ORDL-AMF · Apr 20, 2026 · Apr 20, 2026 · Apr 20, 2026 · Apr 20, 2026 · Apr 20, 2026
diff --git a/.gitignore b/.gitignore
@@ -5,6 +5,7 @@ build/
 dist/
 wheels/
 *.egg-info
+.claude/
 
 # Virtual environments
 .venv
@@ -40,11 +41,21 @@ result-*
 # macOS files
 .DS_Store
 
+# Development artifacts
+build.sh
+cleanup.sh
+handoffs/
+kimi.log
+.codex
+PYEOF
+
 # Rust files
 target/
 
 node_modules/
 static/
 .memo/
 .entire
-.claude
+.watch_merge.sh
+.watch_merge.log
+docs/AGENT-SPEC-REFERENCE.md
diff --git a/AGENTS.md b/AGENTS.md
@@ -16,6 +16,70 @@ If running tools directly, use `uv run ...`.
 Kimi Code CLI is a Python CLI agent for software engineering workflows. It supports an interactive
 shell UI, ACP server mode for IDE integrations, and MCP tool loading.
 
+## LLM Provider Configuration
+
+### Default model
+
+The default model is `kimi-for-coding` (beta preview), served through the **agent gateway**.
+
+### Agent Gateway
+
+When the provider `base_url` contains `agent-gw.kimi.com`, the CLI automatically switches from
+OpenAI Chat Completions format to **Anthropic Messages API** format. This is handled transparently
+in `src/kimi_cli/llm.py`.
+
+```toml
+[models."kimi-code/kimi-for-coding"]
+provider = "managed:kimi-code"
+model = "kimi-for-coding"
+max_context_size = 2000000
+capabilities = ["video_in", "image_in", "thinking"]
+
+[providers."managed:kimi-code"]
+type = "kimi"
+base_url = "https://agent-gw.kimi.com/coding/v1"
+api_key = "sk-kimi-..."
+```
+
+### API format details
+
+| Endpoint | Method | Format |
+|----------|--------|--------|
+| `/coding/v1/messages` | POST | Anthropic Messages API |
+| `/coding/v1/models` | GET | OpenAI-compatible list |
+| `/coding/v1/usages` | GET | Platform quota info |
+| `/coding/v1/feedback` | POST | Feedback submission |
+
+The Anthropic SDK appends `/v1/messages` to the base URL. The CLI strips the trailing `/v1`
+from `base_url` before passing it to the Anthropic client when `agent-gw` is detected.
+
+### Environment overrides
+
+| Variable | Effect |
+|----------|--------|
+| `KIMI_BASE_URL` | Overrides provider `base_url` |
+| `KIMI_API_KEY` | Overrides provider `api_key` |
+| `KIMI_MODEL_NAME` | Overrides model name |
+| `KIMI_MODEL_MAX_CONTEXT_SIZE` | Overrides `max_context_size` |
+| `KIMI_MODEL_TEMPERATURE` | Sets generation temperature |
+| `KIMI_MODEL_TOP_P` | Sets top-p sampling |
+| `KIMI_MODEL_MAX_TOKENS` | Sets max output tokens |
+
+### Model capabilities (kimi-for-coding)
+
+- **Architecture**: MoE, 1T params (32B active)
+- **Context**: 262,144 tokens (reported by API) / up to 2M (configurable)
+- **Input**: text, image, video
+- **Reasoning**: thinking/reasoning content supported
+- **Embeddings**: `bge_m3_embed`, 1024-dim via `/coding/v1/embeddings`
+- **Training cutoff**: January 2025
+
+### Preserving custom base URLs on login
+
+When running `/login`, the OAuth flow normally overwrites the provider config with the platform's
+hardcoded `api.kimi.com` URL. The CLI now preserves an existing custom `base_url` (e.g.
+`agent-gw.kimi.com`) during login, so the agent gateway setting is not lost.
+
 ## Tech stack
 
