Kild CLI

kild CLI

kild runs pi coding agents in your projects. It is the scriptable interface an agent can drive over the Bash tool. Installed globally: run kild <args> from anywhere (bun link from engine/; source engine/src/cli.ts).

kild run is one-shot: it starts an agent, lets it work to completion, and prints the result. For live, multi-agent, steerable work, drive a kild (kild new --detach / ls / log / send / stop) — see "Driving kilds" below.

Vocabulary

One word per concept. See docs/VOCABULARY.md.

• kild — a git worktree and the agents working in it. The unit of isolated, parallel work. Also the name of the tool; "a kild" is the unit.
• agent — one running agent process. The only inhabitant type.
• persona — a markdown file that shapes what an agent is. .pi/agents/*.md, .claude/agents/*.md. The dir is called agents/ by upstream convention; the contents are personas.
• handle — an agent's @name, addressable inside its kild.

Command reference

There is no kild worktree group: a kild is a worktree, so it has one set of verbs. kild worktree ls → kild ls, kild worktree rm → kild rm, and kild worktree prune → kild ls --state reclaimable (then kild rm what you mean to reclaim — nothing is removed behind your back). All of these take a live kild's id or an orphan tree's worktree name; kild ls prints whichever applies.

Add --json to any command for machine-readable output on stdout.

Flags

--project <name|path> · --persona <name> · --model <ref> · --worktree <name> · --base <branch> · --agents a,b,c (for kild new) · --to a,b (for kild send) · --as <handle> (for attach/inbox/spawn, and stop to stop one agent) · --task <text> (for kild spawn) · --state live|orphan|reclaimable · --git (both for kild ls) · --since <seq> (for kild log) · --execute (for kild land) · --detach · --force · --json

Driving kilds (multi-agent units of parallel work)

A kild is an isolated workspace where one or more agents collaborate. Unlike run (one-shot, single agent), a kild is steerable and multi-agent:

ID=$(kild new "Build feature X" --detach --project myproj --worktree feat-x)
kild ls                       # the glance: kilds + rosters (fast — no git)
kild ls --git                 # + branch, ahead/behind, dirty, conflicts, collisions
kild log "$ID"                # the full conversation (each line prefixed by its seq)
kild log "$ID" --since 12     # only what arrived after seq 12
kild send "$ID" "use the prp-implement skill to implement plan Y"
kild send "$ID" --to reviewer "check the auth path"   # address one agent
kild stop "$ID"               # end it

Agents come from the project's own personas (--agents a,b,c; with none, one general agent). Each gets kild's default prompt plus any skills the project's config plugs in — so you can send use the prp-X skill to … and the agent loads it.

A harness kild does not spawn — the Claude Code session you are in — can also hold a handle: kild attach "$ID" --as claude registers it, and it collects mail by draining at its own turn boundary rather than being pushed to.

Config — plug in skills/personas, base branch, and a model catalog

.kild/config.json (project) and $KILD_HOME/config.json (global, merged; project wins):

{
  "plugins": ["./prp-core"],
  "baseBranch": "dev",
  "models": {
    "openai-codex/gpt-5.6-sol": "Strongest model. Hard reasoning, orchestration, synthesis. ($$$)",
    "openai-codex/gpt-5.6-terra": "Workhorse. Research, exploration, implementation. ($$)",
    "minimax/MiniMax-M3": "Cheap/fast. Mechanical/bulk work. ($)"
  }
}

• plugins — dirs laid out like a Claude Code plugin (agents/ + skills/); absolute or ~/… paths load from anywhere. Also agentPaths / skillPaths for explicit dirs.
• baseBranch — default base for worktrees + git status.
• models — a provider/model → description catalog, appended to a delegating session's system prompt so it knows which model to pass to spawn per fan-out agent.

Delegation is asynchronous

Inside a kild, delegating is fire-and-forget: you keep going, and a delegate's message wakes you automatically. Don't busy-wait re-asking an agent that already replied.

An agent in a kild has one verb for both: send. Addressing a handle nobody holds creates that agent and delivers your message to it, so delegation is "pick a handle, say who it should be with persona, write the assignment". Naming several new handles in one to gives them all the same brief — that is a fan-out. There is no spawn step and no idle just-created agent to remember to brief.

From outside (this CLI, the REST API, the pi extension) the two stay separate on purpose: a typo'd handle must not quietly cost a process, so kild send refuses an unknown recipient and kild spawn <id> --as <handle> [--task <text>] is the explicit way to add one.

Seeing who is in a kild: kild show <id> prints the roster (with each agent's persona, model, and invitedBy), kild ls prints it per kild in one line. An agent inside a kild does not need a roster tool — its send result names anyone it just created, and a to that cannot be resolved comes back listing both the current roster and the available personas.

