/** * Trellis Context Injection Plugin * * Injects context when Task tool is called with supported subagent types. * Uses OpenCode's tool.execute.before hook. * * Compatibility: * - If oh-my-opencode handles via .claude/hooks/, this plugin skips * - Otherwise, this plugin handles injection */ import { existsSync, writeFileSync } from "fs" import { join } from "path" import { TrellisContext, debugLog } from "../lib/trellis-context.js" // Supported subagent types const AGENTS_ALL = ["implement", "check", "debug", "research"] const AGENTS_REQUIRE_TASK = ["implement", "check", "debug"] // Agents that don't update phase (can be called at any time) const AGENTS_NO_PHASE_UPDATE = ["debug", "research"] /** * Update current_phase in task.json based on subagent_type */ function updateCurrentPhase(ctx, taskDir, subagentType) { if (AGENTS_NO_PHASE_UPDATE.includes(subagentType)) { return } const taskJsonPath = join(ctx.directory, taskDir, "task.json") const content = ctx.readFile(taskJsonPath) if (!content) return try { const taskData = JSON.parse(content) const currentPhase = taskData.current_phase || 0 const nextActions = taskData.next_action || [] // Map action names to subagent types const actionToAgent = { "implement": "implement", "check": "check", "finish": "check" // finish uses check agent } // Find the next phase that matches this subagent_type let newPhase = null for (const action of nextActions) { const phaseNum = action.phase || 0 const actionName = action.action || "" const expectedAgent = actionToAgent[actionName] // Only consider phases after current_phase if (phaseNum > currentPhase && expectedAgent === subagentType) { newPhase = phaseNum break } } if (newPhase !== null) { taskData.current_phase = newPhase writeFileSync(taskJsonPath, JSON.stringify(taskData, null, 2)) debugLog("inject", "Updated current_phase to:", newPhase) } } catch (e) { debugLog("inject", "Error updating phase:", e.message) } } /** * Get context for implement agent */ function getImplementContext(ctx, taskDir) { const parts = [] // 1. Read implement.jsonl (or fallback to spec.jsonl) let jsonlPath = join(ctx.directory, taskDir, "implement.jsonl") let entries = ctx.readJsonlWithFiles(jsonlPath) if (entries.length === 0) { // Fallback to spec.jsonl jsonlPath = join(ctx.directory, taskDir, "spec.jsonl") entries = ctx.readJsonlWithFiles(jsonlPath) } if (entries.length > 0) { parts.push(ctx.buildContextFromEntries(entries)) } // 2. Requirements document const prd = ctx.readProjectFile(join(taskDir, "prd.md")) if (prd) { parts.push(`=== ${taskDir}/prd.md (Requirements) ===\n${prd}`) } // 3. Technical design const info = ctx.readProjectFile(join(taskDir, "info.md")) if (info) { parts.push(`=== ${taskDir}/info.md (Technical Design) ===\n${info}`) } return parts.join("\n\n") } /** * Get context for check agent */ function getCheckContext(ctx, taskDir) { const parts = [] // 1. Read check.jsonl const jsonlPath = join(ctx.directory, taskDir, "check.jsonl") const entries = ctx.readJsonlWithFiles(jsonlPath) if (entries.length > 0) { parts.push(ctx.buildContextFromEntries(entries)) } else { // Fallback: hardcoded check files + spec.jsonl const checkFiles = [ [".opencode/commands/trellis/finish-work.md", "Finish work checklist"], [".opencode/commands/trellis/check-cross-layer.md", "Cross-layer check spec"], [".opencode/commands/trellis/check-backend.md", "Backend check spec"], [".opencode/commands/trellis/check-frontend.md", "Frontend check spec"], ] for (const [f, description] of checkFiles) { const content = ctx.readProjectFile(f) if (content) { parts.push(`=== ${f} (${description}) ===\n${content}`) } } // Add spec.jsonl const specJsonlPath = join(ctx.directory, taskDir, "spec.jsonl") const specEntries = ctx.readJsonlWithFiles(specJsonlPath) for (const entry of specEntries) { parts.push(`=== ${entry.path} (Dev spec) ===\n${entry.content}`) } } // 2. Requirements document const prd = ctx.readProjectFile(join(taskDir, "prd.md")) if (prd) { parts.push(`=== ${taskDir}/prd.md (Requirements - for understanding intent) ===\n${prd}`) } return parts.join("\n\n") } /** * Get context for finish phase (final check before PR) */ function getFinishContext(ctx, taskDir) { const parts = [] // 1. Try finish.jsonl first const jsonlPath = join(ctx.directory, taskDir, "finish.jsonl") const entries = ctx.readJsonlWithFiles(jsonlPath) if (entries.length > 0) { parts.push(ctx.buildContextFromEntries(entries)) } else { // Fallback: only finish-work.md (lightweight) const finishWork = ctx.readProjectFile(".opencode/commands/trellis/finish-work.md") if (finishWork) { parts.push(`=== .opencode/commands/trellis/finish-work.md (Finish checklist) ===\n${finishWork}`) } } // 2. Spec update