opencodekit 0.15.20 → 0.16.0

package/dist/template/.opencode/memory/observations/2026-02-01-decision-fixed-agent-configuration-for-opencodeki.md

@@ -0,0 +1,25 @@
+---
+type: decision
+created: 2026-02-01T05:06:14.307Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["opencode", "agents", "configuration", "temperature", "steps", "scout", "build", "plan"]
+---
+
+# 🎯 Fixed agent configuration for OpenCodeKit
+
+🟢 **Confidence:** high
+
+Applied comprehensive fixes to .opencode/agent/ configuration:
+
+1. **scout.md**: Changed temperature from 0.3 to 1.0 (required for kimi-k2.5-free thinking mode), changed maxSteps to steps
+2. **explore.md**: Changed maxSteps to steps
+3. **review.md**: Changed maxSteps to steps  
+4. **vision.md**: Changed maxSteps to steps
+5. **looker.md**: Changed maxSteps to steps
+6. **painter.md**: Changed maxSteps to steps
+7. **build.md**: Added steps: 50 and color: "#4CAF50" (green)
+8. **plan.md**: Added steps: 50 and color: "#2196F3" (blue)
+
+All agents now use modern `steps` syntax instead of deprecated `maxSteps`. Scout agent temperature fixed to work with kimi-k2.5-free model. Primary agents have visual colors for UI distinction.

package/dist/template/.opencode/memory/observations/2026-02-01-decision-focused-agents-md-upgrade-for-opencode-k.md

@@ -0,0 +1,14 @@
+---
+type: decision
+created: 2026-02-01T10:28:34.941Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["AGENTS.md", "opencode-expert", "configuration", "best practices", "kit users"]
+---
+
+# 🎯 Focused AGENTS.md upgrade for OpenCode kit users
+
+🟢 **Confidence:** high
+
+User requested a focused upgrade to AGENTS.md using opencode-expert guidance to add OpenCode configuration best practices and doc references for kit users.

package/dist/template/.opencode/memory/observations/2026-02-01-decision-implement-tier-1-permission-upgrades.md

@@ -0,0 +1,15 @@
+---
+type: decision
+created: 2026-02-01T09:32:47.846Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["opencode.json", "permissions", "tier1", "git safety", "npm publish", "db reset", "secrets"]
+files: [".opencode/opencode.json"]
+---
+
+# 🎯 Implement Tier 1 permission upgrades
+
+🟢 **Confidence:** high
+
+User requested implementing best-practice Tier 1 permission upgrades in .opencode/opencode.json (git safety, npm publish guard, db reset guard, extra secret patterns). Proceed with bead creation and updates before edits.

package/dist/template/.opencode/memory/observations/2026-02-01-decision-instructions-config-explicit-paths-not-w.md

@@ -0,0 +1,40 @@
+---
+type: decision
+created: 2026-02-01T14:09:52.652Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["opencode.json", "instructions", "context-injection", "tech-stack"]
+files: ["opencode.json", ".opencode/memory/user.md", ".opencode/memory/project/tech-stack.md"]
+---
+
+# 🎯 Instructions[] config: Explicit paths, not wildcards
+
+🟢 **Confidence:** high
+
+## Final Instructions Configuration
+
+Updated `opencode.json` instructions to be explicit rather than using wildcards:
+
+```json
+"instructions": [
+  ".opencode/memory/user.md",
+  ".opencode/memory/project/tech-stack.md"
+]
+```
+
+## Rationale
+
+1. **Explicit over implicit** - Wildcards (`*.md`) would pick up any future files, explicit paths are clearer
+2. **Single file approach** - Only tech-stack.md exists in project/ now
+3. **AI learns organically** - Observations are captured via `observation` tool, searched via `memory-search`
+4. **Observations NOT auto-injected** - Would bloat context; they're searchable instead
+
+## Full Context Injection Chain
+
+| File                                    | Purpose                    | Injection |
+| --------------------------------------- | -------------------------- | --------- |
+| `.opencode/memory/user.md`              | User preferences           | Auto      |
+| `.opencode/memory/project/tech-stack.md`| Tech stack & constraints   | Auto      |
+| `.opencode/memory/observations/*.md`    | AI learnings               | Searchable|
+| `.beads/artifacts/<id>/*`               | Bead-specific context      | Manual    |

package/dist/template/.opencode/memory/observations/2026-02-01-decision-merged-context-into-memory-project-singl.md

