claudeye 0.2.2 → 0.3.0

package/.next/standalone/README.md

@@ -23,11 +23,16 @@ This starts a local server on port 8020 and opens the dashboard in your browser.
 ### Options
 
 ```
---projects-path <path>  Path to Claude projects directory (default: ~/.claude/projects)
---port <number>         Preferred port (default: 8020)
---evals <path>          Path to an evals file (see Custom Evals below)
---no-open               Don't auto-open the browser
--h, --help              Show help
+--projects-path, -p <path>  Path to Claude projects directory (default: ~/.claude/projects)
+--port <number>             Preferred port (default: 8020)
+--host <address>            Host to bind to (default: localhost)
+                            Use 0.0.0.0 for LAN access
+--evals <path>              Path to an evals/enrichments file (see Custom Evals below)
+--cache <on|off>            Enable/disable result caching (default: on)
+--cache-path <path>         Custom cache directory (default: ~/.claudeye/cache)
+--cache-clear               Clear all cached results and exit
+--no-open                   Don't auto-open the browser
+-h, --help                  Show help
 ```
 
 ### Examples
@@ -38,6 +43,21 @@ claudeye --projects-path /path/to/projects
 
 # Run on a different port without opening the browser
 claudeye --port 3000 --no-open
+
+# Bind to all interfaces for LAN access
+claudeye --host 0.0.0.0
+
+# Load custom evals and enrichments
+claudeye --evals ./my-evals.js
+
+# Disable caching
+claudeye --cache off
+
+# Use a custom cache directory
+claudeye --cache-path /tmp/claudeye-cache
+
+# Clear cached results
+claudeye --cache-clear
 ```
 
 ## Features
@@ -60,6 +80,7 @@ Open a session to get a full breakdown of the agent execution:
 - **Subagent logs** — nested subagent executions load on demand
 - **Thinking blocks** — collapsible reasoning traces
 - **Virtual scrolling** — handles large logs without performance issues
+- **Download** — export session logs as JSONL files
 
 ### Filtering
 
@@ -67,10 +88,73 @@ Open a session to get a full breakdown of the agent execution:
 - **Keyword search**: add multiple keywords as chips, AND logic, case-insensitive
 - Filters combine — both time and keyword conditions must match
 
+### Auto-Refresh
+
+The navbar includes a refresh button with an auto-refresh dropdown (Off, 5s, 10s, 30s). It works globally across all pages — projects, sessions, and log views all refresh when triggered.
+
 ### Theme Support
 
 Light and dark mode with system preference detection. Toggle from the navbar. Preference is saved to `localStorage`.
 
+---
+
+## Caching
+
+Claudeye caches eval and enrichment results to avoid recomputing them on every page load. Caching is **on by default** and requires zero configuration.
+
+### How It Works
+
+- Results are cached to `~/.claudeye/cache/` as JSON files, keyed by project and session
+- Cache invalidation is **content-hash-based**, not TTL-based:
+  - The session file's modification time and size are hashed (fast, no full file read)
+  - The evals module file content is hashed
+  - The set of registered eval/enricher names is tracked
+- A cache hit is only valid when **all three** match — if the session log grows, the evals file changes, or evals are added/removed, the cache is automatically invalidated
+- Clicking **Re-run** in the dashboard always bypasses the cache and runs fresh computation
+- Cached results show a subtle "cached" badge next to the duration in the results panel
+
+### Running Without Cache
+
+```bash
+# Disable caching entirely
+claudeye --cache off
+
+# Or during development
+npm run dev -- --cache off
+```
+
+When caching is disabled, no cache files are created or read. The startup log will show `Cache: disabled`.
+
+### Custom Cache Location
+
+```bash
+claudeye --cache-path /path/to/custom/cache
+```
+
+### Clearing the Cache
+
+```bash
+# Clear the default cache directory and exit
+claudeye --cache-clear
+
+# Clear a custom cache directory
+claudeye --cache-clear --cache-path /path/to/custom/cache
+```
+
+### Startup Log
+
+When the server starts, the cache status is shown alongside other configuration:
+
+```
+Starting Claudeye dashboard...
