feat: JIT Skill generation

.claude/tsc-cache/5d0c2956-0d3c-4ab8-a101-011a94d2ccaa/affected-repos.txt

@@ -0,0 +1 @@
+root

.gitignore

@@ -51,3 +51,6 @@ fixtures/runner_crate/target
 .snfoundry_cache
 
 python/starklings_results/
+
+
+smithers.db*

.smithers/agents.ts

@@ -0,0 +1,32 @@
+import { ClaudeCodeAgent, CodexAgent } from "smithers-orchestrator";
+import { REPO_ROOT } from "./config";
+
+const SYSTEM_PROMPT_BASE = `You are working on Cairo Coder, a RAG-based Cairo code generation service.
+Stack: Python/FastAPI backend, DSPy for LLM orchestration, PostgreSQL/pgvector.
+Package manager: uv (NEVER pip/poetry).
+Read CLAUDE.md for full project conventions before making any changes.
+
+CRITICAL OUTPUT REQUIREMENT:
+When you have completed your work, you MUST end your response with a JSON object
+wrapped in a code fence. The JSON format is specified in your task prompt.`;
+
+export const planner = new ClaudeCodeAgent({
+  model: "opus",
+  systemPrompt: SYSTEM_PROMPT_BASE,
+  cwd: REPO_ROOT,
+  permissionMode: "default",
+});
+
+export const implementer = new CodexAgent({
+  model: "gpt-5.3-codex",
+  dangerouslyBypassApprovalsAndSandbox: true,
+  systemPrompt: SYSTEM_PROMPT_BASE,
+  cwd: REPO_ROOT,
+});
+
+export const reviewer = new ClaudeCodeAgent({
+  model: "opus",
+  systemPrompt: SYSTEM_PROMPT_BASE,
+  cwd: REPO_ROOT,
+  permissionMode: "default",
+});

.smithers/config.ts

@@ -0,0 +1,8 @@
+export const MAX_PASSES = 5;
+
+export const REPO_ROOT = process.cwd();
+
+export const VERIFICATION_COMMANDS = {
+  test: "cd python && uv run pytest -v",
+  lint: "trunk check --fix",
+} as const;

.smithers/preload.ts

@@ -0,0 +1,2 @@
+import { mdxPlugin } from "smithers-orchestrator/mdx-plugin";
+mdxPlugin();

.smithers/prompts/final-review.mdx

@@ -0,0 +1,38 @@
+# Final Review — STRICT GATE — Pass {props.pass}
+
+You are the final quality gate. REFUSE to approve unless ALL criteria are met.
+
+## Completed Tasks
+
+{props.completedTasks || "None"}
+
+## Latest Review
+
+{props.latestReview ?? "No review yet"}
+
+## Verification
+
+Run ALL checks:
+
+```bash
+cd python && uv run pytest -v
+trunk check --fix
+```
+
+## Criteria — ALL must pass
+
+- [ ] ALL tests pass (`uv run pytest` exits 0)
+- [ ] ALL lint/type checks pass (`trunk check` exits 0)
+- [ ] Implementation matches SPEC.md requirements
+- [ ] No dead code left behind (old McpGenerationProgram removed if replaced)
+- [ ] New code has unit tests
+- [ ] DSPy patterns followed correctly
+- [ ] No hardcoded paths or credentials
+
+Set `readyToMoveOn: true` ONLY if you genuinely cannot find ANYTHING to improve.
+
+If `readyToMoveOn: false`, explain exactly what must be fixed — this feeds into the next pass's implement step.
+
+## REQUIRED OUTPUT
+
+{props.schema}

