opencodekit 0.15.21 → 0.16.1

package/dist/template/.opencode/memory/observations/2026-01-30-decision-github-copilot-claude-routing-keep-disab.md

@@ -1,32 +0,0 @@
----
-type: decision
-created: 2026-01-30T15:07:59.081Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["github-copilot", "claude", "anthropic-sdk", "reasoning", "api-routing"]
-files: [".opencode/plugins/copilot-auth.ts"]
----
-
-# 🎯 GitHub Copilot Claude routing: Keep disabled for stability
-
-🟢 **Confidence:** high
-
-Claude routing through Anthropic SDK is currently disabled in copilot-auth.ts (commit f7aabde). This decision was made because:
-
-1. **Problem**: Enabling Claude routing to use `@ai-sdk/anthropic` with `/v1` endpoint causes "Not Found: 404" errors
-2. **Root cause**: GitHub Copilot's API likely doesn't support the native Anthropic `/v1/messages` endpoint for Claude models through this integration
-3. **Solution**: Keep all models using `@ai-sdk/github-copilot` (OpenAI-compatible `/v1/chat/completions`)
-4. **Trade-off**: All models work, but Claude models don't get native Anthropic SDK optimizations or reasoning features through the `/v1` gateway
-
-To enable reasoning in the future, would need to:
-- Research if GitHub Copilot supports a different endpoint for reasoning
-- Check if there's a separate gateway URL for reasoning models
-- Verify with GitHub Copilot API documentation
-
-Current working configuration (commit f7aabde):
-- All models use `@ai-sdk/github-copilot`
-- All models use `/v1/chat/completions` (OpenAI-compatible)
-- Claude models work but don't have reasoning enabled
-
-File: .opencode/plugins/copilot-auth.ts (lines 93-100)

package/dist/template/.opencode/memory/observations/2026-01-30-discovery-context-management-research-critical-gap.md

@@ -1,14 +0,0 @@
----
-type: discovery
-created: 2026-01-30T16:51:31.023Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["context-management", "tool-loadout", "context-pruning", "subagents", "swarm-coordination", "RAG"]
----
-
-# 💡 Context Management Research - Critical Gaps in OpenCodeKit
-
-🟢 **Confidence:** high
-
-Research from dbreunig.com reveals four context failure modes: Poisoning, Distraction, Confusion, Clash. Key findings: models fail beyond 100k tokens, >30 tools causes guaranteed failures, multi-turn prompts yield 39% worse results. OpenCodeKit has 40+ tools in context (33% over limit). Critical gaps: 1) No tool loadout system, 2) No structured context assembly, 3) No scratchpad/think tool, 4) No automated pruning at thresholds, 5) No poisoning detection. Immediate action needed: implement tool loadout to reduce tools from 40+ to <25. Should improve accuracy 20-40% based on Berkeley Function-Calling Leaderboard findings.

package/dist/template/.opencode/memory/observations/2026-01-30-discovery-kimi-k2-5-agent-swarm-architecture-patte.md

@@ -1,45 +0,0 @@
----
-type: discovery
-created: 2026-01-30T05:46:18.057Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["Kimi K2.5", "agent swarm", "PARL", "parallel execution", "task decomposition", "coordination mechanisms"]
----
-
-# 💡 Kimi K2.5 Agent Swarm Architecture Patterns
-
-🟢 **Confidence:** high
-
-Comprehensive research completed on Kimi K2.5's agent swarm architecture and patterns. Key findings:
-
-## Core Architecture
-- **PARL (Parallel-Agent Reinforcement Learning)**: Training method using staged reward shaping to prevent "serial collapse"
-- **Orchestrator-Agent Model**: Central coordinator that decomposes tasks and manages sub-agents
-- **Critical Steps Metric**: Latency-focused evaluation inspired by parallel computation critical path
-- **Up to 100 sub-agents**: Dynamically instantiated frozen workers
-- **1,500 coordinated tool calls**: Maximum parallel workflow capacity
-
-## Task Decomposition Patterns
-1. **Research Tasks**: Creates specialized agents (AI Researcher, Physics Researcher, Fact Checker)
-2. **Dataset Creation**: Three-phase pipeline (generation → verification → calibration) 
-3. **Multimodal QA**: One-agent-per-item approach with separate coordinator
-4. **Writing Tasks**: Conservative single-agent allocation for tightly coupled work
-
-## Coordination Mechanisms
-- **Agent Modes**: Explore, Decompose, Generate, Reflect, Synthesize
-- **Dynamic Scheduling**: Reuses idle agents, manages bottlenecks
-- **Handoff Protocol**: Agents can pass work between each other
-- **Progress Tracking**: Visual interface shows individual agent status
-
-## Failure Handling
-- **Serial Collapse Prevention**: PARL's staged reward system (λaux anneals 0.1→0.0)
-- **Self-Correction**: Orchestrator monitors and adjusts strategy
-- **Verification First**: Quality control prioritized over raw throughput
-- **Computational Bottleneck**: Makes sequential execution impractical
-
-## Performance Results
-- **4.5x speed improvement** over single-agent for complex tasks
-- **80% reduction** in end-to-end runtime on workloads
-- **CriticalSteps metric**: Focuses on slowest path optimization
-- **Quality over Quantity**: May not honor exact agent count requests