process (for active spec sync) const updateSpec = ctx.readProjectFile(".opencode/commands/trellis/update-spec.md") if (updateSpec) { parts.push(`=== .opencode/commands/trellis/update-spec.md (Spec update process) ===\n${updateSpec}`) } // 3. Requirements document (for verifying requirements are met) const prd = ctx.readProjectFile(join(taskDir, "prd.md")) if (prd) { parts.push(`=== ${taskDir}/prd.md (Requirements - verify all met) ===\n${prd}`) } return parts.join("\n\n") } /** * Get context for debug agent */ function getDebugContext(ctx, taskDir) { const parts = [] // 1. Read debug.jsonl (or fallback to spec.jsonl + check files) const jsonlPath = join(ctx.directory, taskDir, "debug.jsonl") const entries = ctx.readJsonlWithFiles(jsonlPath) if (entries.length > 0) { parts.push(ctx.buildContextFromEntries(entries)) } else { // Fallback: use spec.jsonl + hardcoded check files const specJsonlPath = join(ctx.directory, taskDir, "spec.jsonl") const specEntries = ctx.readJsonlWithFiles(specJsonlPath) for (const entry of specEntries) { parts.push(`=== ${entry.path} (Dev spec) ===\n${entry.content}`) } const checkFiles = [ [".opencode/commands/trellis/check-backend.md", "Backend check spec"], [".opencode/commands/trellis/check-frontend.md", "Frontend check spec"], [".opencode/commands/trellis/check-cross-layer.md", "Cross-layer check spec"], ] for (const [f, description] of checkFiles) { const content = ctx.readProjectFile(f) if (content) { parts.push(`=== ${f} (${description}) ===\n${content}`) } } } // 2. Codex review output (if exists) const codex = ctx.readProjectFile(join(taskDir, "codex-review-output.txt")) if (codex) { parts.push(`=== ${taskDir}/codex-review-output.txt (Codex Review Results) ===\n${codex}`) } return parts.join("\n\n") } /** * Get context for research agent */ function getResearchContext(ctx, taskDir) { const parts = [] parts.push(`## Project Spec Directory Structure \`\`\` .trellis/spec/ ├── shared/ # Cross-project common specs ├── frontend/ # Frontend standards ├── backend/ # Backend standards └── guides/ # Thinking guides .trellis/big-question/ # Known issues and pitfalls \`\`\` ## Search Tips - Spec files: \`.trellis/spec/**/*.md\` - Known issues: \`.trellis/big-question/\` - Code search: Use Glob and Grep tools - Tech solutions: Use mcp__exa__web_search_exa or mcp__exa__get_code_context_exa`) if (taskDir) { const jsonlPath = join(ctx.directory, taskDir, "research.jsonl") const entries = ctx.readJsonlWithFiles(jsonlPath) if (entries.length > 0) { parts.push("\n## Additional Search Context\n") parts.push(ctx.buildContextFromEntries(entries)) } } return parts.join("\n\n") } /** * Build enhanced prompt with context */ function buildPrompt(agentType, originalPrompt, context, isFinish = false) { const templates = { implement: `# Implement Agent Task You are the Implement Agent in the Multi-Agent Pipeline. ## Your Context ${context} --- ## Your Task ${originalPrompt} --- ## Workflow 1. **Understand specs** - All dev specs are injected above 2. **Understand requirements** - Read requirements and technical design 3. **Implement feature** - Follow specs and design 4. **Self-check** - Ensure code quality ## Important Constraints - Do NOT execute git commit - Follow all dev specs injected above - Report list of modified/created files when done`, check: isFinish ? `# Finish Agent Task You are performing the final check before creating a PR. ## Your Context ${context} --- ## Your Task ${originalPrompt} --- ## Workflow 1. **Review changes** - Run \`git diff --name-only\` to see all changed files 2. **Verify requirements** - Check each requirement in prd.md is implemented 3. **Spec sync** - Analyze whether changes introduce new patterns, contracts, or conventions - If new pattern/convention found: read target spec file → update it → update index.md if needed - If infra/cross-layer change: follow the 7-section mandatory template from update-spec.md - If pure code fix with no new patterns: skip this step 4. **Run final checks** - Execute lint and typecheck 5. **Confirm ready** - Ensure code is ready for PR ## Important Constraints - You MAY update spec files when gaps are detected (use update-spec.md as guide) - MUST read the target spec file BEFORE editing (avoid duplicating existing content) - Do NOT update specs for trivial changes (typos, formatting, obvious fixes) - If critical CODE issues found, report them clearly (fix specs, not code) - Verify all acceptance criteria in prd.md are met` : `# Check Agent Task You are the Check Agent in the Multi-Agent Pipeline. ## Your Context ${context} --- ## Your Task ${originalPrompt} --- ## Workflow 1. **Get changes** - Run \`git diff --name-only\` and \`git diff\` 