[PR #903] [MERGED] ci(bonk): bump opencode to 1.4.11 to unblock OAI through cf-ai-gateway #935

New issue

Closed

opened 2026-05-06 13:10:55 +02:00 by BreizhHardware · 0 comments

BreizhHardware commented

2026-05-06 13:10:55 +02:00

Owner

📋 Pull Request Information

Original PR: https://github.com/cloudflare/vinext/pull/903
Author: @NathanDrake2406
Created: 4/25/2026
Status: ✅ Merged
Merged: 4/25/2026
Merged by: @james-elicx

Base: main ← Head: nathan/bump-opencode-1.4.11

📝 Commits (1)

7f8c91e ci(bonk): bump opencode to 1.4.11 to unblock OAI through cf-ai-gateway

📊 Changes

2 files changed (+2 additions, -2 deletions)

View changed files

📝 .github/workflows/bigbonk.yml (+1 -1)
📝 .github/workflows/bonk.yml (+1 -1)

📄 Description

What

Bump opencode_version from 1.4.6 to 1.4.11 in both bonk.yml and bigbonk.yml. Model + variant unchanged (cloudflare-ai-gateway/openai/gpt-5.4, xhigh).

Why

Two stacked regressions in opencode have made bonk on OAI through cf-ai-gateway uniquely broken on this repo. #902 dodged one and reintroduced the other.

Current state (1.4.6 + gpt-5.4):

APIError: Unsupported parameter: 'max_tokens' is not supported with this model.
Use 'max_completion_tokens' instead.
URL: https://gateway.ai.cloudflare.com/v1/compat/chat/completions

Failing run after #902: https://github.com/cloudflare/vinext/actions/runs/24939826275

opencode 1.4.6 predates the cf-ai-gateway plugin's chat.params hook (anomalyco/opencode#22864, merged 2026-04-16, first release 1.4.7). The hook strips maxOutputTokens for OpenAI reasoning models so @ai-sdk/openai-compatible no longer emits the rejected max_tokens field. Without it, every gpt-5.x request through the unified gateway fails at the OAI edge.

Why not 1.14.x: ProviderInitError on 1.14.22 / 1.14.25 (#898, #900) is a separate regression in opencode's npm install path. The 1.14 major bump added loadOptions(dir) via @npmcli/config inside Npm.add. When that throws on a CI runner, it surfaces as InstallFailedError, gets re-wrapped at packages/opencode/src/provider/provider.ts:1522 as InitError({providerID}, {cause: e}), and the session error path drops the original cause when emitting NamedError.Unknown({message: err.message}). Result: bots see UnknownError: ProviderInitError with no actionable detail. Same issue hit workers-sdk, workerd, kumo, etc., which all pin 1.4.6 / 1.2.27 for that reason.

Why 1.4.11: last release on the 1.4 line. Has the chat.params max_tokens patch (1.4.7+). Pre-dates the 1.14.x loadOptions install regression. ai-gateway-provider@3.1.2 and the @ai-sdk/* pins are identical to 1.4.6, so the install path stays the same arborist.reify shape that works for workers-sdk + anthropic on 1.4.6. vinext is the only Cloudflare repo running OAI through cf-ai-gateway, which is why we need the gpt-5.x plugin patch and they don't.

Approach

bonk.yml: opencode_version: 1.4.6 -> 1.4.11
bigbonk.yml: opencode_version: 1.4.6 -> 1.4.11
No model, variant, or env changes

Validation

Reproduced both failure modes locally with opencode 1.14.25 + cloudflare-ai-gateway/openai/gpt-5.4 + fake creds. Confirmed:

1.14.x: Npm.add short-circuits on a stale empty cache dir (one path) and loadOptions fails on fresh install (other path), both surface as silent ProviderInitError.
1.4.6 with gpt-5.4: install succeeds, request reaches OAI, OAI rejects max_tokens.
1.4.7+ adds the chat.params hook that drops maxOutputTokens for OAI reasoning models.

Will comment /bigbonk review on this PR before merging to confirm bots come back up. If 1.4.11 still throws ProviderInitError for any other reason, fall back to anthropic + 1.4.6 (matches workers-sdk known-good).

Risks / follow-ups

1.4 line is no longer maintained upstream. We should roll forward once loadOptions is fixed (no upstream issue tracking it yet, since it surfaces as the masked ProviderInitError).
The cause-swallowing in opencode's session error path is a real diagnostic bug worth filing upstream.