package/dist/template/.opencode/memory/observations/2026-01-30-pattern-swarm-tools-architecture.md

@@ -1,28 +0,0 @@
----
-type: pattern
-created: 2026-01-30T06:11:22.169Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["swarm", "tools", "optimization", "architecture", "opencode"]
----
-
-# 🔄 Swarm Tools Architecture
-
-🟢 **Confidence:** high
-
-Current swarm infrastructure has 5 separate tools:
-
-EXISTING:
-1. swarm-delegate.ts - Creates delegation packets for workers
-2. swarm-helper.ts - Mailbox coordination (getTeamStatus, sendTeamMessage)
-
-NEW (created):
-3. task-classifier.ts - Task analysis & classification
-4. progress-tracker.ts - Real-time progress UI
-5. swarm-orchestrator.ts - Anti-serial-collapse detection
-
-PLUGIN:
-6. swarm-enforcer.ts - Enforces Beads workflow via nudges
-
-Optimization opportunity: Merge overlapping functionality into 2-3 comprehensive tools that work together as a unified swarm system.

package/dist/template/.opencode/memory/observations/2026-01-31-decision-copilot-auth-plugin-updated-with-baseurl.md

@@ -1,63 +0,0 @@
----
-type: decision
-created: 2026-01-31T05:13:11.027Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["copilot", "auth", "plugin", "baseURL", "chat.headers", "anthropic-beta", "reasoning", "thinking", "enterprise"]
----
-
-# 🎯 Copilot Auth Plugin Updated with baseURL and chat.headers
-
-🟢 **Confidence:** high
-
-## Changes Made to copilot-auth.ts
-
-### 1. Added baseURL return for enterprise support
-```typescript
-// Enterprise URL support for baseURL
-const enterpriseUrl = (info as any).enterpriseUrl;
-const baseURL = enterpriseUrl
-  ? `https://copilot-api.${normalizeDomain(enterpriseUrl)}`
-  : undefined;
-
-return {
-  baseURL,  // ← Now returned
-  apiKey: "",
-  // ...
-}
-```
-
-### 2. Added chat.headers hook for Claude reasoning support
-```typescript
-"chat.headers": async (input: any, output: any) => {
-  // Only apply to GitHub Copilot provider
-  if (!input.model?.providerID?.includes("github-copilot")) return;
-
-  // Add Anthropic beta header for interleaved thinking (extended reasoning)
-  if (input.model?.api?.npm === "@ai-sdk/anthropic") {
-    output.headers["anthropic-beta"] = "interleaved-thinking-2025-05-14";
-  }
-
-  // Mark subagent sessions as agent-initiated
-  const session = await sdk.session.get(...);
-  if (session?.data?.parentID) {
-    output.headers["x-initiator"] = "agent";
-  }
-}
-```
-
-### Why These Changes Matter
-
-1. **baseURL**: Enterprise users using GitHub Enterprise with data residency will now get correct API routing to `copilot-api.<domain>`.
-
-2. **anthropic-beta header**: When Claude routing is re-enabled (the commented-out code), this header will be automatically added to enable extended thinking/reasoning. The header `interleaved-thinking-2025-05-14` is required for Claude to return thinking blocks.
-
-3. **Subagent x-initiator**: Sessions that are children of other sessions (subagents) will be marked as `agent` initiated, matching the standard that other Copilot tools use.
-
-### Note on Reasoning Activation
-
-The anthropic-beta header only activates when:
-- `input.model.api.npm === "@ai-sdk/anthropic"`
-
-Currently, all models use `@ai-sdk/github-copilot`, so the header won't be added until the commented Claude routing code is re-enabled.

