school-core Architecture Review

Problem: The CURRICULUM dict (autonomous_loop.py:58-197) is 30 static toy problems. Agents have run 50+ rounds producing is_palindrome, factorial, and add(a, b). The EMA score converges to ~70 because every task is trivial and the hardcoded score was 70. The gate system is decorative — no agent ever crosses a gate from real capability growth.

Solution: The issue_bridge already provides real-task drive (GitHub issues). The curriculum should be a fallback, not the primary path. Wire the bridge as the default task source; keep curriculum only for warm-up or when no issues are available. The real fix: agents need to read the codebase before writing code — the issue prompt should include repo context (file paths, existing code), not just the issue title.

• Agents produce real fixes, not toy functions
• Verification scores reflect actual output quality
• Gate crossings become meaningful capability signals
• Delete 140 lines of static CURRICULUM dict

Problem: routing.py defines route_task() (lines 18-76) with candidate selection and gate logic. director.run_task (lines 228-247) implements the same logic inline. routing.py is never imported by any module. This is dead code that creates confusion about where routing decisions actually live.

Solution: Delete routing.py. The routing logic in director.run_task is the real implementation — it's already wired into both the autonomous loop and the issue bridge. Keeping two copies guarantees future divergence.

• Delete routing.py (76 lines of dead code)
• Single source of truth for agent selection
• Zero risk of routing logic diverging

Problem: director.run_task enriches prompts with vault context via context_orchestrator.enrich_prompt() (line 204). But verify_task_output in issue_bridge sends the raw prompt to the verifier with no context. The verifier has no idea what the codebase looks like, so it scores based on surface structure — bullet points and confident language get high scores regardless of substance.

Solution: Pass the same vault context to the verifier. Add a context parameter to verify_task_output and include it in the verification prompt template. The verifier should see: repo tree, relevant file paths, the agent's system prompt, and the original issue — then judge whether the response actually addresses the real codebase.

• Verifier scores substance over structure
• Hallucinated "Step 1: Identify..." gets low scores
• Responses referencing real files get high scores

Problem: triage_classifier.py implements a rule-based classifier (label matching, keyword heuristics, body-length analysis) that categorizes issues as bug/enhancement and assigns a workflow state. But issue_bridge.py never calls it — line 188 fetches issues and line 196 iterates them, but the triage step is skipped entirely. The bridge assumes every fetched issue is already ready-for-agent. The classifier only exists as a helper in director.triage_issue() (line 139) which nothing calls.

Solution: Add a triage step in bridge_issues() before dispatch. Call classify_issue(issue["title"], issue["labels"], issue.get("body", "")) and skip issues that return needs-info or needs-triage. This makes the pipeline self-healing — new issues get classified automatically instead of requiring manual labeling.

• New issues auto-classified on fetch
• needs-info issues skipped (no wasted agent runs)
• Classifier's state machine actually exercised

Problem: director.py imports StaffSandbox, StaffLoader, and StaffContext at the top level (lines 13-15) but only uses them inside run_staff() (line 431+). These imports create hard dependencies on the staff package for every director call, even though staff plugins are never invoked in the main flow. Similarly, _queue_for_confirmation in scoring.py is a stub that does nothing.

Solution: Move staff imports inside run_staff() (lazy imports already used at line 439-440). Remove the top-level stub imports. Delete _queue_for_confirmation or implement it. This eliminates 3 unnecessary module-level dependencies.

• 3 unused top-level imports removed
• Staff package only loaded when staff tasks run
• Cleaner dependency graph

Problem: autonomous_loop.py (562 lines) and issue_bridge.py (318 lines) both orchestrate the same pipeline: fetch tasks → dispatch to agents → score results. The autonomous loop has two modes: curriculum (synthetic) and issue-driven (delegates to bridge_issues). The issue bridge has its own polling loop (bridge_poll). Two entry points, two polling implementations, two CLI argument parsers — for what is fundamentally one pipeline.

Solution: Merge into one entry point with mode selection: python -m school --mode issues --repo user/repo (issue-driven, the primary path) or python -m school --mode curriculum (synthetic fallback). One polling loop, one CLI, one place to add features. The issue_bridge flow becomes the default; curriculum becomes a legacy fallback.

• One entry point, one polling loop
• Issue-driven mode as default
• Curriculum mode as fallback (or delete entirely)