claude-code

prompts/10-context-and-prompts.md

@@ -0,0 +1,100 @@
+# Prompt 10: Wire Up System Prompt, Context Gathering & Memory System
+
+## Context
+
+You are working in `/workspaces/claude-code`. The CLI constructs a detailed system prompt before each conversation. This prompt includes:
+1. **Static instructions** — core behavior rules (from `src/constants/prompts.ts`)
+2. **Dynamic context** — OS, shell, git status, working directory (from `src/context.ts`)
+3. **Tool descriptions** — auto-generated from tool schemas
+4. **Memory** — persistent `.claude.md` files (from `src/memdir/`)
+5. **User context** — config, preferences, project settings
+
+## Key Files
+
+- `src/constants/prompts.ts` — System prompt construction
+- `src/constants/system.ts` — System identity strings
+- `src/context.ts` — OS/shell/git context collection
+- `src/context/` — Additional context modules
+- `src/memdir/` — Memory directory system (reads `.claude.md`, `CLAUDE.md` files)
+- `src/utils/messages.ts` — Message construction helpers
+
+## Task
+
+### Part A: Trace the system prompt construction
+
+Read `src/constants/prompts.ts` and map:
+1. What is `getSystemPrompt()`'s signature and return type?
+2. What sections does the system prompt contain?
+3. How are tools described in the prompt?
+4. What model-specific variations exist?
+5. Where does the `MACRO.ISSUES_EXPLAINER` reference resolve to?
+
+### Part B: Fix the context gathering
+
+Read `src/context.ts` and:
+1. Understand `getSystemContext()` and `getUserContext()`
+2. These collect OS info, shell version, git status, etc.
+3. Verify they work on Linux (this codebase was likely developed on macOS, so some paths may be macOS-specific)
+4. Fix any platform-specific issues
+
+### Part C: Wire up the memory system
+
+Read `src/memdir/` directory:
+1. How does it find `.claude.md` / `CLAUDE.md` files?
+2. How is memory content injected into the system prompt?
+3. Does it support project-level, user-level, and session-level memory?
+
+Verify it works by:
+1. Creating a test `CLAUDE.md` in the project root
+2. Running the system prompt builder
+3. Checking the memory appears in the output
+
+### Part D: Create a prompt inspection script
+
+Create `scripts/test-prompt.ts`:
+```ts
+// scripts/test-prompt.ts
+// Dump the full system prompt that would be sent to the API
+// Usage: bun scripts/test-prompt.ts
+
+import './src/shims/preload.js'
+
+async function main() {
+  // Import the prompt builder
+  const { getSystemPrompt } = await import('./src/constants/prompts.js')
+  
+  // May need to pass tools list and model name
+  // Check the function signature
+  const prompt = await getSystemPrompt([], 'claude-sonnet-4-20250514')
+  
+  console.log('=== SYSTEM PROMPT ===')
+  console.log(prompt.join('\n'))
+  console.log('=== END ===')
+  console.log(`\nTotal length: ${prompt.join('\n').length} characters`)
+}
+
+main().catch(err => {
+  console.error('Prompt test failed:', err)
+  process.exit(1)
+})
+```
+
+### Part E: Fix MACRO references in prompts
+
+The prompt system references `MACRO.ISSUES_EXPLAINER`. Make sure our `MACRO` global (from `src/shims/macro.ts`) provides this value. If the prompt references other `MACRO` fields, add them too.
+
+### Part F: Context module audit
+
+Check `src/context/` for additional context modules:
+- Project detection (language, framework)
+- Git integration (branch, status, recent commits)
+- Environment detection (CI, container, SSH)
+
+Verify these work in our dev environment.
+
+## Verification
+
+1. `bun scripts/test-prompt.ts` dumps a complete system prompt
+2. The prompt includes: tool descriptions, OS context, memory content
+3. No `undefined` or `MACRO.` references in the output
+4. Memory system reads `.claude.md` from the project root