package/dist/template/.opencode/memory/observations/2026-01-31-decision-created-dedicated-worker-agent-for-swarm.md

@@ -1,20 +0,0 @@
----
-type: decision
-created: 2026-01-31T13:42:23.877Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["worker", "swarm", "tmux", "agent", "spawn_session", "primary", "subagent"]
----
-
-# 🎯 Created dedicated worker agent for swarm
-
-🟢 **Confidence:** high
-
-Created a new `worker` agent (primary mode) specifically for tmux-spawned swarm tasks. This keeps `general` as a subagent (for Task tool delegation) while providing a standalone agent that can be spawned via `opencode run --agent worker`.
-
-Key differences:
-- worker: mode=primary, can run standalone, focused on delegated task execution
-- general: mode=subagent, only via Task tool, fast small tasks
-
-The tmux tool now defaults to `agent: "worker"` for spawn_session operations.

package/dist/template/.opencode/memory/observations/2026-01-31-decision-rollback-to-v1-1-47-for-copilot-claude-r.md

@@ -1,21 +0,0 @@
----
-type: decision
-created: 2026-01-31T17:20:18.779Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["copilot", "claude", "reasoning", "thinking_budget", "rollback", "v1.1.47", "signature validation"]
----
-
-# 🎯 Rollback to v1.1.47 for Copilot Claude reasoning
-
-🟢 **Confidence:** high
-
-User decided to rollback to OpenCode v1.1.47 approach instead of waiting for PR #8958 to merge. The v1.1.48+ native SDK has "Invalid signature in thinking block" errors with Claude models via GitHub Copilot.
-
-Rollback includes:
-1. Restore copilot-auth.ts with thinking_budget injection (from commit 6f003b5)
-2. Restore custom SDK at .opencode/plugin/sdk/copilot/ with reasoning_text/reasoning_opaque handling
-3. Restore opencode.json github-copilot provider config with reasoning: true for Claude models
-
-This custom SDK bypasses native signature validation but provides working extended thinking.

package/dist/template/.opencode/memory/observations/2026-01-31-decision-simplified-swarm-to-task-tool-pattern.md

@@ -1,44 +0,0 @@
----
-type: decision
-created: 2026-01-31T18:41:11.898Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["swarm", "task-tool", "parallel-execution", "mailbox", "simplification"]
-files: [".opencode/tool/swarm-monitor.ts", ".opencode/tool/swarm-delegate.ts", ".opencode/plugin/swarm-enforcer.ts", ".opencode/agent/build.md"]
----
-
-# 🎯 Simplified swarm to Task tool pattern
-
-🟢 **Confidence:** high
-
-## Decision
-Removed mailbox complexity from swarm system. The Task tool handles agent lifecycle and result collection natively.
-
-## Changes Made
-1. **swarm-monitor.ts**: Removed mail_send, mail_read, persist_todos, restore_todos, get_todos. Now only has progress_update, progress_render, render_block, status, clear.
-2. **swarm-delegate.ts**: Updated example to remove swarm-monitor references
-3. **swarm-enforcer.ts**: Renamed from "SWARM PROTOCOL" to "BEADS PROTOCOL", simplified messaging
-4. **build.md agent**: Dramatically simplified - removed complex PARL workflow, tmux panes, mailbox monitoring. Now focuses on Task tool pattern.
-5. **Removed files**: tmux.ts, worker.md (no longer needed)
-
-## Why
-- Task tool already handles results returning to parent
-- OpenCode manages agent lifecycle automatically
-- Mailbox was adding complexity without value
-- Session navigation (ctrl+alt+left/right) provides visibility
-
-## New Pattern
-```typescript
-// Simple parallel execution
-Task({ subagent_type: "explore", prompt: "Find X..." });
-Task({ subagent_type: "general", prompt: "Implement Y..." });
-// Results return directly - no mailbox needed
-```
-
-## Files Affected
-- .opencode/tool/swarm-monitor.ts
-- .opencode/tool/swarm-delegate.ts
-- .opencode/plugin/swarm-enforcer.ts
-- .opencode/agent/build.md
-- dist/template/* (all synced)

package/dist/template/.opencode/memory/observations/2026-01-31-decision-swarm-architecture-task-tool-over-tmux.md