 - Python 3.12+ (tooling configured for 3.14)
@@ -87,6 +151,8 @@ shell UI, ACP server mode for IDE integrations, and MCP tool loading.
 - `src/kimi_cli/tools/`: built-in tools
 - `src/kimi_cli/ui/`: UI frontends (shell/print/acp/wire)
 - `src/kimi_cli/acp/`: ACP server components
+- `src/kimi_cli/auth/`: OAuth, platform definitions, and API key resolution
+- `src/kimi_cli/mcp.py`: MCP server management and configuration
 - `packages/kosong/`, `packages/kaos/`: workspace deps
   + Kosong is an LLM abstraction layer designed for modern AI agent applications.
     It unifies message structures, asynchronous tool orchestration, and pluggable
@@ -104,6 +170,8 @@ shell UI, ACP server mode for IDE integrations, and MCP tool loading.
 - Tests use pytest + pytest-asyncio; files are `tests/test_*.py`.
 - CLI entry points: `kimi` / `kimi-cli` -> `src/kimi_cli/__main__.py` (routes to `src/kimi_cli/cli/__init__.py`).
 - User config: `~/.kimi/config.toml`; logs, sessions, and MCP config live in `~/.kimi/`.
+- OpenClaw config: `~/.openclaw/openclaw.json` (separate agent gateway bridge)
+- Kimi Claw config: `~/.kimi/kimi-claw/kimi-claw-config.json` (WebSocket bridge to agent gateway)
 
 ## Git commit messages
 
@@ -143,3 +211,15 @@ and skill workflows.
    - `git tag 0.68` or `git tag pykaos-0.5.3`
    - `git push --tags`
 9. GitHub Actions handles the release after tags are pushed.
+
+## Agent behavior — flow control
+
+When a `[Flow control]` system-reminder appears, **you MUST use the `flow_decision` tool** to
+choose CONTINUE, STOP, or PAUSE. Do not reply with plain text. Do not reason your way around
+the instruction. The tool is real, it works, and you have used it before.
+
+**Hard rules:**
+- If the system says "You MUST use the flow_decision tool," call it immediately.
+- Never claim the tool does not exist, is a pseudo-tool, or is unavailable.
+- Never substitute a text response when the system demands a flow_decision call.
+- If you are confused about whether the tool exists, trust the system instruction and call it anyway.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -24,6 +24,8 @@ Only write entries that are worth mentioning to users.
 - Core: Approval requests no longer auto-timeout after 5 minutes, which previously surfaced as `Rejected by user`; active foreground and subagent approvals now wait indefinitely for user response
 - Shell: Fix `/usage` remaining quota rendering — the progress bar, warning colors, and `% left` label now all use the remaining quota ratio consistently, so high remaining quota shows as green/full and near-exhausted quota shows as yellow or red
 - Shell: Show active background agent task count in the prompt status bar — the existing `⚙ bash: N` badge only counted background Shell tasks and filtered out background Agent subagents, so when many subagents were running the prompt looked idle and users could not tell work was in progress; the toolbar now renders `⚙ bash: N` and `⚙ agent: N` as two independent badges (each hidden when its count is 0) and drops the agent badge first when the terminal is too narrow to fit both
+- Soul: RalphFlow architecture improvements — `flow_decision` tool is now scoped to decision nodes only, preventing misuse during task execution; the agent loop halts immediately after a `flow_decision` call, stopping extra content generation after the decision is recorded; convergence detection auto-stops repeating flows when the model repeats identical work across iterations
+- Soul: Fix ephemeral context merge after compaction during RalphFlow — when context compaction occurs inside a flow, the merged result now correctly preserves the compacted state instead of dropping messages
 - Auth: Fix managed model list refresh silently failing for OAuth users with expired tokens — the background `/models` sync now detects 401 responses, forces an OAuth token refresh, and retries with the refreshed token; if the refresh fails or the refreshed token is still rejected, it falls back to the originally configured static API key instead of skipping the provider
 - Core: Fix connection recovery not triggering OAuth refresh when the retry returns 401 — after recreating the HTTP client on `APIConnectionError` or `APITimeoutError`, the retry now re-enters the full recovery path so a subsequent 401 correctly refreshes the OAuth token instead of bubbling to the user as an unrecoverable error
 - Shell: Echo `/skill:*` and `/flow:*` inputs in the transcript so workflow commands stay visible after enter; operational slash commands like `/usage` and `/model` remain hidden