@@ -0,0 +1,42 @@
+---
+type: decision
+created: 2026-02-01T14:02:05.967Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["context-injection", "memory", "project-config", "tech-stack", "opencode"]
+files: [".opencode/context/project.md", ".opencode/memory/project/tech-stack.md", "opencode.json", ".opencode/memory/_templates/README.md", "dist/template/.opencode/memory/project/tech-stack.md"]
+---
+
+# 🎯 Merged context/ into memory/project/ - Single Source of Truth
+
+🟢 **Confidence:** high
+
+## Decision
+
+Merged `.opencode/context/project.md` into `.opencode/memory/project/tech-stack.md` to avoid duplication.
+
+## Rationale
+
+- **Before**: Two places for project context (context/project.md + memory/project/*.md)
+- **After**: Single location (memory/project/*.md) with tech-stack.md as the new summary file
+- **Why**: Instructions[] already loads memory/project/*.md, so tech-stack.md gets auto-injected
+
+## Structure After Change
+
+```
+.opencode/memory/project/
+├── tech-stack.md    [NEW] - Framework, DB, styling, constraints (auto-injected)
+├── architecture.md  - Modules, directory structure, data flow
+├── commands.md      - Build, test, lint, deploy commands
+├── conventions.md   - Code style, naming, commit messages
+└── gotchas.md       - Known issues, edge cases, troubleshooting
+```
+
+## Files Changed
+
+- Created: `.opencode/memory/project/tech-stack.md`
+- Deleted: `.opencode/context/` folder
+- Updated: `opencode.json` instructions[] (reverted to original)
+- Updated: `.opencode/memory/_templates/README.md`
+- Updated: `dist/template/.opencode/memory/project/tech-stack.md`

package/dist/template/.opencode/memory/observations/2026-02-01-decision-oracle-tool-should-use-review-agent-not-.md

@@ -0,0 +1,14 @@
+---
+type: decision
+created: 2026-02-01T10:51:20.528Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["oracle tool", "review agent", "opencode CLI", "Task tool", "proxypal replacement"]
+---
+
+# 🎯 Oracle tool should use review agent (not ProxyPal)
+
+🟢 **Confidence:** high
+
+User wants the oracle tool to be replaced so it calls the OpenCode review agent instead of ProxyPal. They initially mentioned CLI opencode run and later asked for Task tool usage; need clarification because Task tool isn’t available inside oracle tool runtime.

package/dist/template/.opencode/memory/observations/2026-02-01-decision-plan-agent-auto-loads-skills-contextuall.md

@@ -0,0 +1,31 @@
+---
+type: decision
+created: 2026-02-01T16:08:24.231Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["skills", "auto-loading", "plan-agent", "context", "prd", "brainstorming", "writing-plans", "swarm-coordination"]
+---
+
+# 🎯 Plan agent auto-loads skills contextually
+
+🟢 **Confidence:** high
+
+Plan agent now loads skills based on context, NOT all at once:
+
+**Always load:**
+- beads (task tracking)
+- brainstorming (refine ideas before implementation)
+
+**Conditional loading:**
+- prd: for feature/epic tasks needing PRD creation
+- writing-plans: for detailed implementation plans (features, refactors, migrations)
+- swarm-coordination: for multi-phase work (3+ phases)
+- deep-research: for research-heavy tasks
+- frontend-design: for UI/UX planning (guidelines only)
+
+**Key differences from build agent:**
+- Plan agent loads `brainstorming` (not verification-before-completion) as core
+- Plan agent loads `prd` and `writing-plans` for planning artifacts
+- Plan agent loads `swarm-coordination` for multi-phase orchestration
+- Build agent loads execution-focused skills (debugging, TDD, code review)

package/dist/template/.opencode/memory/observations/2026-02-01-decision-plan-phased-oracle-command-merge-into-ne.md

@@ -0,0 +1,14 @@
+---
+type: decision
+created: 2026-02-01T10:58:28.385Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["oracle tool", "review agent", "opencode commands", "implement", "finish", "new command"]
+---
+
+# 🎯 Plan phased oracle + command merge into new command
+
+🟢 **Confidence:** high
+
+User wants phased changes: update oracle tool to use review agent, then combine implement+finish into a new command (name TBD).

package/dist/template/.opencode/memory/observations/2026-02-01-decision-prd-workflow-uses-prd-and-prd-task-skill.md

