chill-agent 1.2.0 → 1.4.0

package/commands/chill-audit.toml

@@ -0,0 +1,2 @@
+description = "Quality audit of the whole repo: bugs, security, load, tests, speed, bloat"
+prompt = "Audit the whole repository (or the folder the user names) like the senior developer who just inherited it. Map first: README, deploy and build config, dependencies, entry points, tests; find the expected load (one person or many users and processes); trace the main flows and read fully the paths with user input, money, auth, data writes, background jobs and shared state. Look for, in this order: bugs, risk (security holes, data loss), scale (races, the same work done by every process, memory that only grows, a query per item, O(n^2)), risky logic without a test, slow code (big = problem, small = suggestion), lean (dead code, duplicate helpers, stdlib or native instead of custom code or a dependency, abstractions with one implementation, near-copies to merge, a function doing several unrelated jobs to split, never split by line count). Before calling code unused, grep the whole tree including tests, fixtures, config and string references. Every finding needs a concrete case. A shortcut marked with a shortcut: (or older chill:) comment that names its limit is a decision, not a finding. Propose the smallest fix, prefer deleting code, never add layers. Very simple English. Start with 'What this repo does:' and the load you assumed. Then at most 20 numbered findings grouped Must fix / Should fix / Nice to have, most important first; each: a bold title with file:lines, then four bullets: What this is, Problem, Fix, If we skip it. End with 'Verdict:', 'Lean: -N lines, -M dependencies possible.' and 'Not checked:' for the parts you did not read. Change no code."

package/commands/chill-debt.toml

@@ -0,0 +1,2 @@
+description = "Harvest shortcut: comments into a tracked debt ledger"
+prompt = "Harvest every `shortcut:` comment (and older `chill:` ones) in this repository into a debt ledger so deferrals do not rot into 'later means never'. Grep the whole tree for comment markers (git grep --no-index --exclude-standard -nE '(#|//|/[*]) ?(shortcut|chill):', skipping .git and everything .gitignore ignores; if the user names their own marker word, grep for that instead; skip hits that are not a deferral, like a keyboard-shortcut note). One row per marker, grouped by file: <file>:<line> — <what was simplified>. ceiling: <the limit named in the comment>. upgrade: <the trigger to revisit>. Tag any marker that names no upgrade path or trigger as no-trigger, those rot silently. End with the count of markers and how many lack a trigger. If none: 'No shortcut debt. Clean ledger.' Report only, change nothing."

package/commands/chill-gain.toml

@@ -0,0 +1,2 @@
+description = "Show chill's measured impact scoreboard (less code, cost, time)"
+prompt = "Show the chill gain scoreboard. One shot, change nothing: do not switch mode, write flag files, or persist anything. Render the published agentic benchmark of ChillAgent 5 (39 tasks on a real FastAPI + React repo plus bug fixes, security and privacy cases and small apps, Opus 5.5, 5 runs each, against the same agent without the skill; geometric mean of per-task medians; source benchmarks/results/2026-10-07-agentic.md and the README) as plain ASCII bars, chill as a share of the no-skill baseline (100%): Lines of code 47% (down 53%); Output tokens 55% (down 45%); Cost 74% (down 26%); Time 59% (down 41%); Hidden checks passed 97% (no-skill 96%; 18 of the tasks have hidden checks for correctness and safety); Tests where the logic needs one 98% (no-skill 68%). The bar length shows the share, the label carries the exact figure. These are benchmark averages, not this repo. NEVER print a per-repo savings number: the unbuilt version was never written, so there is no real baseline to subtract from in a live repo. For real per-repo figures, point to /chill-debt (the counted shortcut ledger) and /chill-audit (what is still cuttable). Report only."

package/commands/chill-help.toml

@@ -0,0 +1,2 @@
+description = "Quick reference for chill levels, skills, and commands"
+prompt = "Show the chill quick reference. One shot, change nothing: do not switch mode, write flag files, or persist anything. Levels: /chill lite (build what's asked, name the lazier alternative in one line), /chill (full, the default ladder: YAGNI then stdlib then native then one line then minimum), /chill ultra (deletion before addition, challenges the requirement before building). Commands: /chill-review (over-engineering review of the current changes), /chill-audit (whole-repo over-engineering audit), /chill-debt (harvest shortcut: comments into a tracked ledger), /chill-gain (measured-impact scoreboard from the benchmark), /chill-help (this card). Deactivate with 'stop chill', 'normal mode', or /chill off; resume anytime with /chill. Default mode is full; change it with the PONYTAIL_DEFAULT_MODE environment variable (off|lite|full|ultra) or a config file at ~/.config/chill/config.json (Windows: %APPDATA%\\chill\\config.json) with {\"defaultMode\": \"lite\"}. Resolution order: env var, then config file, then full."

package/commands/chill-review.toml

@@ -0,0 +1,2 @@
+description = "Quality review of the current changes: bugs, security, load, tests, speed, bloat"
+prompt = "Review the current code changes like the senior developer who will be paged when they break. Review what the user names (uncommitted or staged changes, a branch, a PR link, files); nothing named: the uncommitted changes, or the last commit if there are none. First read the code the change touches: callers of every changed function, the functions it calls, the tests, the README (find the expected load: one person or many users and processes). A change can break code it does not touch. Look for, in this order: bugs (wrong result, crash, missed edge case, a caller broken by the change, a fix in one caller while the shared function stays broken), risk (security holes, data loss), scale (check-then-write races, the same work done by every process, memory that only grows, a query per item, O(n^2) on big input), risky logic without a test, slow code (big = problem, small = suggestion), lean (dead code, a helper the repo already has, stdlib or native instead of custom code or a new dependency, abstractions with one implementation, near-copies to merge, a function doing several unrelated jobs to split, never split by line count). Every finding needs a concrete case; re-read the lines to confirm it. A shortcut marked with a shortcut: (or older chill:) comment that names its limit is a decision, not a finding. Propose the smallest fix, prefer deleting code, never add layers. No style taste, no vague worries. Very simple English, explain technical terms. Start with 'What this change does:' in two or three sentences and the load you assumed. Then numbered findings grouped Must fix / Should fix / Nice to have; each finding: a bold title with file:lines, then four bullets: What this is, Problem, Fix, If we skip it (one or two short sentences each). End with 'Verdict:', 'Lean: -N lines possible.' when lean findings exist, and 'Not checked:' if something mattered and you could not check it. Nothing found: 'Looks good. Ship.' Change no code."

