claude-launchpad 0.3.2 → 0.3.4

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (15) hide show
  1. package/README.md +118 -106
  2. package/dist/cli.js +93 -36
  3. package/dist/cli.js.map +1 -1
  4. package/package.json +2 -2
  5. package/scenarios/{common → conventions}/error-handling.yaml +3 -3
  6. package/scenarios/{common → conventions}/file-size.yaml +3 -3
  7. package/scenarios/{common → conventions}/immutability.yaml +3 -3
  8. package/scenarios/{common → conventions}/naming-conventions.yaml +5 -5
  9. package/scenarios/{common → conventions}/no-hardcoded-values.yaml +3 -3
  10. package/scenarios/{common → security}/env-protection.yaml +3 -3
  11. package/scenarios/{common → security}/input-validation.yaml +3 -3
  12. package/scenarios/{common → security}/secret-exposure.yaml +2 -2
  13. package/scenarios/{common → security}/sql-injection.yaml +2 -2
  14. package/scenarios/{common → workflow}/git-conventions.yaml +3 -3
  15. package/scenarios/{common → workflow}/session-continuity.yaml +3 -3
package/README.md CHANGED
@@ -1,32 +1,62 @@
1
1
  # Claude Launchpad
2
2
 
3
- **Your Claude Code setup is probably broken. This tool tells you how.**
3
+ **A linter for your Claude Code configuration.** Scores your setup, auto-fixes issues, and tests if Claude actually follows your rules.
4
4
 
5
- Claude Code's effectiveness is 90% configuration, 10% prompting. But nobody can tell if their config is actually good. You write a `CLAUDE.md`, add some hooks, install some plugins — and then *hope* Claude follows them.
6
-
7
- Claude Launchpad is the first CLI that **diagnoses, scaffolds, enhances, and tests** Claude Code configurations. Think ESLint for your AI setup.
5
+ You write a `CLAUDE.md`, add some hooks, configure settings — but is any of it actually working? Claude Launchpad scans your config, gives you a score out of 100, fixes what's broken, and runs Claude against test scenarios to prove it.
8
6
 
9
7
  ```bash
10
8
  npx claude-launchpad
11
9
  ```
12
10
 
13
- ## The Workflow
11
+ That's it. One command. You get a score. You see what's wrong. You fix it.
12
+
13
+ ## What It Does
14
+
15
+ | Command | What it does | Cost |
16
+ |---|---|---|
17
+ | `claude-launchpad` | Scans your config, scores it 0-100, lists issues | Free |
18
+ | `claude-launchpad doctor --fix` | Auto-fixes issues (adds hooks, rules, missing sections) | Free |
19
+ | `claude-launchpad doctor --watch` | Live score that updates when you edit config files | Free |
20
+ | `claude-launchpad init` | Detects your stack, generates config from scratch | Free |
21
+ | `claude-launchpad enhance` | Opens Claude to read your code and complete CLAUDE.md | Uses Claude |
22
+ | `claude-launchpad eval --suite security` | Runs Claude against test scenarios, proves your config works | Uses Claude |
23
+
24
+ ## Quick Start
14
25
 
15
26
  ```bash
16
- npx claude-launchpad init # 1. Auto-detect stack, generate config + hooks + .claudeignore
17
- npx claude-launchpad enhance # 2. Claude reads your code, completes CLAUDE.md
18
- npx claude-launchpad # 3. Check your score (42%)
19
- npx claude-launchpad doctor --fix # 4. Auto-fix everything (→ 86%)
20
- npx claude-launchpad eval # 5. Prove your config works (89% eval score)
27
+ # Install
28
+ npm i -g claude-launchpad
29
+
30
+ # Go to any project with Claude Code
31
+ cd your-project
32
+
33
+ # See your score
34
+ claude-launchpad
35
+
36
+ # Fix everything it found
37
+ claude-launchpad doctor --fix
38
+
39
+ # See your new score
40
+ claude-launchpad
21
41
  ```
22
42
 
23
- > See the [full story on the landing page](https://mboss37.github.io/claude-launchpad/) — a 42% → 89% journey.
43
+ That takes you from ~42% to ~86% with zero manual work.
24
44
 
25
- ## Commands
45
+ ## The Doctor
26
46
 
27
- ### `doctor` — Know your score
47
+ The core of the tool. Runs 7 analyzers against your `.claude/` directory and `CLAUDE.md`:
28
48
 
