feat(harness-kimi): kimi-code terminal agent + local MLX preset

packages/harness-kimi/README.md

@@ -0,0 +1,45 @@
+<!-- lineage
+role: package-readme
+conforms_to: ../../README.md
+consumes: config/config.local-mlx.toml
+-->
+
+# harness-kimi
+
+[kimi-code](https://moonshotai.github.io/kimi-code/) — Moonshot's terminal
+coding agent (binary `kimi`) — at the repo's curated pin (brew formula
+`kimi-code`, currently 0.29.1; see root CLAUDE.md upstream-tracking).
+
+```bash
+./install.sh   # brew install kimi-code + copy local-endpoint config preset
+```
+
+## What it wires
+
+- **`config/config.local-mlx.toml`** → `~/.kimi-code/config.toml` (copy-if-absent).
+  Two OpenAI-type providers + model aliases:
+  - `gpt-oss-local` → **rapid-mlx :11437** (harmony) — the default model
+  - `qwen-fast-local` → **mlx_lm.server :11436** (standard template)
+
+## Config schema notes (verified via `kimi doctor`, 0.29.1)
+
+- Provider `type` is an enum: `openai | openai_responses | anthropic | kimi | google-genai | vertexai`. Local MLX = `openai` (chat).
+- Each model alias needs `max_context_size`, which **doubles as kimi's output
+  budget** — so it's set to 32768, well under gpt-oss's 131072 window, to avoid
+  `APIContextOverflowError` (kimi otherwise requests completion = max_context_size).
+- Adding providers via `kimi provider catalog add <id>` needs a Moonshot
+  `KIMI_REGISTRY_API_KEY`; this preset hand-authors the config instead (no key).
+
+```bash
+mlx-serve flagship   # rapid-mlx :11437 (gpt-oss, harmony)
+kimi -m gpt-oss-local -p "your prompt"
+```
+
+Verified 2026-07-26: `kimi -m gpt-oss-local` drives local gpt-oss correctly.
+
+## herdr
+
+`herdr integration install kimi` installs a state reporter to
+`~/.kimi-code/hooks/herdr-agent-state.sh` (requires kimi-code ≥ 0.14.0).
+
+No secrets — localhost endpoints only, per repo convention.

packages/harness-kimi/config/config.local-mlx.toml

@@ -0,0 +1,31 @@
+# harness-kimi local-endpoint preset → ~/.kimi-code/config.toml.
+# Schema verified against kimi-code 0.29.1 via `kimi doctor` (provider.type is an
+# enum: openai|openai_responses|anthropic|kimi|google-genai|vertexai; models need
+# max_context_size). No secrets: localhost, api_key = "none".
+#
+# gpt-oss needs rapid-mlx (harmony) on :11437; standard-template models use
+# mlx_lm.server on :11436. max_context_size doubles as kimi's output budget, so
+# it's kept well under the model's real window to avoid context overflow.
+default_model = "gpt-oss-local"
+
+[providers.local-mlx-rapid]
+name = "Local MLX — rapid-mlx (harmony)"
+base_url = "http://localhost:11437/v1"
+type = "openai"
+api_key = "none"
+
+[providers.local-mlx]
+name = "Local MLX — mlx_lm.server"
+base_url = "http://localhost:11436/v1"
+type = "openai"
+api_key = "none"
+
+[models.gpt-oss-local]
+provider = "local-mlx-rapid"
+model = "mlx-community/gpt-oss-120b-MXFP4-Q8"
+max_context_size = 32768
+
+[models.qwen-fast-local]
+provider = "local-mlx"
+model = "mlx-community/Qwen3.6-35B-A3B-8bit"
+max_context_size = 32768

packages/harness-kimi/install.sh

@@ -0,0 +1,27 @@
+#!/usr/bin/env bash
+# Install kimi-code (Moonshot's terminal coding agent) + wire this package's
+# local-endpoint preset. Idempotent. No curl|bash — brew-native install.
+set -euo pipefail
+
+PKG_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+command -v brew >/dev/null 2>&1 || { echo "brew not found — install Homebrew first" >&2; exit 1; }
+
+# kimi-code is a Homebrew formula; binary is `kimi`. Curated pin tracked in root
+# CLAUDE.md upstream-tracking.
+brew list kimi-code >/dev/null 2>&1 || brew install kimi-code
+kimi --version
+
+# Config preset: copy if absent; never clobber a live config (auth/providers).
+mkdir -p "${HOME}/.kimi-code"
+if [ ! -f "${HOME}/.kimi-code/config.toml" ]; then
+  cp "${PKG_DIR}/config/config.local-mlx.toml" "${HOME}/.kimi-code/config.toml"
+  echo "installed config.toml (local-endpoint providers) -> ~/.kimi-code/config.toml"
+else
+  echo "${HOME}/.kimi-code/config.toml exists — NOT overwritten. Merge providers/models from ${PKG_DIR}/config/config.local-mlx.toml manually."
+fi
+
+kimi doctor >/dev/null 2>&1 && echo "kimi config valid" || echo "note: run 'kimi doctor' to validate config"
+echo "kimi-code ready — local: kimi -m gpt-oss-local -p 'your prompt'"
+echo "  (serve first: mlx-serve flagship  → rapid-mlx :11437, harmony parsers)"
+echo "  herdr: herdr integration install kimi"

packages/harness-kimi/moon.yml

@@ -0,0 +1,13 @@
+$schema: 'https://moonrepo.dev/schemas/project.json'
+
+layer: 'tool'
+
+project:
+  name: 'harness-kimi'
+  description: 'kimi-code (Moonshot terminal agent) harness — install + local-endpoint preset'
+
+tasks:
+  lint:
+    command: 'shellcheck install.sh'
+    inputs:
+      - 'install.sh'