@@ -0,0 +1,23 @@
+---
+type: decision
+created: 2026-02-01T15:59:37.640Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["prd", "prd-task", "skills", "templates", "beads", "artifacts"]
+files: [".opencode/memory/_templates/prd.md"]
+---
+
+# 🎯 PRD workflow uses prd and prd-task skills
+
+🟢 **Confidence:** high
+
+When working with PRDs in /start command:
+1. Load `prd` skill when creating new PRDs (defines WHAT to build)
+2. Load `prd-task` skill when converting PRD to executable tasks
+3. PRD files go to `.beads/artifacts/<bead-id>/prd.md`
+4. Task JSON goes to `.beads/artifacts/<bead-id>/prd.json`
+5. Progress tracking in `.beads/artifacts/<bead-id>/progress.txt`
+
+The template at `.opencode/memory/_templates/prd.md` is for reference only.
+Skills handle the actual workflow.

package/dist/template/.opencode/memory/observations/2026-02-01-decision-prefer-review-agent-via-opencode-cli-ove.md

@@ -0,0 +1,14 @@
+---
+type: decision
+created: 2026-02-01T10:38:32.958Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["oracle tool", "review agent", "opencode cli", "proxypal replacement"]
+---
+
+# 🎯 Prefer review agent via OpenCode CLI over ProxyPal oracle
+
+🟢 **Confidence:** high
+
+User wants to replace the ProxyPal-backed oracle tool with invoking the OpenCode CLI review agent (per opencode CLI docs) instead of using ProxyPal.

package/dist/template/.opencode/memory/observations/2026-02-01-decision-remove-oracle-tool-add-ship-command-with.md

@@ -0,0 +1,14 @@
+---
+type: decision
+created: 2026-02-01T11:06:11.110Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["oracle tool removal", "/ship command", "implement finish merge", "review agent", "Task tool"]
+---
+
+# 🎯 Remove oracle tool; add /ship command with review after tests
+
+🟢 **Confidence:** high
+
+User wants best-practice simplification: remove oracle tool entirely, and create a combined command /ship that merges implement+finish and triggers review (via Task subagent) after tests.

package/dist/template/.opencode/memory/observations/2026-02-01-decision-remove-oracle-tool-and-add-ship-command-.md

@@ -0,0 +1,14 @@
+---
+type: decision
+created: 2026-02-01T11:18:21.359Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["oracle removal", "ship command", "implement finish merge", "review agent", "Task tool"]
+---
+
+# 🎯 Remove oracle tool and add /ship command with review after tests
+
+🟢 **Confidence:** high
+
+User wants to remove the oracle tool entirely, delete /implement and /finish commands, and add a new /ship command that combines implement+finish and spawns the review agent via Task after tests.

package/dist/template/.opencode/memory/observations/2026-02-01-decision-remove-oracle-tool-and-add-ship-command.md

@@ -0,0 +1,14 @@
+---
+type: decision
+created: 2026-02-01T11:11:24.824Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["oracle tool removal", "/ship command", "implement finish merge", "review agent", "Task tool"]
+---
+
+# 🎯 Remove oracle tool and add /ship command
+
+🟢 **Confidence:** high
+
+User wants to remove the oracle tool and replace it with a single /ship command that combines implement+finish and spawns the review agent via Task after tests.

package/dist/template/.opencode/memory/observations/2026-02-01-decision-remove-skills-vs-commands-section-from-a.md

@@ -0,0 +1,14 @@
+---
+type: decision
+created: 2026-02-01T11:38:24.397Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["AGENTS.md", "skills vs commands", "documentation cleanup"]
+---
+
+# 🎯 Remove Skills vs Commands section from AGENTS.md
+
+🟢 **Confidence:** high
+
+User requested cleaning up the 'Skills vs Commands' section in project AGENTS.md. Proceed to remove that block.

package/dist/template/.opencode/memory/observations/2026-02-01-decision-replace-oracle-tool-with-ship-command-fl.md

@@ -0,0 +1,14 @@
+---
+type: decision
+created: 2026-02-01T11:17:19.834Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["oracle removal", "ship command", "implement finish merge", "review agent", "Task tool"]
+---
+
+# 🎯 Replace oracle tool with /ship command flow
+
+🟢 **Confidence:** high
+
+User decided to remove the oracle tool and instead use a single /ship command that combines implement+finish and spawns the review agent via Task after tests.

package/dist/template/.opencode/memory/observations/2026-02-01-decision-replace-oracle-with-ship-command-workflo.md