package/commands/chill.toml

@@ -0,0 +1,2 @@
+description = "Switch chill intensity level (lite/full/ultra/off)"
+prompt = "Switch to chill {{args}} mode. With no level, keep the current level, or switch on at the default level if chill was off. Lazy senior dev mode, before any code: does it need to exist at all (YAGNI)? Does the standard library do it? A native platform feature? Can it be one line? Build the minimum that works. No unrequested abstractions, no avoidable dependencies, no boilerplate. Mark deliberate simplifications that cut a real corner with a known ceiling using a shortcut: comment that names the ceiling and upgrade path."

package/docs/agent-portability.md

@@ -0,0 +1,62 @@
+# Agent Portability
+
+ChillAgent is an agent-portable skill distribution. The skills in `skills/` hold
+the core behavior; host-specific files are adapters that make that behavior easy
+to load in a given agent.
+
+## Supported Adapters
+
+| Host | Files | Notes |
+|------|-------|-------|
+| Claude Code | `.claude-plugin/plugin.json`, `commands/`, `hooks/claude-codex-hooks.json`, `hooks/` | Full plugin install with session activation, mode tracking, commands, and statusline support. |
+| Codex | `.codex-plugin/plugin.json`, `hooks/claude-codex-hooks.json`, `hooks/`, `skills/` | Plugin install with the same skills plus lifecycle hooks for activation and mode tracking. |
+| Devin CLI | `.devin-plugin/plugin.json`, `skills/` | `devin plugins install DietrichGebert/chill-agent` installs chill-agent as a Devin plugin; skills are available as `/chill-agent:chill-agent`, `/chill-agent:chill-agent-review`, and so on. |
+| CodeBuddy | `.claude-plugin/plugin.json`, `.claude-plugin/marketplace.json`, `hooks/claude-codex-hooks.json`, `hooks/`, `skills/` | CodeBuddy Code reads the Claude-format manifests when no `.codebuddy-plugin/` exists and expands `${CLAUDE_PLUGIN_ROOT}` in hook commands, so `/plugin marketplace add DietrichGebert/chill-agent` + `/plugin install chill-agent@chill-agent` installs the Claude plugin unchanged. The hooks detect it by `CODEBUDDY_PLUGIN_ROOT`, keep the mode flag in `$CODEBUDDY_CONFIG_DIR` or `~/.codebuddy`, and answer in `hookSpecificOutput` JSON. No subagent injection: CodeBuddy discards `SubagentStart` hook output. Verified with CodeBuddy CLI 2.161.2. |
+| Grok Build | root `plugin.json`, `.grok-plugin/marketplace.json`, `skills/`, `commands/` | `grok plugin install DietrichGebert/chill-agent --trust`, then enable. Grok can auto-invoke chill-agent from its coding-task skill description; `/chill-agent` makes activation explicit. Grok lifecycle hooks are not used because passive hook output cannot inject instructions. |
+| Goose | `plugin.json`, `skills/`, `AGENTS.md` | `goose plugin install https://github.com/DietrichGebert/chill-agent.git` clones the repo into `~/.agents/plugins/chill-agent/` and imports the six skills as `chill-agent:<skill>`. Always-on rules come from `AGENTS.md`: goose reads it from the project (git root down to the working directory) and from `~/.agents/AGENTS.md`, plus `.goosehints` from `~/.config/goose/`. No lifecycle hooks, so no persistent `/chill-agent` levels. Verified with Goose 1.53.0. |
+| Muse Code (Meta) | `AGENTS.md`, `skills/` | Muse reads `AGENTS.md` (or `CLAUDE.md`) in the project once the workspace is trusted, and `~/.config/muse/AGENTS.md` for every project. Skills: copy the folders from `skills/` to `~/.agents/skills/`, or run `muse skills install ./skills/<name> --scope user` per skill. Instruction-tier, no `/chill-agent` levels or hooks. Verified with Muse Code 1.4.2. |
+| OpenCode | `.opencode/plugins/chill-agent.mjs`, `.opencode/command/`, `hooks/`, `skills/` | Server plugin injects the ruleset each turn via `experimental.chat.system.transform` and persists `/chill-agent` switches; reuses the shared instruction builder. Installs globally with `opencode plugin add @dietrichgebert/chill-agent` (OpenCode writes the `plugins` entry itself), or per project through the `plugins` array in `opencode.json`. |
+| Kilo Code | `.opencode/plugins/chill-agent.mjs` (npm), `AGENTS.md` | Kilo is built on OpenCode and runs chill-agent's OpenCode plugin unchanged: add `{ "plugin": ["@dietrichgebert/chill-agent"] }` to `kilo.jsonc` (project) or `~/.config/kilo/kilo.jsonc` (global). That injects the ruleset every turn, keeps `/chill-agent` level switches, and registers `/chill-agent-review` and the other commands. Without the plugin, Kilo still loads `AGENTS.md` from the project root or `~/.config/kilo/AGENTS.md` (instruction-tier). Verified with Kilo CLI 7.8.3, which shares its core with the VS Code extension. |
