Route ALL OpenClaw LLM calls through AgentWeave proxy, not just delegated subagent ones

[arniesaha]

Problem

In Nexus's lakehouse, AgentWeave traces only capture ~6% of OpenClaw (Nix) sessions. The other 94% are visible via OpenClaw's own JSONL shipper (lakehouse.agent_events) but never reach Tempo.

Numbers (from nexus-context-engine-26.lakehouse, time window Apr 14-26)

Source	Distinct Nix sessions	Total rows
OpenClaw native shipper → agent_events	63	1,945
AgentWeave proxy → spans	4	1,732

Same time window. The AgentWeave side has comparable row counts (each Nix LLM call → one span) but only 4 distinct session.id values, vs 63 distinct session UUIDs the native OpenClaw JSONL shipper recorded.

Furthermore, OpenClaw native data goes back to March 3, 2026 (54 days). AgentWeave traces for any agent only start April 14, 2026 — that's roughly when AgentWeave was introduced, but even so the bridge plugin appears to only intercept LLM calls when Nix is delegated to as a subagent from another agent (Claude Code's session). Direct user-driven Nix sessions (e.g., openclaw invoked from the CLI) bypass the proxy.

Hypothesis

Looking at openclaw-agentweave-bridge plugin: it likely hooks into the subagent-spawn / delegated-invocation code path but not the top-level LLM-call path. So:

• claude-code → invokes Nix as subagent → bridge active → AgentWeave records → ✅
• User runs openclaw "do something" directly → bridge inactive → straight to LLM → ❌

Asks

1. Audit the bridge plugin's hook points and ensure ALL outbound LLM calls from OpenClaw flow through the AgentWeave proxy, not just the delegated-subagent path.
2. Set prov.session.id to OpenClaw's local session UUID (the same value that ends up in the JSONL) so AgentWeave spans can be joined back to the native shipper's sessions table by id.
3. Bonus: emit prov.repository / prov.cwd so we can answer "which repo was this session on" from AgentWeave alone (today this only exists in OpenClaw's raw_data.cwd).

Verification

Once fixed, this query against Tempo should approach parity with the native shipper for any 24h window:

# count distinct Nix sessions in last 24h via Tempo
curl ".../api/search?q=%7B%20prov.agent.id%20%3D%20%22nix-v1%22%20%7D&limit=1000&start=$START&end=$END" \
  | jq '[.traces[] | .traceID] | length'

Linked

• arniesaha/nexus#22 → arniesaha/nexus#23 (already promoted prov.wasAssociatedWith etc.)
• Emit prov.session.id, prov.parent.session.id, prov.project, prov.task.label for non-Claude agents #176 (this issue is the next layer of the same instrumentation theme)

[arniesaha]

Quick audit update after digging through the current repo state and trying to land a fix here.

What is already implemented in AgentWeave

There has already been meaningful progress on the AgentWeave side since this issue was opened:

• d0015c5 wires OpenClaw bridge /session pushes with session_id, task_label, parent_session_id, etc.
• b368426 improves session_id propagation and parent-span linking.
• 392464c / PR fix(proxy): isolate session context by session_key and task-label header #185 moves the bridge + proxy onto the concurrent-safe session_key path and adds a request-level path for prov.task.label.

Concretely, the bridge now POSTs /session with session metadata, and the proxy now supports per-session forced context keyed by session_key.

What I found trying to fix #180 in this repo

I tried to turn this into an AgentWeave-only code fix twice. After isolating away some local worktree/setup noise, I still did not find an honest additional AgentWeave patch that would plausibly move coverage from "only delegated/subagent flows" to "all OpenClaw LLM calls".

The important distinction is:

• AgentWeave can only observe calls that actually reach the bridge/proxy path.
• If direct top-level OpenClaw turns never invoke the bridge hook (or never get their provider base URLs rewritten to the proxy), there is nothing left inside AgentWeave itself to capture.

Current hypothesis

The remaining gap now looks primarily OpenClaw-side, not AgentWeave-side:

• the bridge/plugin instrumentation appears to cover delegated / bridge-aware flows
• but direct top-level OpenClaw LLM calls still appear able to bypass that path entirely

That would explain why the repo already has the session/context plumbing, but session coverage still does not approach parity with OpenClaw's native JSONL shipper.

Suggested next step

I think this issue should either:

1. be re-scoped as an OpenClaw integration gap (where the top-level provider call path is not consistently routed through the AgentWeave bridge/proxy), or
2. be split into:
   • AgentWeave: done/mostly done for session propagation + per-session attribution plumbing
   • OpenClaw: ensure all top-level outbound LLM calls actually pass through the bridge/proxy path

Happy to keep digging, but I no longer think the missing 94% is fixable by proxy-only changes in this repo.

[arniesaha]

Closing in favor of three more focused issues. The hypothesis in the original report ("the bridge appears to only intercept LLM calls when Nix is delegated to as a subagent") doesn't hold up against the bridge source — plugins/openclaw-agentweave-bridge/src/service.ts subscribes to message.queued for every sessionKey and starts a root span regardless of main vs. subagent. The agent-type detection there only chooses span attributes; it doesn't gate span creation.

The empirical 4-of-63 gap is real, but the cause is operational, not a missing hook. The NAS install (~/.openclaw/openclaw.json) has the bridge enabled and pointing at our OTLP/proxy — sessions originating there are observed. The 59 missing sessions originate on OpenClaw installs that never had the plugin registered.

Splitting into:

• Distribute agentweave-bridge plugin to all OpenClaw installs in the fleet #186 — distribute the bridge plugin to all OpenClaw installs in the fleet
• Bridge: set prov.session.id to OpenClaw session UUID, not sessionKey #187 — set `prov.session.id` to OpenClaw's session UUID (not `sessionKey`) so AgentWeave spans can join back to `lakehouse.agent_events`
• Bridge: emit prov.repository and prov.cwd on root spans #188 — emit `prov.repository` / `prov.cwd` on root spans