.smithers/prompts/fix.mdx

@@ -0,0 +1,41 @@
+# Fix Review Issues — Pass {props.pass}
+
+You are a senior Python engineer fixing code review issues in Cairo Coder.
+
+## Task
+
+### {props.taskName}
+
+## Issues to Fix
+
+{props.issues?.map((issue, i) => `${i + 1}. ${issue}`).join("\n") ?? "No issues listed"}
+
+{props.reviewFeedback ? `## Reviewer Notes\n${props.reviewFeedback}` : ""}
+
+## Instructions
+
+1. Read CLAUDE.md for project conventions
+2. Fix each issue listed above
+3. Do NOT introduce new features — only fix the reported issues
+
+## Verification
+
+After fixing, run:
+
+```bash
+cd python && uv run pytest -v
+trunk check --fix
+```
+
+## GIT COMMIT RULES
+
+- Make atomic commits — one fix per commit
+- Format: `🐛 fix(scope): what was fixed`
+- Examples:
+  - `🐛 fix(dspy): add missing type hint on SkillGenerationProgram.forward()`
+  - `🐛 fix(pipeline): correct skill field name in MCP branch`
+- `git add` the specific files changed, then `git commit`
+
+## REQUIRED OUTPUT
+
+{props.schema}

.smithers/prompts/implement.mdx

@@ -0,0 +1,54 @@
+# Implement — Pass {props.pass}
+
+You are a senior Python engineer implementing a change in Cairo Coder.
+
+## Task
+
+### {props.taskName}
+
+{props.implementationPrompt}
+
+{props.previousWork
+? `## Previous Work\nPrevious implementation did: ${props.previousWork.summary}\nNext smallest unit: ${props.previousWork.nextSmallestUnit}`
+: ""}
+
+{props.reviewFixes ? `## Review Fixes Applied\n${props.reviewFixes}` : ""}
+
+## Files
+
+- **Create:** {JSON.stringify(props.filesToCreate ?? [])}
+- **Modify:** {JSON.stringify(props.filesToModify ?? [])}
+
+## Instructions
+
+1. Read CLAUDE.md for project conventions
+2. Read the existing files before modifying them
+3. Implement the SMALLEST atomic unit described above
+4. Follow DSPy patterns: Signatures → Modules → Programs
+5. Use type hints everywhere (enforced by mypy)
+6. Use `structlog` for logging: `get_logger(__name__)`
+
+## Verification
+
+After implementing, run:
+
+```bash
+cd python && uv run pytest -v
+trunk check --fix
+```
+
+## GIT COMMIT RULES
+
+- Make atomic commits — one logical change per commit
+- Commit EACH smallest unit of work separately, do NOT batch
+- Use emoji prefixes: 🐛 fix, ♻️ refactor, 🧪 test, ⚡ perf, ✨ feat
+- Format: `EMOJI type(scope): description`
+- Examples:
+  - `✨ feat(dspy): add SkillGeneration signature`
+  - `🧪 test(dspy): add unit tests for SkillGenerationProgram`
+  - `♻️ refactor(pipeline): replace McpGenerationProgram with SkillGenerationProgram`
+- `git add` the specific files changed, then `git commit` with the emoji message
+
+## REQUIRED OUTPUT
+
+{props.schema}

.smithers/prompts/plan.mdx

@@ -0,0 +1,32 @@
+# Plan — Pass {props.pass}
+
+You are a senior Python/DSPy architect. Read the spec and examine the codebase to plan the next atomic unit of work.
+
+## Context
+
+- **Spec file:** {props.specPath}
+- **Completed tasks:** {props.completedTasks || "None yet"}
+- **Codebase areas:** `python/src/cairo_coder/dspy/`, `python/src/cairo_coder/core/`, `python/tests/`
+
+{props.previousFeedback ? `## Feedback from Previous Pass\n${props.previousFeedback}` : ""}
+
+## Instructions
+
+1. Read CLAUDE.md for project conventions
+2. Read the spec at {props.specPath}
+3. Examine the codebase — especially the files mentioned in the spec
+4. Identify the NEXT smallest atomic unit of work (one signature, one module, one test file). Balance properly so as not to pick tasks too small or too large: organize logically for reviewers.
+5. Research what exists and what needs to change
+6. Write a detailed implementation prompt
+
+If the SPEC mentions _CRITICAL_ or _REORIENTATION_ or _STEERING_ sections, make sure to follow them; make sure those tasks are prioritized. This is what we are using to steer to the right direction.
+
+## Constraints
+
+- Pick ONE atomic unit, not the entire feature
+- The implementation prompt must be specific enough for another agent to execute without ambiguity
+- List exact file paths for files to create/modify
+
+## REQUIRED OUTPUT
+
+{props.schema}

