openclaw - 💡(How to fix) Fix [Bug]: agentRuntime=claude-cli silently ignored on cold-start 2026.4.26 install — dispatcher routes to openrouter/gpt-oss-120b:free with no failover log [1 comments, 2 participants]

Bug type

Behavior bug (incorrect output/state without crash) — likely regression vs. 2026.4.24

Beta release blocker

No

Summary

On a cold-start install of OpenClaw 2026.4.26, configuring agents.defaults.agentRuntime.id = "claude-cli" plus agents.defaults.cliBackends.claude-cli and an auth.profiles["anthropic:claude-cli"] OAuth profile is not honored by the dispatcher. The agent silently routes to openrouter/openai/gpt-oss-120b:free (last-resort fallback) without attempting agents.defaults.model.primary (anthropic/claude-opus-4-7) and without emitting any [model-fallback/decision] or [agent/cli-backend] events. A side-by-side host on 2026.4.24 with the identical config routes correctly through claude-cli.

Steps to reproduce

1. Cold-start install: npm install -g [email protected] on a clean host (no prior ~/.openclaw/agents/main/agent/auth-profiles.json).
2. Install Claude Code, run claude /login, complete OAuth. Verify claude --print "hi" returns a response (~5s).
3. Edit ~/.openclaw/openclaw.json:

   {
     "agents.defaults.agentRuntime": { "id": "claude-cli" },
     "agents.defaults.cliBackends": {
       "claude-cli": {
         "command": "claude",
         "output": "jsonl",
         "jsonlDialect": "claude-stream-json"
       }
     },
     "agents.defaults.model.primary": "anthropic/claude-opus-4-7",
     "auth.profiles": {
       "anthropic:claude-cli": { "provider": "claude-cli", "mode": "oauth" }
     },
     "auth.order": { "anthropic": ["anthropic:claude-cli"] }
   }

4. kill -SIGUSR1 <gateway-pid> (or restart gateway).
5. Send any Telegram message to the configured agent.

Expected behavior

Trajectory shows provider=claude-cli; logs include [agent/cli-backend] cli exec: provider=claude-cli model=.... Confirmed working on a side-by-side host running 2026.4.24 with the identical openclaw.json.

Actual behavior

Trajectory shows session.started directly with provider=openrouter, modelId=openai/gpt-oss-120b:free. No [model-fallback/decision] events. No [agent/cli-backend] events. The primary model anthropic/claude-opus-4-7 is never attempted. Setting agents.defaults.model.fallbacks = [] does not change behavior — dispatcher still chooses openrouter/gpt-oss-120b:free directly.

OpenClaw version

2026.4.26 (entrypoint /home/linuxbrew/.linuxbrew/lib/node_modules/openclaw/dist/index.js)

Operating system

Linux 6.17.0-22-generic (Ubuntu-derivative)

Install method

npm install -g openclaw under linuxbrew node

Model

Configured primary: anthropic/claude-opus-4-7 via claude-cli runtime Effective (observed): openrouter/openai/gpt-oss-120b:free

Provider / routing chain

Configured: openclaw -> claude-cli (subprocess) -> Anthropic OAuth Observed: openclaw -> openrouter -> openai/gpt-oss-120b:free (skipped primary entirely)

Additional provider/model setup details

• Plugin anthropic is loaded (boot log: 3 plugins: anthropic, browser, telegram).
• claude binary at /home/linuxbrew/.linuxbrew/bin/claude (v2.1.119), authenticated standalone via ~/.claude/.credentials.json. claude --print "hi" returns in ~4.8s.
• Per-agent state divergence: On the broken .26 host, ~/.openclaw/agents/main/agent/auth-profiles.json does NOT contain the anthropic:claude-cli profile after cold start (despite global config defining it). On the working .24 host (silver, nvm install), the same path contains a fully-populated profile with OAuth tokens.
• Manually injecting {provider:"claude-cli", mode:"oauth"} into the per-agent auth-profiles.json and reloading does NOT fix the routing — the dispatcher still picks gpt-oss-120b:free.
• No imperative bootstrap command exists for this state. openclaw migrate-claude only imports Claude Desktop settings, not runtime CLI auth.

Logs, screenshots, and evidence

# Working host (silver, 2026.4.24):
[reload] config change detected — agentRuntime=claude-cli registered
[agent/cli-backend] cli exec: provider=claude-cli model=claude-opus-4-7 promptChars=...

# Broken host (carbon, 2026.4.26) — same config, same Telegram input:
session.started provider=openrouter modelId=openai/gpt-oss-120b:free
(no [reload], no [model-fallback/decision], no [agent/cli-backend])

Impact and severity

• Affected: Operators upgrading to 2026.4.26 who configure claude-cli runtime via documented config keys. Likely 100% of cold-start installs on .26 (reproduced 1/1 on this host; 0/1 on .24 side-by-side).
• Severity: High — silently routes paid/private workloads to a free fallback model, defeating user's explicit provider choice and OAuth subscription. No error surfaces to operator.
• Frequency: Always (every turn, every channel).
• Consequence: Wrong model serving responses; subscription unused; no diagnostic trail to recover from.

Additional information

• Last known good: 2026.4.24 (verified working with identical config on a different host).
• First known bad: 2026.4.26 (this host).
• Adjacent / possibly related:
  • #70279 — same silent-skip symptom, different root cause (systemd-root)
  • #73095 — cliBackends-only installs hit uninitialized registries on cold start
  • #73682 / #71026 — auth-profiles.json not populated for anthropic:claude-cli at runtime (closed; symptom may have regressed in .26)
  • #57326 — helper paths still bypass runCliAgent()
  • #72576 — no integration test catches the registration regression class
• Suggested mitigations (in order of cost):
  1. Surface failover decisions in the trajectory whenever the configured agentRuntime cannot be resolved at dispatch time (currently silent — most actionable fix).
  2. Document the canonical cold-start setup flow for claude-cli as agentRuntime on .26+.
  3. Add an explicit openclaw setup-claude-cli command (or extend migrate-claude) that hydrates per-agent auth-profiles.json from ~/.claude/.credentials.json.