| #789 |
Orchestrator capped at per-LEAF token budget (100k) + budget check on resume drops an already-approved action
run_goal (orchestrator) defaults token_budget to max_tokens_per_leaf (100k) when None -- but that's ...
|
closed |
high |
2026-06-30 23:04 |
- |
|
| #788 |
Agent loop can 'complete' at turn 0 with zero tool calls on a structured/constrained goal
A run_goal_deferred with a 'Step 1: delegate... do NOT research yourself; Step 2: do A/B/C' goal com...
|
closed |
high |
2026-06-30 00:48 |
- |
|
| #787 |
HITL gates reads same as side-effects -> approval spam on research (research->approve->act)
run_goal/_deferred with approval gates tools.http.request regardless of whether it's a READ (GET to ...
|
closed |
high |
2026-06-30 00:48 |
- |
|
| #786 |
Switch agent LLM route to ollama/glm-5:cloud (2x faster than qwen3.5:397b)
Benchmarked Ollama Cloud models for the [llm.routes] agent route (controlled: identical prompt, num_...
|
closed |
medium |
2026-06-30 00:08 |
- |
|
| #785 |
schedule_goal crons collide: all share job_name agent_sched_<def_id[:8]> (fixed SCHEDULED_AGENT_DSL)
Same root cause as #784 but for schedule_goal (recurring crons). job_name = 'agent_sched_'+def_id[:8...
|
closed |
high |
2026-06-29 23:19 |
- |
|
| #784 |
watch_gmail poll: stop+recreate leaves watch inactive (durable_cron idempotent on shared job_name)
All poll watches share job_name 'agent_sched_'+def_id[:8] where def_id is the FIXED SCHEDULED_AGENT_...
|
closed |
high |
2026-06-29 22:53 |
- |
|
| #783 |
Gmail auto-responder dedup is non-deterministic (LLM-judged, sender-per-day) -> 26-min delay + skipped/duplicate replies
Watch auto-responder (_watch_goal in highway-mcp/server.py) enforces reply-idempotency via an LLM IN...
|
closed |
high |
2026-06-29 20:19 |
- |
|
| #782 |
Agent eval/regression harness: trace-based assertions (routing, over-fetch, delivery, cost) + selftest
Built the regression harness (highway-mcp/eval_agent.py) the strategy called for. Upgrades the prior...
|
closed |
high |
2026-06-29 18:30 |
- |
|
| #781 |
Wire up web-search provider: self-hosted SearXNG backend for web_search (was keyless DDG, blocked on datacenter IP)
web_search used keyless DuckDuckGo HTML which throttles datacenter IPs -> empty results -> thin leaf...
|
closed |
high |
2026-06-29 03:04 |
- |
|
| #780 |
Unified agent over-researches after a delegate fan-out (redundant direct fetches = extra approvals)
After a successful delegate fan-out, the orchestrator sometimes does a 2nd pass of direct http fetch...
|
closed |
medium |
2026-06-29 02:35 |
- |
|
| #779 |
Unify research_goal + run_goal into a single MCP goal tool (customer no longer picks research vs act)
Bad UX: customers had to choose research_goal (delegate-only, no outbound) vs run_goal (act+email, n...
|
closed |
high |
2026-06-29 02:35 |
- |
|
| #778 |
Demo/research tenant Gmail token stale -> send fails ('Gmail token refresh failed'); offer SMTP fallback for agent email
run_goal research+email path works end-to-end (HITL approval + dispatch) but apps.platform.gmail.sen...
|
open |
low |
2026-06-29 01:53 |
- |
|
| #777 |
Dashboard UX: agent_leaf child runs show as 'leaf' and clutter the run list (should show run name + nest under parent)
Research campaigns spawn N child agent_leaf runs. Backend data is correct: each child workflow_run.w...
|
open |
medium |
2026-06-29 01:53 |
- |
|
| #772 |
Live activity worker emits completion event without tenant_id (durable-leaf child runs hang for non-default tenants)
hwe activity-worker uses enterprise/services/activity_worker.py. Its _complete_activity/_fail_activi...
|
closed |
high |
2026-06-29 00:09 |
- |
|
| #771 |
Agentic durable leaves crash on activity worker: ActivityContext has no jumper_client
durable_leaves=true routes tools.agent.execute_leaf_agent to the activity worker (agent_leaf DSL, 60...
|
closed |
critical |
2026-06-29 00:04 |
- |
|