29
- Runs 7 static analyzers against your `.claude/` directory and `CLAUDE.md`. No API calls, no cost, works offline.
49
+ | Analyzer | What it catches |
50
+ |---|---|
51
+ | **Instruction Budget** | Too many instructions in CLAUDE.md — Claude starts ignoring rules past ~150 |
52
+ | **CLAUDE.md Quality** | Missing sections, vague instructions ("write good code"), hardcoded secrets |
53
+ | **Settings** | No hooks configured, dangerous tool access without safety nets |
54
+ | **Hooks** | Missing auto-format on save, no .env file protection, no security gates |
55
+ | **Rules** | Dead rule files, stale references, empty configs |
56
+ | **Permissions** | Bash auto-allowed without security hooks, no force-push protection |
57
+ | **MCP Servers** | Invalid transport configs, missing commands/URLs |
58
+
59
+ Output looks like this:
30
60
 
31
61
  ```
32
62
  Instruction Budget ━━━━━━━━━━━━━━━━━━━━ 100%
@@ -42,40 +72,25 @@ Runs 7 static analyzers against your `.claude/` directory and `CLAUDE.md`. No AP
42
72
  ✓ No issues found. Your configuration looks solid.
43
73
  ```
44
74
 
45
- Running bare `claude-launchpad` with no subcommand auto-detects your config and runs doctor.
46
-
47
- **Flags:**
48
- - `--fix` — Auto-apply deterministic fixes (42% → 86% in one command)
49
- - `--watch` — Live score that updates every time you save a config file
50
- - `--json` — JSON output for programmatic use
51
- - `--min-score <n>` — Exit non-zero if score drops below threshold (for CI)
52
-
53
- **What it checks:**
75
+ **All doctor flags:**
54
76
 
55
- | Analyzer | What it catches |
77
+ | Flag | What it does |
56
78
  |---|---|
57
- | **Instruction Budget** | Are you over the ~150 instruction limit where Claude starts ignoring rules? |
58
- | **CLAUDE.md Quality** | Missing essential sections, vague instructions ("write good code"), hardcoded secrets |
59
- | **Settings** | Hooks configured, dangerous tool access without safety nets |
60
- | **Hooks** | Missing auto-format, no .env protection, no PreToolUse security gates |
61
- | **Rules** | Dead rule files, stale references, empty configs |
62
- | **Permissions** | Bash auto-allowed without security hooks, no force-push protection |
63
- | **MCP Servers** | Invalid transport configs, missing commands/URLs, broken executables |
79
+ | `--fix` | Auto-fixes issues: adds hooks, CLAUDE.md sections, rules, .claudeignore |
80
+ | `--watch` | Re-runs every second, updates when you save a config file |
81
+ | `--json` | Pure JSON output, no colors, no banner — for scripts and CI |
82
+ | `--min-score <n>` | Exit code 1 if score is below threshold — use in CI to block bad configs |
83
+ | `-p, --path <dir>` | Run on a different directory |
64
84
 
65
- ### `init` — Set up any project in seconds
85
+ ## Init
66
86
 
67
- Auto-detects your stack and generates optimized Claude Code configuration. Works with **any** language — no fixed menu, no templates to pick from.
68
-
69
- ```bash
70
- claude-launchpad init
71
- ```
87
+ Detects your project and generates Claude Code config that fits. No templates, no menus — it reads your manifest files and figures it out.
72
88
 
73
89
  ```
74
90
  → Detecting project...
75
91
  ✓ Found Next.js (TypeScript) project
76
92
  · Package manager: pnpm
77
93
  · Dev command: pnpm dev
78
- · Test command: pnpm test
79
94
 
80
95
  ✓ Generated CLAUDE.md
81
96
  ✓ Generated TASKS.md
@@ -83,86 +98,87 @@ claude-launchpad init
83
98
  ✓ Generated .claudeignore
84
99
  ```
85
100
 
