Repository navigation
Description
🤖 Generated by the Agentic Engineer (interactive Claude Code session)
Why
The maintainer now receives a $200 monthly Claude API credit and asked, on 2026-11-05, to use it for
another scheduled Agentic Engineer: "we should set up KAgent with the Agentic Engineering plugin to have
another scheduled agent to work on the portfolio."
The earlier kagent track (#2074, closed 2026-09-10) failed on economics: pay-per-token at today's run shape
cost about 73× its cap, because spend grew with cadence. A fixed monthly credit changes that. The lane's
cadence and run size are derived from the credit, so spend cannot outgrow it.
It is also the first concrete step of #4853 (agents in sandboxed, fully tooled runners cut off from the
local network) and of #4601 (agents act through the Devantler Tech Engineering GitHub App).
Decisions (maintainer, 2026-11-07)
- Model:
claude-opus-5-5only. No fallback model and no Fable-family model. - Budget: one run a day, each capped at about $6, so at most about $190 a month. A workspace spend limit equal to the credit stops the key before any paid overage. Running out of credit stops the lane; it never buys more.
- Identity: the Devantler Tech Engineering GitHub App (registered 2026-11-07, bot
devantler-tech-engineering[bot]). No personal credential enters the cluster. The lane stays suspended until the App is installed and its key is stored. - Runtime: kagent 1.0's native Claude Code harness on the platform cluster. kagent was re-checked against the alternatives on 2026-11-07 and is still the best fit: the model key is injected by its gateway and never enters the agent's sandbox, egress is allow-listed, and plugins are pinned to a git commit.
- One brain: the lane loads the
agentic-engineeringplugin at the reviewedlibraries/agent-pluginsgitlink. It is never a forked copy of the role. - Lane: instance
kagent-cluster, writer namespacekagent/*, Agentic Engineer role only. The Agent Improver stays on the host lanes.
Outcome
A scheduled Agentic Engineer runs once a day inside the platform cluster, under its own GitHub App identity, on the monthly credit. It works the same ladder as the host lanes and ships draft PRs that the normal review and merge gates accept.
Acceptance criteria
- An ADR records the decision, and the lane is registered in the contract (instance registry, cadence, trust gate, accepted billing route).
- kagent and its Substrate runtime run on the platform cluster with least-privilege RBAC, Cilium-compatible networking and an egress allow-list.
- The agent authenticates to GitHub with short-lived App installation tokens injected at egress. The App's private key never enters the agent's sandbox, and every commit it pushes is signed.
- The agent's runtime image carries the tools a normal run needs, pinned by digest and built in CI.
- The plugin's agents, hooks and MCP servers load on the Claude harness, not only its skills (upstream [FEATURE] Plugins on the claude and codex harnesses: mcp.json, dual manifests, Claude-format on codex kagent-dev/kagent#3388).
- A daily schedule, the per-run cap and the workspace spend limit are in place and verified, and one canary run completes end to end (survey, work, draft PR).
- The Agent Improver can read the lane's transcripts and telemetry.
Constraints
- Every contract guardrail holds: trust gate, untrusted-input model, draft PRs, no pushes to protected branches, signed commits.
- Keep credentials, credential scopes and cluster topology out of public artifacts.
- Spend is bounded by the credit. A spend or quota failure stops dependent work; it never widens billing.
🤖 Generated with Claude Code
Metadata
Metadata
Assignees
Labels
Type
Fields
Projects
- StatusShow more project fields🏃🏻♂️ In Progress
Activity
docs(contract): gate the kagent lane's resume on every #4866 sub-issue
feat(contract): register a budget-capped kagent Agentic Engineer lane (…
docs(kagent-lane): record the daily run as scheduled for a trial (#5077)