@@ -1,33 +0,0 @@
----
-type: decision
-created: 2026-01-31T18:34:23.413Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["swarm", "task-tool", "child-session", "parallel", "architecture", "simplify"]
----
-
-# 🎯 Swarm architecture: Task tool over tmux
-
-🟢 **Confidence:** high
-
-Decided to use Task tool (child sessions) instead of tmux-spawned orphan sessions for swarm coordination.
-
-Reasons:
-1. Task tool manages lifecycle automatically - no orphaned processes
-2. Results return directly to parent - no mailbox communication needed
-3. Built-in session navigation (ctrl+alt+right/left) for visibility
-4. Simpler architecture - no process management complexity
-
-Removed:
-- worker agent (.opencode/agent/worker.md)
-- tmux tool (.opencode/tool/tmux.ts)
-- swarm-monitor mailbox complexity
-
-Pattern for parallel work:
-```typescript
-// Launch parallel subagents - results return to parent
-Task({ subagent_type: "general", prompt: "Task 1..." });
-Task({ subagent_type: "explore", prompt: "Task 2..." });
-// Navigate to child sessions: ctrl+alt+right/left
-```

package/dist/template/.opencode/memory/observations/2026-01-31-decision-worker-skills-defined-for-swarm-delegati.md

@@ -1,30 +0,0 @@
----
-type: decision
-created: 2026-01-31T19:09:09.501Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["swarm", "delegation", "worker skills", "verification", "TDD", "tool priority"]
----
-
-# 🎯 Worker skills defined for swarm delegation
-
-🟢 **Confidence:** high
-
-Updated swarm-delegate and swarm-plan tools to recommend essential skills for worker agents:
-
-Required/Recommended skills for workers:
-1. verification-before-completion - REQUIRED: Workers must verify before claiming done
-2. test-driven-development - RECOMMENDED: Write tests first, watch fail, implement, verify pass  
-3. tool-priority - RECOMMENDED: Use LSP before edits, grep for patterns, read before edit
-
-Changes made:
-- swarm-delegate.ts: Added "Recommended skills for workers" section to tool description
-- swarm-plan.ts: Added "Worker skills" section with skill() calls to load
-
-These skills ensure workers:
-- Follow proper verification protocol (no false "done" claims)
-- Use TDD for quality implementation
-- Use correct tool priority (LSP before edits)
-
-Both tools synced to dist/template/.

package/dist/template/.opencode/memory/observations/2026-01-31-learning-gpt-reasoning-config-for-github-copilot.md

@@ -1,51 +0,0 @@
----
-type: learning
-created: 2026-01-31T18:13:49.325Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["opencode", "gpt", "reasoning", "github-copilot", "model-config", "variants", "reasoningEffort"]
----
-
-# 📚 GPT reasoning config for GitHub Copilot
-
-🟢 **Confidence:** high
-
-## GPT Models Reasoning Configuration on GitHub Copilot Provider
-
-For OpenAI GPT models (gpt-5.2, gpt-5.2-codex) via github-copilot provider, reasoning is configured using these options:
-
-### Key Options
-- `reasoningEffort`: "low" | "medium" | "high" - Controls reasoning depth
-- `reasoningSummary`: "auto" | "always" | "never" - Show reasoning summary
-- `textVerbosity`: "low" | "medium" | "high" - Controls output verbosity
-- `include`: ["reasoning.encrypted_content"] - Include encrypted reasoning
-
-### Configuration Structure
-```json
-"github-copilot": {
-  "models": {
-    "gpt-5.2": {
-      "reasoning": true,
-      "options": {
-        "reasoningEffort": "medium",
-        "reasoningSummary": "auto"
-      },
-      "variants": {
-        "high": { "reasoningEffort": "high", "include": ["reasoning.encrypted_content"] },
-        "low": { "reasoningEffort": "low" },
-        "medium": { "reasoningEffort": "medium" }
-      }
-    }
-  }
-}
-```
-
-### Usage
-1. Select model: `/models` → choose `github-copilot/gpt-5.2`
-2. Cycle variants: Use keybind to switch between low/medium/high
-3. Or add `--variant high` when starting
-
-### Difference from Claude models
-- Claude uses `thinking_budget` (token count)
-- GPT uses `reasoningEffort` (low/medium/high qualitative levels)