.smithers/prompts/review.mdx

@@ -0,0 +1,44 @@
+# Review — Pass {props.pass}
+
+You are a senior code reviewer for Cairo Coder, a Python/FastAPI/DSPy project.
+
+## What Changed
+
+**Task:** {props.taskName}
+**Summary:** {props.summary}
+**Files changed:** {JSON.stringify(props.filesChanged ?? [])}
+**Test output:** {props.testOutput ?? "No test output provided"}
+
+## Instructions
+
+1. Read CLAUDE.md for project conventions
+2. Read ALL changed files listed above — do not review blindly
+3. Run verification checks in order:
+
+```bash
+cd python && uv run pytest -v
+trunk check --fix
+```
+
+4. Check for:
+   - Correct DSPy patterns (Signature → Module → factory function)
+   - Type hints on all functions
+   - No hardcoded paths
+   - No duplicated fixtures (all shared fixtures in `tests/conftest.py`)
+   - Async/await for I/O operations
+   - Proper structlog usage
+
+## Decision
+
+Set `lgtm: true` ONLY if:
+
+- ALL tests pass
+- ALL lint/type checks pass
+- Code follows CLAUDE.md conventions
+- No security issues (OWASP top 10)
+
+If ANY check fails, set `lgtm: false` and list every specific issue.
+
+## REQUIRED OUTPUT
+
+{props.schema}

.smithers/schemas.ts

@@ -0,0 +1,55 @@
+import { z } from "zod";
+
+export const PlanSchema = z.object({
+  taskName: z.string().describe("Name of the next task to implement"),
+  research: z
+    .string()
+    .describe("What was discovered by examining the codebase"),
+  implementationPrompt: z
+    .string()
+    .describe("Detailed prompt for the implementer"),
+  filesToCreate: z.array(z.string()).describe("Files that need to be created"),
+  filesToModify: z.array(z.string()).describe("Files that need to be modified"),
+  nextSmallestUnit: z
+    .string()
+    .nullable()
+    .describe("Next smallest atomic unit after this task"),
+});
+
+export const ImplementSchema = z.object({
+  summary: z.string().describe("What was implemented"),
+  filesChanged: z
+    .array(z.string())
+    .describe("Files that were created or modified"),
+  testOutput: z.string().describe("Output from running pytest"),
+  commitMessage: z.string().describe("Git commit message for this change"),
+  nextSmallestUnit: z
+    .string()
+    .nullable()
+    .describe("Next smallest atomic unit to implement"),
+});
+
+export const ReviewSchema = z.object({
+  lgtm: z.boolean().describe("true ONLY if ALL checks pass"),
+  review: z.string().describe("Summary of the review findings"),
+  issues: z.array(z.string()).describe("Specific issues found"),
+});
+
+export const FixSchema = z.object({
+  summary: z.string().describe("What was fixed"),
+  filesChanged: z.array(z.string()).describe("Files that were modified"),
+});
+
+export const FinalReviewSchema = z.object({
+  readyToMoveOn: z.boolean().describe("true ONLY if all criteria met"),
+  reasoning: z
+    .string()
+    .describe("Why ready or not ready — feeds back into next pass"),
+  qualityScore: z.number().min(1).max(10).describe("Overall quality score"),
+});
+
+export const PassTrackerSchema = z.object({
+  totalIterations: z.number(),
+  tasksCompleted: z.array(z.string()),
+  summary: z.string(),
+});