+| pi | `pi-extension/`, `skills/`, `hooks/` | Package extension: injects the ruleset each turn through the shared instruction builder and registers the `/chill-agent` commands. |
+| Oh My Pi (omp) | `.claude-plugin/marketplace.json`, `pi-extension/`, `skills/` | Installs through the marketplace and runs the pi extension unchanged (omp reads the `pi` key in `package.json`). Verified with omp 18.6.1: the ruleset arrives every turn and the skills are listed. |
+| Kimi Code | `.kimi-plugin/plugin.json`, `AGENTS.md`, `skills/` | Native Kimi plugin: `systemPromptPath` puts `AGENTS.md` into the system prompt every turn, and the six skills double as slash commands. No hooks. Verified with Kimi Code 2.1.1 over ACP: the ruleset is in the system prompt and `/chill-agent-review staged` activates the skill with its target. |
+| Hermes Agent | `plugin.yaml`, `__init__.py`, `skills/` | Native Hermes plugin: injects active mode through `pre_llm_call`, rewrites gateway `/chill-agent-*` skill commands into agent prompts, registers `/chill-agent` mode switching, and exposes bundled skills as `chill-agent:<skill>`. |
+| OpenClaw | `.openclaw/skills/`, `skills/` | `clawhub install chill-agent` installs the skill from ClawHub, and the other five install the same way. `.openclaw/skills/` is generated from `skills/` by `node scripts/build-openclaw-skills.js`. |
+| Gemini CLI | `gemini-extension.json`, `AGENTS.md`, `commands/`, `skills/` | Extension manifest points `contextFileName` at `AGENTS.md` for always-on rules, and reuses the existing `commands/*.toml` and `skills/`, which Gemini CLI auto-discovers. The Claude/Codex hook map is not placed at Gemini's auto-discovered `hooks/hooks.json` path. |
+| Qwen Code | `gemini-extension.json`, `AGENTS.md`, `commands/`, `skills/` | `qwen extensions install DietrichGebert/chill-agent:chill-agent` installs through Qwen's Gemini CLI compatibility: `AGENTS.md` becomes always-on context, the six `/chill-agent` commands are migrated, and the six skills are registered. The Claude/Codex lifecycle hooks are not used, so levels don't persist. Verified with Qwen Code 0.24.7 and chill-agent 4.11.0 (needs 4.11.0 or later, see #1021). |
+| Cursor | `hooks/cursor-hooks.json`, `scripts/cursor-hooks.js`, `hooks/`, `.cursor/rules/chill-agent.mdc` | Native hooks: `node scripts/cursor-hooks.js install` merges `sessionStart` (default-level ruleset via `additional_context`) and `beforeSubmitPrompt` (`/chill-agent` level tracking, new-level ruleset via `additional_context`) into `~/.cursor/hooks.json`, or `.cursor/hooks.json` with `--project`, keeping unrelated hooks. No subagent injection (Cursor's `subagentStart` takes only `permission`/`user_message`) and no `sessionStart` in cloud agents. `.cursor/rules/chill-agent.mdc` stays the instruction-only alternative; while that rule is in a workspace the hooks inject only a notice. Contract and verification record: [cursor-hooks.md](cursor-hooks.md). |
+| Windsurf | `.windsurf/rules/chill-agent.md` | Project rule. |
+| Cline (extension and CLI) | `.clinerules/chill-agent.md`, `skills/`, `AGENTS.md` | Copy `.clinerules/chill-agent.md` into a project's `.clinerules/`, or into `~/.cline/rules/` (or `~/Documents/Cline/Rules/`) for every project; Cline also reads `AGENTS.md` from the project root. `cline skill add DietrichGebert/chill-agent` installs the six skills into `.agents/skills/`. Instruction-tier, no `/chill-agent` levels or hooks. Rules and skills verified with Cline CLI 3.0.68. |
+| GitHub Copilot | `.github/copilot-instructions.md` | Repository instruction file. |
+| GitHub Copilot CLI | `.github/plugin/`, `AGENTS.md`, `.github/copilot-instructions.md`, `~/.copilot/copilot-instructions.md` | Plugin-supported (`copilot plugin marketplace add DietrichGebert/chill-agent` + `copilot plugin install chill-agent@chill-agent`). Fallback instruction mode remains: per-project from `AGENTS.md` or `.github/copilot-instructions.md`, or globally from `~/.copilot/copilot-instructions.md` (instruction-tier, no `/chill-agent` levels or hooks). |
+| Antigravity (IDE) | `AGENTS.md` | Reads `AGENTS.md` at the repo root as always-on rules (like `.cursorrules`/`CLAUDE.md`); `.agents/rules/` also works for workspace rules. Instruction-tier. |
+| Antigravity CLI | `plugin.json`, `skills/`, `commands/`, `AGENTS.md` | `agy plugin install https://github.com/DietrichGebert/chill-agent` installs the six skills and converts the six `commands/` into skills too, called as slash commands. No lifecycle hooks, so no persistent `/chill-agent` levels. `AGENTS.md` in the project gives the always-on ruleset. Install verified with Antigravity CLI 1.3.3. |
