Verification

KHAELOR never claims completion because the model produced a confident sentence. It runs your checks itself — and repairs what fails, before you ever see the answer.

How the loop works#

The model finishes a batch of edits (write/edit tool calls).
The Tool Runtime launches the configured checks in parallel through the workspace seam (autofix checks like eslint --fix run last, serially).
Each check lands as a durable verify.result event: name, command, real exit code, duration, and output truncated errors-first.
Failing results are injected back into the conversation — the model repairs them before the turn can complete.
The loop is bounded: after maxRepairLoops failing rounds (default 3) per user turn, checks stop re-running and the honest failure report stands.

the repair loop

 ├─ edit src/context/compaction.ts  +14 −3
 ⟳ verify   typecheck ✓ 1.2s · tests ✗ 4.8s · lint ✓
   FAIL src/context/engine.test.ts — compaction preserves running processes
 KHAELOR is repairing the failure…
 ├─ edit src/context/compaction.ts  +6 −2
 ⟳ verify   typecheck ✓ · tests ✓ 5.3s · lint ✓
 ✓ verified 5.3s

Configuration#

// .khaelor/verify.json
{
  "typecheck": { "cmd": "npx tsc --noEmit", "timeout": 60 },
  "test":      { "cmd": "npx vitest run --changed", "timeout": 120 },
  "lint":      { "cmd": "npx eslint --fix", "timeout": 30, "autofix": true },
  "policy": "after-each-edit-batch",
  "maxRepairLoops": 3
}

Field	Meaning
`<name>.cmd`	The command; its exit code is the verdict. Run through your shell, in the project root.
`<name>.timeout`	Seconds before the check is killed (recorded as exit `null`, i.e. failed).
`<name>.autofix`	The check mutates files — run it after the parallel read-only checks.
`policy`	`after-each-edit-batch` (default) · `before-final-answer` (model-run, nudged) · `off`.
`maxRepairLoops`	Failing rounds per user turn before KHAELOR stops looping and reports.

Auto-detection#

Without a verify.json, KHAELOR detects conservatively:

Found	Checks
`package.json` scripts	`npm run typecheck` / `npm test` / `npm run lint` when the scripts exist; `npx tsc --noEmit` when only `tsconfig.json` exists.
`Cargo.toml`	`cargo check`, `cargo test`.
`pyproject.toml`	`python3 -m pytest -x -q`.

Honesty rules#

Passing checks stay out of the model's context (evidence only) — failures enter it verbatim.
Output is truncated errors-first: lines matching failure patterns are kept ahead of noise, with an explicit truncation marker.
/verify runs the whole suite on demand; the daemon refuses to escalate a goal run whose checks failed.