opencode-go-stub-400-retry

A single-file OpenCode V2 plugin that retries the opaque HTTP 400 that opencode-go intermittently returns for DeepSeek Flash, so the turn no longer dies and needs a manual "continue".

Not affiliated with OpenCode. Client-side workaround, not a gateway fix.

The problem

POST https://opencode.ai/zen/go/v1/chat/completions intermittently answers with an HTTP 400 whose entire body is a 31-byte stub:

{"model":"deepseek-v4.1-flash"}

No error, no message, no code. The failure is not a verdict on the request: replaying the identical bytes usually returns 200 (observed 200/400/200 across three byte-faithful attempts in #51201). It happens mid-session after ordinary tool-call turns, at a few percent of requests, and "continue" normally recovers it.

OpenCode's error extractor finds no error.message/message in the body and falls back to the literal Provider request failed with HTTP 400, classifying it provider.invalid-request. Every 4xx is deliberately non-retryable (OpenCode PR #49195), so the turn ends there and the raw body is not persisted. Hence: no retry, no diagnosis, manual continue.

What it does

Registers OpenCode's session.hook("retry") for provider opencode-go and turns that one opaque signature into a bounded retry: two extra attempts, 3 s then 9 s. Everything else keeps the built-in policy.

Matched only when all hold:

Condition	Value
provider	opencode-go
model	deepseek-v4-flash / deepseek-v4.1-flash (any variant)
error type	provider.invalid-request
status	400
message	opaque — no upstream reason

A structured 400 carries the real reason and is left alone, so deterministic failures (content filter, reasoning_content contract, input over the route limit) are not retried pointlessly.

Install from source

Requires OpenCode V2 (the retry session hook). Verified on 2.0.19 and 2.0.21.

git clone https://github.com/SamSpiri/opencode-go-stub-400-retry.git
./opencode-go-stub-400-retry/install.sh

Or on a machine without a checkout:

curl -fsSL https://raw.githubusercontent.com/SamSpiri/opencode-go-stub-400-retry/main/install.sh | sh

Nothing to install, no dependencies. OpenCode discovers .ts files in the global plugins directory automatically and reloads on change, so there is no config to edit. ~/.config/opencode is $XDG_CONFIG_HOME/opencode when that is set.

To keep it out of the config directory, point at the file from opencode.json(c) instead:

{
  "plugins": ["/absolute/path/to/opencode-go-stub-retry.ts"]
}

The file deliberately exports a plain object rather than calling Plugin.define, so it loads without @opencode/plugin being resolvable — the global plugins directory has no node_modules for that package.

Verify

opencode api get /api/plugin | jq '.data[] | select(.id=="opencode-go.stub-400-retry")'
# { "state": { "status": "active" }, ... }

A forced retry shows up in the session UI as a retry countdown.

End-to-end test

The stub cannot be induced on the real gateway, so point a throwaway project at a local server that answers 400 with {"model":"deepseek-v4.1-flash"} for the first N requests and a normal completion afterwards, then run one prompt. The result on OpenCode 2.0.19, failing the first three requests:

Plugin	Requests seen	Turn
disabled	3, all 400, within the same second	fails: Provider request failed with HTTP 400
enabled	4 — the three, then one 9.1 s later returning 200	completes: ok

Without the plugin the turn dies exactly as reported. With it, the retry carries the turn through.

The delivered message shape

Worth knowing if you write a similar matcher: the message is not stable across versions. On 2.0.19 a bodyless 400 persisted as the bare Provider request failed with HTTP 400; a 2.0.21 node produced

Provider request failed with HTTP 400: {"model":"deepseek-v4.1-flash"}

with the raw body appended, plus an error.response.body holding the stub. A matcher anchored with $ on the first form silently never fires on the second. This plugin strips the prefix and inspects whatever follows.

Tune or disable

DELAYS at the top of the file is the whole policy; one entry per extra attempt, in milliseconds.

const DELAYS = [3000, 9000]

Delete the file (or remove the plugins entry) to disable.

What it does not fix

• reasoning_content contract failures. Echoed reasoning_content on historical assistant tool-call turns combined with any reasoning_effort value is a deterministic rejection (#48180, bisected: removing either side returns 200). Retrying will not help. Drop the #max variant, or the model.
• Genuinely oversized sessions. The gateway is session-pinned to upstreams with different ceilings; one rejects around 582k tokens while the catalog advertises 1,000,000, so auto-compaction (ceiling 968k) never fires (#50761, #50574). Retrying wastes two attempts. Lower the model's limit.context (e.g. 800000) to make compaction fire earlier, or /compact.
• The 403 error code: 1010 is unrelated: Cloudflare denying a request by its browser/UA signature, typically on a hand-made replay.

Evidence

• The opaque stub, mid-session, after tool turns: #51434, #51477, #51990
• Byte-faithful replay flipping 200/400/200; retry usually recovers: #51201
• Route-dependent ceilings behind the same stub: #50761, #50574, #50446
• 4xx deliberately non-retryable: OpenCode PR #49195
• The message is the client fallback, not the gateway's: OpenCode 2.0.19 bundle, hm() in the provider error path

Verified end-to-end against a mock gateway (see above), and against real persisted records: provider.invalid-request, status 400, with the message in both the bare and body-appending forms.