fix: use kiro-cli-chat directly to prevent orphan child processes

[vokako]

Problem

kiro-cli is a wrapper that forks kiro-cli-chat as the actual ACP server process. When acpx sends SIGTERM to kiro-cli on session close, only the wrapper is killed — the child process kiro-cli-chat continues running as an orphan.

These orphaned processes accumulate over time and cause resource conflicts, resulting in ACP_TURN_FAILED errors (exit code 1) on subsequent messages in persistent sessions. The first message in a session works fine, but follow-up messages fail because the stale kiro-cli-chat process is still holding resources.

Fix

Point the built-in kiro agent directly at kiro-cli-chat acp, bypassing the wrapper entirely. This ensures clean process lifecycle management without needing process group kill logic.

One-line change in src/agent-registry.ts:

-  kiro: "kiro-cli acp",
+  kiro: "kiro-cli-chat acp",

Related

• feat: add Kiro CLI as built-in agent with proper process group cleanup #42 proposed a more complex fix using process group cleanup (detached:true + process.kill(-pgid)). This PR provides a simpler alternative by avoiding the wrapper altogether.

Testing

• Created persistent kiro session with acpx kiro sessions new
• Sent multiple consecutive messages — all succeeded with session reuse
• Verified only 1 kiro-cli-chat acp process running (no orphans)
• Previously: 6+ orphan processes accumulated after a few session cycles

[dutifulbob]

Triage result

Human attention: ⚠️ Required
Recommendation: 🏁 escalate to a human
Human decision needed: decide whether to require manual Kiro validation before landing, or approve based on code review and normal CI alone.

Quick read

This PR is a focused bug fix: it changes the built-in kiro command from kiro-cli acp to kiro-cli-chat acp so acpx talks to the real ACP server directly instead of a wrapper.

The initial merge conflict against origin/main was limited to skills/acpx/SKILL.md and had a clear mechanical resolution path.

The remaining blocker in this flow is validation: the fix was not proven locally here because no targeted repro/test command could be derived from the repo state for this Kiro-specific behavior.

Intent

Change acpx's built-in kiro agent to launch Kiro's real ACP server directly so Kiro sessions shut down cleanly and persistent follow-up messages keep working.

Underlying problem: kiro-cli acp is a wrapper that spawns kiro-cli-chat; acpx only terminates the direct child, so stale kiro-cli-chat processes remain and later persistent-session turns can fail.

Why

• src/client.ts shuts agents down by ending stdin and then sending SIGTERM/SIGKILL to the spawned child only; it does not do process-tree cleanup.
• The PR changes src/agent-registry.ts from kiro-cli acp to kiro-cli-chat acp, which directly matches that failure mode.
• The repo already uses agent-specific built-in command mappings, and the PR updates the related docs and skill docs consistently.

Codex review

• Solution verdict: good_enough
• Validation path: bug
• Validation status: fix_not_proven
• No targeted repro/test command could be derived from the local repo context for this Kiro-specific lifecycle issue.
• No deeper Codex review findings were produced in this handoff path.

CI/CD

• CI status in this run: not evaluated here.
• Final conflict-gate status in this run: not available.
• Initial conflict status: detected against origin/main in skills/acpx/SKILL.md, with a clear resolution path.

Recommendation

Escalate to a human. The code change is small and plausibly correct, but someone needs to decide whether normal repo review/CI is sufficient here or whether this PR should get manual Kiro-specific verification before landing:

• reproduce the prior failure in a persistent Kiro session if possible
• confirm follow-up turns now succeed
• confirm kiro-cli-chat does not remain orphaned after session shutdown