[BUG] gpt 5.6 sol doesn't work

[nxtkofi]

Bug Description
Terra and luna work fine, Sol has problems

Steps to Reproduce
Open opencode and try querying gpt 5.6 sol

Expected Behavior
Sol answers the question

Actual Behavior
Terra and Luna work, Sol returns:

All 3 account(s) returned 'model not supported' for the requested model. The backend rejected 'gpt-5.6-sol' as not entitled for Codex OAuth on every pooled account. Codex model access is account/workspace gated; default gpt-5.5/gpt-5-codex selectors auto-fallback through the GPT-5.4 family when possible. Set `unsupportedCodexPolicy: "fallback"` for the full manual fallback chain, or see `codex-health` for per-account details.

Tried without other plugins and could pinpoint the issue to oc-codex-multi-auth

Environment

• opencode version: 1.17.20
• Plugin version: 6.8
• Operating System:arch linux, kernel 6.18.37-1-lts
• Node.js version: 22.21.1

[coderabbitai]

Coding Plan

Summary

root cause is not a sol-specific data/slug bug — code treats sol/terra/luna symmetrically everywhere except the auto-fallback allowlist in lib/request/fetch-helpers.ts:125-133, which lists only gpt-5.5/gpt-5-codex, so sol never degrades under the default strict policy despite docs promising it does.

fix by adding the three 5.6 tiers to DEFAULT_AUTO_FALLBACK_ENTRY_OPT_OUT_ENV behind a new CODEX_AUTH_DISABLE_GPT56_AUTO_FALLBACK opt-out, unblocking the already-correct sol→terra→luna→gpt-5.5 chain.

close the test gap that let this ship: add resolveUnsupportedCodexFallbackModel regression cases for the 5.6 tiers in test/fetch-helpers.test.ts, with env teardown to avoid cross-test bleed.

realign docs/configuration.md, docs/troubleshooting.md, CHANGELOG.md, and the user-facing messages in index.ts/fetch-helpers.ts so documented and shipped behavior finally match.

Design Choices

Design Choice 1: should gpt-5.6-sol degrade automatically by default, or should we just document that `unsupportedCodexPolicy: "fallback"` is required?

Options Considered:

1. add the 5.6 tiers to the default-selector auto-fallback allowlist so sol auto-degrades under the default strict policy, with an opt-out env var (mirrors gpt-5.5/gpt-5-codex).
2. leave code as-is and only rewrite docs/changelog to say unsupportedCodexPolicy: "fallback" is required for 5.6 tiers.

Chosen Option: 1

Rationale: the changelog (CHANGELOG.md 6.7.0) and docs/configuration.md:61 both state an account without access "degrades sol → terra → luna → gpt-5.5 through the unsupported-model fallback chain," implying automatic behavior. the code (lib/request/fetch-helpers.ts:319-327) only auto-degrades models on the allowlist, and 5.6 tiers are absent, so the documented behavior is not what ships. option 1 makes shipped behavior match the documented "limited preview auto-degrades" promise and directly resolves the reported bug, whereas option 2 leaves flagship-tier users hard-failing by default.

💡 User Tips

Regenerate the plan with different choices with @coderabbitai <feedback>.

Implementation Steps

Phase 1: make the 5.6 tiers auto-degrade by default

the core fix. bring gpt-5.6-sol/terra/luna onto the same default-selector auto-fallback path that gpt-5.5/gpt-5-codex already enjoy, so a sol request that hits model_not_supported_with_chatgpt_account on every pooled account degrades to terra → luna → gpt-5.5 without requiring the user to flip unsupportedCodexPolicy: "fallback". keep an opt-out escape hatch so strict-entitlement users can still force a hard failure.