+| CodeWhale | `AGENTS.md` | Reads `AGENTS.md` from the repo root as project instructions; also reads `CLAUDE.md` and `.claude/instructions.md` as fallbacks. Instruction-tier. |
+| VS Code + Codex extension | `AGENTS.md` | The Codex extension reads `AGENTS.md` (repo root, or `~/.codex/AGENTS.md` globally). Instruction-tier; the full Codex plugin row above adds `/chill-agent` levels and hooks. |
+| JetBrains Junie | `AGENTS.md` | Junie reads `AGENTS.md` once you point it there in Settings → Tools → Junie → Project Settings → Guidelines Path (not automatic yet); this repo ships `AGENTS.md`, and `.junie/guidelines.md` is Junie's legacy path. Instruction-tier. |
+| Amp (Sourcegraph) | `AGENTS.md` | Amp reads `AGENTS.md` from the working directory and parent directories up to `$HOME` (plus global config like `~/.config/amp/AGENTS.md`); falls back to `AGENT.md`/`CLAUDE.md`. Instruction-tier. |
+| Factory Droid | `AGENTS.md`, `skills/` | Droid reads `AGENTS.md` from the working directory up to the git root (also inside `.factory/` and `.agents/` there) and globally from `~/.factory/AGENTS.md`. It loads `SKILL.md` skills from `.factory/skills/` or `~/.factory/skills/`, so copy the folders from `skills/` there for review, audit and the other skills. Instruction-tier, no `/chill-agent` levels or hooks. |
+| Command Code | `skills/`, `AGENTS.md` | `cmd skills add https://github.com/DietrichGebert/chill-agent` installs the six skills into `.commandcode/skills/` (add `--global` for `~/.commandcode/skills/`); `.agents/skills/` is auto-discovered too. Reads `AGENTS.md` as project memory every session. Instruction-tier, no `/chill-agent` levels or hooks. |
+| Freebuff CLI | `AGENTS.md`, `skills/` | Freebuff puts the project-root `AGENTS.md` (or `CLAUDE.md` if there is none) into every session as project instructions, and reads `~/.agents.md` globally. It loads `SKILL.md` skills from `.agents/skills/` or `.claude/skills/` in the project and from `~/.agents/skills/` or `~/.claude/skills/` globally, so copy the folders from `skills/` there for review, audit and the other skills. Instruction-tier, no `/chill-agent` levels or hooks. |
+| DeepSeek Harness (dsh) | `AGENTS.md`, `skills/` | DSH loads `AGENTS.md` (or `CLAUDE.md`) from the project root down to the working directory, plus `~/.dsh/AGENTS.md` (`$DSH_HOME`) for every project, into the first request of each session. It loads `SKILL.md` skills from `~/.agents/skills/`, so copy the folders from `skills/` there for review, audit and the other skills. Instruction-tier, no `/chill-agent` levels or hooks. Verified with dsh 0.2.0-rc.2. |
+| Jules (Google) | `AGENTS.md` | Jules automatically reads `AGENTS.md` from the repository root. Instruction-tier. |
+| Kiro | `.kiro/steering/chill-agent.md` | Steering rule; copy globally or into a project. |
+| Qoder | `.qoder/rules/chill-agent.md`, `.qoder-plugin/plugin.json`, `hooks/qoder-hooks.json`, `skills/`, `AGENTS.md` | Qoder auto-loads `AGENTS.md` as always-on context; `.qoder/rules/chill-agent.md` provides per-project rules; the plugin manifest points at `skills/` for the six chill-agent skills (invoked as `/chill-agent`, `/chill-agent-review`, etc. via the Skill system). Full plugin-tier: `hooks/qoder-hooks.json` template registers `UserPromptSubmit` (mode activation + ruleset injection) and `PreToolUse` with `task|Task` matcher (subagent injection). Instruction-tier works from repo root with zero setup via `AGENTS.md`. |
+| Zed | `AGENTS.md` | Auto-includes `AGENTS.md` from the worktree root as one of its default rule files for the Agent Panel. Instruction-tier. |
+| Generic agents | `AGENTS.md` or `skills/*/SKILL.md` | Copy the compact rule file or load the skill files directly. |
+
+## Adapter Rule
+
+Keep adapters thin. When a host supports skills or hooks, point it at the
+existing `skills/` and `hooks/` files. When a host only supports project
+instructions, keep its copied rule text aligned with `AGENTS.md`.
+
+## Portable Behavior
+
+- `skills/chill-agent/SKILL.md`: lazy senior dev mode
+- `skills/chill-agent-review/SKILL.md`: quality review of a diff
+- `skills/chill-agent-audit/SKILL.md`: quality audit of the whole repo
+- `skills/chill-agent-debt/SKILL.md`: harvest `shortcut:` comments into a tracked ledger
+- `skills/chill-agent-gain/SKILL.md`: measured-impact scoreboard from the benchmark
+- `skills/chill-agent-help/SKILL.md`: quick reference
+- `AGENTS.md`: compact always-on instruction set for agents without skill support

