feat: upstream PR triage — 6 fixes + e2e tests

[wende]

Summary

Triaged all 14 open PRs on atalovesyou/claude-max-api-proxy and cherry-picked the valuable fixes into our fork.

Implemented (6 changes)

Change	Upstream PR	Author	Details
Fix normalizeModelName crash on undefined model	#7	@wende	P0 regression — our OpenClaw integration commit (bbeb4c7) accidentally reverted this fix. When modelUsage is {} (rate limits), normalizeModelName receives undefined and crashes. Restored the string | undefined signature + guard.
Pass prompt via stdin instead of CLI arg	#12	@kevinfealey	P1 — large prompts (conversation history, system prompts, base64) can exceed OS ARG_MAX (128KB–2MB), causing spawn() to fail with E2BIG. Now writes prompt to stdin instead.
Increase subprocess timeout to 15 minutes	#20	@Max-shipper	P1 — agentic tasks with multiple tool calls routinely run 8–12 minutes. Previous 5-minute timeout killed them silently.
Add Claude 4.5/4.6 model IDs	#10, #20	@jamshehan, @Max-shipper	P2 — added claude-opus-4-6, claude-sonnet-4-5, claude-sonnet-4-6, claude-haiku-4-5 to MODEL_MAP and /v1/models. Also added claude-max/* provider prefix and opus-max/sonnet-max aliases.
Include usage in streaming SSE response	#16	@smartchainark	P2 — OpenAI spec includes usage in the final streaming chunk. Downstream consumers (session compaction, cost tracking) need it.
Conditional debug logging	#5, #16	@alexrudloffBot, @smartchainark	P3 — all subprocess console.error calls now gated behind DEBUG_SUBPROCESS env var. Production logs are no longer noisy.

Additional fix (not from upstream)

• Strip CLAUDECODE env var from subprocesses — Claude CLI refuses to start when CLAUDECODE=1 is set (nested session protection). The proxy was inheriting this from the parent environment, breaking it when launched from within a Claude Code session.

Already implemented in our fork (skipped)

Upstream PR	What	Author	Why skipped
#3	Fix [object Object] serialization	@grassX1998	Our extractText() already handles both string and array content
#11	Fix [object Object] serialization	@kevinfealey	Duplicate of atalovesyou#3
#17	Handle array content format	@lukedd312	Duplicate of atalovesyou#3
#19	Handle array-style content blocks	@superdav42	Duplicate of atalovesyou#3
#13	CLAUDE_DANGEROUSLY_SKIP_PERMISSIONS env var	@kevinfealey	We hardcode --dangerously-skip-permissions always

Deferred

Upstream PR	What	Author	Why deferred
#4	Docker image	@Ihor-Porokhnia	Nice-to-have; our fork targets macOS LaunchAgent
#14	Vision/image support	@rjabalosiii	Conflicts with OpenClaw tool mapping (CLI isolation with cwd: /tmp and disabled tools)
#18	Multi-CLI support (Cursor + Gemini)	@jsmjsm	Out of scope — 1,688 additions, different project vision

Test plan

• npm run build compiles cleanly
• E2e test suite added (7 tests, 3 hit real Claude CLI with haiku):
  • GET /health returns ok
  • GET /v1/models lists all 7 model IDs
  • 404 for unknown routes
  • 400 for empty messages
  • Non-streaming completion returns valid OpenAI response with usage
  • Non-streaming with array-style content blocks works
  • Streaming completion returns valid SSE chunks with usage in final chunk
• Manual curl verification of non-streaming, streaming, and new model IDs

[sourcery-ai]

Hey - I've found 2 issues, and left some high level feedback:

• The hard-coded model ID list in handleModels and the MODEL_MAP/related logic in openai-to-cli.ts now need to be kept in sync manually; consider extracting a shared source of truth (e.g., a models config module) so adding or renaming models is less error-prone.
• In normalizeModelName, silently defaulting undefined to "claude-sonnet-4" may mask upstream issues; consider either surfacing an explicit error or at least logging/handling the undefined case earlier so unintended model selection is easier to detect.
• The usage field added to the final streaming chunk is typed via an any cast on doneChunk; it would be more robust to extend the response/chunk type definition so this shape is enforced by TypeScript instead of bypassing type checking.

Prompt for AI Agents

Please address the comments from this code review:

## Overall Comments
- The hard-coded model ID list in `handleModels` and the `MODEL_MAP`/related logic in `openai-to-cli.ts` now need to be kept in sync manually; consider extracting a shared source of truth (e.g., a models config module) so adding or renaming models is less error-prone.
- In `normalizeModelName`, silently defaulting `undefined` to `"claude-sonnet-4"` may mask upstream issues; consider either surfacing an explicit error or at least logging/handling the undefined case earlier so unintended model selection is easier to detect.
- The `usage` field added to the final streaming chunk is typed via an `any` cast on `doneChunk`; it would be more robust to extend the response/chunk type definition so this shape is enforced by TypeScript instead of bypassing type checking.