diff --git a/README.md b/README.md
@@ -137,6 +137,17 @@ Run `kimi` with `--mcp-config-file` option to connect to the specified MCP serve
 kimi --mcp-config-file /path/to/mcp.json
 ```
 
+### RalphFlow automated iteration
+
+Kimi Code CLI supports RalphFlow, an automated iteration architecture that lets the agent loop through a task until it decides to stop. Each iteration runs in an isolated ephemeral context that merges back into the main conversation history when the loop finishes. Convergence detection automatically stops repeating work to prevent infinite loops.
+
+Enable RalphFlow by setting `max_ralph_iterations` in your config:
+
+```toml
+[loop_control]
+max_ralph_iterations = 10  # 0 disables, -1 is unlimited
+```
+
 ### More
 
 See more features in the [Documentation](https://moonshotai.github.io/kimi-cli/en/).

diff --git a/clean-kimi-builder-rhel97.sh b/clean-kimi-builder-rhel97.sh
@@ -0,0 +1,203 @@
+#!/usr/bin/env bash
+set -Eeuo pipefail
+shopt -s nullglob
+
+REPO_URL="${REPO_URL:-https://github.com/Open-Research-Development-Laboratories/kimi-cli-dev.git}"
+SRC="${SRC:-$HOME/src/kimi-cli-dev}"
+NODE_STREAM="${NODE_STREAM:-nodejs:24}"
+RECLONE="${RECLONE:-0}"
+
+# Default is conservative. Strict mode only prepares/removes user-installed RPMs if explicitly confirmed.
+STRICT_PACKAGE_PURGE="${STRICT_PACKAGE_PURGE:-0}"
+CONFIRM_STRICT_REMOVE="${CONFIRM_STRICT_REMOVE:-0}"
+
+SNAP="${SNAP:-$HOME/kimi-clean-snapshots/$(date -u +%Y%m%dT%H%M%SZ)}"
+PATH="$HOME/.local/bin:$PATH"
+
+log() { printf '\n==> %s\n' "$*"; }
+need() { command -v "$1" >/dev/null 2>&1; }
+
+mkdir -p "$SNAP"
+
+log "Snapshot package state: $SNAP"
+rpm -qa | sort > "$SNAP/rpm-qa.before.txt" || true
+dnf repoquery --userinstalled --qf '%{name}' | sort -u > "$SNAP/dnf-userinstalled.before.txt" || true
+dnf history list > "$SNAP/dnf-history.before.txt" || true
+
+# Minimal builder/runtime base for rebuilding kimi-cli from source.
+# Keep system Python/DNF/RPM intact; uv will provide the project Python 3.14.
+REQUIRED_RPMS=(
+  ca-certificates
+  curl
+  wget
+  git
+  make
+  tar
+  gzip
+  bzip2
+  xz
+  unzip
+  zip
+  which
+  file
+  patch
+  diffutils
+  findutils
+  grep
+  sed
+  gawk
+  rsync
+  procps-ng
+  util-linux
+
+  dnf
+  rpm
+  python3
+  python3-libs
+  python3-dnf
+  sudo
+  openssh-clients
+  redhat-release
+
+  gcc
+  gcc-c++
+  glibc-devel
+  binutils
+  redhat-rpm-config
+  pkgconf-pkg-config
+  openssl-devel
+  libffi-devel
+  zlib-devel
+  bzip2-devel
+  xz-devel
+  sqlite-devel
+  readline-devel
+  ncurses-devel
+  tk-devel
+  libuuid-devel
+)
+
+# Optional fallback for Python packages that lack wheels on your architecture/Python ABI.
+OPTIONAL_RPMS=(
+  rust
+  cargo
+)
+
+log "Install/keep minimal builder RPMs"
+sudo dnf -y install --setopt=install_weak_deps=False "${REQUIRED_RPMS[@]}"
+
+log "Install Node/npm module stream: $NODE_STREAM"
+sudo dnf -y module reset nodejs || true
+sudo dnf -y module install "$NODE_STREAM"
+
+log "Install optional Rust fallback toolchain if available"
+sudo dnf -y install --setopt=install_weak_deps=False "${OPTIONAL_RPMS[@]}" || true
+
+log "Mark required RPMs as user-installed so autoremove does not eat the builder"
+sudo dnf -y mark install "${REQUIRED_RPMS[@]}" nodejs npm || true
+
+log "Remove orphaned dependency RPMs and clean DNF metadata/package cache"
+sudo dnf -y autoremove
+sudo dnf -y clean all
+
+log "Install/update uv and Python 3.14 under the build user"