+  Projects: /home/user/.claude/projects
+  Evals:    /home/user/my-evals.js
+  Cache:    local (/home/user/.claudeye/cache/)
+  URL:      http://localhost:8020
+```
+
+---
+
 ## Custom Evals & Enrichments
 
 Claudeye supports two types of custom extensions you can define in a single JS/TS file and load with the `--evals` flag:
@@ -165,107 +249,510 @@ node my-evals.js
 
 This spawns the Claudeye dashboard as a child process. When the same file is loaded via `--evals`, `listen()` automatically becomes a no-op so you won't get a duplicate server.
 
-### API Reference
+---
+
+## Conditional Evals & Enrichments
+
+Conditions let you control when evals and enrichments run. There are two levels:
+
+1. **Global condition** — gates _all_ evals and enrichments at once
+2. **Per-eval / per-enrichment conditions** — gate individual functions
 
-#### `createApp()`
+### Global Condition
 
-Returns a `ClaudeyeApp` instance.
+Use `app.condition()` to set a global condition. If it returns `false` (or throws), every registered eval and enrichment is skipped.
+
+```js
+import { createApp } from 'claudeye';
+
+const app = createApp();
 
-#### `app.eval(name, fn)`
+// Only run evals/enrichments for sessions with actual content
+app.condition(({ entries }) => entries.length > 0);
 
-Register an eval function. Chainable.
+// Only run for non-test projects
+app.condition(({ projectName }) => !projectName.includes('test'));
+```
+
+The global condition receives the same `EvalContext` as evals and enrichments, so you can inspect entries, stats, project name, and session ID.
+
+> **Note:** Calling `app.condition()` multiple times replaces the previous condition — only the last one is active.
+
+### Per-Eval Conditions
+
+Pass a `condition` function in the options object (third argument) to `app.eval()`:
+
+```js
+app.eval('budget-check',
+  ({ stats }) => ({
+    pass: stats.toolCallCount <= 50,
+    score: 1 - Math.min(stats.toolCallCount / 100, 1),
+    message: `${stats.toolCallCount} tool call(s)`,
+  }),
+  {
+    condition: ({ stats }) => stats.toolCallCount > 0,
+  }
+);
+```
+
+If the condition returns `false`, the eval is marked as **skipped** in the results panel. If the condition throws, the eval is marked as **errored** with the message `Condition error: <message>`.
+
+### Per-Enrichment Conditions
+
+Same pattern — pass a `condition` in the options to `app.enrich()`:
+
+```js
+app.enrich('error-analysis',
+  ({ entries }) => {
+    const errors = entries.filter(e => e.raw?.is_error === true);
+    return {
+      'Total Errors': errors.length,
+      'Error Rate': `${((errors.length / entries.length) * 100).toFixed(1)}%`,
+    };
+  },
+  {
+    condition: ({ entries }) => entries.some(e => e.raw?.is_error === true),
+  }
+);
+```
+
+### Evaluation Order
+
+When a session is loaded, conditions are evaluated in this order:
+
+```
+1. Global condition checked
+   ├─ Returns false or throws → ALL evals/enrichments marked "skipped"
+   └─ Returns true → proceed to step 2
+
+2. For each eval/enrichment:
+   ├─ Has per-item condition?
+   │   ├─ Returns false → that item marked "skipped"
+   │   ├─ Throws → that item marked "errored" (not skipped)
+   │   └─ Returns true → run the function
+   └─ No condition → run the function
+
+3. Function executes
+   ├─ Returns result → recorded normally
+   └─ Throws → marked "errored", other items still run
+```
+
+### Condition Function Signature
+
+Both global and per-item conditions use the same type:
+
+```ts
+type ConditionFunction = (context: EvalContext) => boolean | Promise<boolean>;
+```
+
+Conditions can be async — the runner awaits them.