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/cloudflare/vinext/pull/903 **Author:** [@NathanDrake2406](https://github.com/NathanDrake2406) **Created:** 4/25/2026 **Status:** ✅ Merged **Merged:** 4/25/2026 **Merged by:** [@james-elicx](https://github.com/james-elicx) **Base:** `main` ← **Head:** `nathan/bump-opencode-1.4.11` --- ### 📝 Commits (1) - [`7f8c91e`](https://github.com/cloudflare/vinext/commit/7f8c91e48208d9728b166420ba0ef3af17633f17) ci(bonk): bump opencode to 1.4.11 to unblock OAI through cf-ai-gateway ### 📊 Changes **2 files changed** (+2 additions, -2 deletions) <details> <summary>View changed files</summary> 📝 `.github/workflows/bigbonk.yml` (+1 -1) 📝 `.github/workflows/bonk.yml` (+1 -1) </details> ### 📄 Description ## What Bump `opencode_version` from `1.4.6` to `1.4.11` in both `bonk.yml` and `bigbonk.yml`. Model + variant unchanged (`cloudflare-ai-gateway/openai/gpt-5.4`, `xhigh`). ## Why Two stacked regressions in opencode have made bonk on OAI through cf-ai-gateway uniquely broken on this repo. #902 dodged one and reintroduced the other. **Current state (1.4.6 + gpt-5.4):** ``` APIError: Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead. URL: https://gateway.ai.cloudflare.com/v1/compat/chat/completions ``` Failing run after #902: https://github.com/cloudflare/vinext/actions/runs/24939826275 opencode 1.4.6 predates the cf-ai-gateway plugin's `chat.params` hook (anomalyco/opencode#22864, merged 2026-04-16, first release 1.4.7). The hook strips `maxOutputTokens` for OpenAI reasoning models so `@ai-sdk/openai-compatible` no longer emits the rejected `max_tokens` field. Without it, every gpt-5.x request through the unified gateway fails at the OAI edge. **Why not 1.14.x:** `ProviderInitError` on 1.14.22 / 1.14.25 (#898, #900) is a separate regression in opencode's npm install path. The 1.14 major bump added `loadOptions(dir)` via `@npmcli/config` inside `Npm.add`. When that throws on a CI runner, it surfaces as `InstallFailedError`, gets re-wrapped at `packages/opencode/src/provider/provider.ts:1522` as `InitError({providerID}, {cause: e})`, and the session error path drops the original `cause` when emitting `NamedError.Unknown({message: err.message})`. Result: bots see `UnknownError: ProviderInitError` with no actionable detail. Same issue hit workers-sdk, workerd, kumo, etc., which all pin `1.4.6` / `1.2.27` for that reason. **Why 1.4.11:** last release on the 1.4 line. Has the `chat.params` max_tokens patch (1.4.7+). Pre-dates the 1.14.x `loadOptions` install regression. `ai-gateway-provider@3.1.2` and the `@ai-sdk/*` pins are identical to 1.4.6, so the install path stays the same `arborist.reify` shape that works for workers-sdk + anthropic on 1.4.6. vinext is the only Cloudflare repo running OAI through cf-ai-gateway, which is why we need the gpt-5.x plugin patch and they don't. ## Approach - `bonk.yml`: `opencode_version: 1.4.6` -> `1.4.11` - `bigbonk.yml`: `opencode_version: 1.4.6` -> `1.4.11` - No model, variant, or env changes ## Validation Reproduced both failure modes locally with opencode 1.14.25 + `cloudflare-ai-gateway/openai/gpt-5.4` + fake creds. Confirmed: - 1.14.x: `Npm.add` short-circuits on a stale empty cache dir (one path) and `loadOptions` fails on fresh install (other path), both surface as silent `ProviderInitError`. - 1.4.6 with gpt-5.4: install succeeds, request reaches OAI, OAI rejects `max_tokens`. - 1.4.7+ adds the `chat.params` hook that drops `maxOutputTokens` for OAI reasoning models. Will comment `/bigbonk review` on this PR before merging to confirm bots come back up. If 1.4.11 still throws `ProviderInitError` for any other reason, fall back to anthropic + 1.4.6 (matches workers-sdk known-good). ## Risks / follow-ups - 1.4 line is no longer maintained upstream. We should roll forward once `loadOptions` is fixed (no upstream issue tracking it yet, since it surfaces as the masked `ProviderInitError`). - The `cause`-swallowing in opencode's session error path is a real diagnostic bug worth filing upstream. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>