Opus 4.6 persistent "overloaded_error" on Max OAuth — Claude Code works fine on same machine

[ripe escarp]

Environment:
▎ - OpenClaw 2026.3.28 (f9b1079)
▎ - Auth: Max OAuth session (not API key)
▎ - Model: anthropic/claude-opus-4-6
▎ - Windows 11

▎ What I've tried:
▎ - Full restart (not just service restart)
▎ - openclaw doctor --fix — no auth issues found
▎ - Updated from 2026.3.24 to 2026.3.28
▎ - Cooldown cleared, retried — same error

▎ Error log:
▎ [agent/embedded] embedded run agent end: isError=true model=claude-opus-4-6 provider=anthropic error=The AI service
is temporarily overloaded.
▎ rawError={"type":"error","error":{"type":"overloaded_error","message":"Overloaded"}}

▎ Claude Code on the same machine, same Max subscription, same model — no issues at all. Seems like something in how
OpenClaw sends the Max OAuth token to Anthropic's API.

[remote nymph]

Yep — this usually means OpenClaw is hitting a different Anthropic lane/policy than Claude Code, so one can overload while the other keeps working.

First, can you paste this (version first):

pnpm openclaw --version
pnpm openclaw status
pnpm openclaw models status --json

Then quick high-signal test in the same thread/session:

1. /fast on
2. /new
3. send a tiny prompt like “ping”

If that works, it was likely service_tier/queue behavior (fast off maps to standard_only, which is more likely to choke during load).

If it still fails, paste:

pnpm openclaw config get agents.defaults.models
pnpm openclaw logs --follow

…and include one fresh failing request line (with request_id).
I can tell you exactly whether it’s overload-only, auth profile cooldown, or config (e.g. context1m/service tier override).

Docs:

• https://docs.openclaw.ai/providers/anthropic
• https://docs.openclaw.ai/concepts/model-failover

Super useful repro details — thank you.
Given this combo (Claude Code works, OpenClaw anthropic/... returns overloaded_error), this is very likely not your Windows host and more likely Anthropic lane/tier behavior on the OpenClaw API path.

Try this exact order:

# 1) Force fast tier in the affected chat session
/fast on
/new
/status

/status should show Fast as on.

Then refresh Anthropic subscription auth using the recommended token path:

claude setup-token
pnpm openclaw models auth paste-token --provider anthropic
pnpm openclaw gateway restart

If you want to match Claude Code transport more closely as a workaround, switch OpenClaw to Claude CLI backend:

pnpm openclaw models auth login --provider anthropic --method cli --set-default

⚠️ caveat: claude-cli/... backend is text-in/text-out (OpenClaw tool use is limited there).

If it still reproduces after the above, paste:

pnpm openclaw models status --json
pnpm openclaw config get agents.defaults.models

plus one fresh error block from pnpm openclaw logs --follow.

Docs:

• https://docs.openclaw.ai/providers/anthropic
• https://docs.openclaw.ai/concepts/oauth
• https://docs.openclaw.ai/concepts/model-failover

[ripe escarp]

11.12.0

● +----------------------+----------------------------------------------------------------------------------------------
-+
| Dashboard | http://127.0.0.1:18789/
|
| OS | windows 10.0.26200 (x64) · node 24.13.1
|
| Tailscale | off
|
| Channel | stable (default)
|
| Update | pnpm · up to date · npm latest 2026.3.28
|
| Gateway | local · ws://127.0.0.1:18789 (local loopback) · reachable 28ms · auth token · RSX (192.168.x.
|
| | x) app 2026.3.28 windows 10.0.26200
|
| Gateway service | Scheduled Task installed · registered · unknown (state Bereit, Task status is locale-
|
| | dependent and no numeric Last Run Result was available.)
|
| Node service | Scheduled Task not installed
|
| Agents | 2 · no bootstrap files · sessions 29 · default main active 7m ago
|
| Memory | 31 files · 81 chunks · sources memory · plugin memory-core · vector ready · fts ready ·
|
| | cache on (508)
|
| Plugin compatibility | none
|
| Probes | skipped (use --deep)
|
| Events | none
|
| Heartbeat | 1h (main), disabled (agent2)
|
| Sessions | 29 active · default claude-opus-4-6 (200k ctx) · 2 stores
|
+----------------------+

---

-+

Sessions
+--------------------------------------------+--------+---------+-----------------+-----------------------------------
-+
| Key | Kind | Age | Model | Tokens
|
+--------------------------------------------+--------+---------+-----------------+-----------------------------------
-+
| agent:main:main | direct | 7m ago | claude-opus-4-6 | unknown/1000k (?%) · 🗄️ 98% cached
|
| agent:main:whatsapp:group:xxxxx… | group | 1h ago | claude-opus-4-6 | unknown/200k (?%)
|
| agent:agent2:cron:xxxxx… | direct | 1h ago | claude-opus-4-6 | 22k/1000k (2%) · 🗄️ 485% cached
|
| agent:agent2:cron:xxxxx… | direct | 1h ago | claude-opus-4-6 | 22k/1000k (2%) · 🗄️ 485% cached
|
| agent:agent2:main | direct | 1h ago | claude-opus-4-6 | 37k/1000k (4%) · 🗄️ 99% cached
|
| agent:agent2:cron:xxxxx… | direct | 1h ago | claude-opus-4-6 | 23k/1000k (2%) · 🗄️ 469% cached
|
| agent:agent2:cron:xxxxx… | direct | 1h ago | claude-opus-4-6 | 23k/1000k (2%) · 🗄️ 469% cached
|
| agent:agent2:cron:xxxxx… | direct | 1h ago | claude-opus-4-6 | 18k/1000k (2%) · 🗄️ 485% cached
|
| agent:agent2:cron:xxxxx… | direct | 13h ago | claude-opus-4-6 | 18k/1000k (2%) · 🗄️ 392% cached
|
| agent:agent2:cron:xxxxx… | direct | 14h ago | claude-opus-4-6 | 24k/1000k (2%) · 🗄️ 578% cached
|
+--------------------------------------------+--------+---------+-----------------+-----------------------------------
-+