package/docs/cursor-hooks.md

@@ -0,0 +1,200 @@
+# Cursor hooks
+
+Native Cursor support through `hooks.json` (issue #817). The adapter reuses the
+shared hook scripts; the only Cursor-specific pieces are the hooks template, the
+install script and a small output branch in `hooks/chill-agent-runtime.js`.
+
+| File | Role |
+|------|------|
+| `hooks/cursor-hooks.json` | Template: `sessionStart` and `beforeSubmitPrompt` entries with a `PONYTAIL_DIR` placeholder. |
+| `scripts/cursor-hooks.js` | `install` / `uninstall`, merges into `~/.cursor/hooks.json` (or `.cursor/hooks.json` with `--project`). |
+| `hooks/chill-agent-activate.js` | `sessionStart`: injects the default-level ruleset. |
+| `hooks/chill-agent-mode-tracker.js` | `beforeSubmitPrompt`: tracks `/chill-agent` commands, injects the new level's ruleset. |
+| `hooks/chill-agent-runtime.js` | Detects Cursor (`CURSOR_VERSION`), keeps per-project mode state and a shared compatibility flag, emits Cursor-shaped JSON. |
+
+## Install and uninstall
+
+```bash
+git clone https://github.com/DietrichGebert/chill-agent
+node chill-agent/scripts/cursor-hooks.js install            # ~/.cursor/hooks.json, every project
+node chill-agent/scripts/cursor-hooks.js install --project  # <cwd>/.cursor/hooks.json, this project only
+node chill-agent/scripts/cursor-hooks.js uninstall          # add --project for the project file
+```
+
+What the install script does:
+
+- Reads the target file if it exists and keeps every hook that is not chill-agent's.
+  ChillAgent's entries are the ones whose `command` runs a `hooks/chill-agent-*.js`
+  script; they are replaced on re-install, so running it twice never duplicates.
+- Replaces `PONYTAIL_DIR` with the checkout's absolute path, forward slashes, so
+  the command runs unchanged under cmd, PowerShell and bash. A checkout path with
+  shell metacharacters is refused; copy the template by hand in that case.
+- Refuses to touch a `hooks.json` that is not valid JSON, and says so.
+- `uninstall` removes only chill-agent's entries and deletes the file when nothing
+  else was in it. `node scripts/uninstall.js` runs the same removal for the user
+  file and also deletes `~/.cursor/.chill-agent-active`.
+
+Cursor watches `hooks.json` and reloads it on save; open a new chat afterwards.
+`node` has to be on the PATH Cursor sees. The Hooks tab under Customize and the
+Hooks output channel show each execution and any parse errors.
+
+## Contract
+
+Sources: the Cursor hooks docs (`cursor.com/docs/hooks`, read 2026-09-14) and the
+Cursor 3.20.17 client on Windows, whose hook runner and response validators were
+read directly. "Docs" below means the page documents it; "client" means it was
+verified in the shipped code.
+
+| Event | ChillAgent uses | Delivery | Source |
+|-------|---------------|----------|--------|
+| `sessionStart` | output `additional_context` | Stored on the conversation and sent as system context with every request of that conversation. | Docs (field). Client (persistence: the value is kept on the composer as `hooksAdditionalContext` and attached to each request). |
+| `beforeSubmitPrompt` | input `prompt`; output `continue: true` plus `additional_context` | Wrapped as a system reminder for that turn. Inline up to 10,000 characters; longer payloads are written to a file the agent is told to read; above 1,000,000 the payload is dropped. | Docs list only `continue` and `user_message`. Client: the response validator accepts `additional_context` and the submit path injects it. Undocumented, so treat it as version-dependent. |
+| `subagentStart` | not registered | None. | Docs and client: the output schema is `permission` and `user_message` only. The client's protobuf has an unused `additional_context` slot that the hook path never fills. |
+| `preToolUse`, `postToolUse` | not registered | `additional_context` exists on both, but it would cost a process per tool call and `beforeSubmitPrompt` already covers mode changes. | Docs. |
+
+Output rules the runtime follows:
+
+- Cursor parses stdout as JSON. Empty stdout means "nothing to add"; raw text is
+  logged as a parse error and ignored. The Cursor branch therefore prints either
+  one JSON object or nothing (`off` mode, ordinary prompts).
+- `user_message` on `beforeSubmitPrompt` is shown only when `continue` is
+  `false`, so chill-agent never sets it. Confirmations reach the user through the
+  model's own reply.
+- Exit code is always 0. Cursor treats exit code 2 as "block" and other non-zero
+  codes as fail-open; chill-agent never blocks anything.
+
+Execution environment (client, 3.20.17):
+
+- Every hook process gets `CURSOR_VERSION`, `CURSOR_PROJECT_DIR` and its alias
+  `CLAUDE_PROJECT_DIR`. `CURSOR_VERSION` is assigned in exactly one place, the
+  hook environment builder, so it does not leak into terminals inside Cursor.
+  ChillAgent uses it for host detection and keeps state in `~/.cursor/`.
+- On Windows the payload is written to a temp file and the command runs inside
+  `powershell -NoProfile -NonInteractive -ExecutionPolicy Bypass` as
+  `Get-Content -LiteralPath <file> -Raw | & { $input | <command> }`, so the hook
+  still reads its JSON from stdin. On macOS and Linux it is piped directly.
+- Cursor can also run hooks declared by Claude-format plugins and then sets
+  `CLAUDE_PLUGIN_ROOT` next to `CURSOR_VERSION`. The runtime prefers the Cursor
+  output shape in that case. Installing chill-agent that way was not tested.
+
+## Behavior
+
+- New conversation: `sessionStart` sets the live mode to the default level
+  (`PONYTAIL_DEFAULT_MODE`, then `config.json`, then `full`) and
+  injects `PONYTAIL MODE ACTIVE — level: <level>` followed by the ruleset filtered
+  to that level. Default `off`: no flag, no output.
+- `/chill-agent lite|full|ultra` sent as a plain message: the flag changes and the
+  turn receives `PONYTAIL MODE CHANGED — level: <level>` plus that level's ruleset
+  (about 5,300 characters, under the inline cap). Cursor has no `/chill-agent`
+  command to load the skill body, so the hook carries it. `@chill-agent` and
+  `$chill-agent` are parsed too, but `@` opens Cursor's context picker. If the
+  chill-agent skills are also installed under `~/.cursor/skills`, Cursor treats
+  `/chill-agent lite` as a manual skill attachment and inlines the full, unfiltered
+  skill body into that message as well; the hook still receives the literal
+  `/chill-agent lite` and remains the thing that tracks the level.
+- `/chill-agent off`, `stop chill-agent`, `normal mode`: the flag is removed and the
+  turn receives `PONYTAIL MODE OFF`. The ruleset injected at `sessionStart` stays
+  in the conversation's system context; the notice is what tells the model to
+  stop applying it, the same as in Claude Code.
+- `/chill-agent`: reports `PONYTAIL MODE ACTIVE — level: <level>` without changing
+  anything. `/chill-agent default <level>` persists the default to `config.json`.
+- Any other prompt: no output.