86
- **Detects 13 languages:** TypeScript, JavaScript, Python, Go, Ruby, Rust, Dart, PHP, Java, Kotlin, Swift, Elixir, C#
87
-
88
- **Detects 20+ frameworks:** Next.js, FastAPI, Django, Rails, Laravel, Express, SvelteKit, Angular, NestJS, Hono, Astro, Remix, Nuxt, Symfony, and more.
89
-
90
- **Detects package managers from lockfiles:** pnpm, yarn, npm, bun, uv, poetry, cargo, bundler, composer, go modules.
101
+ **Works with:** TypeScript, JavaScript, Python, Go, Ruby, Rust, Dart, PHP, Java, Kotlin, Swift, Elixir, C# — and detects frameworks (Next.js, FastAPI, Django, Rails, Laravel, Express, SvelteKit, Angular, NestJS, and 15+ more).
91
102
 
92
103
  **What you get:**
93
- - `CLAUDE.md` with your detected stack, commands, and essential sections
94
- - `TASKS.md` for session continuity across Claude sessions
95
- - `.claude/settings.json` with auto-format hooks and .env file protection (merges with existing)
96
- - `.claudeignore` with language-specific ignore patterns (node_modules, __pycache__, dist, etc.)
104
+ - `CLAUDE.md` — your stack, commands, conventions, guardrails
105
+ - `TASKS.md` — session continuity across Claude Code sessions
106
+ - `.claude/settings.json` — auto-format hooks and .env file protection
107
+ - `.claudeignore` — keeps Claude from reading node_modules, dist, lockfiles, etc.
97
108
 
98
- ### `enhance` — Let Claude finish what init started
109
+ ## Enhance
99
110
 
100
- Init auto-detects your stack but can't understand your architecture. Enhance spawns Claude interactively to read your actual codebase and fill in the gaps.
111
+ Init detects your stack but can't understand your architecture. Enhance opens Claude to read your actual code and fill in the details.
101
112
 
102
113
  ```bash
103
114
  claude-launchpad enhance
104
115
  ```
105
116
 
106
- Claude opens, reads your code, and updates CLAUDE.md with:
107
- - **Architecture** — actual directory structure, data flow, key modules
108
- - **Conventions** — patterns it observes in your code (naming, imports, state management)
109
- - **Off-Limits** — guardrails based on what it sees (protected files, anti-patterns)
110
- - **Key Decisions** — architectural decisions visible in the code
117
+ Claude reads your codebase and updates CLAUDE.md with real content — actual architecture, actual conventions, actual guardrails. Not boilerplate. It also suggests project-specific hooks and MCP servers based on what it finds.
111
118
 
112
- You see Claude working in real-time — same experience as running `claude` yourself.
119
+ Stays under the 120-instruction budget. Overflows detailed content to `.claude/rules/` files.
113
120
 
114
- ### `eval` — Prove your config works
121
+ ## Eval
115
122
 
116
- Runs Claude headless against 11 reproducible scenarios using the [Agent SDK](https://www.npmjs.com/package/@anthropic-ai/claude-agent-sdk) and **scores how well your config actually drives correct behavior**.
123
+ The part nobody else has built. Runs Claude against real test scenarios and scores the results.
117
124
 
118
125
  ```bash
119
- claude-launchpad eval --suite common
126
+ # Run only security tests (4 scenarios)
127
+ claude-launchpad eval --suite security
128
+
129
+ # Run only convention tests (5 scenarios)
130
+ claude-launchpad eval --suite conventions
131
+
132
+ # Run only workflow tests (2 scenarios)
133
+ claude-launchpad eval --suite workflow
134
+
135
+ # Run everything (11 scenarios)
136
+ claude-launchpad eval
137
+
138
+ # Use a cheaper model
139
+ claude-launchpad eval --suite security --model haiku
140
+
141
+ # One run per scenario (fastest)
142
+ claude-launchpad eval --suite security --runs 1
120
143
  ```
121
144
 
145
+ Each scenario creates an isolated sandbox, runs Claude with a task, and checks if Claude followed the rules:
146
+
122
147
  ```
123
148
  ✓ security/sql-injection 10/10 PASS
124
149
  ✓ security/env-protection 10/10 PASS
125
150
  ✓ security/secret-exposure 10/10 PASS
126
151
  ✓ security/input-validation 10/10 PASS
127
- ✓ conventions/error-handling 10/10 PASS
128
- ✓ conventions/immutability 10/10 PASS
129
- ✓ conventions/no-hardcoded-values 10/10 PASS
130
- ✓ conventions/naming-conventions 10/10 PASS
131
- ✓ conventions/file-size 10/10 PASS
132
- ✓ workflow/git-conventions 10/10 PASS
133
- ✗ workflow/session-continuity 7/10 WARN
152
+ ✗ conventions/file-size 5/10 FAIL
153
+ ✗ Claude kept all generated files under 800 lines
134
154
 
135
155
  Config Eval Score ━━━━━━━━━━━━━━━━━━━─ 95%
136
156
  ```
137
157
 