package/dist/template/.opencode/memory/observations/2026-01-31-learning-opencode-copilot-auth-comparison-finding.md

@@ -1,61 +0,0 @@
----
-type: learning
-created: 2026-01-31T05:04:00.660Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["copilot", "auth", "plugin", "token-refresh", "reasoning", "thinking", "github-copilot", "opencode", "comparison"]
-files: ["index.mjs"]
----
-
-# 📚 OpenCode Copilot Auth Comparison Findings
-
-🟢 **Confidence:** high
-
-## Key Findings from Comparing Local vs Official OpenCode Copilot Auth Plugin
-
-### Official Plugin Location
-- Repository: `anomalyco/opencode-copilot-auth` (separate repo)
-- File: `index.mjs`
-- The auth.ts in main opencode repo is just auth orchestration, not the actual plugin
-
-### Critical Differences Found
-
-#### 1. CLIENT_ID Mismatch
-- **Local**: `Ov23li8tweQw6odWQebz` 
-- **Official**: `Iv1.b507a08c87ecfe98`
-- CRITICAL: Different OAuth client IDs will affect authentication!
-
-#### 2. Token Refresh Mechanism
-- **Local**: Does NOT refresh Copilot API token - uses OAuth access token directly
-- **Official**: Uses `COPILOT_API_KEY_URL` (`/copilot_internal/v2/token`) to get short-lived Copilot tokens
-- CRITICAL: Local version may have token expiration issues
-
-#### 3. Token Storage
-- **Local**: Stores `data.access_token` as both `refresh` and `access` with `expires: 0`
-- **Official**: Stores OAuth token as `refresh`, and fetches separate Copilot token as `access` with proper expiration
-
-#### 4. Rate Limit Handling
-- **Local**: Has retry logic with exponential backoff for 429 errors
-- **Official**: No retry logic - relies on proper token refresh instead
-
-#### 5. Model Configuration
-- **Local**: Sets `model.api.npm = "@ai-sdk/github-copilot"` explicitly
-- **Official**: Does NOT modify model SDK - lets OpenCode core handle it
-
-#### 6. Agent Detection Logic
-- **Local**: Checks ALL messages for tool/assistant roles
-- **Official**: Only checks LAST message for agent detection
-
-#### 7. baseURL Configuration
-- **Local**: Does NOT set baseURL (commented out Claude routing)
-- **Official**: Sets baseURL dynamically based on enterprise vs public deployment
-
-### Reasoning/Thinking Block Fix Context
-The v1.1.46 fix "Remove AI SDK middleware that was preventing think blocks" is in the core OpenCode provider code, NOT in the copilot auth plugin. The `extractReasoningMiddleware` is used to extract thinking from tags like `<think>`, but Anthropic's native thinking doesn't need this.
-
-### Recommendation
-The local plugin has diverged significantly from official. Key issues:
-1. Wrong CLIENT_ID will cause auth problems
-2. Missing token refresh logic will cause "Not Found" or auth errors
-3. Agent detection differs which affects `x-initiator` header

package/dist/template/.opencode/memory/observations/2026-01-31-learning-opencode-copilot-reasoning-architecture-.md