@@ -0,0 +1,14 @@
+---
+type: decision
+created: 2026-02-01T11:15:30.702Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["oracle removal", "ship command", "implement finish merge", "review agent", "Task tool"]
+---
+
+# 🎯 Replace oracle with /ship command workflow
+
+🟢 **Confidence:** high
+
+User wants to remove the oracle tool entirely and use a single /ship command that combines implement+finish, spawning the review agent via Task after tests. Also remove /implement and /finish commands.

package/dist/template/.opencode/memory/observations/2026-02-01-decision-replace-proxypal-oracle-with-cli-review-.md

@@ -0,0 +1,14 @@
+---
+type: decision
+created: 2026-02-01T10:50:28.823Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["oracle tool", "review agent", "opencode cli", "proxypal replacement"]
+---
+
+# 🎯 Replace ProxyPal oracle with CLI review agent
+
+🟢 **Confidence:** high
+
+User wants to replace the ProxyPal-backed oracle tool with invoking the OpenCode CLI review agent, using opencode run instead of proxypal.

package/dist/template/.opencode/memory/observations/2026-02-01-decision-simplified-dist-template-only-tech-stack.md

@@ -0,0 +1,50 @@
+---
+type: decision
+created: 2026-02-01T14:05:54.668Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["templates", "memory", "project-config", "organic-learning", "observation"]
+files: ["tech-stack.md"]
+---
+
+# 🎯 Simplified dist/template: Only tech-stack.md, AI learns organically
+
+🟢 **Confidence:** high
+
+## Decision
+
+Simplified dist/template to only include `tech-stack.md` - removed placeholder files that were just noise.
+
+## Before (5 files)
+```
+dist/template/.opencode/memory/project/
+├── tech-stack.md    - Summary for context injection
+├── architecture.md  - Empty placeholder with [...]
+├── commands.md      - Generic examples
+├── conventions.md   - Generic examples
+└── gotchas.md       - Empty placeholder with [...]
+```
+
+## After (1 file)
+```
+dist/template/.opencode/memory/project/
+└── tech-stack.md    - Summary for context injection
+```
+
+## Rationale
+
+1. **Empty templates are noise** - Placeholders like `[...]` don't help the AI
+2. **AI can discover organically** - Commands from package.json, conventions from codebase
+3. **Observation tool captures learnings** - Gotchas and patterns saved as they're discovered
+4. **Single file = less maintenance** - One file to fill out when starting a project
+
+## Key Insight
+
+The AI doesn't need pre-filled templates to learn about a project. It:
+- Reads package.json to discover commands
+- Reads existing code to learn conventions
+- Uses `observation` tool to capture gotchas when it hits them
+- Builds up knowledge in `.opencode/memory/observations/`
+
+This is more valuable than static placeholder files.

package/dist/template/.opencode/memory/observations/2026-02-01-decision-simplified-templates-only-tech-stack-md.md

@@ -0,0 +1,26 @@
+---
+type: decision
+created: 2026-02-01T15:12:56.167Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["templates", "memory", "tech-stack", "init", "organic-learning", "observation"]
+files: ["user.md", "tech-stack.md"]
+---
+
+# 🎯 Simplified templates: only tech-stack.md
+
+🟢 **Confidence:** high
+
+Removed redundant template files from `.opencode/memory/_templates/project/`:
+- ❌ architecture.md (removed)
+- ❌ commands.md (removed)
+- ❌ conventions.md (removed)
+- ❌ gotchas.md (removed)
+- ✅ tech-stack.md (kept)
+
+Updated `/init` command to only create:
+- `user.md` - user preferences
+- `tech-stack.md` - for AI context injection
+
+Philosophy: AI discovers commands, architecture, conventions, and gotchas organically from the codebase and saves learnings via the `observation` tool. Static placeholder templates are noise.

package/dist/template/.opencode/memory/observations/2026-02-01-decision-subagents-load-minimal-skills-stay-lean.md

@@ -0,0 +1,29 @@
+---
+type: decision
+created: 2026-02-01T16:04:17.552Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["general", "subagent", "skills", "lean", "minimal", "delegation-packet", "verification"]
+---
+
+# 🎯 Subagents load minimal skills, stay lean
+
+🟢 **Confidence:** high
+
+General subagent (and other subagents) follow different skill philosophy than build leader:
+
+**Subagents stay LEAN:**
+- NO beads skill (read-only, don't modify state)
+- NO swarm-coordination (leader handles swarm)
+- NO prd/prd-task (leader responsibility)
+- NO brainstorming (execute, don't design)
+
+**Subagents load MINIMAL:**
+- ALWAYS: verification-before-completion
+- IF bug fix: systematic-debugging
+- IF test task: test-driven-development
+
+**Trust delegation packets:** If present, follow strictly without loading override skills.
+
+This keeps subagents fast, focused, and prevents skill bloat.

