Kiro CLI goal mode

LoopX adapter for Kiro CLI (binary kiro-cli) — a terminal coding agent with skills, steering files, agent configs, hooks, MCP, and a native goal loop. Kiro CLI ships both halves of a goal-mode host: a goal primitive and a host-enforced iteration budget, so LoopX binds the objective to the host's own loop instead of pretending the agent merely drives itself.

Native goal primitive

/goal [--max N] <description> | clear — the built-in command as the host contracts it. The host re-dispatches turns toward the stated objective, verifies each iteration against the acceptance criteria it derives from the goal statement, and the model must prove completion through the built-in goal tool before the loop ends.

• Iteration budget: the host default is 5, raised with --max. The flag precedes the description, and everything after it is the goal statement.
• Acceptance criteria travel inside the description. The host derives them from the goal statement, so LoopX appends the criteria its todo already owns (after Done when:) instead of passing a separate flag. There is no criteria flag to pass: an unrecognised flag after the description would be read as more goal text, polluting the objective and silently dropping --max.
• /goal clear cancels the active goal. A running goal is steered in place rather than read back — the host advertises no goal status subcommand.
• The goal tool's complete command enforces a completion contract: each success criterion needs cited tool output, and belief or narrative confidence is explicitly not evidence — the same standard LoopX writeback wants.

The contracted surface is checked against the host itself, not transcribed. The installed binary advertises its own command registry over ACP, so the shape can be replayed on any machine that has the CLI:

python3 examples/kiro-cli-goal-command-contract-probe.py

It asserts that /goal exists and that its advertised subcommands are exactly the ones LoopX projects. An earlier revision of this document claimed --validate, --agent and /goal status and cited a strings probe of the binary; that probe returns nothing on the shipped app bundle, and neither the published command reference nor the host's own registry contracts those arguments, so they are no longer projected.

Native hook seam

Agent configs (~/.kiro/agents/<name>.json or .kiro/agents/<name>.json) carry hooks for agentSpawn, userPromptSubmit, preToolUse, postToolUse, and stop; a preToolUse hook can block a tool call by exiting 2. That is a real machine-enforceable seam, but LoopX installs no Kiro hook today, so this surface does not claim enforcement. The trigger list is recorded in __init__.py so a future enforced-quota binding has one place to start from rather than rediscovering it from host docs.

What this surface is

Kiro CLI discovers global skills from <KIRO_HOME>/skills/<name>/SKILL.md (~/.kiro when KIRO_HOME is unset) and workspace skills from .kiro/skills/<name>/SKILL.md; the default agent carries both as skill:// resources, and every discovered skill is invocable as a /<skill-name> slash command with $ARGUMENTS expansion. LoopX reaches a Kiro CLI session through the generated /loopx skill facade, and the activation binds the objective with the native /goal --max <N> <task_body> Done when: <criteria>.

Three honest limits, stated in the activation packet:

• Quota pacing is advisory. LoopX installs no Kiro hook that intercepts a native goal iteration, so quota should-run entry is facade guidance the agent is instructed to follow — not a host-enforced gate.
• The loop lives and dies with the session. The /goal loop runs only while the CLI session is alive; there is no cross-session daemon, so it bounds a live session's segments, not an unattended host loop.
• A same-named file prompt wins. Kiro resolves .kiro/prompts/*.md and ~/.kiro/prompts/*.md before skills, so a user prompt named loopx shadows the managed skill. The installer never touches the prompt directories.

Install

loopx slash-commands --install --surface kiro-cli

Writes the managed LoopX skill facades (loopx/SKILL.md, loopx-global-*/SKILL.md, …) into <KIRO_HOME>/skills/ using Kiro's per-skill directory layout. KIRO_HOME is the host's own override for that global root, so LoopX resolves it and falls back to ~/.kiro when it is unset; install and uninstall always target the same resolved root. Managed files carry the loopx-managed-slash-command marker and are refreshed by rerunning the installer; user-owned files are never overwritten.

Use

From a Kiro CLI session in a connected project, run /loopx <complex task>. The facade instructs the agent to run:

loopx start-goal --guided --project . --slash-command-arguments="<task>" --host-surface kiro-cli

After todo writeback, bind the generated heartbeat task body with /goal --max <N> <task_body> Done when: <criteria> — stating the criteria the todo already names inside the goal statement, because the host derives its acceptance criteria from that statement rather than from a flag, and N taken from the remaining quota slots (host default is 5) — steer a running goal in place rather than expecting a status readback, start every turn and native goal iteration with quota should-run (advisory guidance; LoopX does not intercept native host iterations), and settle through the built-in goal tool only after LoopX writeback so the cited evidence matches what LoopX recorded.

Kiro CLI exports KIRO_SESSION_ID for every session; it is the stable value a LoopX thread binding should key on instead of prose.

Dashboard and Chat agent

Kiro CLI also ships an ACP agent (kiro-cli acp), so loopx dashboard and loopx chat list Kiro CLI as a built-in Agent alongside Codex and Claude Code. It reuses loopx.chat_acp.ACPStdioAdapter rather than a second transport: probed on 2.21.1, the host answers initialize with protocolVersion: 1 and loadSession: true, and session/new returns a session id, which is exactly what that adapter expects.

loopx dashboard                                  # Kiro CLI appears when it is on PATH
loopx dashboard --kiro-cli-bin /path/to/kiro-cli # explicit executable

Two boundaries this does not cross:

• Owner-managed host permissions. LoopX Chat answers every interactive ACP session/request_permission with cancelled, exposes no client host tools, and launches without --trust-all-tools. Kiro can still execute a tool without asking when its user, workspace, or agent permission rules already say allow; LoopX cannot turn those persistent host rules into a read-only sandbox. The capability therefore advertises workspace_write, and the owner must configure Kiro permissions for the desired boundary. A planning prompt or a cancelled request is not an authority gate.
• Not the governed loop. A dashboard Chat session is one bounded conversation. The /goal loop above is entered from a Kiro CLI session through the installed skill facade; the two surfaces share the host, not the loop.

The built-in id kiro-cli is reserved in the owner-local endpoint registry so a hand-registered endpoint cannot silently shadow it.

Control-plane identity

Serving an Agent row is not the same as being reachable. Three surfaces resolve a host by identity, and each one needs Kiro CLI in its table:

• Endpoint to Goal agent. A durable Goal agent id is operator-chosen, so chat_actions collapses both onto a host family: Endpoint kiro-cli resolves a registered kiro-worker-1, the same way codex resolves codex-main-control. Without the row, selecting Kiro CLI in the workspace raised agent_binding_required for an agent the user did register.
• Host thread binding. KIRO_SESSION_ID is read for the kiro-cli host surface, so start-goal binds the live session instead of leaving the thread unbound.
• Project skills. loopx project-skill --surface kiro-cli delivers into .kiro/skills, the workspace skills root Kiro discovers per project.

Layout

• __init__.py — host facts: install surface id, fixed skills root resolution, the agent-type catalog entry, the activation extras (native goal command, host iteration budget, completion tool, advisory quota boundary), and the Chat/ACP launch facts the dashboard's built-in Agent row is built from.