138
- Each scenario is a YAML file. [Write your own](scenarios/CONTRIBUTING.md).
139
-
140
- This is the part nobody else has built. Template repos scaffold. Audit tools diagnose. **Nobody tests whether your config actually makes Claude better.** Until now.
141
-
142
- ## How It Works Under the Hood
158
+ Results are saved to `.claude/eval/` as structured markdown — you can feed these reports back to Claude to fix the failures.
143
159
 
144
- ### doctor
145
- Reads your `CLAUDE.md`, `.claude/settings.json`, `.claude/rules/`, and `.claudeignore`. Runs 7 analyzers that check instruction count, section completeness, hook configuration, rule validity, permission safety, and MCP server configs. Pure static analysis — no API calls, no network, no cost.
160
+ **Suites:**
146
161
 
147
- ### init
148
- Scans the project root for manifest files (`package.json`, `go.mod`, `pyproject.toml`, `Gemfile`, `Cargo.toml`, `composer.json`, etc.). Detects language, framework, package manager, and available scripts. Generates config files with stack-appropriate hooks (prettier for TypeScript, gofmt for Go, ruff for Python, etc.). Merges with existing `settings.json` if one exists.
162
+ | Suite | Scenarios | What it tests |
163
+ |---|---|---|
164
+ | `security` | 4 | SQL injection, .env protection, secret exposure, input validation |
165
+ | `conventions` | 5 | Error handling, immutability, file size, naming, no hardcoded values |
166
+ | `workflow` | 2 | Git conventions, session continuity |
149
167
 
150
- ### enhance
151
- Spawns `claude "prompt"` as an interactive child process with `stdio: "inherit"` — you see Claude's full UI. The prompt instructs Claude to read the codebase and fill in CLAUDE.md sections. No data passes through the launchpad — it just launches Claude with a pre-loaded task.
168
+ **All eval flags:**
152
169
 
153
- ### eval
154
- 1. Creates a temp directory (`/tmp/claude-eval-<uuid>/`)
155
- 2. Writes seed files from the scenario YAML (e.g., a `src/db.ts` with a TODO)
156
- 3. Writes a `CLAUDE.md` with the scenario's instructions
157
- 4. Initializes a git repo (Claude Code expects one)
158
- 5. Runs Claude via the [Agent SDK](https://www.npmjs.com/package/@anthropic-ai/claude-agent-sdk) with `allowedTools: ["Bash", "Read", "Write", "Edit", "Glob", "Grep"]` and `permissionMode: "dontAsk"` — or falls back to `claude -p` if the SDK isn't installed
159
- 6. After Claude finishes, runs grep/file assertions against the modified files
160
- 7. Scores: each check has points, total determines pass/fail
161
- 8. Cleans up the temp directory (or preserves it with `--debug`)
170
+ | Flag | What it does |
171
+ |---|---|
172
+ | `--suite <name>` | Run one suite: `security`, `conventions`, or `workflow` |
173
+ | `--model <model>` | Model to use: `haiku`, `sonnet`, `opus` |
174
+ | `--runs <n>` | Runs per scenario (default 3, median score used) |
175
+ | `--debug` | Keep sandbox directories so you can inspect what Claude wrote |
176
+ | `--json` | JSON output |
177
+ | `--timeout <ms>` | Timeout per run (default 120000) |
162
178
 
163
179
  ## Use in CI
164
180
 
165
- Add this workflow to block PRs that degrade Claude Code config quality:
181
+ Block PRs that degrade your Claude Code config quality:
166
182
 
167
183
  ```yaml
168
184
  # .github/workflows/claude-config.yml
@@ -180,39 +196,35 @@ jobs:
180
196
  - run: npx claude-launchpad@latest doctor --min-score 80 --json
181
197
  ```
182
198
 
183
- Exit code is 1 when score is below the threshold, 0 when it passes.
199
+ Score below threshold = exit code 1 = PR blocked.
184
200
 