Task 1: add 5.6 tiers to the auto-fallback allowlist

extend DEFAULT_AUTO_FALLBACK_ENTRY_OPT_OUT_ENV in lib/request/fetch-helpers.ts:125 to include GPT_56_SOL_MODEL_ID, GPT_56_TERRA_MODEL_ID, and GPT_56_LUNA_MODEL_ID, each keyed to a single new opt-out env var, e.g. CODEX_AUTH_DISABLE_GPT56_AUTO_FALLBACK.

because each tier is registered as its own entry point, resolveAutoFallbackEntryModel (lib/request/fetch-helpers.ts:135-151) returns the tier itself and shouldAutoFallbackForDefaultSelector (lib/request/fetch-helpers.ts:319-320) becomes true for each hop, so resolveUnsupportedCodexFallbackModel walks the existing DEFAULT_UNSUPPORTED_CODEX_FALLBACK_CHAIN (lib/request/fetch-helpers.ts:~99-109) sol→terra→luna→gpt-5.5 without the options.fallbackOnUnsupportedCodexModel gate blocking it.

no change needed to the static chain map itself — it already has correct sol/terra/luna entries; we are only unblocking traversal.

add a short comment near the allowlist explaining the 5.6 preview rationale so the next reader understands why the flagship tier is opt-out rather than opt-in.

Task 2: refresh the user-facing exhaustion and error messages

update the account-exhaustion message in index.ts:~2889 so it no longer implies only gpt-5.5/gpt-5-codex auto-fallback; it should reflect that 5.6 tiers now degrade down their tier chain too.

update normalizeErrorPayload guidance in lib/request/fetch-helpers.ts:~1091-1100 similarly, and mention the new CODEX_AUTH_DISABLE_GPT56_AUTO_FALLBACK opt-out for users who genuinely want strict failures on 5.6.

keep the unsupportedCodexPolicy: "fallback" hint for the full manual chain (5.4 family tail) since that still applies.

🤖 Prompt for AI agents

implementing the core fix for gpt-5.6 auto-fallback: bring the sol/terra/luna
tiers onto the same default-selector auto-fallback path that gpt-5.5/gpt-5-codex
already use, so a sol request degrades down the chain without requiring a manual
policy flip, while keeping an opt-out for strict-entitlement users.

- extend `DEFAULT_AUTO_FALLBACK_ENTRY_OPT_OUT_ENV` in
`lib/request/fetch-helpers.ts:125` to add `GPT_56_SOL_MODEL_ID`,
`GPT_56_TERRA_MODEL_ID`, and `GPT_56_LUNA_MODEL_ID`, all keyed to one new
opt-out env var: `CODEX_AUTH_DISABLE_GPT56_AUTO_FALLBACK`.
- confirm `resolveAutoFallbackEntryModel`
(`lib/request/fetch-helpers.ts:135-151`) returns the tier itself for each of
these entries, and that `shouldAutoFallbackForDefaultSelector`
(`lib/request/fetch-helpers.ts:319-320`) evaluates true for each hop, so
`resolveUnsupportedCodexFallbackModel` traverses the existing
`DEFAULT_UNSUPPORTED_CODEX_FALLBACK_CHAIN`
(`lib/request/fetch-helpers.ts:~99-109`) sol→terra→luna→gpt-5.5 without being
blocked by the `options.fallbackOnUnsupportedCodexModel` gate.
- do not touch the static chain map itself — sol/terra/luna entries are already
correct there; this task only unblocks traversal.
- add a short comment next to the allowlist entry explaining the 5.6 preview
rationale, so future readers understand why the flagship tier is opt-out rather
than opt-in.
- update the account-exhaustion message in `index.ts:~2889` so it reflects that
5.6 tiers now auto-degrade down their chain too, not just gpt-5.5/gpt-5-codex.
- update the `normalizeErrorPayload` guidance in
`lib/request/fetch-helpers.ts:~1091-1100` to match, and mention the new
`CODEX_AUTH_DISABLE_GPT56_AUTO_FALLBACK` opt-out for users wanting strict
failures on 5.6.
- keep the existing `unsupportedCodexPolicy: "fallback"` hint in messaging for
the full manual chain covering the 5.4 family tail, since that behavior is
unchanged.

