Skip to content

implementation_plan.md

Latest commit

 

History

History
96 lines (81 loc) · 3.78 KB

implementation_plan.md

File metadata and controls

96 lines (81 loc) · 3.78 KB
Raw

🧠 AI Code Agent – Progress Summary

✅ Phase 1: CLI Agent Core

  • Modular CLI via Typer
  • Configurable prompts using Jinja2
  • Claude via AWS Bedrock
  • Git integration: commit diff + message
  • generate, review-commit commands
  • Prompt debugging: --debug-prompt
  • JSON memory log for each run

🧠 Phase 2: Memory + Retry Graph

  • Persistent MemoryGraph stores prompt/response DAG
  • history, describe-node, retry <node_id> commands
  • Node lineage tracking for retries/fixes

📂 Phase 3: Code Review + Coverage

  • review-commit: code reviewer using Git diff + Claude
  • analyze-coverage: suggest tests from coverage JSON
  • suggest_tests.py module added
  • Executor-based modular pipeline

🔄 Phase 4: Smart Thread (Agent Loop)

  • smart-thread: one-shot task handler (generate → test → fix)
  • plan: multi-step planner using LLM (plan_task.j2)
  • Modular executors/:
    • run_codegen.py
    • run_tests.py
    • fix_errors.py
    • suggest_tests.py

🧭 Phase 5: Planner Loop System

  • planner-loop <plan_id> executes plan steps in order
  • Steps are parsed from LLM JSON output
  • Executor dispatch is now LLM-based via classify_step.j2
  • Goal completion is checked via task_done.j2
  • Final result: agent can plan, execute, adapt, and decide when done

🔄 Phase 6: Self-Correction Workflow

  • diagnose_failure.j2 prompt to explain test/code failure
  • self-correct <node_id> --file X.py command:
    • Diagnose → Fix → Re-test (retry loop)
    • Uses fix_code_error.j2 + run_tests.py
  • Integrated into planner-loop: auto self-correct after failed test
  • Supports retry budget (max_retries)
  • Fully documented in README.md

📦 Phase 7: Packaging and CLI Distribution

  • 🔁 Renamed project to CyberMule
  • 🧩 Package structure reorganized under cybermule/
  • ✅ Added setup.py and pyproject.toml
  • ✅ Exposed CLI via entry point: cybermule=cybermule.cli.main:app
  • 📄 Added __init__.py with versioning
  • 🧪 Project can now be installed with pip install .
  • 📘 CLI command is available as: cybermule <subcommand>

🧪 CLI Commands

Command Description
generate Generate code from prompt
review-commit Review a Git commit with Claude
retry <node_id> Retry or fix a failed node
plan Break down a task into executable steps
planner-loop <plan_id> Execute steps, LLM-mapped executors, check goal
smart-thread --file X.py Run code → test → fix → suggest in one go
analyze-coverage --file Suggest missing tests from coverage.json
history, describe-node View memory graph nodes

🧠 Prompts Used

  • generate_code.j2
  • review_git_commit.j2
  • plan_task.j2
  • classify_step.j2
  • task_done.j2
  • fix_code_error.j2
  • suggest_test_cases.j2

✅ Status Summary

Feature Status
Modular CLI ✅ Done
Bedrock Claude integration ✅ Done
Git diff + review ✅ Done
MemoryGraph ✅ Done
Retry/fix loop ✅ Done
Codegen + test executor ✅ Done
Planner + classifier loop ✅ Done
Task completion detection ✅ Done
UI (TUI or web) ⏸️ On hold
GitHub PR integration ⏸️ On hold
PyPI / packaging 🔜 Next