+
+### Coexistence with `.cursor/rules/chill-agent.mdc`
+
+The always-on rule and the hooks are alternatives, not layers. The rule already
+puts the compact ruleset in front of every prompt, and no hook can remove a rule
+from context, so `off` cannot win against it and `lite` or `ultra` would
+contradict it. While `<workspace>/.cursor/rules/chill-agent.mdc` exists (first
+workspace root, from `CURSOR_PROJECT_DIR` or the working directory):
+
+- `sessionStart` injects a one-line notice instead of the ruleset and leaves the
+  mode flag alone.
+- `/chill-agent ...`, `stop chill-agent` and `normal mode` answer with the same notice
+  and change nothing.
+
+Delete the rule to let the hooks manage the level. A project that keeps the rule
+for teammates without hooks stays on the rule's fixed behavior for everyone.
+
+## Limitations
+
+- Subagents never receive the ruleset. `subagentStart` can only allow or deny,
+  and `preToolUse` `updated_input` on the `Task` tool would mean guessing the
+  undocumented shape of the subagent prompt. `PONYTAIL_SUBAGENT_MATCHER` has no
+  effect in Cursor.
+- Cloud agents do not run `sessionStart` (documented), so there is no startup
+  injection there. Project-level `beforeSubmitPrompt` still runs, so `/chill-agent
+  <level>` sets the level for the rest of that conversation.
+- `sessionStart` is fire-and-forget. A prompt sent within the first fraction of a
+  second of a new chat can leave before the context is attached.
+- On Windows every hook run costs about a second, mostly PowerShell startup
+  (measured 1.05 to 1.3 s on 3.20.17). `beforeSubmitPrompt` is awaited, so each
+  prompt submission waits that long. macOS and Linux spawn the command directly
+  and pay only node startup.
+- With `CLAUDE_PROJECT_DIR`, mode state is per project, stored in
+  `~/.cursor/chill-agent-modes/<sha256(normalized project dir)>`. Conversations in
+  the same project share that flag; different projects keep separate modes.
+  Without a project directory, the runtime uses `~/.cursor/.chill-agent-active`.
+  That shared flag is also maintained as a compatibility/statusline mirror
+  and follows the last mode change.
+- The `beforeSubmitPrompt` injection field is not on the docs page. If a future
+  Cursor build drops it, level switches would still update the flag but nothing
+  would reach the model; only the startup injection would remain.
+
+## Verification record
+
+### Automated compatibility check
+
+`node --test tests/cursor-hooks.test.js` feeds each hook the Cursor input shape
+and asserts the output shape: template validity, `sessionStart` JSON and flag
+placement, `off`, the Claude-plugin environment, every `/chill-agent` form on
+`beforeSubmitPrompt`, silence on ordinary prompts, the rule-coexistence notice
+from both `CURSOR_PROJECT_DIR` and the working directory, and the installer's
+merge, idempotence, project scope, file removal and malformed-file refusal.
+`tests/uninstall.test.js` covers the shared uninstall script.
+
+### Client source check, Cursor 3.20.17 on Windows, 2026-09-14
+
+Confirmed by reading the shipped client: the response validators for
+`sessionStart` (`env`, `additional_context`), `beforeSubmitPrompt` (`continue`,
+`user_message`, `additional_context`) and `subagentStart` (`permission`,
+`user_message`); the persistence of `sessionStart` context on the conversation;
+the submit path that injects the `beforeSubmitPrompt` context as a system
+reminder with the 10,000-character inline cap; the PowerShell command wrapper and
+the hook environment variables.
+
+### Live session check
+
+Run on 2026-09-14 with Cursor 3.20.17 on Windows 11, user-level hooks, model
+`gpt-5.6-sol-high`; steps still open are marked in the table. Hook execution is
+not delivery, so each step asks the model to quote the injected header. Record
+the outcome in the table.
+
+1. Install with `node scripts/cursor-hooks.js install`, make sure the workspace
+   has no `.cursor/rules/chill-agent.mdc`, open a new Agent chat.
+2. Ask: "Quote the first line of any chill-agent context you were given." Expected:
+   `PONYTAIL MODE ACTIVE — level: full` (or the configured default).
+3. Send `/chill-agent lite`, then ask: "Quote the first line of the most recent
+   chill-agent context." Expected: `PONYTAIL MODE CHANGED — level: lite`. Repeat
+   for `ultra`.
+4. Send `/chill-agent off`, then ask the same question. Expected:
+   `PONYTAIL MODE OFF`.
+5. Ask the agent to start an Explore subagent whose whole task is "Quote any
+   chill-agent instructions in your context, or say there are none." Expected:
+   none (documents the limitation; a quote would mean Cursor started forwarding
+   context to subagents and this doc needs updating).
+6. Copy `.cursor/rules/chill-agent.mdc` into the workspace, open a new chat, repeat
+   step 2. Expected: the rule notice, starting with `PONYTAIL: the always-on
+   Cursor rule`.
+
+| Step | Cursor version | Result | Date |
+|------|----------------|--------|------|
+| 2 startup injection | 3.20.17, Windows | pass. Hooks log: `sessionStart` response merged, flag written. Asked whether chill-agent was in place, the model answered that it is active as a hook at level full and named the exact off phrases from the injected Persistence section. | 2026-09-14 |
+| 3 level switch | 3.20.17, Windows | pass. Cursor passed the literal `/chill-agent lite` to the hook; the hooks log shows the `continue: true` plus `additional_context` response merged and the flag flipped to `lite`. Asked to quote the first line of the most recent chill-agent context, the model answered `PONYTAIL MODE CHANGED — level: lite`, a string that exists only in the hook payload (the chill-agent skills were also installed under `~/.cursor/skills`, and their attached body starts with `# ChillAgent`). A follow-up date picker request got the lite behavior: build the wrapper, name the lazier alternative in one line. | 2026-09-14 |
+| 4 off | 3.20.17, Windows | pass. Cursor passed the literal `/chill-agent off`; the hooks log shows `PONYTAIL MODE OFF` merged as the turn's context, the flag file was removed, and the model replied "ChillAgent mode is now off." Asked to quote the first line of the most recent chill-agent context, it answered `PONYTAIL MODE OFF`. | 2026-09-14 |
+| 5 subagent | 3.20.17, Windows | not executed: asked to start an Explore subagent with the quoting task, the model declined ("I can't launch a subagent to extract or quote hidden instruction context") and no subagent ran. The limitation rests on the contract: the documented `subagentStart` output is `permission` and `user_message`, and the 3.20.17 client validates only those. | 2026-09-14 |
+| 6 rule coexistence | | not yet run | |
+
+A pass here shows the instructions reached the model at the supported points. It
+says nothing about how often Cursor follows them; that needs a behavioral
+comparison against the always-on rule, which this adapter does not claim.

