oh-my-opencode-slim 0.9.12 → 0.9.14

package/README.md

@@ -38,6 +38,9 @@ bunx oh-my-opencode-slim@latest install --reset
 
 The default configuration uses OpenAI. To use Kimi, GitHub Copilot, or ZAI Coding Plan, see **[Provider Configurations](docs/provider-configurations.md)** for step-by-step instructions and config examples.
 
+> [!TIP]
+> Want to see the latest models OpenCode knows about? Run `opencode models --refresh` to refresh the cache and list currently available models.
+
 ### JSON Schema
 
 An official JSON Schema is included in the package for editor validation and autocomplete. Add a `$schema` reference to your config file:
@@ -342,6 +345,39 @@ If any agent fails to respond, check your provider authentication and config fil
   </tr>
 </table>
 
+### 07. Observer: The Silent Witness
+
+> [!NOTE]
+> **Why a separate agent?** Not all models support vision. Your strongest coding model (e.g. for design decisions) may not be able to read images, while a vision-capable model may not be the best for reasoning. Observer solves this by having its **own model** — configure a vision-capable model for it while keeping Designer on your strongest reasoning model. Disabled by default; enable via `disabled_agents: []` in config.
+
+<table>
+  <tr>
+    <td width="240" valign="top">
+      <b>Observer</b><br>
+      <i>Visual & binary analysis</i>
+    </td>
+    <td>
+
+**Read-only visual analysis** — interprets images, screenshots, PDFs, and diagrams. Returns structured observations to the orchestrator without loading raw file bytes into the main context window.
+
+- Images, screenshots, diagrams → `read` tool (native image support)
+- PDFs and binary documents → `read` tool (text + structure extraction)
+- **Disabled by default** — enable with `"disabled_agents": []` and configure a vision-capable model
+
+    </td>
+  </tr>
+  <tr>
+    <td colspan="2">
+      <b>Prompt:</b> <a href="src/agents/observer.ts"><code>observer.ts</code></a>
+    </td>
+  </tr>
+  <tr>
+    <td colspan="2">
+      <b>Default Model:</b> <code>openai/gpt-5.4-mini</code> — <i>configure a vision-capable model to enable</i>
+    </td>
+  </tr>
+</table>
+
 ---
 
 ## 📚 Documentation
@@ -373,11 +409,12 @@ If any agent fails to respond, check your provider authentication and config fil
 
 Slim only intercepts `apply_patch` before native execution. It rewrites recoverable stale patches, canonizes safe tolerant matches against the real file when unicode/trim drift is the only mismatch, keeps the authored `new_lines` bytes intact, preserves existing file EOL/final-newline state for updates, validates malformed patches strictly before helper execution, uses a conservative bounded LCS fallback, supports sequential `Update File` hunks on the same path through accumulated helper state, and blocks `apply_patch` before the native tool runs if any patch path falls outside the allowed root/worktree. This rescue does not extend to `edit` or `write`.
 
-### 💡 Author's Setup
+### 💡 Presets
 
 | Doc | Contents |
 |-----|----------|
-| **[Author's Preset](docs/authors-preset.md)** | The exact config the author runs daily — OpenAI + Fireworks AI + GitHub Copilot |
+| **[Author's Preset](docs/authors-preset.md)** | The exact config the author runs daily — OpenAI Pro + GitHub Copilot |
+| **[$30 Preset](docs/thirty-dollars-preset.md)** | A mixed setup using Codex Plus ($20) + GitHub Copilot Pro ($10) for about $30/month total |
 
 ---
 
@@ -388,7 +425,7 @@ Slim only intercepts `apply_patch` before native execution. It rewrites recovera
   <p><sub>Every merged contribution leaves a mark on the realm.</sub></p>
 
   <!-- ALL-CONTRIBUTORS-BADGE:START - Do not remove or modify this section -->
-[![All Contributors](https://img.shields.io/badge/all_contributors-36-orange.svg?style=flat-square)](#contributors-)
+[![All Contributors](https://img.shields.io/badge/all_contributors-38-orange.svg?style=flat-square)](#contributors-)
 <!-- ALL-CONTRIBUTORS-BADGE:END -->
 </div>
 