Reaching another agent requires an explicit send. Your normal output is private — narration is not delivered to anyone. If you finish delegated work without sending your result, whoever delegated is left waiting. The engine does not chase you: reporting is your responsibility (and your persona's), not a mechanism that nudges.

Worktrees — every kild is one (you name them)

--worktree <name> runs the kild (or kild run) in an isolated git worktree on branch kild/<name>; kild adds the kild/ prefix, you choose the suffix. There is no auto-naming — omit --worktree and the work happens in the project's main checkout with no isolation.

Conventions when you drive kild:

• One kild = one worktree. Name it for the task: --worktree fix-2247, --worktree feat-dark-mode. Never run two independent kilds in the same tree.
• Reuse the name to share a tree. Two sessions naming the SAME --worktree attach to one tree (e.g. a reviewer joining a coder's kild). Different names split.
• Base branch. New worktrees fork from — and git status/collisions are measured against — the base: --base <branch> wins, else .kild/config.json baseBranch, else the checkout's current branch. On a repo whose trunk is dev, set {"baseBranch":"dev"} so ahead/behind and collisions reflect only this kild's work.
• Observe & land. kild ls --git shows each kild's branch, ahead/behind, dirty, conflicts, and cross-kild file collisions (plain kild ls skips git and stays fast). kild land <id> previews the merge without touching anything; kild land <id> --execute performs it and prints the sha (recorded on the kild, so the ledger names the commit).
• Clean up. kild rm <id> frees a tree. It refuses while the branch carries commits the base does not have — that is unlanded work, and you are told so rather than the tree quietly surviving. Uncommitted files are not a refusal (provisioning litter is not work): they are discarded and listed. The kild/<name> branch always survives, which is why --force costs no commits. kild ls --state reclaimable is the list kild rm would accept.
• Stranded trees are addressable. kild ls enumerates worktrees from git, so a tree from an earlier engine run appears as an orphan (no agents, no log) under its worktree name and can be landed or removed like any other kild.

The output contract

• stdout — the result. Plain text by default; with --json, a JSON value.
• stderr — progress (tool activity), the model line, and stats. Ignore for parsing (2>/dev/null).
• exit code — 0 on success, non-zero on failure. Always check it.

kild run

kild run [--project <name>] [--persona <name>] [--model <ref>] [--worktree <name>] <prompt…>

• cwd — defaults to the current directory. --project <name> instead runs in a registered project's path. Common pattern: cd <some-dir> && kild run ….
• --persona <name> — layer a named system prompt. Omit for the built-in general persona: no specialisation, for when no other persona fits. List options with kild persona ls.
• --model <ref> — e.g. claude-opus-4-8, claude-haiku-4-5. Omit for pi's default.
• --worktree <name> — run in an isolated kild/<name> worktree. Created if missing, attached if it exists. The worktree persists after the run.

--json result shape

{
  "model": "anthropic/claude-haiku-4-5",
  "text": "the agent's full reply",
  "tokens": 8092,
  "cost": 0.00176
}

Examples

Run a one-shot task and capture just the answer:

kild run --json "Summarize what this repo does in two sentences." 2>/dev/null | jq -r .text

Delegate to a specialized persona in a registered project:

kild run --project myapp --persona planner --json \
  "Draft an implementation plan for adding OAuth login." 2>/dev/null | jq -r .text

Two agents on one repo in isolation, then review + clean up:

kild run --project myapp --worktree fix-auth --json "Fix the auth bug." 2>/dev/null &
kild run --project myapp --worktree add-logs --json "Add request logging." 2>/dev/null &
wait
kild ls --project myapp --git --json     # both trees, with their git state
kild land fix-auth                       # dry run: what would merge, what collides
kild land fix-auth --execute             # merge it, print the sha
kild rm fix-auth                         # free the tree (the branch stays)

Register a project, then list its personas:

kild project add myapp ~/projects/myapp
kild persona ls --project ~/projects/myapp --json | jq -r '.[].name'

Notes

• Personas are files. A persona named <name> is a <name>.md file (its body is the system prompt) in a project's .kild/agents/, .claude/agents/, or .pi/agents/, or globally in ~/.config/kild/agents/ or ~/.claude/agents/. The built-in general uses pi's own prompt. To add one, drop a file — kild only reads them.
• kild run is one-shot. It blocks until the agent finishes; tool progress streams to stderr. To steer an agent mid-flight, use a kild and kild send.
• Errors are explicit. An unknown project/persona, an unreadable persona file, or a pi spawn failure prints to stderr and exits non-zero — never silently degrades.