2g and indicated that going forward this will be globally solved: https://bsky.app/profile/sethkarten.ai/post/3mw5bjkkx5s2c

…

On Wed, Sep 23, 2026 at 5:47 AM Felipe Lima ***@***.***> wrote: as a temporary workaround, Prime Agent lets models.json headers override the hard-coded header. Add this to ~/.prime/agent/models.json: { "providers": { "anthropic": { "headers": { "user-agent": "claude-cli/2.1.280" } } } } — Reply to this email directly, view it on GitHub <#2544?email_source=notifications&email_token=A3XJBJ6PK2F2X3RL3PXLZ5L5QPA6NA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOBVGY3DQMZZUZZGKYLTN5XKMYLVORUG64VFMV3GK3TUVRTG633UMVZF6Y3MNFRWW#discussioncomment-18566839>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A3XJBJ3MBNITZKP6TQT27TD5QPA6NAVCNFSNUABJKJSXA33TNF2G64TZHMYTEMZSGQ4TGNBQGY5UI2LTMN2XG43JN5XDWMJQHA3DOMJXHCQXMAQ> . You are receiving this because you authored the thread.Message ID: <PrimeIntellect-ai/prime-agent/repo-discussions/2544/comments/18566839@ github.com>

[cornagraphy]

Has Anthropic said anything about 3rd party usage on prime? 1-100 what are chances I'll get cooked?

[felipecsl]

i've been using it fine for the past several weeks with prime-agent, but no idea if they will keep allowing it in the long run

[chrispatil]

Root cause for the GPT-6 Sol / Luna half, verified against OpenAI directly with my Codex OAuth token on 0.9.6:

• Codex model discovery (GET chatgpt.com/backend-api/codex/models?client_version=...) is gated on client_version, the same way Anthropic gates on the Claude Code version. Prime Agent 0.9.6 pins it to 0.153.4. Sweeping the parameter: 0.153.4 and 0.154.0 return only gpt-6-astra; 0.155.0 and up also return gpt-6-sol and gpt-6-luna.
• A direct POST .../codex/responses with model: gpt-6-sol (and gpt-6-luna) returns 200 with a completed response on the same token, so entitlement isn't the issue. Re-logging in doesn't change anything.
• Unlike the Anthropic case there's no models.json workaround, since it's a query parameter on discovery rather than a header, and getExecutableModels filters openai-codex entries against that list.

So a bump of the pinned Codex client_version (current codex CLI is 0.156.1) unblocks both; reading it from the installed CLI, or dropping the pin, would stop this recurring.

[jasoncarreira]

Seeing the same on 0.9.6: discovery with client_version=0.153.4 returns only gpt-6-astra for my account, while 0.156.1 also returns gpt-6-sol and gpt-6-luna.

A partial workaround for RLM subagents until the pin is bumped: _resolveRlmSubagentModel skips the discovery filter when the requested model is the parent session's own model. So if the parent session runs on openai-codex/gpt-6-sol (selecting it via /model works, since the picker shows the full catalog), subagents spawned without a model argument, or with model="openai-codex/gpt-6-sol", inherit it and run fine. Only spawning a different model than the parent (e.g. gpt-6-luna from a Sol session, or a subagentDefaultModel that differs from the parent) still fails, because that path is filtered against the discovery list.

[chrispatil]

Status check, 2026-09-25, since the thread has gone quiet on the Codex half:

• main still pins OPENAI_CODEX_CLIENT_VERSION = "0.153.4" (packages/coding-agent/src/core/model-registry.ts:395), and no open PR touches it. models: add claude-opus-5-5, gpt-6-sol, gpt-6-luna across the providers that serve them #2531/models: add claude-opus-5-5, gpt-6-sol, gpt-6-luna exporter knowledge #2532, which would have added Sol 6 / Luna 6 to the catalogs, were closed unmerged.
• This is the second time the pin has fallen behind: fix(coding-agent): report a supported Codex client version for model discovery #1375 (Aug 13) bumped it for the same reason and noted that the backend's response is graded, so a low pin yields a silently partial catalog rather than an error. @jasoncarreira's confirmation above matches: 0.153.4 → Astra only; 0.155.0+ → Astra, Sol, Luna.

Would a PR bumping the constant to 0.156.1 (the current Codex CLI release) be welcome as a stopgap? Happy to open one if invited. Longer term, reading the version from an installed codex binary when present, or refreshing it with the catalog fetch that landed in #2507, would stop this recurring with each model launch.

[rain-lee1999]

I independently reproduced the Codex discovery version gate with same-account, read-only catalog requests: client_version=0.153.4 returned 8 models without GPT-6 Sol/Luna; changing only the version to 0.157.0 returned 10 models including both. These results are account/time-specific, not a claim about universal model entitlement.

I prepared the small fix in #2784, including a regression test. npm run check, npm run check:test-policy, and all 55 model-registry tests passed. Restoring the old version caused the strengthened regression to fail.

#2784 was automatically closed because my account is not vouched, not following a technical rejection of the fix. I understand the Discussion-first policy. Would a maintainer be willing to accept this patch or invite a PR? I will wait for an invitation rather than open another unsolicited PR.

[awhite0030]

Investigated this report and produced a candidate fix, validated on a fork of this repository.

Root cause, fix and validation from the working notes:

Root cause: Anthropic Claude Opus 5.5 and OpenAI Codex GPT-6 Sol discovery/spawning are failing because the pinned client versions in prime-agent are too old for the backends.
Fix: Updated claudeCodeVersion to 2.1.280 in packages/ai/src/providers/anthropic.ts and OPENAI_CODEX_CLIENT_VERSION to 0.156.1 in packages/coding-agent/src/core/model-registry.ts. Also updated regression tests to check for these minimal versions.
Validation: Ran npm run check which passed. Executed npx tsx ../../node_modules/vitest/dist/cli.js --run test/suite/regressions/702-codex-client-version.test.ts for coding-agent and npx tsx ../../node_modules/vitest/dist/cli.js --run test/suite/regressions/2544-anthropic-client-version.test.ts for ai package. Both passed successfully.
Upstream discussion: #2544

Diff: awhite0030/prime-agent@main...jules-14127925222576234241-f4f88f27
Full pull request: awhite0030#92

The change passes npm run check and the focused regression tests against current main. If a fix for this is still wanted, I can open a pull request.

[kevinjosethomas]

Hi all, we pushed a fix for this is v0.9.8