Phase 2: regression coverage and docs alignment

exploration flagged that resolveUnsupportedCodexFallbackModel has zero test coverage for the 5.6 tiers (test/gpt56-models.test.ts:241-258 only asserts the static map shape; test/fetch-helpers.test.ts:362-780 covers 5.3/5.4/5.5/codex only). this is exactly the gap that let the bug ship, so lock the new default behavior down and correct the docs that over-promised.

Task 1: add regression tests for 5.6 default auto-fallback

in test/fetch-helpers.test.ts, add cases invoking resolveUnsupportedCodexFallbackModel({ requestedModel: "gpt-5.6-sol", fallbackOnUnsupportedCodexModel: false, ... }) and assert it returns gpt-5.6-terra, then gpt-5.6-luna, then gpt-5.5 as attempted grows — i.e. auto-fallback works with the default (strict) policy.

add a parallel case entering at gpt-5.6-terra and gpt-5.6-luna to confirm they degrade correctly, and one asserting the bare gpt-5.6 alias (normalized to sol) behaves identically.

add an opt-out test setting CODEX_AUTH_DISABLE_GPT56_AUTO_FALLBACK=1 and asserting no auto-fallback occurs unless fallbackOnUnsupportedCodexModel is true — mirror the existing honors the default selector auto-fallback opt-out pattern. flag: this env-driven test must restore process.env in teardown to avoid cross-test bleed given the suite runs cases in the same process (concurrency risk if parallelized).

concurrency note: the fallback path only mutates the per-request attempted set in index.ts; no shared/global mutable state is introduced, so no locking concern — but the tests should still construct a fresh attempted set per case.

windows note: the opt-out relies on process.env["CODEX_AUTH_DISABLE_GPT56_AUTO_FALLBACK"] !== "1" string compare, which is platform-neutral; no path/separator concerns here, but confirm the test doesn't assume case-insensitive env keys (windows env keys are case-insensitive, our reads use exact-case constants — fine).

Task 2: correct changelog and docs

update docs/configuration.md:56-63 ("GPT-5.6 notes") to state explicitly that sol/terra/luna auto-degrade down the tier chain by default (like gpt-5.5/gpt-5-codex), and document the CODEX_AUTH_DISABLE_GPT56_AUTO_FALLBACK opt-out.

update the "auto-fallback" list in docs/troubleshooting.md:359-402 to include the 5.6 tiers and their chain (sol → terra → luna → gpt-5.5 → gpt-5.4 family) alongside the existing gpt-5.5/gpt-5-codex entries, plus the new opt-out env var next to the existing CODEX_AUTH_DISABLE_* ones.

add a CHANGELOG.md entry describing the fix: 5.6 tiers now auto-degrade under the default strict policy, resolving sol hard-failing when only terra/luna are entitled.

🤖 Prompt for AI agents

closing the test gap that let this bug ship and realigning docs/changelog with
the new default auto-fallback behavior for gpt-5.6 tiers.

