Ql Chat

---
name: ql-chat
description: 'Start or continue a QualityLayer Chat in Claude Code or Codex when the person explicitly asks for it, for example "Using QualityLayer, chat about …". Product discussion, quoted examples and ordinary software requests do not activate Chat. Use ql-workflow for the three-step workflow (Plan, Implement, Verify).'
---

The binary: `{{QUALITYLAYER_BIN}}`

When earlier project decisions matter, search and read relevant lessons again after a fresh
start or compaction. Use `memory_search` and `memory_read`; save reusable corrections with
`memory_save` and `replaces`, and confirm, retire or restore with `memory_keep`. Without those
tools, use the matching `qualitylayer memory` commands described by `ql-manage`. A client without
an automatic lesson hook uses this same CLI fallback.

## Start or continue

For a new Chat (the prompt may say "start chat: <request>"), run
`{{QUALITYLAYER_BIN}} task start --type chat --json "<the person's request>"`.
For an existing Chat, select its exact task with `--task <slug>` on each command.
Read the result, then work with the person in the conversation. A Chat has no workflow
steps or approval gates. Keep the agent's native task list current and use its native
picker when a choice needs the person's answer. After your first reply, give the Chat
a short title with `{{QUALITYLAYER_BIN}} chat rename "<title>" --json`.

## Optional tools

Use these only when the person requests the corresponding work:

- **Save the plan:** `{{QUALITYLAYER_BIN}} chat plan save --file <local-markdown> --json`.
  Alternatively use `--text <markdown>`, exactly one input. Save ordinary Markdown;
  retain its wording and line endings. The result names `chat-plan.md` for the Reader.
- **Build in slices:** derive complete vertical slices from that saved plan in the
  same slices-and-cards grammar a task's build uses, with the complete slice file example below.
  Save this local Markdown file, then run `{{QUALITYLAYER_BIN}} chat build --file <cards-markdown> --json`.
  Start each returned helper as a fresh agent with only its `brief_file` path and
  "Read it in full and follow it". Respect file ownership and preserve other helpers'
  edits. Save every result with `helper result <id> --file <report>` and finish with
  `progress done --agent <id>`. Resume with `chat build`; check existing helpers before
  launching returned offers, inspect saved reports and tick only verified completed
  cards in `chat-slices.md`. With no helper running, it resumes unchecked slices.
  Completed and stopped reports remain on disk.
- **Checks:** `{{QUALITYLAYER_BIN}} chat check [--slice <n>] --json` prints the actual
  validation commands. Run those commands yourself under the session's existing
  permissions and retain each exit status. Printing a command does not prove it passed.
- **Independent review:** `{{QUALITYLAYER_BIN}} chat review --json` returns a fresh
  reviewer brief. Start that independent reviewer with only its `brief_file` path.
  Record the returned report with `{{QUALITYLAYER_BIN}} chat review --done --id <id> --report <local-report> --json`.
  If no fresh reviewer is available, state that limitation. Address concrete findings.
- **Artifacts:** use the `ql-artifact` skill for creating, changing and opening artifacts.

### Slice file example

Use this complete shape in the local file passed to `chat build --file`. Replace
the task, paths and commands to match the saved plan; retain the labels and nesting.
This example is for a Bun CLI project.

```markdown
## Slice 1: Add a version flag

- [ ] T1 Print the version
  - **Objective:** Print the current version when --version is passed.
  - **Files:** `src/cli.ts`, `tests/cli.test.ts`
  - **Test first:** Add a failing CLI test for --version.
  - **Done when:**
    - --version prints the current version and exits successfully.
    - verify: `bun test tests/cli.test.ts`

**Validation:** automated: `bun test tests/cli.test.ts` · manual: none
```

For more slices, repeat the block with card IDs T1, T2, … unique across the whole
file. Give every card its own backticked Files paths and `verify:` command, and
every slice its own Validation command.

## Convert to task

Only when the person requests it (the App's Convert to task), run
`{{QUALITYLAYER_BIN}} chat promote --type feature|bugfix --json`, choosing the kind
they asked for. Running slice helpers must finish or be explicitly stopped first;
stop the native helper through the agent's own controls, then record it with
`{{QUALITYLAYER_BIN}} chat build --stop <slice> --json`. Promotion keeps this task and conversation,
records the Chat plan, artifacts and changed files as the Plan's context, and is allowed
once. Follow its instruction to run `{{QUALITYLAYER_BIN}} next --json` and continue
with `ql-workflow` at the Plan. The Plan's approval governs subsequent work.