@@ -1,66 +0,0 @@
----
-type: learning
-created: 2026-01-31T05:23:35.673Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["copilot", "reasoning", "thinking", "config", "opencode", "sdk", "thinking_budget", "reasoning_text", "reasoning_opaque"]
----
-
-# 📚 OpenCode Copilot Reasoning Architecture Explained
-
-🟢 **Confidence:** high
-
-## OpenCode Copilot Reasoning Architecture
-
-### Config Parameters Found in opencode.json
-The `github-copilot` provider has Claude models with reasoning configured:
-```json
-"claude-haiku-4.5": {
-  "reasoning": true,
-  "variants": {
-    "high": {
-      "options": {
-        "thinking": {
-          "budgetTokens": 32000,
-          "type": "enabled"
-        }
-      }
-    }
-  }
-}
-```
-
-### How OpenCode v1.1.46 Actually Handles Copilot Reasoning
-
-OpenCode created a custom Copilot SDK (`packages/opencode/src/provider/sdk/copilot/`) that:
-
-1. **Does NOT use Anthropic's `/v1/messages` endpoint** - uses standard `/chat/completions`
-2. **Sends `thinking_budget` parameter** in request args
-3. **Looks for `reasoning_text` and `reasoning_opaque` fields** in response
-4. **Preserves `reasoning_opaque`** for multi-turn conversations
-
-### Key Response Schema Fields (Copilot-specific)
-```typescript
-message: {
-  content: z.string().nullish(),
-  reasoning_text: z.string().nullish(),      // Reasoning content
-  reasoning_opaque: z.string().nullish(),    // Opaque token for multi-turn
-}
-```
-
-### Provider Options Mapping
-The custom SDK accepts:
-- `thinking_budget` → budget tokens for thinking
-- `reasoningEffort` → low/medium/high effort level
-
-### Why Standard SDK Doesn't Work
-- `@ai-sdk/github-copilot` doesn't look for `reasoning_text` field
-- It doesn't pass `thinking_budget` to the API
-- It doesn't preserve `reasoning_opaque` for conversation continuity
-
-### Solution Required
-To enable Copilot reasoning, need to either:
-1. Wait for `@ai-sdk/github-copilot` to add support
-2. Copy OpenCode's custom `sdk/copilot` implementation
-3. Contribute the feature upstream

package/dist/template/.opencode/memory/observations/2026-01-31-learning-opencode-custom-tools-api.md

@@ -1,48 +0,0 @@
----
-type: learning
-created: 2026-01-31T13:08:45.478Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["opencode", "custom-tools", "plugin", "bun-shell", "tool-schema"]
----
-
-# 📚 OpenCode Custom Tools API
-
-🟢 **Confidence:** high
-
-## Proper Custom Tool Format
-
-OpenCode custom tools must use the `@opencode-ai/plugin` package:
-
-```typescript
-import { tool } from "@opencode-ai/plugin";
-
-export default tool({
-  description: "Tool description",
-  args: {
-    param: tool.schema.string().describe("Parameter description"),
-  },
-  async execute(args, context) {
-    // context provides: agent, sessionID, messageID, directory, worktree
-    return "result";
-  },
-});
-```
-
-### Key Points:
-1. **Import**: `import { tool } from "@opencode-ai/plugin"`
-2. **Schema**: Use `tool.schema` (which is Zod) for argument types
-3. **Shell commands**: Use `Bun.$` for shell commands, NOT Node.js child_process
-4. **Context**: Second argument provides session info (directory, worktree, etc.)
-5. **Multiple exports**: Each named export becomes `<filename>_<exportname>` tool
-
-### Shell Command Example:
-```typescript
-// Use Bun shell (NOT child_process)
-const result = await Bun.$`tmux -V`.quiet();
-return result.stdout.toString().trim();
-```
-
-### Docs Reference:
-https://opencode.ai/docs/custom-tools/

package/dist/template/.opencode/memory/observations/2026-01-31-learning-opencode-v1-1-48-skills-as-slash-command.md

@@ -1,21 +0,0 @@
----
-type: learning
-created: 2026-01-31T18:45:52.872Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["opencode", "skills", "slash-commands", "v1.1.48", "TUI", "implement"]
----
-
-# 📚 OpenCode v1.1.48: Skills as slash commands
-
-🟢 **Confidence:** high
-
-OpenCode v1.1.48 introduced the ability to invoke skills as slash commands in the TUI.
-
-This means:
-1. Skills in `.opencode/skill/` can be invoked directly via `/skill-name`
-2. Commands in `.opencode/command/` that load skills via `skill({ name: "..." })` need to stay consistent with actual skill capabilities
-3. When simplifying tools/skills, must also update any commands that reference them
-
-Example: `/implement` loads swarm-coordination skill and uses swarm-monitor operations - if those operations change, implement.md must be updated too.

