auto review fails with 400 MissingSessionID on opencode-go routes (reviewer omits x-opencode-session)

[intellipu]

Summary

With DSH 0.2.1-rc.2 and an opencode-go/* model route, selecting the auto (Auto review) permission preset makes every tool call fail closed:

Auto review of tool "<name>" failed: reviewer ended with error INVALID_REQUEST: 400:
{"type":"MissingSessionID","message":"Request is missing x-opencode-session ..."}

The tool body is never executed. Non-auto presets are unaffected.

Environment

• DSH 0.2.1-rc.2 (desktop), Windows 11
• @deepseek-ai/dsh-experimental-auto-review (bundled), @deepseek-ai/dsh-permission-presets
• Model route: opencode-go/deepseek-v4.1-flash (the default here)

Reproduction

1. Use any opencode-go/* model as the session route.
2. /permission auto (or pick 自动审查 / Auto review in the picker).
3. Let the agent call any tool (e.g. a shell command). The call is rejected with the 400 above instead of being reviewed.

Root cause

dsh-experimental-auto-review/lib/index.js → classifyRisk() builds the reviewer request with only the provider/model route:

const options = deepFreeze({
    provider: snapshot.provider,
    model: snapshot.model,
    system: REVIEW_POLICY,
    messages: [{ role: "user", content: [{ type: "text", text: reviewUserText(snapshot) }] }],
    temperature: 0,
    signal,
});
return readDecision(ctx.llm.stream(options));

There is no sessionId. The opencode-go provider only attaches its mandatory per-conversation header when that field exists:

const OPENCODE_SESSION_HEADER = "x-opencode-session";
function withSessionHeader(options) {
    if (!options?.sessionId || hasHeader(options.headers, OPENCODE_SESSION_HEADER)) return options;
    return { ...options, headers: { ...options.headers, [OPENCODE_SESSION_HEADER]: options.sessionId } };
}

So the reviewer's POST https://opencode.ai/zen/go/v1/chat/completions goes out without the header and the gateway answers 400 MissingSessionID.

Direct API evidence (same key, same endpoint)

• no header → 400 {"type":"error","error":{"type":"MissingSessionID","message":"Request is missing x-opencode-session ..."}}
• x-opencode-session: <any uuid> → 200 chat.completion

So the credential and endpoint are fine; only the reviewer's request shape is wrong.

Suggested fixes (any one is enough)

1. Pass the owning session's id (sessionId) into classifyRisk's ctx.llm.stream options.
2. Or have the opencode-go provider mint a fallback per-call session id when none is supplied.
3. Or let the auto-review preset pin a provider route that does not require this header.

Impact

The auto preset is unusable on every opencode-go/* route; users must either fall back to full access (no reviewer) or move to another provider. Happy to test a build.

[zgw2]

Confirmed on macOS with DSH 0.2.1-rc.5 — still present in rc.5, same route (opencode-go/deepseek-v4.1-flash).

Extra data point from probing the gateway directly (POST https://opencode.ai/zen/go/v1/chat/completions):

request	result
no x-opencode-session	400 MissingSessionID
x-opencode-session: dsh-unknown-session	200

So the reviewer call does not need the real session id to succeed — any stable value satisfies routing. That makes the one-line fix safe even if the session identity is not fully plumbed through.

Workaround (no ASAR patching, survives app updates)

A small user plugin on the llm/stream waterfall that defaults sessionId only when the caller omitted it. Main-turn calls keep their real id, so per-session routing and caching are untouched.

export const name = "opencode-session-fallback";
export const inject = ["llm"];
export function apply(ctx) {
  ctx.on("llm/stream", (options, next) =>
    next(options.sessionId === undefined ? { ...options, sessionId: "dsh-unknown-session" } : options));
}

Register it from the profile layer ($DSH_HOME/profiles/<profile>/cordis.patch.yml):

- insert:
  - id: opencode-session-fallback
    name: /absolute/path/to/index.mjs

With this in place the auto preset works again on opencode-go routes.

Fix

 const options = deepFreeze({
   provider: snapshot.provider,
   model: snapshot.model,
+  sessionId: agent.session.id,
   system: REVIEW_POLICY,

Suggestion: when a provider declares that it needs a routing identity, an llm.stream call that omits sessionId should fail loudly at dispatch time instead of surfacing as a 400 that the preset turns into "deny every tool call".

Also, this looks reported several times (#9972, #8764, #10643, #10381) — perhaps worth consolidating into one canonical thread so it can be tracked.

[zgw2]

Correction to the workaround in my comment above: the llm/stream waterfall snippet does not work. cordis waterfall() continues with the args it captured (const next = () => (cbs.shift() ?? inner)(...args)), so a modified options object passed to next(...) is silently dropped — and the reviewer's options are deepFreezed, so they cannot be mutated in place either.

A working user-side version (wrap the llm service's stream() so the object is replaced before it enters the waterfall) is in #11986, together with the 0.2.1-rc.5 verification:

#11986

[intellipu]

Received, thank you Nemo ***@***.***