🦞 OpenClaw 2026.3.28 (f9b1079)

Config : ~.openclaw\openclaw.json
Default : anthropic/claude-opus-4-6
Fallbacks (0) : -
Aliases (1) : opus -> anthropic/claude-opus-4-6
Configured models (1): anthropic/claude-opus-4-6

Auth overview
Providers w/ OAuth/tokens (2): anthropic (3), openai-codex (1)

• anthropic: 3 profiles (all token-based, static)
• openai-codex: 1 profile (OAuth, ok, expires in 3d)

OAuth/token status

• anthropic
  • anthropic:profile1 static
  • anthropic:profile2 static
  • anthropic:profile3 static
• openai-codex usage: 5h 100% left
  • openai-codex:default ok expires in 3d

[remote nymph]

Got it — 11.12.0 looks like your Claude CLI version (good sign).

Let’s do the quickest bypass (uses the same Claude CLI path that already works for you):

pnpm openclaw models auth login --provider anthropic --method cli --set-default
pnpm openclaw gateway restart

Then test Opus again in this thread.

If you prefer staying on anthropic/... API path instead, do token refresh:

claude setup-token
pnpm openclaw models auth paste-token --provider anthropic

pnpm openclaw gateway restart

Docs: https://docs.openclaw.ai/providers/anthropic#option-b-claude-cli-as-the-message-provider

Nice — this status output is clean (gateway/service/model plumbing looks healthy), so this is almost certainly Anthropic OAuth API lane behavior, not your local setup.

Let’s do one decisive check/fix:

pnpm openclaw config get agents.defaults.model.primary

If that still shows anthropic/claude-opus-4-6, switch to Claude CLI backend (same path that currently works for you in Claude Code):

pnpm openclaw models auth login --provider anthropic --method cli --set-default
pnpm openclaw gateway restart
pnpm openclaw config get agents.defaults.model.primary

You should now see claude-cli/claude-opus-4-6.

Then retry in chat.
If it still fails after primary is claude-cli/..., paste:

pnpm openclaw models status --json

At that point it’s likely a real regression and worth opening an issue with that JSON + one failing log block:
https://github.com/openclaw/openclaw/issues

Docs: https://docs.openclaw.ai/providers/anthropic#option-b-claude-cli-as-the-message-provider

This is the first real clue 👀
Your main session is at 1000k context (claude-opus-4-6 | unknown/1000k), while others are 200k.

That strongly suggests context1m is enabled for the failing session/model, and that combo on Max OAuth can absolutely behave worse (including overloads) than Claude Code defaults.

Try this:

1. In the affected chat:
   /new
2. Then:
   /status
   (check if context goes back to ~200k)
3. If it still shows 1000k, disable context1m for Opus in config (agents.defaults.models["anthropic/claude-opus-4-6"].params.context1m = false) and restart gateway.

If you want immediate stability without chasing Anthropic API behavior, use Claude CLI backend:

pnpm openclaw models auth login --provider anthropic --method cli --set-default
pnpm openclaw gateway restart

Refs:

• https://docs.openclaw.ai/gateway/troubleshooting#anthropic-429-extra-usage-required-for-long-context
• https://docs.openclaw.ai/providers/anthropic#1m-context-window-anthropic-beta

[ripe escarp]

⚠️ Agent failed before reply: Unknown model: claude-cli/claude-opus-4-6.
Logs: openclaw logs --follow

● ▎ Context1M is now disabled (agents.defaults.models.anthropic/claude-opus-4-6.params.context1m = false), model is back
on anthropic/claude-opus-4-6. Started a /new session.

▎ Still getting intermittent overloaded errors, but some requests are going through now — tool calls are firing
between failures:
▎ 11:21:31 [agent/embedded] embedded run agent end: isError=true model=claude-opus-4-6 error=The AI service is
temporarily overloaded.
▎ 11:21:35 [agent/embedded] read tool called without path: toolCallId=toolu_01VbnXhZWJjfXNj9WYTXeFhd
▎ 11:21:35 [agent/embedded] read tool called without path: toolCallId=toolu_01VrzD3unuuts28mFa49tXvd

▎ So it's partially working now. Seems like Anthropic Opus capacity is flapping. Claude Code (same machine, same Max
sub) still works fine in parallel.

[remote nymph]

Great debugging — this narrows it a lot.

You just confirmed:

• context1m was a major amplifier (disabling it improved behavior)
• remaining failures are likely capacity flaps on Anthropic token lane
• read tool called without path is a secondary malformed tool-call symptom that can happen when replies degrade under overload

What I’d do now for stability:

1. Keep /fast on in the active session (or set fast mode default in config).
2. Pin one Anthropic profile (avoid rotating across 3 during flaky periods):

pnpm openclaw models auth order set --provider anthropic anthropic:<best-profile>

3. Add a fallback model so runs continue when Opus overloads (instead of hard-failing).
4. If you need consistent tool reliability, use Anthropic API key auth (subscription token lanes are more variable).