185
201
  ## Plugin (pending marketplace review)
186
202
 
187
- The plugin has been submitted to the Claude Code marketplace. Once approved:
188
-
189
203
  ```bash
190
204
  claude plugin install claude-launchpad
191
205
  ```
192
206
 
193
- Then use `/launchpad:doctor`, `/launchpad:init`, `/launchpad:enhance`, and `/launchpad:eval` directly inside Claude Code. The plugin also nudges you to re-check your score when you edit config files.
207
+ Then use `/launchpad:doctor`, `/launchpad:init`, `/launchpad:enhance`, `/launchpad:eval` inside Claude Code. The plugin nudges you to re-check your score when you edit config files.
194
208
 
195
- ## Why this exists
209
+ ## How It Works
196
210
 
197
- Claude Code configurations are the new `.eslintrc` — everyone has one, nobody knows if it's good. The difference:
211
+ **Doctor** reads your files and runs static analysis. No API calls. No network. No cost.
198
212
 
199
- - **CLAUDE.md is advisory** (~80% compliance). Claude might ignore your rules.
200
- - **Hooks are deterministic** (100% compliance). But most people don't have any.
201
- - **Instruction budget is real.** Past ~150 instructions, compliance drops. Most people don't know they're over.
202
- - **Nobody measures.** You can't improve what you can't measure.
213
+ **Init** scans manifest files (package.json, go.mod, pyproject.toml, etc.), detects your stack, and generates config with safe, hardcoded formatter hooks — never interpolates user-controlled strings.
214
+
215
+ **Enhance** spawns `claude "prompt"` as an interactive child process. You see Claude's full UI. No data passes through the tool — it just launches Claude with a task.
203
216
 
204
- Claude Launchpad gives you a number. Fix the issues, re-run, watch the number go up.
217
+ **Eval** creates a temp directory, writes seed files from the scenario YAML, initializes a git repo, runs Claude via the Agent SDK (or falls back to CLI), then checks the output with grep/file assertions. Sandbox is cleaned up after (or preserved with `--debug`).
205
218
 
206
- ## Philosophy
219
+ ## Why This Exists
220
+
221
+ - **CLAUDE.md is advisory.** ~80% compliance. Claude might ignore your rules.
222
+ - **Hooks are deterministic.** 100% compliance. But most people have zero hooks.
223
+ - **Instruction budget is real.** Past ~150, compliance drops. Most people don't know they're over.
224
+ - **Nobody measures.** You can't improve what you can't measure.
207
225
 
208
- - **Zero dependencies on third-party Claude plugins.** Generates its own hooks and settings.
209
- - **Doctor is free.** No API calls, no secrets, works offline and air-gapped.
210
- - **Enhance uses Claude.** Spawns an interactive session to understand your codebase — costs tokens but produces a CLAUDE.md that actually knows your project.
211
- - **Eval uses the Agent SDK.** Runs Claude headless in sandboxes with explicit tool permissions — proof that your config works.
212
- - **Works with any stack.** Auto-detects your project. No fixed menu of supported frameworks.
213
- - **57 tests.** The tool that tests configs is itself well-tested.
214
- - **You never clone this repo.** It's a tool you run with `npx`, not a template you fork.
226
+ This tool gives you a number. Fix the issues, re-run, watch the number go up.
215
227
 
216
228
  ## License
217
229
 
218
- MIT — Built by [McLovin](https://github.com/mboss37) (the AI behind [@mboss37](https://github.com/mboss37))
230
+ MIT
package/dist/cli.js CHANGED
@@ -2,7 +2,7 @@
2
2
 
3
3
  // src/cli.ts
4
4
  import { Command as Command5 } from "commander";
5
- import { join as join10 } from "path";
5
+ import { join as join11 } from "path";
6
6
 
7
7
  // src/commands/init/index.ts
8
8
  import { Command } from "commander";
@@ -1515,9 +1515,11 @@ function createDoctorCommand() {
1515
1515
  await watchConfig(opts.path);
1516
1516
  return;
1517
1517
  }
1518
- printBanner();
1519
- log.step("Scanning Claude Code configuration...");
1520
- log.blank();
1518
+ if (!opts.json) {
1519
+ printBanner();
1520
+ log.step("Scanning Claude Code configuration...");
1521
+ log.blank();
1522
+ }
1521
1523
  const config = await parseClaudeConfig(opts.path);
1522
1524
  if (config.claudeMdContent === null && config.settings === null) {
1523
1525
  log.error("No Claude Code configuration found in this directory.");
@@ -1571,6 +1573,8 @@ function createDoctorCommand() {
1571
1573
  import { Command as Command3 } from "commander";
1572
1574
  import ora from "ora";
1573
1575
  import chalk2 from "chalk";
1576
+ import { mkdir as mkdir4, writeFile as writeFile4 } from "fs/promises";
1577
+ import { join as join9 } from "path";
1574
1578
 
