openclaw - 💡(How to fix) Fix [Bug]: slug-generator HTTP 400 misclassified as profile-wide billing failure (5h cooldown), kills all agents on profile [1 participants]

Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

A transient HTTP 400 returned to the internal generateSlugViaLLM helper is misclassified as a profile-wide billing failure: OpenClaw marks the entire anthropic:claude-cli auth profile as disabled:billing with a 5-hour cooldown. Every other agent that uses the same profile then silently fails with Provider claude-cli has billing issue (skipping all models) despite the underlying subscription being healthy.

Steps to reproduce

1. Configure anthropic:claude-cli as the primary auth profile (Max subscription, healthy).
2. Allow OpenClaw default hooks.internal.entries.session-memory.llmSlug: true to remain on.
3. Trigger any session that ends up in the slug-generator path; let the underlying claude invocation return any 400 (we observed it as LLM request rejected: You're out of extra usage).
4. Observe ~/.openclaw/agents/main/agent/auth-state.json — usageStats["anthropic:claude-cli"] gains disabledReason: "billing", disabledUntil set 5 hours forward.
5. Send any prompt to a Cognitor agent on the same profile (Telegram, CLI, anything). It immediately fails with the cooldown message — no upstream Anthropic call is made.

Expected behavior

A failure inside the slug-generation lane (a low-priority, internal helper) should be lane-local:

• It should not flip the auth profile to disabled:billing for any non-billing-coded HTTP error.
• The classifier should require a real billing signal (HTTP 402, explicit billing/quota body) before marking the profile down — same direction as #56053 (the 402 fallback fix). HTTP 400 with arbitrary text is not a billing signal.
• At minimum, an internal-helper failure should be scoped to the helper, not propagate to the profile-wide failover state used by user-facing agents.

Actual behavior

• Real incident 2026-04-14 22:19–22:37 UTC.
• Lane log: lane task error: lane=session:temp:slug-generator ... error="FailoverError: LLM request rejected: You're out of extra usage..."
• Auth-state log: event=auth_profile_failure_state_updated provider=claude-cli reason=billing disabledReason=billing disabledUntil=2026-04-15T03:19:..Z
• Subsequent traffic: reason=billing errorHash=sha256:52c819aaad70 "Provider claude-cli has billing issue (skipping all models)" — same errorHash on every request for the next 5 hours.
• Direct claude -p --model claude-opus-4-6 "ping" returned a normal response throughout the cooldown — confirming Anthropic itself is fine.
• User-visible symptom: every Cognitor agent (Telegram bot, CLI sessions) returned ⚠️ Something went wrong… with no diagnostic for the operator.

OpenClaw version

2026.4.23 (incident reproduced on the version current on 2026-04-14; behavior path still present in 2026.4.23 — the workaround flag is the only thing keeping it dormant)

Operating system

Ubuntu 24.04.4 LTS (kernel 6.8.0-107-generic)

Install method

npm global (/home/ubuntu/.npm-global/lib/node_modules/openclaw)

Model

claude-cli/claude-opus-4-7 (primary)

Provider / routing chain

openclaw -> auth-profile anthropic:claude-cli -> claude (CLI) -> anthropic.com

Additional provider/model setup details

Hot-fix workaround applied locally and recommended for any user with this profile shape:

"hooks": {
  "internal": {
    "entries": {
      "session-memory": {
        "enabled": true,
        "llmSlug": false
      }
    }
  }
}

This disables the slug-generator entirely (sessions then named by UUID/timestamp), removing the offending lane. Hot-reload picks it up: [reload] config hot reload applied (hooks.internal.entries.session-memory.llmSlug). Functional, but leaves the default behaviour exposed — every new install with llmSlug: true (the default) is one transient internal 400 away from a 5-hour silent outage.

Manual recovery once the cooldown is set, until OpenClaw classifies this correctly:

python3 <<'PY'
import json
p = '/home/ubuntu/.openclaw/agents/main/agent/auth-state.json'
d = json.load(open(p))
cli = d['usageStats']['anthropic:claude-cli']
for k in ('disabledUntil','disabledReason','failureCounts','lastFailureAt'):
    cli.pop(k, None)
cli['errorCount'] = 0
json.dump(d, open(p,'w'), indent=2)
PY
systemctl --user restart openclaw-gateway.service

Suggested direction

Mirror what #56053 / PR #56069 did for HTTP 402: tighten the failover classifier so that only billing-coded HTTP responses (402, or 4xx with an explicit billing/quota body marker) flip a profile to disabled:billing. For internal helper lanes (session:temp:slug-generator and similar), additionally scope the classifier so a helper-lane failure cannot escalate to a profile-wide cooldown — at most disable the helper.

Related (same family of misclassification): #56053, #14272, #25191, #33962.

Reported by @nikolaykazakovvs-ux via Cognitor (claude-opus-4-7 substrate).