BUG: tracker dispatch mode fails before any worker HTTP — claims task, sets up workspace, never contacts tagged workers, retries exhaust silently #53

Environment

• nullboiler v2026.5.29 (released binary), nulltickets v2026.3.2, nullhub v2026.5.29
• Single alpine container, one boiler process, tracker pull mode
• Workers (config.json workers array, per multi-bot-integration.md):
  • nullclaw-1 → http://127.0.0.1:3000/webhook, protocol webhook, tags ["coder"]
  • zeroclaw-1 → http://zeroclaw-test:42617/webhook, protocol webhook, tags ["reviewer"] (ZeroClaw v0.8.5, reachable — direct POST returns a valid LLM response)
• tracker: { "url": "http://127.0.0.1:7700", "agent_id": "nullboiler", "poll_interval_ms": 5000 }

Setup

• Pipeline (nulltickets POST /pipelines): FSM todo → done, states.todo.agent_role = "reviewer", states.done.terminal = true
• Workflow file in workflows/ (loaded and listed by GET /workflows):

{
  "id": "hetero-demo",
  "pipeline_id": "<uuid of pipeline above>",
  "claim_roles": ["reviewer"],
  "execution": "dispatch",
  "dispatch": { "worker_tags": ["reviewer"], "protocol": "webhook" },
  "prompt_template": "Task: {{task.title}}\nDescription: {{task.description}}",
  "on_success": { "transition_to": "done" },
  "on_failure": { "transition_to": "failed" }
}

Observed

1. GET /tracker/status → {"mode":"pull", ...} — tracker connected, polling every 5s
2. Tasks created in the pipeline are claimed: latest_run.status = "running", agent_id = "nullboiler", agent_role = "reviewer" ✓
3. /tracker/status shows the run in "workspace_setup", then "failed" (current_turn: 1, max_turns: 20), failed_count increments on each retry cycle
4. The tagged worker's gateway logs show zero inbound requests — POST /webhook count unchanged across the entire experiment; tcpdump-level confirmation unnecessary (worker-side request log is quiet)
5. Task stage never transitions (stays todo); on_failure.transition_to ("failed") never applied — possibly the same root cause as fix(tracker): apply the workflow's on_failure transition when a dispatch fails #52
6. Cycle repeats until retries exhaust; no log line names the dispatch failure reason

Direct evidence the failure is boiler-side

• POST http://127.0.0.1:7700/leases/claim with {"agent_id":"nullboiler","agent_role":"reviewer","lease_ttl_ms":60000} returns the task immediately — the tracker's claim-by-role works
• Direct POST to the ZeroClaw worker's /webhook returns a valid model response — the worker works
• The failure is strictly inside nullboiler between claim and worker HTTP: no request is ever emitted

Expected

• execution: "dispatch" with dispatch.worker_tags matching a configured worker's tags should select the worker, POST the rendered prompt_template, and apply on_success/on_failure transitions — or at minimum log a named, actionable failure (e.g. "no eligible worker for tags [...]") instead of failing silently.

Related

• BUG: run with no eligible worker waits silently forever (no fail-fast on permanent dead-end) #44 (no eligible worker waits silently forever — same silent-failure family)
• fix(tracker): apply the workflow's on_failure transition when a dispatch fails #52 (on_failure transition not applied — observed here too)
• BUG: /tracker/* endpoints block for the duration of tick() network I/O #41/fix(tracker): serve /tracker/* from tick-independent snapshot #50 (/tracker/* snapshot staleness made diagnosis harder)

Note on ZeroClaw A2A

ZeroClaw v0.8.5 serves A2A (agent cards at /.well-known/agent-card.json, per-alias /a2a/{alias} JSON-RPC — confirmed from compiled binary and git history), so once dispatch-to-worker works, an A2A dispatch protocol option would let boiler drive ZeroClaw's native task lifecycle, not just its webhook. Happy to test.