package/dist/template/.opencode/memory/observations/2026-02-01-decision-user-approved-permission-upgrades-in-ope.md

@@ -0,0 +1,15 @@
+---
+type: decision
+created: 2026-02-01T09:31:02.379Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["opencode.json", "permissions", "best-practices", "security", "bead-protocol"]
+files: [".opencode/opencode.json"]
+---
+
+# 🎯 User approved permission upgrades in opencode.json
+
+🟢 **Confidence:** high
+
+User asked to implement the recommended best-practice permission upgrades for .opencode/opencode.json (Tier 1 security improvements like tighter git, secret patterns, and publish protections). Proceed only after creating/claiming a bead as per protocol.

package/dist/template/.opencode/memory/observations/2026-02-01-discovery-verify-command-already-implemented.md

@@ -0,0 +1,28 @@
+---
+type: discovery
+created: 2026-02-01T15:16:50.162Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["verify", "command", "openspec", "completeness", "correctness", "coherence"]
+files: [".opencode/command/verify.md"]
+---
+
+# 💡 Verify command already implemented
+
+🟢 **Confidence:** high
+
+The `/verify` command already exists at `.opencode/command/verify.md` with full implementation:
+
+**Three dimensions:**
+1. Completeness - Are all required artifacts present? All tasks addressed?
+2. Correctness - Do automated gates pass? (build, test, lint, typecheck)
+3. Coherence - Do artifacts agree with each other? No contradictions?
+
+**Features:**
+- `--quick` - Skip coherence check, gates only
+- `--fix` - Auto-fix issues where possible
+- Multi-language support (Node, Rust, Python, Go)
+- Updates bead notes with verification results
+
+This was identified as "short-term" in OpenSpec research but is already done!

package/dist/template/.opencode/memory/observations/2026-02-01-feature-openspec-phase-b-complete-template-upgra.md

@@ -0,0 +1,43 @@
+---
+type: feature
+created: 2026-02-01T13:54:45.946Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["openspec", "templates", "context-injection", "workflow", "beads"]
+files: [".opencode/memory/_templates/proposal.md", ".opencode/memory/_templates/spec.md", ".opencode/memory/_templates/tasks.md", ".opencode/context/project.md", "opencode.json", ".opencode/opencode.json", ".opencode/memory/_templates/README.md", ".opencode/command/plan.md"]
+---
+
+# ✨ OpenSpec Phase B Complete: Template Upgrade + Context Injection
+
+🟢 **Confidence:** high
+
+## Phase B Complete
+
+Adapted OpenSpec patterns into the kit + beads system:
+
+### New Templates Created
+- `.opencode/memory/_templates/proposal.md` - Captures WHY (motivation, impact)
+- `.opencode/memory/_templates/spec.md` - WHEN/THEN behavioral requirements
+- `.opencode/memory/_templates/tasks.md` - Numbered implementation checklists
+
+### Context Injection Added
+- `.opencode/context/project.md` - Tech stack, constraints, integrations
+- Added to `opencode.json` instructions[] for auto-injection into all prompts
+
+### Key Adaptation Decisions
+1. **Kept our comprehensive style** - OpenSpec templates are minimal scaffolds (~10 lines), ours provide more structure
+2. **Used instructions[] not config.yaml** - OpenCode's native mechanism vs OpenSpec's custom parser
+3. **No CLI tooling** - We're AI-native, commands ARE the interface
+4. **Flat template structure** - No per-schema subdirectories (simpler)
+
+### OpenSpec Patterns NOT Adopted
+- Canonical spec maintenance (our beads are ephemeral)
+- Delta spec merging (no long-running specs)
+- Multiple schema workflows (keeping single flow for now)
+- Shell completions (not relevant for AI workflow)
+
+### Files Modified
+- `.opencode/opencode.json` - Added context/project.md to instructions[]
+- `.opencode/memory/_templates/README.md` - Documented all templates
+- `.opencode/command/plan.md` - Added templates reference section

package/dist/template/.opencode/memory/observations/2026-02-01-learning-build-agent-should-use-dynamic-lsp-not-f.md