+
+### Combining Global and Per-Item Conditions
+
+Global and per-item conditions stack. The global condition runs first; if it passes, per-item conditions are checked individually:
+
+```js
+const app = createApp();
+
+// Global: skip everything for empty sessions
+app.condition(({ entries }) => entries.length > 0);
+
+// Per-eval: only check turn count for sessions with tool calls
+app.eval('efficient-tools',
+  ({ stats }) => ({
+    pass: stats.toolCallCount <= stats.turnCount * 2,
+    score: Math.max(0, 1 - (stats.toolCallCount / (stats.turnCount * 4))),
+  }),
+  { condition: ({ stats }) => stats.toolCallCount > 0 }
+);
+
+// Per-enrichment: only compute cost for sessions that used a model
+app.enrich('cost',
+  ({ entries }) => {
+    const tokens = entries.reduce((s, e) => s + (e.raw?.usage?.total_tokens || 0), 0);
+    return { 'Total Tokens': tokens, 'Est. Cost': `$${(tokens * 0.00001).toFixed(4)}` };
+  },
+  { condition: ({ stats }) => stats.models.length > 0 }
+);
+```
+
+### UI Behavior
+
+In the dashboard, conditional results appear as follows:
+
+| Status | Evals Panel | Enrichments Panel |
+|--------|-------------|-------------------|
+| **Skipped** | Grayed-out row with "skipped" label | Grayed-out row with "skipped" label |
+| **Condition error** | Row with warning icon and error message | Row with warning icon and error message |
+| **Passed / Data** | Green check with score bar | Key-value pairs grouped by enricher |
+| **Failed** | Red X with score bar | N/A |
+
+Skipped items are counted separately in the summary bar (e.g. "2 passed, 1 skipped").
+
+### Full Example with Conditions
+
+```js
+import { createApp } from 'claudeye';
+
+const app = createApp();
+
+// ── Global condition: require at least one user message ──
+app.condition(({ entries }) =>
+  entries.some(e => e.type === 'user')
+);
+
+// ── Evals ──
+
+// Always runs (when global condition passes)
+app.eval('has-completion', ({ entries }) => {
+  const lastAssistant = [...entries].reverse().find(e => e.type === 'assistant');
+  const hasText = lastAssistant?.message?.content?.some?.(b => b.type === 'text');
+  return {
+    pass: !!hasText,
+    score: hasText ? 1.0 : 0,
+    message: hasText ? 'Session completed with text response' : 'No final text response',
+  };
+});
+
+// Only runs for sessions with tool calls
+app.eval('tool-success-rate',
+  ({ entries }) => {
+    const toolResults = entries.filter(e => e.type === 'tool');
+    const errors = toolResults.filter(e => e.raw?.is_error === true);
+    const rate = toolResults.length > 0 ? 1 - (errors.length / toolResults.length) : 1;
+    return {
+      pass: rate >= 0.9,
+      score: rate,
+      message: `${errors.length}/${toolResults.length} tool errors`,
+    };
+  },
+  { condition: ({ stats }) => stats.toolCallCount > 0 }
+);
+
+// Only runs for longer sessions
+app.eval('under-budget',
+  ({ stats }) => ({
+    pass: stats.turnCount <= 30,
+    score: Math.max(0, 1 - stats.turnCount / 60),
+    message: `${stats.turnCount} turns`,
+  }),
+  { condition: ({ stats }) => stats.turnCount >= 5 }
+);
+
+// ── Enrichments ──
+
+// Always runs
+app.enrich('overview', ({ stats }) => ({
+  'Turns': stats.turnCount,
+  'Tool Calls': stats.toolCallCount,
+  'Duration': stats.duration,
+  'Models': stats.models.join(', ') || 'none',
+}));
+
+// Only when subagents were spawned
+app.enrich('subagent-info',
+  ({ entries, stats }) => {
+    const subagentEntries = entries.filter(e => e.type === 'assistant' && e.parentUuid);
+    return {
+      'Subagent Count': stats.subagentCount,
+      'Subagent Messages': subagentEntries.length,
+    };
+  },
+  { condition: ({ stats }) => stats.subagentCount > 0 }
+);
+
+// Async condition example
+app.enrich('advanced-metrics',
+  ({ entries }) => ({
+    'Entry Count': entries.length,
+    'Avg Entry Size': Math.round(
+      entries.reduce((s, e) => s + JSON.stringify(e).length, 0) / entries.length
+    ),
+  }),
+  {
+    condition: async ({ entries }) => {
+      // Conditions can be async
+      return entries.length > 10;
+    },
+  }
+);
+```
+
+---
+
+## Subagent-Scoped Evals & Enrichments
+
+By default, evals and enrichments run at the session level. You can also register evals and enrichments that run against individual subagent logs using the `scope` option.