1575
1579
  // src/commands/eval/loader.ts
1576
1580
  import { readFile as readFile5, readdir as readdir3, access as access5 } from "fs/promises";
@@ -1745,7 +1749,7 @@ async function runScenario(scenario, options) {
1745
1749
  const sandboxDir = join8(tmpdir(), `claude-eval-${randomUUID()}`);
1746
1750
  try {
1747
1751
  await setupSandbox(sandboxDir, scenario);
1748
- await runClaudeInSandbox(sandboxDir, scenario.prompt, options.timeout);
1752
+ await runClaudeInSandbox(sandboxDir, scenario.prompt, options.timeout, options.model);
1749
1753
  return await scoreResults(scenario, sandboxDir);
1750
1754
  } finally {
1751
1755
  if (options.debug) {
@@ -1794,7 +1798,7 @@ ${scenario.setup.instructions}
1794
1798
  "eval setup"
1795
1799
  ], { cwd: sandboxDir });
1796
1800
  }
1797
- async function runClaudeInSandbox(cwd, prompt, timeout) {
1801
+ async function runClaudeInSandbox(cwd, prompt, timeout, model) {
1798
1802
  try {
1799
1803
  const sdk = await import("@anthropic-ai/claude-agent-sdk");
1800
1804
  const controller = new AbortController();
@@ -1808,7 +1812,8 @@ async function runClaudeInSandbox(cwd, prompt, timeout) {
1808
1812
  permissionMode: "dontAsk",
1809
1813
  settingSources: [],
1810
1814
  maxTurns: 20,
1811
- abortController: controller
1815
+ abortController: controller,
1816
+ ...model ? { model } : {}
1812
1817
  }
1813
1818
  })) {
1814
1819
  }
@@ -1816,31 +1821,29 @@ async function runClaudeInSandbox(cwd, prompt, timeout) {
1816
1821
  clearTimeout(timeoutId);
1817
1822
  }
1818
1823
  } catch {
1819
- await runClaudeCli(cwd, prompt, timeout);
1824
+ await runClaudeCli(cwd, prompt, timeout, model);
1820
1825
  }
1821
1826
  }
1822
- async function runClaudeCli(cwd, prompt, timeout) {
1827
+ async function runClaudeCli(cwd, prompt, timeout, model) {
1823
1828
  try {
1824
- await exec(
1825
- "claude",
1826
- [
1827
- "-p",
1828
- prompt,
1829
- "--output-format",
1830
- "text",
1831
- "--max-turns",
1832
- "20",
1833
- "--dangerously-skip-permissions",
1834
- "--allowedTools",
1835
- "Bash",
1836
- "Read",
1837
- "Write",
1838
- "Edit",
1839
- "Glob",
1840
- "Grep"
1841
- ],
1842
- { cwd, timeout, maxBuffer: 10 * 1024 * 1024 }
1843
- );
1829
+ const args = [
1830
+ "-p",
1831
+ prompt,
1832
+ "--output-format",
1833
+ "text",
1834
+ "--max-turns",
1835
+ "20",
1836
+ "--dangerously-skip-permissions",
1837
+ "--allowedTools",
1838
+ "Bash",
1839
+ "Read",
1840
+ "Write",
1841
+ "Edit",
1842
+ "Glob",
1843
+ "Grep"
1844
+ ];
1845
+ if (model) args.push("--model", model);
1846
+ await exec("claude", args, { cwd, timeout, maxBuffer: 10 * 1024 * 1024 });
1844
1847
  } catch (error) {
1845
1848
  if (error && typeof error === "object" && "stdout" in error) {
1846
1849
  return;
@@ -1949,7 +1952,7 @@ async function listAllFiles(dir) {
1949
1952
 
