KeiSeiKit-1.0/docs/encyclopedia/multi-cli-agents.md
KeiSei84 3099a58dd9 feat(phase-C): cross-CLI hook enforcement + v0.40.0 release (#48)
Mirror of keigit 596e0b20. Phase C cross-CLI hook enforcement (kei_bash/kei_edit/kei_write MCP tools + 3-tier model). Release v0.40.0.
2026-05-26 17:10:14 +07:00

225 lines
8.8 KiB
Markdown

# Multi-CLI agent invocation
> *Cross-LLM agent execution. Same agent definition, different backend.*
> *Same DNA, swap the brain. KeiSeiKit is no longer Claude-Code-only.*
KeiSeiKit agents are markdown files. Any LLM CLI that takes a prompt can
host them. Three call shapes:
```bash
kei agent <name> "<task>" # DNA-resolved (manifest → primary → claude)
kei agent --on=<backend> <name> "<task>" # override DNA
kei run-via <backend> <name> "<task>" # explicit backend (no DNA lookup)
```
## Backends — smoke-tested 2026-05-26
| Backend | CLI | Flag | Smoke | Notes |
|----------|-----------|--------------|-------|-------|
| claude | `claude` | `-p` | ✅ | Claude Code, native `--agent` flag |
| grok | `grok` | `--print` | ✅ | xAI Grok Build TUI, native `--agent` flag |
| agy | `agy` | `--print` | ✅ | Google Antigravity (Gemini models). Alias: `antigravity` |
| copilot | `copilot` | `--prompt` | ✅ | GitHub Copilot CLI (`@github/copilot`) |
| kimi | `kimi` | TUI-only | ⚠ | No print mode — launcher saves prompt to tmpfile + opens TUI for paste. `kimi acp` JSON-RPC integration is future work. |
| codex | `codex` | `-p` | — | OpenAI Codex (register-only; not installed locally) |
Run `kei run-via list` to see installed backends, current primary, and agent names.
## DNA — agent prefers a provider
Add `provider` to the agent manifest:
```toml
# _manifests/my-agent.toml
name = "my-agent"
provider = "grok" # preferred backend; optional
model = "grok-2" # advisory; informs choice but not yet sent through
```
The assembler emits it into frontmatter:
```yaml
---
name: my-agent
provider: grok
---
```
Resolution order (each falls through if previous returns nothing):
1. `--on=<backend>` flag on the command line
2. `provider:` field in agent manifest
3. `~/.claude/config/primary.toml` (set via `kei primary <backend>`)
4. Default: `claude`
## Primary — your default LLM
```bash
kei primary # show current primary (and fallback)
kei primary grok # set default to Grok
kei primary claude # back to Claude Code
```
`kei primary` writes `~/.claude/config/primary.toml`. Any agent without
its own `provider:` field will resolve to this. This is the lever to
"swap out Claude Code as the primary shell" — set primary to grok, and
every `kei agent <name>` runs on Grok.
## Usage examples
```bash
# DNA mode (manifest's provider, or primary, or claude):
kei agent critic "review src/auth.rs"
# Override DNA — try the same agent on a different model for a second opinion:
kei agent --on=grok critic "review src/auth.rs"
kei agent --on=agy critic "review src/auth.rs"
kei agent --on=copilot critic "review src/auth.rs"
# Explicit backend, no DNA lookup (legacy):
kei run-via grok critic "review src/auth.rs"
# Point at an arbitrary agent file:
kei agent --on=grok --file=/tmp/my-agent.md "do the thing"
# Native --agent flag (grok/claude only):
KEI_NATIVE_AGENT=1 kei agent critic "review src/auth.rs"
```
## How it works
1. Resolves backend from DNA (see above).
2. Reads `~/.claude/agents/<agent-name>.md` (assembler-generated prompt).
3. Strips YAML frontmatter.
4. Composes with task: `<agent prompt>\n\n---\n\nTASK FOR THIS RUN:\n<task>`.
5. Execs the backend's non-interactive CLI with the composed prompt.
No agent file is modified. No new tokens are issued — subscription
authentication is whatever each CLI uses (its own login / config dir).
## When to use each
This is a tool, not a recommendation. Each backend has different
strengths; the substrate is agnostic about which you pick. Pick by:
- **Familiarity** — the CLI you already use day-to-day.
- **Subscription cost** — burn the one with cheaper marginal cost first.
- **Specific feature** — e.g. `grok --agent` for native sub-agent
switching mid-conversation; `agy --sandbox` for terminal restriction.
- **Independent second opinion** — same agent, different model, see if
conclusions diverge.
## Orchestrator picker — `kei` no longer hardcodes claude
Without args, `kei` reads `~/.claude/config/primary.toml` and execs that CLI.