+
+### Scope Options
+
+The `scope` option controls when an eval or enrichment runs:
+
+| Scope | Runs at session level | Runs at subagent level |
+|-------|----------------------|----------------------|
+| `'session'` (default) | Yes | No |
+| `'subagent'` | No | Yes |
+| `'both'` | Yes | Yes |
+
+### Basic Usage
+
+```js
+import { createApp } from 'claudeye';
+
+const app = createApp();
+
+// Session-only (default, backward-compatible)
+app.eval('no-errors', fn);
+
+// Subagent-only, runs for all subagent types
+app.eval('agent-quality', fn, { scope: 'subagent' });
+
+// Subagent-only, only runs for Explore subagents
+app.eval('explore-depth', fn, { scope: 'subagent', subagentType: 'Explore' });
+
+// Runs at both session and subagent level
+app.eval('general-check', fn, { scope: 'both' });
+
+// Enrichments work the same way
+app.enrich('agent-cost', fn, { scope: 'subagent' });
+app.enrich('agent-tokens', fn, { scope: 'subagent', subagentType: 'Explore' });
+```
+
+### Subagent Context
+
+When running at subagent level, the `EvalContext` includes additional metadata:
+
+```js
+app.eval('adaptive-eval', (ctx) => {
+  if (ctx.scope === 'subagent') {
+    // Running against a subagent's log
+    console.log(ctx.subagentId);          // e.g. 'a1b2c3'
+    console.log(ctx.subagentType);        // e.g. 'Explore'
+    console.log(ctx.subagentDescription); // e.g. 'Search for auth code'
+    console.log(ctx.parentSessionId);     // parent session ID
+  }
+  return { pass: true };
+}, { scope: 'both' });
+```
+
+### SubagentType Filtering
+
+When you specify `subagentType`, the eval/enrichment only runs for subagents of that type. Subagents of other types will not see the eval panel at all.
+
+```js
+// Only runs for Explore subagents
+app.eval('explore-thoroughness', ({ entries }) => ({
+  pass: entries.length > 5,
+  score: Math.min(entries.length / 20, 1),
+}), { scope: 'subagent', subagentType: 'Explore' });
+
+// Runs for all subagent types
+app.eval('agent-efficiency', ({ stats }) => ({
+  pass: stats.turnCount <= 10,
+  score: Math.max(0, 1 - stats.turnCount / 20),
+}), { scope: 'subagent' });
+```
+
+### UI Behavior
+
+When a subagent is expanded in the log viewer, eval and enrichment panels appear below the stats bar (only if matching subagent-scoped evals/enrichments are registered). Panels use a compact layout to fit within the nested subagent view. If no subagent-scoped evals match the subagent's type, the panels are not rendered.
+
+### Caching
+
+Subagent eval/enrichment results are cached separately from session results:
+
+```
+~/.claudeye/cache/evals/{project}/{sessionId}.json              # session-level
+~/.claudeye/cache/evals/{project}/{sessionId}/agent-{id}.json   # subagent-level
+```
+
+Cache invalidation works the same way — based on the subagent log file's mtime+size and the evals module content hash.
+
+### Edge Cases
+
+- **No subagents in session**: Subagent-scoped evals never run — panels only mount inside expanded subagent cards
+- **`scope: 'both'` with `subagentType`**: At session level, the `subagentType` filter is ignored. At subagent level, it applies.
+- **Conditions + scope**: Scope filtering happens first (registry level), then conditions run with the full `EvalContext` including scope metadata
+- **Backward compatibility**: Existing evals with no `scope` option default to `'session'` — behavior is unchanged
+
+---
+
+## API Reference
+
+### `createApp()`
+
+Returns a `ClaudeyeApp` instance. All methods are chainable.
