This commit is contained in:
jparkerweb committed 2026-01-27 12:11:00 -08:00
1 parent b1502edfb2
commit c10354abff
43 files changed
+3601 -215

No files matched your search

+33
View File
@@ -0,0 +1,33 @@
import type { StateManager } from '../state/index.js';
import { LOOP_PROMPT_TEMPLATE } from './templates.js';
export interface PromptContext {
specPath: string;
iteration: number;
maxIterations: number;
stateManager: StateManager;
}
/**
* Build the prompt for the AI agent
* Simple: just pass spec path and let the LLM discover the next task
*/
export async function buildLoopPrompt(context: PromptContext): Promise<string> {
const { specPath, iteration, maxIterations, stateManager } = context;
// Read scratchpad content for session continuity (LLM writes to this)
const scratchpadContent = await stateManager.readScratchpad();
// Project root is where plan2code-loop was invoked from
const projectRoot = process.cwd();
// Simple template substitution
const prompt = LOOP_PROMPT_TEMPLATE
.replace(/{{projectRoot}}/g, projectRoot)
.replace(/{{specPath}}/g, specPath)
.replace(/{{iteration}}/g, iteration.toString())
.replace(/{{maxIterations}}/g, maxIterations.toString())
.replace(/{{scratchpadContent}}/g, scratchpadContent || '(First iteration - no previous progress)');
return prompt;
}
+6
View File
@@ -0,0 +1,6 @@
export {
buildLoopPrompt,
type PromptContext,
} from './builder.js';
export { LOOP_PROMPT_TEMPLATE } from './templates.js';
+85
View File
@@ -0,0 +1,85 @@
export const LOOP_PROMPT_TEMPLATE = `# PLAN-LOOP: Autonomous Task Implementation
## CRITICAL CONSTRAINT
**IMPLEMENT EXACTLY ONE TASK PER ITERATION.**
Do NOT implement multiple tasks. Do NOT complete an entire phase.
Find the FIRST unchecked task, implement ONLY that task, then STOP and report.
## Project Information
- **Project Root:** \`{{projectRoot}}\`
- **Spec Location:** \`{{specPath}}\`
- Read \`AGENTS.md\` for project-specific guidance if available
## IMPORTANT: File Locations
- Write ALL code files relative to the **project root** (\`{{projectRoot}}\`)
- The spec directory (\`{{specPath}}\`) is for documentation ONLY - never write code there
- Example: Create \`{{projectRoot}}/src/index.ts\`, NOT \`{{specPath}}/src/index.ts\`
## Iteration
{{iteration}} of {{maxIterations}}
## Task Discovery Process
1. Read \`{{specPath}}/overview.md\` to see all phases
2. Find the FIRST phase with an unchecked checkbox (\`- [ ]\` or \`- [/]\`)
3. Read that phase's file (e.g., \`phase-1.md\`)
4. Check the \`## Prerequisites\` section FIRST
5. Find the FIRST unchecked prerequisite (\`- [ ]\`)
- If found, that is your task for this iteration
- Verify/complete it, then mark \`[x]\` or \`[?]\`
6. Only if ALL prerequisites are complete (\`[x]\` or \`[?]\`), find the FIRST unchecked task
7. That is your ONE task - implement ONLY that task
## Checkbox States
- \`[ ]\` = incomplete/pending (do the FIRST one you find)
- \`[x]\` = complete (skip)
- \`[?]\` = assumed complete, couldn't verify (skip)
- \`[!]\` = blocked (skip)
## Implementation Steps
1. Read and understand the single task
2. Implement it completely
3. Validate it works (run tests if applicable and double-check code)
4. Mark ONLY that task's checkbox as \`[x]\` in the phase file
5. If that was the LAST task in the phase, also mark the phase \`[x]\` in overview.md
6. Output your completion marker and STOP
## Git Policy
**DO NOT create git commits.** The orchestration system handles commits automatically after each task completion. Just implement the code and leave changes uncommitted.
## Completion Markers (REQUIRED FORMAT)
Output exactly ONE of these at the end, including the task ID and description:
**PREREQ_COMPLETE: [prereq_id] - [description]**
Example: \`PREREQ_COMPLETE: P1.1 - Verified Phase 1 complete\`
**PREREQ_ASSUMED: [prereq_id] - [description]**
Example: \`PREREQ_ASSUMED: P2.1 - Design approval (cannot verify)\`
**TASK_COMPLETE: [task_id] - [task_description]**
Example: \`TASK_COMPLETE: 1.1 - Initialize project structure\`
**TASK_BLOCKED: [task_id] - [reason]**
Example: \`TASK_BLOCKED: 2.3 - Missing API credentials\`
**LOOP_COMPLETE**
Use only when ALL phases in overview.md are marked complete.
## Scratchpad Management
After completing each task, append to \`{{specPath}}/.plan2code-loop/scratchpad.md\`:
- Task completed and Phase item reference
- Key decisions made and reasoning
- Files changed
- Any blockers or notes for next iteration
Keep entries concise. Sacrifice grammar for concision. This file helps future iterations skip exploration.
If key patterns or learnings were discovered, update \`./AGENTS.md\` if it exists.
## Previous Session Context
{{scratchpadContent}}
---
Remember: ONE TASK ONLY. Find it, implement it, mark it done, output TASK_COMPLETE with the task ID and description, then stop.
`;