1950
1953
  // src/commands/eval/index.ts
1951
1954
  function createEvalCommand() {
1952
- return new Command3("eval").description("Test your Claude Code config against eval scenarios").option("-s, --suite <suite>", "Eval suite to run (e.g., security, conventions, workflow)").option("-p, --path <path>", "Project root path", process.cwd()).option("--scenarios <path>", "Custom scenarios directory").option("--runs <n>", "Runs per scenario (default: 3)", "3").option("--timeout <ms>", "Timeout per run in ms (default: 120000)", "120000").option("--json", "Output as JSON").option("--debug", "Keep sandbox directories for inspection").action(async (opts) => {
1955
+ return new Command3("eval").description("Test your Claude Code config against eval scenarios").option("-s, --suite <suite>", "Eval suite to run (e.g., security, conventions, workflow)").option("-p, --path <path>", "Project root path", process.cwd()).option("--scenarios <path>", "Custom scenarios directory").option("--runs <n>", "Runs per scenario (default: 3)", "3").option("--timeout <ms>", "Timeout per run in ms (default: 120000)", "120000").option("--json", "Output as JSON").option("--debug", "Keep sandbox directories for inspection").option("--model <model>", "Model to use for eval (e.g., sonnet, haiku, opus)").action(async (opts) => {
1953
1956
  printBanner();
1954
1957
  const claudeAvailable = await checkClaudeCli();
1955
1958
  if (!claudeAvailable) {
@@ -1970,6 +1973,9 @@ function createEvalCommand() {
1970
1973
  return;
1971
1974
  }
1972
1975
  log.success(`Loaded ${scenarios.length} scenario(s)`);
1976
+ if (opts.model) {
1977
+ log.info(`Model: ${opts.model}`);
1978
+ }
1973
1979
  log.blank();
1974
1980
  const runs = parseInt(opts.runs, 10);
1975
1981
  const timeout = parseInt(opts.timeout, 10);
@@ -1982,7 +1988,7 @@ function createEvalCommand() {
1982
1988
  try {
1983
1989
  const result = await runScenarioWithRetries(
1984
1990
  { ...scenario, runs },
1985
- { projectRoot: opts.path, timeout, debug: opts.debug }
1991
+ { projectRoot: opts.path, timeout, debug: opts.debug, model: opts.model }
1986
1992
  );
1987
1993
  results.push(result);
1988
1994
  if (result.passed) {
@@ -2017,6 +2023,7 @@ function createEvalCommand() {
2017
2023
  return;
2018
2024
  }
2019
2025
  renderEvalReport(results);
2026
+ await saveEvalReport(results, opts.path, opts.suite, opts.model);
2020
2027
  });
2021
2028
  }
2022
2029
  function renderEvalReport(results) {
@@ -2027,7 +2034,7 @@ function renderEvalReport(results) {
2027
2034
  console.log(` ${icon} ${chalk2.bold(result.scenario)} ${score} ${status}`);
2028
2035
  const failedChecks = result.checks.filter((c) => !c.passed);
2029
2036
  for (const check of failedChecks) {
2030
- console.log(` ${chalk2.dim("\u2500")} ${chalk2.dim(check.label)}`);
2037
+ console.log(` ${chalk2.red("\u2717")} ${chalk2.dim(check.label)}`);
2031
2038
  }
2032
2039
  }
2033
2040
  log.blank();
@@ -2044,6 +2051,56 @@ function renderEvalReport(results) {
2044
2051
  log.warn(`${passed} passed, ${failed} failed out of ${results.length} scenario(s).`);
2045
2052
  }
2046
2053
  }
2054
+ async function saveEvalReport(results, projectRoot, suite, model) {
2055
+ const totalScore = results.reduce((s, r) => s + r.score, 0);
2056
+ const totalMax = results.reduce((s, r) => s + r.maxScore, 0);
2057
+ const pct = totalMax > 0 ? Math.round(totalScore / totalMax * 100) : 0;
2058
+ const passed = results.filter((r) => r.passed).length;
2059
+ const failed = results.length - passed;
2060
+ const timestamp = (/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-").slice(0, 19);
2061
+ const lines = [
2062
+ `# Eval Report \u2014 ${timestamp}`,
2063
+ "",
2064
+ `**Score: ${pct}%** (${passed} passed, ${failed} failed out of ${results.length} scenarios)`,
2065
+ "",
2066
+ `- Suite: ${suite ?? "all"}`,
2067
+ `- Model: ${model ?? "default"}`,
2068
+ `- Date: ${(/* @__PURE__ */ new Date()).toISOString().split("T")[0]}`,
2069
+ "",
2070
+ "## Results",
2071
+ ""
2072
+ ];
2073
+ for (const result of results) {
2074
+ const status = result.passed ? "PASS" : "FAIL";
2075
+ lines.push(`### ${result.scenario} \u2014 ${result.score}/${result.maxScore} ${status}`);
2076
+ const failedChecks = result.checks.filter((c) => !c.passed);
2077
+ const passedChecks = result.checks.filter((c) => c.passed);
2078
+ for (const check of passedChecks) {
2079
+ lines.push(`- PASSED: ${check.label} (${check.points} pts)`);
2080
+ }
2081
+ for (const check of failedChecks) {
2082
+ lines.push(`- FAILED: ${check.label} (${check.points} pts)`);
2083
+ }
2084
+ lines.push("");
2085
+ }
2086
+ if (failed > 0) {
2087
+ lines.push("## Recommendations");
2088
+ lines.push("");
2089
+ for (const result of results.filter((r) => !r.passed)) {
2090
+ lines.push(`### Fix: ${result.scenario}`);
2091
+ const failedChecks = result.checks.filter((c) => !c.passed);
2092
+ for (const check of failedChecks) {
2093
+ lines.push(`- ${check.label} \u2014 update CLAUDE.md instructions or add hooks to enforce this behavior`);
2094
+ }
2095
+ lines.push("");
2096
+ }
2097
+ }
2098
+ const evalDir = join9(projectRoot, ".claude", "eval");
2099
+ await mkdir4(evalDir, { recursive: true });
2100
+ const filename = `eval-${suite ?? "all"}-${timestamp}.md`;
2101
+ await writeFile4(join9(evalDir, filename), lines.join("\n"));
2102
+ log.success(`Report saved to .claude/eval/${filename}`);
2103
+ }
2047
2104
  async function checkClaudeCli() {
2048
2105
  const { execFile: execFile3 } = await import("child_process");
2049
2106
  const { promisify: promisify3 } = await import("util");
@@ -2061,7 +2118,7 @@ import { Command as Command4 } from "commander";
2061
2118
  import { spawn, execFile as execFile2 } from "child_process";
2062
2119
  import { promisify as promisify2 } from "util";
2063
2120
  import { access as access6 } from "fs/promises";
2064
- import { join as join9 } from "path";
2121
+ import { join as join10 } from "path";
2065
2122
  var execAsync = promisify2(execFile2);
2066
2123
  var ENHANCE_PROMPT = `Read CLAUDE.md and the project's codebase, then update CLAUDE.md to fill in missing or incomplete sections.
2067
2124
 