2. **Check against specs** - Check item by item 3. **Self-fix** - Fix issues directly, don't just report 4. **Run verification** - Run lint and typecheck ## Important Constraints - Fix issues yourself, don't just report - Must execute complete checklist`, debug: `# Debug Agent Task You are the Debug Agent in the Multi-Agent Pipeline. ## Your Context ${context} --- ## Your Task ${originalPrompt} --- ## Workflow 1. **Understand issues** - Analyze issues pointed out 2. **Locate code** - Find positions that need fixing 3. **Fix against specs** - Fix following dev specs 4. **Verify fixes** - Run typecheck ## Important Constraints - Do NOT execute git commit - Run typecheck after each fix`, research: `# Research Agent Task You are the Research Agent in the Multi-Agent Pipeline. ## Core Principle **You do one thing: find and explain information.** ## Project Info ${context} --- ## Your Task ${originalPrompt} --- ## Workflow 1. **Understand query** - Determine search type and scope 2. **Plan search** - List search steps 3. **Execute search** - Run multiple searches in parallel 4. **Organize results** - Output structured report ## Strict Boundaries **Only allowed**: Describe what exists, where it is, how it works **Forbidden**: Suggest improvements, criticize implementation, modify files` } return templates[agentType] || originalPrompt } export default async ({ directory }) => { const ctx = new TrellisContext(directory) debugLog("inject", "Plugin loaded, directory:", directory) return { // ========================================================================== // ⚠️ KNOWN LIMITATION: OpenCode project-level plugins cannot intercept subagents // // This hook will NOT be triggered because: // 1. Project-level plugins (.opencode/plugin/) don't support tool.execute.before // 2. Only global plugins (npm packages) have full hook permissions // 3. This is a known OpenCode architecture limitation (see Issue #5894) // // SOLUTION: Trellis + OpenCode users must install oh-my-opencode (omo) // - omo is a global plugin with full hook permissions // - omo reads .claude/settings.json and executes Python hooks // - .claude/hooks/inject-subagent-context.py handles the actual injection // // References: // - https://github.com/sst/opencode/issues/5894 (plugin hooks don't intercept subagent) // - https://github.com/sst/opencode/issues/2588 (subagent inherit context) // ========================================================================== "tool.execute.before": async (input, output) => { try { debugLog("inject", "tool.execute.before called, tool:", input?.tool) // Only handle Task tool const toolName = input?.tool?.toLowerCase() if (toolName !== "task") { return } const args = output?.args || {} const subagentType = args.subagent_type const originalPrompt = args.prompt || "" debugLog("inject", "Task tool called, subagent_type:", subagentType) // Only handle supported agent types if (!AGENTS_ALL.includes(subagentType)) { debugLog("inject", "Skipping - unsupported subagent_type") return } // Check if we should skip (omo will handle) if (ctx.shouldSkipHook("inject-subagent-context")) { debugLog("inject", "Skipping - omo will handle via .claude/hooks/") return } // Read current task const taskDir = ctx.getCurrentTask() // Agents requiring task directory if (AGENTS_REQUIRE_TASK.includes(subagentType)) { if (!taskDir) { debugLog("inject", "Skipping - no current task") return } const taskDirFull = join(directory, taskDir) if (!existsSync(taskDirFull)) { debugLog("inject", "Skipping - task directory not found") return } // Update current_phase in task.json updateCurrentPhase(ctx, taskDir, subagentType) } // Check for [finish] marker const isFinish = originalPrompt.toLowerCase().includes("[finish]") // Get context based on agent type let context = "" switch (subagentType) { case "implement": context = getImplementContext(ctx, taskDir) break case "check": // Use finish context for [finish] phase (lighter, focused on final verification) // Use check context for regular check (full specs for self-fix loop) context = isFinish ? getFinishContext(ctx, taskDir) : getCheckContext(ctx, taskDir) break case "debug": context = getDebugContext(ctx, taskDir) break case "research": context = getResearchContext(ctx, taskDir) break } if (!context) { debugLog("inject", "No context to inject") return } // Build enhanced prompt const newPrompt = buildPrompt(subagentType, originalPrompt, context, isFinish) // Update the tool input output.args = { ...args, prompt: newPrompt } debugLog("inject", "Injected context for", subagentType, "prompt length:", newPrompt.length) } catch (error) { debugLog("inject", "Error in tool.execute.before:", error.message, error.stack) } } } }