- in `test/fetch-helpers.test.ts`, add regression cases for
`resolveUnsupportedCodexFallbackModel({ requestedModel: "gpt-5.6-sol",
fallbackOnUnsupportedCodexModel: false, ... })` asserting it returns
`gpt-5.6-terra`, then `gpt-5.6-luna`, then `gpt-5.5` as the `attempted` set
grows — this proves auto-fallback works under the default `strict` policy.
- add parallel cases entering the chain at `gpt-5.6-terra` and `gpt-5.6-luna` to
confirm correct degradation, plus a case confirming the bare `gpt-5.6` alias
(normalized to sol) behaves identically.
- add an opt-out test setting `CODEX_AUTH_DISABLE_GPT56_AUTO_FALLBACK=1` and
asserting no auto-fallback happens unless `fallbackOnUnsupportedCodexModel` is
explicitly true — mirror the existing `honors the default selector auto-fallback
opt-out` test pattern.
- flag and handle explicitly: this env-driven test must restore `process.env` in
teardown to avoid cross-test bleed, since the suite runs in the same process.
- concurrency note: fallback resolution only mutates a per-request `attempted`
set in `index.ts`, no shared/global mutable state — still, construct a fresh
`attempted` set per test case.
- windows note: the opt-out compare
(`process.env["CODEX_AUTH_DISABLE_GPT56_AUTO_FALLBACK"] !== "1"`) is
platform-neutral; confirm tests don't assume case-insensitive env keys (windows
env keys are case-insensitive but our reads use exact-case constants, so this is
fine — just verify).
- update `docs/configuration.md:56-63` ("GPT-5.6 notes") to state that
sol/terra/luna auto-degrade down the tier chain by default, matching
`gpt-5.5`/`gpt-5-codex`, and document the
`CODEX_AUTH_DISABLE_GPT56_AUTO_FALLBACK` opt-out.
- update the auto-fallback list in `docs/troubleshooting.md:359-402` to include
the 5.6 tiers and their full chain (`sol → terra → luna → gpt-5.5 → gpt-5.4
family`), plus the new opt-out env var next to existing `CODEX_AUTH_DISABLE_*`
entries.
- add a `CHANGELOG.md` entry describing the fix: 5.6 tiers now auto-degrade
under the default strict policy, resolving sol hard-failing when only terra/luna
are entitled.

Research

the unsupported-model detection and fallback logic lives in lib/request/fetch-helpers.ts (getUnsupportedCodexModelInfo, resolveUnsupportedCodexFallbackModel, and the DEFAULT_UNSUPPORTED_CODEX_FALLBACK_CHAIN / DEFAULT_AUTO_FALLBACK_ENTRY_OPT_OUT_ENV / DEFAULT_AUTO_FALLBACK_CONTINUATION_MODELS tables). the retry/rotation loop in index.ts (~1926–2530, ~2889) calls these per account and emits the user-facing exhaustion message. model ids/aliases are canonicalized in lib/request/helpers/model-map.ts, policy resolution in lib/config.ts, and behavior is documented in docs/configuration.md / docs/troubleshooting.md. exploration confirmed the 5.6 tiers are symmetric everywhere (model-map, families, responses-lite, the static chain map) — the only asymmetry is the missing auto-fallback allowlist entry.

---

🚀 Next Steps

🤖 All AI agent prompts combined

Task: 1

implementing the core fix for gpt-5.6 auto-fallback: bring the sol/terra/luna
tiers onto the same default-selector auto-fallback path that gpt-5.5/gpt-5-codex
already use, so a sol request degrades down the chain without requiring a manual
policy flip, while keeping an opt-out for strict-entitlement users.

- extend `DEFAULT_AUTO_FALLBACK_ENTRY_OPT_OUT_ENV` in
`lib/request/fetch-helpers.ts:125` to add `GPT_56_SOL_MODEL_ID`,
`GPT_56_TERRA_MODEL_ID`, and `GPT_56_LUNA_MODEL_ID`, all keyed to one new
opt-out env var: `CODEX_AUTH_DISABLE_GPT56_AUTO_FALLBACK`.
- confirm `resolveAutoFallbackEntryModel`
(`lib/request/fetch-helpers.ts:135-151`) returns the tier itself for each of
these entries, and that `shouldAutoFallbackForDefaultSelector`
(`lib/request/fetch-helpers.ts:319-320`) evaluates true for each hop, so
`resolveUnsupportedCodexFallbackModel` traverses the existing
`DEFAULT_UNSUPPORTED_CODEX_FALLBACK_CHAIN`
(`lib/request/fetch-helpers.ts:~99-109`) sol→terra→luna→gpt-5.5 without being
blocked by the `options.fallbackOnUnsupportedCodexModel` gate.
- do not touch the static chain map itself — sol/terra/luna entries are already
correct there; this task only unblocks traversal.
- add a short comment next to the allowlist entry explaining the 5.6 preview
rationale, so future readers understand why the flagship tier is opt-out rather
than opt-in.
- update the account-exhaustion message in `index.ts:~2889` so it reflects that
5.6 tiers now auto-degrade down their chain too, not just gpt-5.5/gpt-5-codex.
- update the `normalizeErrorPayload` guidance in
`lib/request/fetch-helpers.ts:~1091-1100` to match, and mention the new
`CODEX_AUTH_DISABLE_GPT56_AUTO_FALLBACK` opt-out for users wanting strict
failures on 5.6.
- keep the existing `unsupportedCodexPolicy: "fallback"` hint in messaging for
the full manual chain covering the 5.4 family tail, since that behavior is
unchanged.
===============================================================================