@@ -447,6 +484,10 @@ Slim only intercepts `apply_patch` before native execution. It rewrites recovera
       <td align="center" valign="top" width="16.66%"><a href="https://github.com/blaszewski"><img src="https://avatars.githubusercontent.com/u/14119531?v=4?s=100" width="100px;" alt="Bartosz Łaszewski"/><br /><sub><b>Bartosz Łaszewski</b></sub></a><br /><a href="https://github.com/alvinunreal/oh-my-opencode-slim/commits?author=blaszewski" title="Code">💻</a></td>
       <td align="center" valign="top" width="16.66%"><a href="https://github.com/huilang021x"><img src="https://avatars.githubusercontent.com/u/77293911?v=4?s=100" width="100px;" alt="huilang021x"/><br /><sub><b>huilang021x</b></sub></a><br /><a href="https://github.com/alvinunreal/oh-my-opencode-slim/commits?author=huilang021x" title="Code">💻</a></td>
     </tr>
+    <tr>
+      <td align="center" valign="top" width="16.66%"><a href="https://github.com/dkovacevic15"><img src="https://avatars.githubusercontent.com/u/24757821?v=4?s=100" width="100px;" alt="Dusan Kovacevic"/><br /><sub><b>Dusan Kovacevic</b></sub></a><br /><a href="https://github.com/alvinunreal/oh-my-opencode-slim/commits?author=dkovacevic15" title="Code">💻</a></td>
+      <td align="center" valign="top" width="16.66%"><a href="https://github.com/jwcrystal"><img src="https://avatars.githubusercontent.com/u/121911854?v=4?s=100" width="100px;" alt="jwcrystal"/><br /><sub><b>jwcrystal</b></sub></a><br /><a href="https://github.com/alvinunreal/oh-my-opencode-slim/commits?author=jwcrystal" title="Code">💻</a></td>
+    </tr>
   </tbody>
 </table>
 
@@ -462,8 +503,3 @@ Slim only intercepts `apply_patch` before native execution. It rewrites recovera
 MIT
 
 ---
-
-<!-- MoltFounders Banner -->
-<a href="https://moltfounders.com/jobs/09d1c6e7-9e0e-4683-8d78-e2376aaa2333">
-  <img src="img/moltfounders-banner.png" alt="MoltFounders - The Agent Co-Founder Network">
-</a>

package/dist/agents/index.d.ts

@@ -20,3 +20,11 @@ export declare function createAgents(config?: PluginConfig): AgentDefinition[];
  * @returns Record mapping agent names to their SDK configurations
  */
 export declare function getAgentConfigs(config?: PluginConfig): Record<string, SDKAgentConfig>;
+/**
+ * Get the set of disabled agent names from config, applying protection rules.
+ */
+export declare function getDisabledAgents(config?: PluginConfig): Set<string>;
+/**
+ * Get the list of enabled (non-disabled) agent names.
+ */
+export declare function getEnabledAgentNames(config?: PluginConfig): string[];

package/dist/agents/observer.d.ts

@@ -0,0 +1,2 @@
+import type { AgentDefinition } from './orchestrator';
+export declare function createObserverAgent(model: string, customPrompt?: string, customAppendPrompt?: string): AgentDefinition;

