You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Schedule the kagent lane daily within the Claude API credit and prove a canary run #4869
🤖 Generated by the Agentic Engineer (interactive Claude Code session)
Part of #4866. Needs the runtime, identity, image and plugin sub-issues first.
Goal
Schedule the lane once a day within the credit and prove one complete run.
Acceptance criteria
A paused kagent scheduled run starts the lane's agent once a day with the same thin prompt the host lanes use, and each run shows in the kagent UI. A Flux-applied Job writes the schedule from the platform repository (feat(agentic-engineer): schedule the daily run with kagent's scheduled runs platform#5999). It is unpaused only after every other sub-issue is done.
A run has no time limit (maintainer decision 2026-11-09, replacing the earlier cap of about $6 a run), and the model is pinned to claude-opus-5-5 with no fallback. The maintainer confirms that the API workspace spend limit equals the credit; it is the lane's only spend cap.
The lane's native memory persists across runs in its own store.
One canary run completes end to end: pre-flight, survey, one piece of work, a signed draft PR under the App identity, and a run report. Its measured cost is recorded against the monthly limit.
The AGENTS.md cadence row and the scheduler match.
🤖 Generated by the Agentic Engineer (interactive Claude Code session)
Smoke canary on prod, 2026-11-09
The maintainer approved it, and I started one run from the suspended CronJob by hand at 15:14:36Z. The trigger sent the canary message and watched the task. The task ended TASK_STATE_COMPLETED after 36 seconds, well inside the 20-minute run limit, so the cancel path was not needed. The Job succeeded.
Step
Result
1. Model
The agent reports Opus 5.5 (claude-opus-5-5), which matches the pinned ModelConfig.
2. git ls-remote through the credential proxy
✅ The agent reported c50e0f547ab1d644b838a6ca07d76b386729fb15, the real main HEAD at that time. I checked it against GitHub separately.
3. Shallow clone
✅ The first line of AGENTS.md was reported correctly.
4. Tools in the sandbox
❌ gh, curl, jq and yq are all missing; only wget exists. This is expected with the stock harness image and is #4867's job.
What this proves: the scheduled trigger, kagent's A2A API, the Substrate sandbox, the Anthropic federation proxy and the GitHub App token all work together on prod. The agent can reach the model and read the portfolio over Git.
What it does not prove yet: the criterion above asks for a complete run (pre-flight, survey, one piece of work, a signed draft PR, a run report). That needs #4867 (tools), #4868 (the whole plugin) and the full run message first. This smoke canary is the first half.
Cost: to be read from the Anthropic console. A 36-second run with four short tool calls should cost a few cents at most, far below the about-$6 cap.
Small trigger defect found: the run report printed one streamed chunk per line, because the trigger joins artifact parts with newlines. Joining artifact parts without a separator fixes it.
🤖 Generated by the Agentic Engineer (interactive Claude Code session)
Per-run cap removed (maintainer decision, 2026-11-09). A run now goes until the agent ends it. Cutting a run off at 20 minutes mid-task would waste what it had already spent, and a real engineering run takes longer than that. The Anthropic workspace spend limit equal to the credit is now the lane's only spend cap. I've updated the acceptance criteria above to match.
Activity
fix(agentic-engineer): ship the lane's daily trigger suspended
devantler commentedon Nov 9, 2026
Smoke canary on prod, 2026-11-09
The maintainer approved it, and I started one run from the suspended CronJob by hand at 15:14:36Z. The trigger sent the canary message and watched the task. The task ended
TASK_STATE_COMPLETEDafter 36 seconds, well inside the 20-minute run limit, so the cancel path was not needed. The Job succeeded.claude-opus-5-5), which matches the pinned ModelConfig.git ls-remotethrough the credential proxyc50e0f547ab1d644b838a6ca07d76b386729fb15, the realmainHEAD at that time. I checked it against GitHub separately.AGENTS.mdwas reported correctly.gh,curl,jqandyqare all missing; onlywgetexists. This is expected with the stock harness image and is #4867's job.What this proves: the scheduled trigger, kagent's A2A API, the Substrate sandbox, the Anthropic federation proxy and the GitHub App token all work together on prod. The agent can reach the model and read the portfolio over Git.
What it does not prove yet: the criterion above asks for a complete run (pre-flight, survey, one piece of work, a signed draft PR, a run report). That needs #4867 (tools), #4868 (the whole plugin) and the full run message first. This smoke canary is the first half.
Cost: to be read from the Anthropic console. A 36-second run with four short tool calls should cost a few cents at most, far below the about-$6 cap.
Small trigger defect found: the run report printed one streamed chunk per line, because the trigger joins artifact parts with newlines. Joining artifact parts without a separator fixes it.
🤖 Generated with Claude Code
devantler commentedon Nov 9, 2026
Per-run cap removed (maintainer decision, 2026-11-09). A run now goes until the agent ends it. Cutting a run off at 20 minutes mid-task would waste what it had already spent, and a real engineering run takes longer than that. The Anthropic workspace spend limit equal to the credit is now the lane's only spend cap. I've updated the acceptance criteria above to match.
🤖 Generated with Claude Code
docs(kagent-lane): record the daily run as scheduled for a trial (#5077)