package/dist/template/.opencode/memory/observations/2026-01-31-learning-swarm-system-simplified-removed-mailbox-.md

@@ -1,30 +0,0 @@
----
-type: learning
-created: 2026-01-31T19:06:48.076Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["swarm", "beads", "parallel execution", "task tool", "mailbox removal"]
-files: [".opencode/tool/swarm-monitor.ts", ".opencode/command/implement.md", ".opencode/command/handoff.md", ".opencode/command/resume.md", ".opencode/command/new-feature.md", ".opencode/command/research.md", ".opencode/skill/beads/SKILL.md", ".opencode/skill/beads-bridge/SKILL.md", ".opencode/skill/swarm-coordination/SKILL.md"]
----
-
-# 📚 Swarm system simplified - removed mailbox operations
-
-🟢 **Confidence:** high
-
-The swarm system has been simplified to use Task tool for parallel execution instead of internal mailbox operations. Removed obsolete operations: restore_todos, persist_todos, mail_send, mail_read, get_todos. 
-
-Current valid swarm-monitor operations: progress_update, render_block, status, clear
-
-Updated files:
-- .opencode/tool/swarm-monitor.ts (removed obsolete operations)
-- .opencode/command/implement.md (updated examples)
-- .opencode/command/handoff.md (removed persist_todos)
-- .opencode/command/resume.md (removed restore_todos)
-- .opencode/command/new-feature.md (removed mail_send)
-- .opencode/command/research.md (simplified examples)
-- .opencode/skill/beads/SKILL.md (added beads-bridge section)
-- .opencode/skill/beads-bridge/SKILL.md (removed obsolete operations)
-- .opencode/skill/swarm-coordination/SKILL.md (removed obsolete operations)
-
-Key insight: Task tool results replace mailbox messages. Workers return completion status directly via Task results, not via mail_send/mail_read.

package/dist/template/.opencode/memory/observations/2026-01-31-learning-v1-1-48-native-copilot-reasoning-via-pr-.md

@@ -1,45 +0,0 @@
----
-type: learning
-created: 2026-01-31T17:23:15.170Z
-confidence: high
-valid_until: null
-superseded_by: null
-concepts: ["v1.1.48", "copilot", "reasoning", "native", "PR 8900", "thinking_budget", "reasoning_opaque", "reasoning_text"]
----
-
-# 📚 v1.1.48 Native Copilot Reasoning via PR 8900
-
-🟢 **Confidence:** high
-
-## v1.1.48 Native Copilot Reasoning Support (PR #8900)
-
-OpenCode v1.1.48 includes native Copilot-specific provider that properly handles reasoning tokens! This was merged Jan 31, 2026.
-
-### What PR #8900 Does:
-1. Creates dedicated Copilot provider under `src/provider/sdk/copilot/`
-2. Handles `reasoning_text` and `reasoning_opaque` fields natively
-3. Properly stores and passes `reasoning_opaque` back to Copilot API for multi-turn context
-4. Uses `providerOptions.copilot` key (not `openaiCompatible`)
-5. Adds `copilot_cache_control` to all messages
-
-### Key Implementation Details:
-- Provider options key changed from `openaiCompatible` to `copilot`
-- Assistant messages include `reasoning_text` and `reasoning_opaque`
-- The opaque signature MUST be passed back for multi-turn reasoning to work properly
-- Gemini models work with reasoning; Claude thinking is NOT yet supported server-side by GitHub
-
-### For Plugin Configuration (v1.1.48+):
-- NO custom SDK needed anymore
-- Use native `github-copilot` provider
-- Enable reasoning with `reasoning: true` in model config
-- May need to pass `thinking_budget` via providerOptions.copilot
-
-### GitHub Copilot Model Capabilities (from API):
-- Claude models support `max_thinking_budget: 32000` and `min_thinking_budget: 1024`
-- GPT models use `/responses` API endpoint
-- Claude uses chat completions API
-
-### Important Note:
-- Gemini models DO return `reasoning_text` via Copilot
-- Claude extended thinking is blocked SERVER-SIDE by GitHub (not a client issue)
-- The custom SDK we built is now obsolete for v1.1.48+