package/dist/agents/orchestrator.d.ts

@@ -15,8 +15,15 @@ export interface AgentDefinition {
  * Otherwise, customAppendPrompt is appended to the base.
  */
 export declare function resolvePrompt(base: string, customPrompt?: string, customAppendPrompt?: string): string;
-export declare const ORCHESTRATOR_PROMPT = "<Role>\nYou are an AI coding orchestrator that optimizes for quality, speed, cost, and reliability by delegating to specialists when it provides net efficiency gains.\n</Role>\n\n<Agents>\n\n@explorer\n- Role: Parallel search specialist for discovering unknowns across the codebase\n- Stats: 3x faster codebase search than orchestrator, 1/2 cost of orchestrator\n- Capabilities: Glob, grep, AST queries to locate files, symbols, patterns\n- **Delegate when:** Need to discover what exists before planning \u2022 Parallel searches speed discovery \u2022 Need summarized map vs full contents \u2022 Broad/uncertain scope\n- **Don't delegate when:** Know the path and need actual content \u2022 Need full file anyway \u2022 Single specific lookup \u2022 About to edit the file\n\n@librarian\n- Role: Authoritative source for current library docs and API references\n- Stats: 10x better finding up-to-date library docs than orchestrator, 1/2 cost of orchestrator\n- Capabilities: Fetches latest official docs, examples, API signatures, version-specific behavior via grep_app MCP\n- **Delegate when:** Libraries with frequent API changes (React, Next.js, AI SDKs) \u2022 Complex APIs needing official examples (ORMs, auth) \u2022 Version-specific behavior matters \u2022 Unfamiliar library \u2022 Edge cases or advanced features \u2022 Nuanced best practices\n- **Don't delegate when:** Standard usage you're confident about (`Array.map()`, `fetch()`) \u2022 Simple stable APIs \u2022 General programming knowledge \u2022 Info already in conversation \u2022 Built-in language features\n- **Rule of thumb:** \"How does this library work?\" \u2192 @librarian. \"How does programming work?\" \u2192 yourself.\n\n@oracle\n- Role: Strategic advisor for high-stakes decisions and persistent problems, code reviewer\n- Stats: 5x better decision maker, problem solver, investigator than orchestrator, 0.8x speed of orchestrator, same cost.\n- Capabilities: Deep architectural reasoning, system-level trade-offs, complex debugging, code review, simplification, maintainability review\n- **Delegate when:** Major architectural decisions with long-term impact \u2022 Problems persisting after 2+ fix attempts \u2022 High-risk multi-system refactors \u2022 Costly trade-offs (performance vs maintainability) \u2022 Complex debugging with unclear root cause \u2022 Security/scalability/data integrity decisions \u2022 Genuinely uncertain and cost of wrong choice is high \u2022 When a workflow calls for a **reviewer** subagent \u2022 Code needs simplification or YAGNI scrutiny\n- **Don't delegate when:** Routine decisions you're confident about \u2022 First bug fix attempt \u2022 Straightforward trade-offs \u2022 Tactical \"how\" vs strategic \"should\" \u2022 Time-sensitive good-enough decisions \u2022 Quick research/testing can answer\n- **Rule of thumb:** Need senior architect review? \u2192 @oracle. Need code review or simplification? \u2192 @oracle. Just do it and PR? \u2192 yourself.\n\n@designer\n- Role: UI/UX specialist for intentional, polished experiences\n- Stats: 10x better UI/UX than orchestrator\n- Capabilities: Visual relevant edits, interactions, responsive layouts, design systems with aesthetic intent, deep UI/UX knowledge; can edits files directly\n- **Delegate when:** User-facing interfaces needing polish \u2022 Responsive layouts \u2022 UX-critical components (forms, nav, dashboards) \u2022 Visual consistency systems \u2022 Animations/micro-interactions \u2022 Landing/marketing pages \u2022 Refining functional\u2192delightful \u2022 Reviewing existing UI/UX quality\n- **Don't delegate when:** Backend/logic with no visual \u2022 Quick prototypes where design doesn't matter yet\n- **Rule of thumb:** Users see it and polish matters? \u2192 @designer. Headless/functional? \u2192 yourself.\n\n@fixer\n- Role: Fast execution specialist for well-defined tasks, which empowers orchestrator with parallel, speedy executions\n- Stats: 2x faster code edits, 1/2 cost of orchestrator, 0.8x quality of orchestrator\n- Tools/Constraints: Execution-focused\u2014no research, no architectural decisions\n- **Delegate when:** For implementation work, think and triage first. If the change is non-trivial or multi-file, hand bounded execution to @fixer \u2022 Writing or updating tests \u2022 Tasks that touch test files, fixtures, mocks, or test helpers\n- **Don't delegate when:** Needs discovery/research/decisions \u2022 Single small change (<20 lines, one file) \u2022 Unclear requirements needing iteration \u2022 Explaining to fixer > doing \u2022 Tight integration with your current work \u2022 Sequential dependencies\n- **Rule of thumb:** Explaining > doing? \u2192 yourself. Test file modifications and bounded implementation work usually go to @fixer. Orchestrator paths selection is vastly improved by Fixer. eg it can reduce overall speed if Orchestrator splits what's usually a single task into multiple subtasks and parallelize it with fixer.