[Bug] 0.2.1-rc.5: auto preset rejects every tool call on opencode-go routes (reviewer call omits sessionId → 400 MissingSessionID)

[zgw2]

Summary

With any opencode-go/* model route, selecting the auto / 自动审查 permission preset makes every tool call fail before its body runs. Non-auto presets are unaffected, and this is still present in 0.2.1-rc.5.

Auto review of tool "" failed; its body was not executed: auto-review: reviewer ended with error INVALID_REQUEST: 400: {"type":"error","error":{"type":"MissingSessionID","message":"Request is missing x-opencode-session ..."}}

Environment

• DSH desktop 0.2.1-rc.5, macOS (Apple Silicon), default profile
• Route: opencode-go/deepseek-v4.1-flash
• Bundled @deepseek-ai/dsh-experimental-auto-review

Root cause

dsh-experimental-auto-review → classifyRisk() builds the reviewer request from the session's provider/model only:

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

sessionId is missing. The opencode-go provider attaches its mandatory routing header only when that field exists (@earendil-works/pi-ai, providers/opencode-headers.js):

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

The agent loop passes sessionId: session.id on normal turns, so chat works; the reviewer call does not, gets a 400, and the preset then fails closed — no tool body ever executes.

Verified against the gateway

POST https://opencode.ai/zen/go/v1/chat/completions with an identical body:

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

So any stable value satisfies routing — the reviewer does not need the real session id for the call to succeed.

Fix

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

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

Suggestion

If a provider declares that it needs a routing identity, an llm.stream call that omits sessionId should fail loudly at dispatch time (or be caught at boot) rather than surfacing later as a 400 that the preset turns into "deny every tool call".

Related

Same defect was reported earlier against rc.2 — #11423, #9972, #8764, #10643, #10381. Posting this one for the 0.2.1-rc.5 confirmation, the gateway-level probe, and a user-side workaround that unblocks the preset today.

[zgw2]

⚠️ Correction to the workaround in the OP — the llm/stream waterfall snippet does not work, please don't copy it.

cordis waterfall() continues the chain with the args it captured, not with whatever a listener passes to next:

const next = () => (cbs.shift() ?? inner)(...args);   // next() takes no parameters

so next({ ...options, sessionId }) silently drops the new object. The reviewer's options are also deepFreezed, so they cannot be mutated in place either.

Working version — wrap the llm service's stream(), which replaces the object before it enters the waterfall:

export const name = "opencode-session-fallback";
export const inject = ["llm"];

const FALLBACK = "dsh-unknown-session";
const PATCHED = Symbol.for("opencode-session-fallback");
const withFallback = (o) =>
  o && typeof o === "object" && o.sessionId === undefined ? { ...o, sessionId: FALLBACK } : o;

export function apply(ctx) {
  const llm = ctx.llm;
  if (!llm) return;
  const wrap = (original) => function (options) {
    return original.call(this, withFallback(options));
  };
  for (const target of [Object.getPrototypeOf(llm), llm]) {
    try {
      if (target && target[PATCHED] !== true && typeof target.stream === "function") {
        target.stream = wrap(target.stream);
        target[PATCHED] = true;
      }
    } catch { /* a proxied cordis service may reject the instance write; the prototype pass covers it */ }
  }
}

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

Verified working on 0.2.1-rc.5 / macOS: reviewer calls now carry the header and the auto preset executes tools normally. Patching both prototype and instance is deliberate — cordis service objects may be proxied.

The one-line sessionId: agent.session.id fix in the OP is still the real solution.