+
+```ts
+import { createApp } from 'claudeye';
+const app = createApp();
+```
+
+### `app.condition(fn)`
+
+Set a global condition that gates all evals and enrichments. Calling this multiple times replaces the previous condition.
+
+```ts
+app.condition(({ entries, stats, projectName, sessionId }) => boolean | Promise<boolean>);
+```
+
+### `app.eval(name, fn, options?)`
+
+Register an eval function.
 
 - **`name`** — unique string identifier for the eval
 - **`fn`** — function receiving an `EvalContext` and returning an `EvalResult`
+- **`options.condition`** — optional condition function to gate this eval
+- **`options.scope`** — `'session'` (default), `'subagent'`, or `'both'`
+- **`options.subagentType`** — only run for subagents of this type (e.g. `'Explore'`)
 
 ```ts
 app.eval('my-eval', (ctx) => ({ pass: true, score: 1.0 }));
+
+app.eval('conditional-eval', evalFn, {
+  condition: (ctx) => ctx.stats.turnCount > 0,
+});
+
+app.eval('subagent-eval', evalFn, {
+  scope: 'subagent',
+  subagentType: 'Explore',
+});
 ```
 
-#### `app.enrich(name, fn)`
+### `app.enrich(name, fn, options?)`
 
-Register an enricher function. Chainable.
+Register an enricher function.
 
 - **`name`** — unique string identifier for the enricher
 - **`fn`** — function receiving an `EvalContext` and returning a `Record<string, string | number | boolean>`
+- **`options.condition`** — optional condition function to gate this enricher
+- **`options.scope`** — `'session'` (default), `'subagent'`, or `'both'`
+- **`options.subagentType`** — only run for subagents of this type (e.g. `'Explore'`)
 
 ```ts
 app.enrich('metrics', (ctx) => ({
   'Turn Count': ctx.stats.turnCount,
   'Models Used': ctx.stats.models.join(', '),
 }));
+
+app.enrich('conditional-enricher', enrichFn, {
+  condition: (ctx) => ctx.entries.length > 0,
+});
+
+app.enrich('agent-cost', enrichFn, {
+  scope: 'subagent',
+});
 ```
 
-#### `app.listen(port?, options?)`
+### `app.listen(port?, options?)`
 
 Start the Claudeye dashboard server.
 
 - **`port`** — port number (default: 8020)
+- **`options.host`** — bind address (default: "localhost", use "0.0.0.0" for LAN)
 - **`options.open`** — auto-open browser (default: true)
 
 When the file is loaded via `--evals` or `CLAUDEYE_EVALS_MODULE`, `listen()` is a no-op.
 
-#### `EvalContext`
+### `EvalContext`
 
 Both evals and enrichers receive the same context object:
 
 ```ts
 interface EvalContext {
-  entries: EvalLogEntry[];  // Parsed log entries for the session
-  stats: EvalLogStats;      // Computed stats (turn count, tool calls, etc.)
-  projectName: string;      // Encoded project folder name
-  sessionId: string;        // Session UUID
+  entries: EvalLogEntry[];        // Parsed log entries for the session or subagent
+  stats: EvalLogStats;            // Computed stats (turn count, tool calls, etc.)
+  projectName: string;            // Encoded project folder name
+  sessionId: string;              // Session UUID
+  scope: 'session' | 'subagent'; // Whether running at session or subagent level
+  subagentId?: string;            // Hex ID of the subagent (subagent scope only)
+  subagentType?: string;          // e.g. 'Explore', 'Bash' (subagent scope only)
+  subagentDescription?: string;   // Short description (subagent scope only)
+  parentSessionId?: string;       // Parent session ID (subagent scope only)
 }
 ```
 
-#### `EvalLogStats`
+### `EvalLogEntry`
+
+```ts
+interface EvalLogEntry {
+  type: string;               // 'user' | 'assistant' | 'tool' | 'queue-operation' | ...