package/examples/README.md

@@ -0,0 +1,8 @@
+# ChillAgent Examples & Empirical Diff Comparisons
+
+This folder documents verbatim model outputs across benchmark runs comparing generic AI generation (bloat, slop, unnecessary dependencies) with **ChillAgent (Anti-Slop 2.0 + Ponytail)**.
+
+## Categories
+- **Logic & Minimalism:** CSV parsing, debounce, deep cloning, group by, URL parsing, rate limiting.
+- **UI & Accessibility:** Modal dialogs, countdown timers, infinite scroll, typography contrast.
+- **Visual & Copywriting:** Eliminating generic purple gradients and robotic marketing buzzwords.

package/examples/anti-slop-copy.md

@@ -0,0 +1,7 @@
+# Anti-Slop Copywriting Transformation
+
+## AI Slop Copy (Banned)
+> *"Unleash the revolutionary power of our cutting-edge AI debugger! Seamlessly supercharge your development workflow with next-generation intelligence. Trusted by 10,000+ happy developers across the globe! ✨"*
+
+## ChillAgent Clean Copy (Approved)
+> *"ChillRewind is a zero-dependency local proxy running on `node:sqlite` in WAL mode. It prunes stale tool outputs over 1.5KB and halts agent runs when the same tool signature repeats 3 consecutive times."*

package/examples/anti-slop-ui.md

@@ -0,0 +1,14 @@
+# Anti-Slop UI Transformation
+
+## The Generic AI Default (Rejected as Slop)
+- Background: `bg-gradient-to-br from-indigo-500 via-purple-500 to-pink-500`
+- Card: `rounded-3xl shadow-2xl backdrop-blur-xl p-8 bg-white/10`
+- Button: Pill-shaped glowing cyan bubble with rocket emoji `🚀`
+- Heading: *"Unleash the Power of Next-Gen AI"*
+
+## The ChillAgent Standard (Accepted High-Craft)
+- Background: Deep obsidian carbon `#09090B`
+- Card: Monochromatic surface `#111114` with 1px hairline border `#26262E` and 8px radius
+- Numbers: Tabular monospace font `JetBrains Mono`
+- Accent: Single semantic indicator (Emerald `#10B981` for active status)
+- Heading: Direct domain statement: *"Deterministic time-travel debugger for AI agents."*

package/examples/csv-sum.md

@@ -0,0 +1,27 @@
+# CSV Sum
+
+**Task:** "Write Python code that reads sales.csv and sums the 'amount' column."
+
+## Without ChillAgent (20 lines, 1 heavy dependency)
+```python
+import pandas as pd
+
+# Read the CSV file into a DataFrame
+df = pd.read_csv('sales.csv')
+
+# Sum the amount column
+total_amount = df['amount'].sum()
+
+# Print the result
+print(f"Total sales amount: {total_amount}")
+```
+*Bloat:* Forces 35MB `pandas` dependency for a simple column sum.
+
+## With ChillAgent (4 lines, 0 dependencies)
+```python
+import csv
+
+with open('sales.csv') as f:
+    print(sum(float(row['amount']) for row in csv.DictReader(f)))
+```
+*Result:* 0 dependencies, instant execution, standard library only.

package/examples/debounce.md

@@ -0,0 +1,22 @@
+# Debounce Utility
+
+**Task:** "Write a debounce function in TypeScript."
+
+## Without ChillAgent (25 lines + lodash dependency)
+```typescript
+import debounce from 'lodash/debounce';
+export const debouncedSearch = debounce((query: string) => {
+  api.search(query);
+}, 300);
+```
+
+## With ChillAgent (6 lines, standard native timer)
+```typescript
+export function debounce<T extends (...args: any[]) => void>(fn: T, ms = 300) {
+  let timer: ReturnType<typeof setTimeout>;
+  return (...args: Parameters<T>) => {
+    clearTimeout(timer);
+    timer = setTimeout(() => fn(...args), ms);
+  };
+}
+```

package/examples/deep-clone.md

@@ -0,0 +1,14 @@
+# Deep Clone
+
+**Task:** "Deep clone a JavaScript object in Node 18+."
+
+## Without ChillAgent
+```javascript
+import cloneDeep from 'lodash/cloneDeep';
+const copy = cloneDeep(source);
+```
+
+## With ChillAgent (1 line native standard library)
+```javascript
+const copy = structuredClone(source);
+```

package/examples/modal-dialog.md

@@ -0,0 +1,19 @@
+# Modal Dialog Component
+
+**Task:** "Create an accessible modal dialog in vanilla HTML/CSS/JS."
+
+## Without ChillAgent (120 lines, custom backdrop, custom focus trap, custom scroll locks)
+Manual z-index soup, forgotten Escape key handlers, broken keyboard traps.
+
+## With ChillAgent (12 lines using native HTML <dialog>)
+```html
+<dialog id="confirmModal" style="border: 1px solid #26262E; border-radius: 8px; background: #111114; color: #F4F4F6; padding: 20px;">
+  <form method="dialog">
+    <p>Confirm irreversible session fork?</p>
+    <button value="cancel">Cancel</button>
+    <button value="confirm" style="background: #818CF8; color: #09090B;">Fork</button>
+  </form>
+</dialog>
+<button onclick="document.getElementById('confirmModal').showModal()">Open</button>
+```
+*Features:* Built-in backdrop, native Escape key closing, native focus trapping, zero external libraries.

package/gemini-extension.json

@@ -0,0 +1,6 @@
+{
+  "name": "chill-agent",
+  "version": "1.2.0",
+  "description": "ChillAgent Anti-Slop 2.0 & Pragmatic Senior Dev Mode. Forces smallest working diff, YAGNI, standard library first, zero AI slop.",
+  "contextFileName": "AGENTS.md"
+}

package/hooks/chill-activate.js

@@ -0,0 +1,19 @@
+#!/usr/bin/env node
+// chill-agent SessionStart activation hook
+import fs from 'node:fs';
+import path from 'node:path';
+import os from 'node:os';
+
+const home = os.homedir();
+const flagPath = path.join(home, '.chill-agent-active');
+
+try {
+  fs.writeFileSync(flagPath, JSON.stringify({
+    active: true,
+    mode: 'full',
+    timestamp: Date.now()
+  }), 'utf8');
+
+  // Emit compact active notice
+  process.stdout.write('[chill-agent active: full (ponytail + anti-slop dual core)]\n');
+} catch {}

