Skip to main content
Version: Next

Add your stack's quality gates

insight-flow ships no language, package-manager, or test-runner assumptions. To make /task-implement and /task-review-fix run your typecheck / lint / test before they mark work done, supply the commands via agents.extend. Each string is appended verbatim to that agent's prompt.

The commands below are examples — they only reach the agent because your taskflow.config.json adds them. insight-flow stays technology-agnostic.

1. Add your gates to agents.extend​

Open taskflow.config.json and add task-implement (gates after the initial implementation) and task-review-fix (gates after every blocker fix).

TypeScript + pnpm:

// taskflow.config.json
{
  "agents": {
    "extend": {
      "task-implement": [
        "Run `pnpm typecheck && pnpm lint && pnpm test` before marking the task implemented. Do not mark it implemented if any gate fails.",
      ],
      "task-review-fix": [
        "Re-run `pnpm typecheck && pnpm lint && pnpm test` after every blocker fix before marking the task fixed.",
      ],
    },
  },
}

Python + uv:

{
  "agents": {
    "extend": {
      "task-implement": [
        "Run `uv run mypy . && uv run ruff check . && uv run pytest` before marking implemented.",
      ],
      "task-review-fix": [
        "Re-run `uv run mypy . && uv run ruff check . && uv run pytest` after every blocker fix.",
      ],
    },
  },
}

Go:

{
  "agents": {
    "extend": {
      "task-implement": [
        "Run `go vet ./... && golangci-lint run && go test ./...` before marking implemented.",
      ],
      "task-review-fix": [
        "Re-run `go vet ./... && golangci-lint run && go test ./...` after every blocker fix.",
      ],
    },
  },
}

2. Apply the change​

Re-scaffold the role files so the extension lands in the prompts:

insight-flow init

Or, to only re-sync agents.extend + the enforcement block into existing role files:

insight-flow prompt-build --apply

3. Verify​

Run /task-implement on a ready task. Before it calls insight-flow implement-end, the agent now runs your gates; a failing gate keeps the task out of the implemented state. The same holds for /task-review-fix after each blocker fix.

Notes​

  • Keep each gate string as a single shell pipeline so the agent runs it as one step and stops on the first failure.
  • An empty array for an agent means it runs the canonical prompt only and notes in its report that the quality-gate step was skipped.

See also​