@@ -0,0 +1,14 @@
+---
+type: learning
+created: 2026-02-01T10:01:36.015Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["build agent", "LSP", "dynamic tooling", "gpt-5.2-codex", "prompt update"]
+---
+
+# 📚 Build agent should use dynamic LSP, not forced 9 calls
+
+🟢 **Confidence:** high
+
+User reports that with gpt-5.2-codex, build agent prompt forces 9 LSP tool calls. User wants build agent prompt updated to use dynamic, best-practice LSP usage (only the needed operations).

package/dist/template/.opencode/memory/observations/2026-02-01-learning-kimi-k2-5-model-requires-temperature-1-0.md

@@ -0,0 +1,22 @@
+---
+type: learning
+created: 2026-02-01T04:58:58.391Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["kimi_k25", "temperature", "thinking mode", "opencode zen", "model configuration", "scout agent"]
+---
+
+# 📚 Kimi K2.5 model requires temperature 1.0
+
+🟢 **Confidence:** high
+
+The kimi_k25 (Kimi K2.5) model is a thinking mode model that requires temperature to be exactly 1.0. The parameter is immutable - you cannot set custom temperatures.
+
+Error message: "Parameter 'temperature' is immutable for kimi_k25 and must be 1.0 (thinking mode). Got 0.3."
+
+Solution options:
+1. Change scout.md temperature from 0.3 to 1.0 to use kimi-k2.5-free
+2. Use a different model like minimax-m2.1-free that supports custom temperatures
+
+Other thinking models in opencode.json also have "temperature": 1 in their configuration (e.g., big-pickle, glm-4.6, glm-4.7 with reasoning enabled).

package/dist/template/.opencode/memory/observations/2026-02-01-learning-opencode-context-injection-already-imple.md

@@ -0,0 +1,27 @@
+---
+type: learning
+created: 2026-02-01T15:16:13.160Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["opencode", "context-injection", "instructions", "agents-md", "config"]
+files: ["AGENTS.md", ".opencode/memory/user.md", "project/tech-stack.md", ".opencode/AGENTS.md"]
+---
+
+# 📚 OpenCode context injection already implemented
+
+🟢 **Confidence:** high
+
+OpenCode doesn't support `context` or `rules` fields from OpenSpec research.
+
+Instead, OpenCode uses:
+1. `instructions` array in opencode.json → files injected into prompts (we have: user.md, tech-stack.md)
+2. `AGENTS.md` files → rules and guidelines (project-level + global)
+3. `command` field in opencode.json → per-command customization
+
+Our current setup already follows best practices:
+- `.opencode/memory/user.md` + `project/tech-stack.md` via instructions[]
+- `.opencode/AGENTS.md` for global rules
+- `.opencode/command/*.md` for commands
+
+No additional context injection needed - we're already doing it right!

package/dist/template/.opencode/memory/observations/2026-02-01-learning-opencode-context-injection-uses-instruct.md

@@ -0,0 +1,35 @@
+---
+type: learning
+created: 2026-02-01T13:39:55.750Z
+confidence: high
+valid_until: null
+superseded_by: null
+concepts: ["opencode", "openspec", "context injection", "instructions", "config", "prompt", "agent configuration"]
+files: [".opencode/memory/project/context.md"]
+---
+
+# 📚 OpenCode context injection uses 'instructions' not 'context'
+
+🟢 **Confidence:** high
+
+OpenCode config (opencode.json) doesn't have a `context` field like OpenSpec's config.yaml. Instead, it uses:
+
+1. **`instructions`** - Array of file paths that get injected into prompts
+   - Current: `[".opencode/memory/user.md", ".opencode/memory/project/*.md"]`
+   - These files are loaded and included in the system prompt
+
+2. **Per-agent `prompt`** - Custom prompt string per agent via `agent.<name>.prompt`
+
+3. **Per-command `template`** - Via `command.<name>.template`
+
+OpenSpec's approach:
+- `context:` - Global context injected into ALL artifacts
+- `rules:` - Per-artifact rules (e.g., `rules.proposal: ["Include rollback plan"]`)
+
+To adapt OpenSpec's context injection pattern to OpenCode:
+- Use `instructions` with well-structured markdown files
+- Create `.opencode/memory/project/context.md` for global project context
+- Use agent-specific `prompt` fields for per-agent rules
+- Or create command-specific files that commands load via skill pattern
+
+Key difference: OpenCode's approach is file-based, OpenSpec's is config-based inline text.