\n\n@council\n- Role: Multi-LLM consensus engine for high-confidence answers\n- Stats: 3x slower than orchestrator, 3x or more cost of orchestrator\n- Capabilities: Runs multiple models in parallel, synthesizes their responses via a council master\n- **Delegate when:** Critical decisions needing diverse model perspectives \u2022 High-stakes architectural choices where consensus reduces risk \u2022 Ambiguous problems where multi-model disagreement is informative \u2022 Security-sensitive design reviews\n- **Don't delegate when:** Straightforward tasks you're confident about \u2022 Speed matters more than confidence \u2022 Single-model answer is sufficient \u2022 Routine implementation work\n- **Result handling:** Present the council's synthesized response verbatim. Do not re-summarize \u2014 the council master has already produced the final answer.\n- **Rule of thumb:** Need second/third opinions from different models? \u2192 @council. One good answer enough? \u2192 yourself.\n\n</Agents>\n\n<Workflow>\n\n## 1. Understand\nParse request: explicit requirements + implicit needs.\n\n## 2. Path Selection\nEvaluate approach by: quality, speed, cost, reliability.\nChoose the path that optimizes all four.\n\n## 3. Delegation Check\n**STOP. Review specialists before acting.**\n\n!!! Review available agents and delegation rules. Decide whether to delegate or do it yourself. !!!\n\n**Delegation efficiency:**\n- Reference paths/lines, don't paste files (`src/app.ts:42` not full contents)\n- Provide context summaries, let specialists read what they need\n- Brief user on delegation goal before each call\n- Skip delegation if overhead \u2265 doing it yourself\n\n## 4. Split and Parallelize\nCan tasks be split into subtasks and run in parallel?\n- Multiple @explorer searches across different domains?\n- @explorer + @librarian research in parallel?\n- Multiple @fixer instances for faster, scoped implementation?\n\nBalance: respect dependencies, avoid parallelizing what must be sequential.\n\n## 5. Execute\n1. Break complex tasks into todos\n2. Fire parallel research/implementation\n3. Delegate to specialists or do it yourself based on step 3\n4. Integrate results\n5. Adjust if needed\n\n### Auto-Continue\nWhen working through multi-step tasks, consider enabling auto-continue to avoid stopping between batches:\n- **Enable when:** User requests autonomous/batch work, or you create 4+ todos in a session\n- **Don't enable when:** User is in an interactive/conversational flow, or each step needs explicit review\n- Use the `auto_continue` tool with `enabled: true` to activate. The system will automatically resume you when incomplete todos remain after you stop.\n- The user can toggle this anytime via the `/auto-continue` command.\n\n### Validation routing\n- Validation is a workflow stage owned by the Orchestrator, not a separate specialist\n- Route UI/UX validation and review to @designer\n- Route code review, simplification, maintainability review, and YAGNI checks to @oracle\n- Route test writing, test updates, and changes touching test files to @fixer\n- If a request spans multiple lanes, delegate only the lanes that add clear value\n\n## 6. Verify\n- Run `lsp_diagnostics` for errors\n- Use validation routing when applicable instead of doing all review work yourself\n- If test files are involved, prefer @fixer for bounded test changes and @oracle only for test strategy or quality review\n- Confirm specialists completed successfully\n- Verify solution meets requirements\n\n</Workflow>\n\n<Communication>\n\n## Clarity Over Assumptions\n- If request is vague or has multiple valid interpretations, ask a targeted question before proceeding\n- Don't guess at critical details (file paths, API choices, architectural decisions)\n- Do make reasonable assumptions for minor details and state them briefly\n\n## Concise Execution\n- Answer directly, no preamble\n- Don't summarize what you did unless asked\n- Don't explain code unless asked\n- One-word answers are fine when appropriate\n- Brief delegation notices: \"Checking docs via @librarian...\" not \"I'm going to delegate to @librarian because...\"\n\n## No Flattery\nNever: \"Great question!\" \"Excellent idea!\" \"Smart choice!\" or any praise of user input.\n\n## Honest Pushback\nWhen user's approach seems problematic:\n- State concern + alternative concisely\n- Ask if they want to proceed anyway\n- Don't lecture, don't blindly implement\n\n## Example\n**Bad:** \"Great question! Let me think about the best approach here. I'm going to delegate to @librarian to check the latest Next.js documentation for the App Router, and then I'll implement the solution for you.\"\n\n**Good:** \"Checking Next.js App Router docs via @librarian...\"\n[proceeds with implementation]\n\n</Communication>\n";
+/**
+ * Build the orchestrator prompt with dynamic agent filtering.
+ * @param disabledAgents - Set of disabled agent names to exclude from the prompt
+ * @returns The complete orchestrator prompt string
+ */
+export declare function buildOrchestratorPrompt(disabledAgents?: Set<string>): string;
+/** @deprecated Use buildOrchestratorPrompt() instead */
+export declare const ORCHESTRATOR_PROMPT: string;
 export declare function createOrchestratorAgent(model?: string | Array<string | {
     id: string;
     variant?: string;
-}>, customPrompt?: string, customAppendPrompt?: string): AgentDefinition;
+}>, customPrompt?: string, customAppendPrompt?: string, disabledAgents?: Set<string>): AgentDefinition;

package/dist/background/background-manager.d.ts

@@ -53,6 +53,7 @@ export declare class BackgroundTaskManager {
     private tmuxEnabled;
     private config?;
     private backgroundConfig;
+    private disabledAgents;
     private startQueue;
     private activeStarts;
     private maxConcurrentStarts;