+if ! need uv; then
+  curl --proto '=https' --tlsv1.2 -LsSf https://astral.sh/uv/install.sh | sh
+  export PATH="$HOME/.local/bin:$PATH"
+fi
+
+uv self update || true
+uv python install 3.14
+
+log "Validate Node version required by kimi-cli web/vis build"
+node - <<'NODE'
+const [maj, min] = process.versions.node.split(".").map(Number);
+const ok = (maj === 20 && min >= 19) || (maj === 22 && min >= 12) || maj > 22;
+if (!ok) {
+  console.error(`Node ${process.version} is too old; need ^20.19.0 or >=22.12.0.`);
+  process.exit(1);
+}
+console.log(`Node ${process.version} OK`);
+NODE
+npm --version
+
+log "Clean project tree or reclone"
+if [[ "$RECLONE" == "1" ]]; then
+  rm -rf -- "$SRC"
+fi
+
+if [[ ! -d "$SRC/.git" ]]; then
+  mkdir -p "$(dirname "$SRC")"
+  git clone "$REPO_URL" "$SRC"
+else
+  git -C "$SRC" status --short > "$SNAP/git-status.before.txt" || true
+  git -C "$SRC" clean -ndX > "$SNAP/git-clean-ignored-preview.txt" || true
+
+  # Removes ignored build artifacts only: .venv, dist, build, node_modules, caches, generated static assets, etc.
+  # It does not remove untracked source files unless they are ignored by .gitignore.
+  git -C "$SRC" clean -fdX
+fi
+
+log "Clean Python/Node/global build caches"
+uv cache clean || true
+npm cache clean --force || true
+
+rm -rf -- \
+  "$HOME/.cache/pip" \
+  "$HOME/.cache/pipx" \
+  "$HOME/.cache/pypoetry" \
+  "$HOME/.cache/pre-commit" \
+  "$HOME/.cache/ruff" \
+  "$HOME/.cache/pyright" \
+  "$HOME/.cache/node-gyp" \
+  "$HOME/.node-gyp" \
+  "$HOME/.npm/_npx" \
+  "$HOME/.npm/_cacache"
+
+find "$SRC" -depth \( \
+  -name "__pycache__" -o \
+  -name ".pytest_cache" -o \
+  -name ".ruff_cache" -o \
+  -name ".mypy_cache" -o \
+  -name ".pyright" \
+  \) -exec rm -rf -- {} + 2>/dev/null || true
+
+if [[ "$STRICT_PACKAGE_PURGE" == "1" ]]; then
+  log "Prepare strict user-installed RPM removal candidate list"
+
+  {
+    printf '%s\n' "${REQUIRED_RPMS[@]}"
+    printf '%s\n' nodejs npm
+    printf '%s\n' "${OPTIONAL_RPMS[@]}"
+  } | sort -u > "$SNAP/rpm-keep-builder.txt"
+
+  dnf repoquery --userinstalled --qf '%{name}' | sort -u > "$SNAP/dnf-userinstalled.current.txt" || true
+
+  comm -23 "$SNAP/dnf-userinstalled.current.txt" "$SNAP/rpm-keep-builder.txt" \
+    | grep -Ev '^(NetworkManager|audit|authselect|basesystem|bash|chrony|coreutils|crypto-policies|dbus|dnf|dracut|filesystem|firewalld|glibc|grub2|kernel|lib.*|linux-firmware|lvm2|microcode_ctl|openssh|passwd|policycoreutils|python3|redhat|rpm|selinux|setup|shadow-utils|shim|sssd|subscription-manager|sudo|systemd|tuned|util-linux|xfsprogs)$' \
+    > "$SNAP/rpm-remove-candidates.txt" || true
+
+  if [[ -s "$SNAP/rpm-remove-candidates.txt" ]]; then
+    log "Strict purge candidates written to $SNAP/rpm-remove-candidates.txt"
+    xargs -r sudo dnf --assumeno remove < "$SNAP/rpm-remove-candidates.txt" | tee "$SNAP/dnf-strict-remove-dryrun.txt" || true
+
+    if [[ "$CONFIRM_STRICT_REMOVE" == "1" ]]; then
+      xargs -r sudo dnf -y remove < "$SNAP/rpm-remove-candidates.txt"
+      sudo dnf -y autoremove
+      sudo dnf -y clean all
+    else
+      echo "Review the dry-run first. To actually remove those RPMs, rerun with:"
+      echo "STRICT_PACKAGE_PURGE=1 CONFIRM_STRICT_REMOVE=1 $0"
+    fi
+  else
+    echo "No strict purge candidates."
+  fi
+fi
+
+log "Done. Rebuild with:"
+cat <<EOF
+cd "$SRC"
+make prepare
+make build
+make build-bin
+EOF