[Bug] OpenClaw runtime: daemon invokes openclaw with Claude Code flags, not openclaw agent protocol

[sh1baaaaa]

Summary

Multica daemon (v0.4.4) invokes OpenClaw using Claude Code CLI flags instead of the openclaw agent protocol. Every task fails with "openclaw returned no parseable output".

Environment

• Multica: v0.4.4 (self-hosted)
• OpenClaw: 2026.5.6 (also tested 2026.5.5, 2026.6.6, 2026.6.11 - same result)
• OS: Ubuntu 24.04 (Linux amd64)

What Multica actually sends to openclaw

Captured via wrapper script logging argv:

openclaw agent -p "hello" --output-format stream-json --model zai/glm-5.1 --workdir /root/multica_workspaces/.../workdir --dangerously-bypass-approvals-and-sandbox --add-dir /root/multica_workspaces/.../.repos/... --strict-mcp-config --mcp-config /tmp/multica-mcp-config-....json --json-input /tmp/openclaw-input-....json --no-tui

What openclaw returns

openclaw 5.6 does not recognize these flags. It prints usage/help text to stdout and exits non-zero. Multica reports: openclaw returned no parseable output

What openclaw agent actually accepts

Usage: openclaw agent [options]
Options:
  -m, --message <text>   Message body for the agent
  --model <id>           Model override
  --json                 Output result as JSON (single object, not NDJSON stream)
  --agent <id>           Agent id
  --session-id <id>      Use an explicit session id
  --local                Run embedded agent locally
  --thinking <level>     Thinking level
  --timeout <seconds>    Timeout override
  --channel <channel>    Delivery channel
  --deliver              Send reply back to channel
  --verbose <on|off>     Verbose level

None of the flags Multica sends exist in any openclaw version tested:

• -p "text" -> openclaw uses -m/--message
• --output-format stream-json -> openclaw uses --json (returns single JSON object, not NDJSON stream)
• --model X -> same in both (only matching flag)
• --json-input FILE -> no equivalent, openclaw uses --session-id for session state
• --workdir DIR -> no equivalent
• --add-dir DIR -> no equivalent
• --mcp-config FILE -> no equivalent, MCP configured per-agent in openclaw.json
• --no-tui -> no equivalent
• --dangerously-bypass-approvals-and-sandbox -> no equivalent
• --strict-mcp-config -> no equivalent

Expected behavior

Multica should invoke openclaw using native openclaw flags, not Claude Code CLI flags. Per server/pkg/agent/openclaw.go and the docs, the expected protocol is openclaw agent with openclaw-native flags and NDJSON from stdout.

Workaround

None that fully works. Partial workaround via wrapper script:

1. Intercept openclaw config file (returns ~/.openclaw/openclaw.json with tilde — daemon cannot resolve it) and return absolute path
2. After fixing config file issue, the Claude Code flags issue surfaces with no clean translation

Additional context

• Tested all openclaw versions from 2026.5.5 to 2026.7.1 - none support these flags
• server/pkg/agent/testdata/openclaw-2026.5.5-stdout.json exists in repo - what protocol does that test expect?
• Hermes and Claude Code runtimes work fine on same setup
• Related: [Bug]: OpenClaw runtime agent always fails: openclaw config file exit status 1 (Windows) #3853 (similar config file issue on Windows, still open)

[vicksiyi]

Thanks for the report. I opened PR #5693 with a fix: #5693

Root cause: OpenClaw must be launched through its native openclaw agent --json --message protocol, but copied Claude/Cursor protocol flags could still be forwarded through OpenClaw custom args and make the CLI print usage instead of JSON.
Fix: The OpenClaw backend now filters those incompatible flags and has a fake-OpenClaw regression test that captures the actual argv from Execute.
Validation: cd server && go test ./pkg/agent -run Openclaw -count=1; cd server && go test ./pkg/agent -count=1; make check-worktree exited 0 locally, with Docker unavailable while checking PostgreSQL.

I will keep an eye on CI and review feedback.