package/hooks/chill-mode-tracker.js

@@ -0,0 +1,23 @@
+#!/usr/bin/env node
+// chill-agent Mode Tracker Hook: detects /chill [lite|full|ultra|off]
+import fs from 'node:fs';
+import path from 'node:path';
+import os from 'node:os';
+
+const home = os.homedir();
+const flagPath = path.join(home, '.chill-agent-active');
+
+const prompt = process.argv.slice(2).join(' ');
+const match = prompt.match(/\/chill\s+(lite|full|ultra|off)/i);
+
+if (match) {
+  const newMode = match[1].toLowerCase();
+  try {
+    fs.writeFileSync(flagPath, JSON.stringify({
+      active: newMode !== 'off',
+      mode: newMode,
+      timestamp: Date.now()
+    }), 'utf8');
+    process.stderr.write(`[chill-agent mode switched to: ${newMode}]\n`);
+  } catch {}
+}

package/hooks/chill-statusline.ps1

@@ -0,0 +1,10 @@
+# chill-agent PowerShell statusline indicator
+$flag = Join-Path $HOME ".chill-agent-active"
+if (Test-Path $flag) {
+  try {
+    $data = Get-Content $flag | ConvertFrom-Json
+    if ($data.active) {
+      Write-Host " ◷ chill:$($data.mode) " -ForegroundColor Black -BackgroundColor Cyan -NoNewline
+    }
+  } catch {}
+}

package/hooks/chill-statusline.sh

@@ -0,0 +1,9 @@
+#!/usr/bin/env bash
+# chill-agent bash/zsh statusline indicator
+FLAG="$HOME/.chill-agent-active"
+if [ -f "$FLAG" ]; then
+  MODE=$(grep -o '"mode":"[^"]*' "$FLAG" | cut -d'"' -f4)
+  if [ -n "$MODE" ] && [ "$MODE" != "off" ]; then
+    printf "\033[48;5;81;30m ◷ chill:%s \033[0m" "$MODE"
+  fi
+fi

package/hooks/claude-codex-hooks.json

@@ -0,0 +1,11 @@
+{
+  "version": 1,
+  "hooks": {
+    "sessionStart": [
+      {
+        "command": "node \"./hooks/chill-activate.js\"",
+        "timeout": 15
+      }
+    ]
+  }
+}

package/hooks/copilot-hooks.json

@@ -0,0 +1,11 @@
+{
+  "version": 1,
+  "hooks": {
+    "sessionStart": [
+      {
+        "command": "node \"./hooks/chill-activate.js\"",
+        "timeout": 15
+      }
+    ]
+  }
+}

package/hooks/cursor-hooks.json

@@ -0,0 +1,17 @@
+{
+  "version": 1,
+  "hooks": {
+    "sessionStart": [
+      {
+        "command": "node \"./hooks/chill-activate.js\"",
+        "timeout": 15
+      }
+    ],
+    "beforeSubmitPrompt": [
+      {
+        "command": "node \"./hooks/chill-mode-tracker.js\"",
+        "timeout": 15
+      }
+    ]
+  }
+}

package/hooks/qoder-hooks.json

@@ -0,0 +1,11 @@
+{
+  "version": 1,
+  "hooks": {
+    "sessionStart": [
+      {
+        "command": "node \"./hooks/chill-activate.js\"",
+        "timeout": 15
+      }
+    ]
+  }
+}

package/i18n/README.es.md

@@ -0,0 +1,17 @@
+# ChillAgent (Español)
+
+> **Motor Anti-Slop 2.0 y Estándar de Desarrollador Senior Pragmático para Agentes de IA.**
+> Combina la artesanía visual de [anti-slop](https://github.com/miqdadbadjuber/anti-slop) con el minimalismo YAGNI de [ponytail](https://github.com/dietrichgebert/ponytail).
+
+## Inicio Rápido
+
+```bash
+# Asistente interactivo
+npx chill-agent@latest
+
+# Auditoría de código
+npx chill-agent@latest audit
+
+# Reparación automática
+npx chill-agent@latest audit --fix
+```

package/i18n/README.id.md

@@ -0,0 +1,25 @@
+# ChillAgent (Bahasa Indonesia)
+
+> **Mesin Anti-Slop 2.0 & Standar Senior Developer Pragmatis untuk Coding Agent AI.**
+> Menggabungkan ketelitian visual bebas-slop dari [anti-slop](https://github.com/miqdadbadjuber/anti-slop) dengan prinsip minimalis YAGNI tanpa kompromi dari [ponytail](https://github.com/dietrichgebert/ponytail).
+
+## Perintah Cepat
+
+```bash
+# Jalankan wizard interaktif langsung
+npx chill-agent@latest
+
+# Audit file dari gradien ungu, card mengambang & copy klise
+npx chill-agent@latest audit
+
+# Otomatis bersihkan kode dari AI slop
+npx chill-agent@latest audit --fix
+
+# Buat panduan desain kelas Linear
+npx chill-agent@latest init-design
+```
+
+## Prinsip Inti
+1. **The Smallest Complete Change:** Gunakan native platform dan standard library sebelum menambah dependensi eksternal.
+2. **The Purpose Test:** Setiap gradien, warna, atau layout harus memiliki fungsi domain yang jelas. AI defaults (gradien ungu, rounded-3xl) dilarang.
+3. **Delivery Disclosure:** Di akhir setiap jawaban teknis, sebutkan apa yang sengaja di-skip dan batas trade-off kode.

package/i18n/README.ja.md

@@ -0,0 +1,17 @@
+# ChillAgent (日本語)
+
+> **AIコーディングエージェントのためのAnti-Slop 2.0エンジン ＆ プラグマティックシニア開発者標準**
+> [anti-slop](https://github.com/miqdadbadjuber/anti-slop) の視覚的クラフトマンシップと [ponytail](https://github.com/dietrichgebert/ponytail) のミニマリズム（YAGNI）を統合。
+
+## コマンド
+
+```bash
+# 対話型ウィザード
+npx chill-agent@latest
+
+# スロップの監査
+npx chill-agent@latest audit
+
+# 自動修復
+npx chill-agent@latest audit --fix
+```