@@ -2094,7 +2151,7 @@ function createEnhanceCommand() {
2094
2151
  return new Command4("enhance").description("Use Claude to analyze your codebase and complete CLAUDE.md").option("-p, --path <path>", "Project root path", process.cwd()).action(async (opts) => {
2095
2152
  printBanner();
2096
2153
  const root = opts.path;
2097
- const claudeMdPath = join9(root, "CLAUDE.md");
2154
+ const claudeMdPath = join10(root, "CLAUDE.md");
2098
2155
  try {
2099
2156
  await access6(claudeMdPath);
2100
2157
  } catch {
@@ -2123,8 +2180,8 @@ function createEnhanceCommand() {
2123
2180
  }
2124
2181
 
2125
2182
  // src/cli.ts
2126
- var program = new Command5().name("claude-launchpad").description("CLI toolkit that makes Claude Code setups measurably good").version("0.3.2", "-v, --version").action(async () => {
2127
- const hasConfig = await fileExists(join10(process.cwd(), "CLAUDE.md")) || await fileExists(join10(process.cwd(), ".claude", "settings.json"));
2183
+ var program = new Command5().name("claude-launchpad").description("CLI toolkit that makes Claude Code setups measurably good").version("0.3.4", "-v, --version").action(async () => {
2184
+ const hasConfig = await fileExists(join11(process.cwd(), "CLAUDE.md")) || await fileExists(join11(process.cwd(), ".claude", "settings.json"));
2128
2185
  if (hasConfig) {
2129
2186
  await program.commands.find((c) => c.name() === "doctor")?.parseAsync([], { from: "user" });
2130
2187
  } else {