diff options
Diffstat (limited to 'help.html')
| -rw-r--r-- | help.html | 50 |
1 files changed, 50 insertions, 0 deletions
@@ -656,6 +656,56 @@ diff a.txt b.txt</pre> <p>A short <strong>"ding"</strong> plays via WebAudio when AI finishes. Toggle with the <strong>🔊 Sound</strong> button in quick-actions. Persists in localStorage.</p> </section> + <!-- ── Agent Loop Studio ── --> + <section id="agent-loops"> + <h2>🔄 Agent Loop Studio</h2> + <div class="callout callout-success"> + <span class="callout-title">Pluggable Agent Loops</span> + Switch between profiles that control max turns, tool sets, temperature, system prompt directives, and loop behavior — without restarting. + </div> + + <h3>Commands</h3> + <div class="table-wrap"><table> + <tr><th>Command</th><th>Description</th></tr> + <tr><td><code>agent-loops</code></td><td>List all profiles with turns, tools, and benchmark results</td></tr> + <tr><td><code>agent-loops <name></code></td><td>Switch to a named profile</td></tr> + <tr><td><code>agent-loops new <name> <desc></code></td><td>Create a new profile from defaults</td></tr> + <tr><td><code>agent-loops edit <name> <key> <val></code></td><td>Edit a profile setting (dot-notation, e.g. <code>loop.maxTurns 12</code>)</td></tr> + <tr><td><code>agent-loops delete <name></code></td><td>Delete a profile (not default)</td></tr> + <tr><td><code>agent-loops clone <src> <dst></code></td><td>Clone a profile</td></tr> + <tr><td><code>agent-loops export <name></code></td><td>Export profile as JSON</td></tr> + <tr><td><code>agent-loops reset <name></code></td><td>Reset profile to defaults</td></tr> + <tr><td><code>bench <profile> [--tasks=ids]</code></td><td>Run benchmark suite on a profile (6 tasks: fs-navigate, fs-create-read, nvfs-resolve, web-search, multi-tool, code-edit)</td></tr> + <tr><td><code>bench-compare <p1> <p2></code></td><td>Compare benchmark results of two profiles</td></tr> + </table></div> + + <h3>Default Profiles</h3> + <div class="table-wrap"><table> + <tr><th>Profile</th><th>Turns</th><th>Tools</th><th>Temp</th><th>Use Case</th></tr> + <tr><td><code>default</code></td><td>8</td><td>all</td><td>inherit</td><td>Balanced general-purpose</td></tr> + <tr><td><code>minimal</code></td><td>4</td><td>4 (custom)</td><td>0.3</td><td>Fast, low-token simple tasks</td></tr> + <tr><td><code>research</code></td><td>6</td><td>7 (custom)</td><td>0.5</td><td>Read-only web + file exploration</td></tr> + <tr><td><code>safe</code></td><td>8</td><td>all (destructive disabled)</td><td>0.5</td><td>No destructive tools</td></tr> + <tr><td><code>creative</code></td><td>12</td><td>all</td><td>0.9</td><td>More turns, higher creativity</td></tr> + <tr><td><code>coder</code></td><td>10</td><td>all</td><td>0.2</td><td>Code-focused with append directive</td></tr> + </table></div> + + <h3>System Prompt Layering</h3> + <div class="table-wrap"><table> + <tr><th>Command</th><th>Effect</th></tr> + <tr><td><code>ai-system +"directive"</code></td><td>Prepend a directive to the active profile's system prompt (persists to profile)</td></tr> + <tr><td><code>ai-system -"directive"</code></td><td>Append a directive to the active profile's system prompt (persists to profile)</td></tr> + <tr><td><code>ai-system "text"</code></td><td>Full replace (persists to localStorage, overrides profile)</td></tr> + <tr><td><code>ai-system reset</code></td><td>Restore default system prompt</td></tr> + </table></div> + + <h3>Self-Improvement</h3> + <p>After each AI run, <code>_postRunReview()</code> detects anomalies: near-max-turns, high token usage, tool failures, and above-benchmark consumption. Lessons are auto-saved as VFS skills under <code>/skills/loop-lesson-*</code>.</p> + + <h3>AI Tools</h3> + <p>The AI can also manage profiles via tools: <code>agent_loop_list</code>, <code>agent_loop_switch</code>, <code>agent_loop_create</code>, <code>agent_loop_benchmark</code>.</p> + </section> + <!-- ── Skills ── --> <section id="skills"> <h2>📚 Skills System</h2> |