## Individual Comments

### Comment 1
<location> `src/server/routes.ts:248-249` </location>
<code_context>
-        // Send final done chunk with finish_reason
+        // Send final done chunk with finish_reason and usage data
         const doneChunk = createDoneChunk(requestId, lastModel);
+        if (result.usage) {
+          (doneChunk as any).usage = {
+            prompt_tokens: result.usage.input_tokens || 0,
+            completion_tokens: result.usage.output_tokens || 0,
</code_context>

<issue_to_address>
**suggestion:** Avoid `as any` by extending the doneChunk type to include `usage`.

Rather than casting to `any`, update the type returned by `createDoneChunk` (or the shared streaming chunk type) to optionally include `usage`. That way the SSE payload shape stays type-safe and any mismatches are caught at compile time.

Suggested implementation:

```typescript
      if (!res.writableEnded) {
        // Send final done chunk with finish_reason and usage data
        const doneChunk = createDoneChunk(requestId, lastModel);
        if (result.usage) {
          doneChunk.usage = {
            prompt_tokens: result.usage.input_tokens || 0,
            completion_tokens: result.usage.output_tokens || 0,
            total_tokens:
              (result.usage.input_tokens || 0) + (result.usage.output_tokens || 0),
          };
        }
        res.write(`data: ${JSON.stringify(doneChunk)}\n\n`);

```

You will also need to update the type definition used by `createDoneChunk`:

1. Locate the type/interface that describes the chunk returned by `createDoneChunk` (for example, something like `SseChunk`, `StreamChunk`, or the explicit return type of `createDoneChunk`).
2. Extend it with an optional `usage` field, e.g.:

```ts
type Usage = {
  prompt_tokens: number;
  completion_tokens: number;
  total_tokens: number;
};

interface DoneChunk {
  // existing fields...
  usage?: Usage;
}
```

3. Ensure `createDoneChunk` is declared to return this updated type (or that the shared streaming chunk type also has `usage?: Usage` if `createDoneChunk` returns that).
4. If you have a discriminated union of chunk types, make sure the `done`/terminal variant supports `usage?: Usage` so the assignment in this handler is type-safe.
</issue_to_address>

### Comment 2
<location> `src/adapter/openai-to-cli.ts:15-46` </location>
<code_context>
   "claude-haiku-4": "haiku",
-  // With provider prefix
+  "claude-haiku-4-5": "haiku",
+  // With provider prefix (claude-code-cli/)
   "claude-code-cli/claude-opus-4": "opus",
+  "claude-code-cli/claude-opus-4-6": "opus",
   "claude-code-cli/claude-sonnet-4": "sonnet",
+  "claude-code-cli/claude-sonnet-4-5": "sonnet",
+  "claude-code-cli/claude-sonnet-4-6": "sonnet",
   "claude-code-cli/claude-haiku-4": "haiku",
-  // Aliases
+  "claude-code-cli/claude-haiku-4-5": "haiku",
+  // With provider prefix (claude-max/)
+  "claude-max/claude-opus-4": "opus",
+  "claude-max/claude-opus-4-6": "opus",
+  "claude-max/claude-sonnet-4": "sonnet",
+  "claude-max/claude-sonnet-4-5": "sonnet",
+  "claude-max/claude-sonnet-4-6": "sonnet",
+  "claude-max/claude-haiku-4": "haiku",
+  "claude-max/claude-haiku-4-5": "haiku",
+  // Bare aliases
   "opus": "opus",
</code_context>

<issue_to_address>
**suggestion:** There’s duplication between explicit provider-prefixed entries and the later prefix-stripping logic.

Because `extractModel` removes `claude-code-cli/` and `claude-max/` before consulting `MODEL_MAP`, the provider-prefixed keys appear redundant. You could keep only the bare IDs (e.g. `claude-opus-4`, `claude-opus-4-6`, etc.) and rely on the prefix-stripping to map all variants, simplifying this map without changing behavior.

```suggestion
const MODEL_MAP: Record<string, ClaudeModel> = {
  // Direct model names (provider prefixes like `claude-code-cli/` and `claude-max/`
  // are stripped by extractModel before consulting this map)
  "claude-opus-4": "opus",
  "claude-opus-4-6": "opus",
  "claude-sonnet-4": "sonnet",
  "claude-sonnet-4-5": "sonnet",
  "claude-sonnet-4-6": "sonnet",
  "claude-haiku-4": "haiku",
  "claude-haiku-4-5": "haiku",

  // Bare aliases
  "opus": "opus",
  "sonnet": "sonnet",
  "haiku": "haiku",
  "opus-max": "opus",
  "sonnet-max": "sonnet",
};
```
</issue_to_address>

---

Sourcery is free for open source - if you like our reviews please consider sharing them ✨

• X
• Mastodon
• LinkedIn
• Facebook

_{Help me be more useful! Please click 👍 or 👎 on each comment and I'll use the feedback to improve your reviews.}