Task: 2

closing the test gap that let this bug ship and realigning docs/changelog with
the new default auto-fallback behavior for gpt-5.6 tiers.

- in `test/fetch-helpers.test.ts`, add regression cases for
`resolveUnsupportedCodexFallbackModel({ requestedModel: "gpt-5.6-sol",
fallbackOnUnsupportedCodexModel: false, ... })` asserting it returns
`gpt-5.6-terra`, then `gpt-5.6-luna`, then `gpt-5.5` as the `attempted` set
grows — this proves auto-fallback works under the default `strict` policy.
- add parallel cases entering the chain at `gpt-5.6-terra` and `gpt-5.6-luna` to
confirm correct degradation, plus a case confirming the bare `gpt-5.6` alias
(normalized to sol) behaves identically.
- add an opt-out test setting `CODEX_AUTH_DISABLE_GPT56_AUTO_FALLBACK=1` and
asserting no auto-fallback happens unless `fallbackOnUnsupportedCodexModel` is
explicitly true — mirror the existing `honors the default selector auto-fallback
opt-out` test pattern.
- flag and handle explicitly: this env-driven test must restore `process.env` in
teardown to avoid cross-test bleed, since the suite runs in the same process.
- concurrency note: fallback resolution only mutates a per-request `attempted`
set in `index.ts`, no shared/global mutable state — still, construct a fresh
`attempted` set per test case.
- windows note: the opt-out compare
(`process.env["CODEX_AUTH_DISABLE_GPT56_AUTO_FALLBACK"] !== "1"`) is
platform-neutral; confirm tests don't assume case-insensitive env keys (windows
env keys are case-insensitive but our reads use exact-case constants, so this is
fine — just verify).
- update `docs/configuration.md:56-63` ("GPT-5.6 notes") to state that
sol/terra/luna auto-degrade down the tier chain by default, matching
`gpt-5.5`/`gpt-5-codex`, and document the
`CODEX_AUTH_DISABLE_GPT56_AUTO_FALLBACK` opt-out.
- update the auto-fallback list in `docs/troubleshooting.md:359-402` to include
the 5.6 tiers and their full chain (`sol → terra → luna → gpt-5.5 → gpt-5.4
family`), plus the new opt-out env var next to existing `CODEX_AUTH_DISABLE_*`
entries.
- add a `CHANGELOG.md` entry describing the fix: 5.6 tiers now auto-degrade
under the default strict policy, resolving sol hard-failing when only terra/luna
are entitled.

💡 Iterate on the plan with: `@coderabbitai` <feedback>

Example Feedback
- `@coderabbitai` You can skip phase 3. Add a simple unit test case for phase 2.
- `@coderabbitai` For design choice 1 go ahead with option 3 and replan.

---

💬 Have feedback or questions? Drop into our discord!

[nxtkofi]

hi I don't think a simple fallback fixes the issue, because gpt 5.6 sol works for me in codex tui, but doesn't work in opencode. Also - works after I uninstalled Your plugin, and doesn't when I installed it back on my friend. Can You look into it?