The picker lets you change it interactively:
```bash
kei pick # interactive menu → set primary → launch it
kei # splash → exec the configured primary
kei --on=grok # one-shot launch of grok (does NOT change primary)
kei primary grok # set default to grok (no launch)
kei primary # show current primary
```
The splash shows `primary CLI: <backend>` so you always know which orchestrator
will start. If the chosen primary isn't installed, `kei` prints the install
command and offers `kei pick` as recovery.
## Cross-CLI sub-agent spawn via MCP — `spawn_agent`
`kei-mcp` exposes a built-in `spawn_agent` MCP tool. Any CLI that connects
to it as an MCP client can invoke KeiSeiKit agents on any backend, no matter
what the orchestrator is:
```jsonrpc
{
"jsonrpc": "2.0",
"method": "tools/call",
"params": {
"name": "spawn_agent",
"arguments": {
"name": "critic",
"task": "review src/auth.rs for race conditions",
"on": "grok"
}
}
}
```
Internally `spawn_agent` shells out to `kei-agent-cli.sh` with the same DNA
resolution as `kei agent`. The `on` argument is optional — without it, the
backend is picked from the agent's manifest, then `primary.toml`, then claude.
**Why this matters:** Claude Code has a native `Agent` tool for sub-agent
spawning. Grok / Antigravity / Copilot / Kimi do NOT have that surface
natively — but they all support MCP. With `spawn_agent` exposed via kei-mcp,
**every backend that speaks MCP gets KeiSeiKit's sub-agent capability**. So
when Grok is your orchestrator, it can still spawn `critic` on Claude (or
`code-implementer` on Antigravity, or anything else) — the orchestrator
choice no longer caps your sub-agent surface.
Wire kei-mcp into the orchestrator's MCP config (each CLI has its own):
| CLI | MCP config |
|---|---|
| claude | `~/.claude/settings.json` `mcpServers` block |
| grok | `~/.grok/config.json` (or check `grok --help`) |
| agy | `~/.antigravity/mcp.json` (check `agy plugin list`) |
| copilot | `~/.copilot/mcp.json` (check `copilot --help`) |
| kimi | `kimi mcp add` subcommand |
Point each at `<kit>/_primitives/_rust/target/release/kei-mcp` (built via
`cargo build -p kei-mcp --release`).
## Rule enforcement — see also: cross-CLI policy
**Phase C delivered**: KeiSeiKit's safety hooks now have a 3-tier enforcement
model across CLIs. See [cross-cli-policy.md](./cross-cli-policy.md) for the
full matrix and `kei mcp-wire` setup. Short version: TIER 1 (full native)
on claude+grok, TIER 2 (MCP-wrapped) on copilot, TIER 3 (advisory) on agy+kimi.
## Rule enforcement caveat (READ THIS — pre-Phase-C view)
KeiSeiKit hooks (`numeric-claims-guard`, `citation-verify`, `no-github-push`,
`safety-guard`, `push-to-main`, etc.) are **Claude Code-side**:
`PreToolUse:Bash` / `:Edit` / `:Write` events that fire inside Claude Code's
process. They do **not** propagate to grok / agy / copilot / kimi.
That means:
- **Prompt-level rules** (the agent's instructions inside the `.md`) DO
carry through — the agent reads Constructor Pattern, Evidence Grading,
No Hallucination, etc. as part of its system prompt on any backend.
- **Tool-level enforcement** (hard-deny on `git push github.com`,
citation guard, etc.) only applies on the **claude** backend. Other
backends' tool surfaces are governed by THEIR own hooks/policies.
If you need true rule-enforcement on a non-claude backend, the path is
the **MCP server** (`_primitives/_rust/kei-mcp/`): registers KeiSeiKit
primitives as MCP tools that the other CLI invokes. Tool-side policies
travel with the MCP wrapper, not with the CLI.
## Adding a new backend
1. Add a `[backend.<name>]` table to `_primitives/cli-backends.toml`.
2. Add a case arm in `scripts/kei-agent-cli.sh` `backend_bin()` and
`backend_invoke()` for the new CLI's print-flag.
3. Add a row to the smoke-test table above (state PASS/FAIL/PARTIAL).
## What it is NOT
- Not a router — picks no backend for you; you (or DNA) ask, it dispatches.
- Not a federation — each backend runs independently with its own
context; there is no cross-backend state.
- Not a rule-enforcement layer — hooks only fire on the claude backend
(see caveat above). For non-claude rule enforcement use MCP server.
- Not a wrapper around the backend's tool surface — what the CLI can
do (Bash, file edits, MCP, etc.) is determined by that CLI, not
KeiSeiKit. The substrate only ships the prompt.
## Related
- `_primitives/_rust/kei-llm-router/` — Beta-posterior router for
*programmatic* model selection inside Rust code (a different layer).
- `_primitives/_rust/kei-mcp/` — MCP server that exposes KeiSeiKit
primitives to ANY MCP-compatible client (Cursor / Continue / Zed /
Aider / Cline / Windsurf / OpenClaw).