+  uuid: string;               // Unique entry ID
+  parentUuid: string | null;  // Parent ID for nested entries (subagents)
+  timestamp: string;          // ISO timestamp
+  timestampMs: number;        // Epoch milliseconds
+  timestampFormatted: string; // Human-readable timestamp
+  message?: {
+    role: string;                          // 'user' | 'assistant'
+    content: string | EvalContentBlock[];  // Text or structured content blocks
+    model?: string;                        // Model ID (assistant messages only)
+  };
+  raw?: Record<string, unknown>;  // Raw JSONL data (usage, errors, etc.)
+  label?: string;                 // Custom label
+}
+```
+
+### `EvalLogStats`
 
 ```ts
 interface EvalLogStats {
-  turnCount: number;
-  userCount: number;
-  assistantCount: number;
-  toolCallCount: number;
-  subagentCount: number;
-  duration: string;        // Formatted duration (e.g. "2m 15s")
-  models: string[];        // Model IDs used in the session
+  turnCount: number;      // Number of conversation turns
+  userCount: number;      // Number of user messages
+  assistantCount: number; // Number of assistant responses
+  toolCallCount: number;  // Total tool invocations
+  subagentCount: number;  // Number of subagent spawns
+  duration: string;       // Formatted duration (e.g. "2m 15s")
+  models: string[];       // Distinct model IDs used
 }
 ```
 
-#### `EvalResult`
+### `EvalResult`
 
 ```ts
 interface EvalResult {
-  pass: boolean;            // Required: did the eval pass?
-  score?: number;           // Optional: 0-1, clamped automatically
-  message?: string;         // Optional: shown in the UI
-  metadata?: Record<string, unknown>;  // Optional: arbitrary data
+  pass: boolean;                          // Did the eval pass?
+  score?: number;                         // 0-1, clamped automatically (default: 1.0)
+  message?: string;                       // Shown in the UI
+  metadata?: Record<string, unknown>;     // Arbitrary data
 }
 ```
 
-#### `EnrichmentResult`
+### `EnrichmentResult`
 
 ```ts
 // Enrichers return a flat key-value map
 type EnrichmentResult = Record<string, string | number | boolean>;
 ```
 
-### How It Works
+### `ConditionFunction`
+
+```ts
+type ConditionFunction = (context: EvalContext) => boolean | Promise<boolean>;
+```
+
+---
+
+## How It Works
 
-1. `createApp()` + `app.eval()` / `app.enrich()` register functions in global registries
-2. When you run `claudeye --evals ./my-file.js`, the server dynamically imports your file, populating both the eval and enricher registries
+1. `createApp()` + `app.eval()` / `app.enrich()` / `app.condition()` register functions in global registries
+2. When you run `claudeye --evals ./my-file.js`, the server dynamically imports your file, populating the eval, enricher, and condition registries
 3. When the dashboard loads a session, server actions run all registered evals and enrichers against the session's parsed log entries
-4. Each function is individually error-isolated — if one throws, the others still run
-5. Results are serialized and displayed in separate panels in the dashboard UI
+4. The global condition is checked first — if it fails, everything is skipped
+5. Per-item conditions are checked individually — skipped items don't block others
+6. Each function is individually error-isolated — if one throws, the others still run
+7. Results are serialized and displayed in separate panels in the dashboard UI
 
-### Tips
+## Tips
 
 - Each eval and enricher is wrapped in a try/catch. If one throws, the others still run and the error is shown in the UI.
-- Eval scores are clamped to 0-1. If you don't provide a score, passing evals default to 1.0 and failing evals default to 0.
+- Eval scores are clamped to 0-1. If you don't provide a score, it defaults to 1.0.
 - Both eval and enricher functions can be async.
+- Condition functions can also be async.
 - Re-registering with the same name replaces the previous function.
-- Click "Re-run" in either panel to re-execute against the current session.
-- You can mix `app.eval()` and `app.enrich()` calls freely in the same file — they're independent systems that share the same context.
+- Click "Re-run" in either panel to re-execute against the current session (always bypasses cache).
+- You can mix `app.eval()`, `app.enrich()`, and `app.condition()` calls freely in the same file.
+- The global condition applies to both evals and enrichments — there's no way to set separate global conditions for each.
+- Per-item condition errors are treated as eval/enrichment errors (not skips), so you'll see the error message in the UI.
 
 ## License