Skip to content
Agent Kanban

Local edition · Routing policies

The five ways a card gets its agent, model and effort, and how escalation works.

This reference describes the standalone local edition started with pnpm start. For the hosted workspace, see Connect a runner.

Every policy implements Policy.choose(card, context) → { armId, reason } (packages/core/src/routing/). Every policy skips arms whose executor is not installed or whose budget is used up, and names the skipped arms in its reason. The reason is printed on the card and on each attempt.

Policy How it picks
fixed Always the arm set in Settings.
rules A task type × difficulty table. No rule: the lowest tier. Rule arm unusable: the next usable arm at its tier or above.
planner The arm the planner proposed for the card, with its reason.
plannerWithLedger Same, but the planner prompt includes a table of attempts, pass rate and median cost per (task type, arm) from the ledger, for this project or else all projects.
cascade The lowest tier with budget left. On failed verification, a new attempt on the next usable tier with the failure evidence, up to the maximum escalations, then Needs Input.

The rules table

{ "bugfix": { "1": "kiro-cheap", "4": "claude-strong" }, "docs": { "1": "kiro-cheap" } }

Keys are task types (bugfix, feature, refactor, test, docs, investigate) and difficulties 1 to 5. A missing entry uses the lowest tier with a usable arm.

Verification and escalation

After the agent exits, the verifier checks, in order:

  1. the executor exited cleanly;
  2. every changed file is inside the card’s files in scope (when set);
  3. no pre-existing test file was modified or deleted;
  4. the acceptance command passes (run through the shell, in the attempt’s worktree, with a timeout).

The verdict is pass only when the acceptance command ran and passed, and unverified when the card has none. Under cascade, a fail starts a new attempt one tier up with the failure evidence in its prompt. Pass rates in Insights count only pass and fail verdicts.