Skip to content

TODO.md

Latest commit

 

History

History
53 lines (42 loc) · 1.73 KB

TODO.md

File metadata and controls

53 lines (42 loc) · 1.73 KB

TODO

Open follow-ups that still appear relevant to the current code.

Agent/runtime

  • Add tests for Agent.run() stop conditions:
    • final text
    • max iterations
    • tool budget exhaustion
    • repeated-call blocking
    • invalid tool-argument JSON recovery
  • Add direct tests for RunResult, ToolResult, and RejectedCall so future refactors do not silently change loop semantics.
  • Consider stronger context-compaction beyond the current workspace summary plus output truncation.

Loop

  • Add tests for PromptLoopRunner early-exit behavior:
    • execution STATUS: COMPLETE
    • empty review pending and blocker
  • Decide whether review should always run after execution, or whether the current “skip review on STATUS: COMPLETE” behavior is final API.
  • Document or implement a clearer crash-resume story for .codescribe/loop/ artifacts.

Tools

  • Tighten bounded bash safety. It still uses shell=True after validation.
  • Add tests for path-bounding edge cases in ReadTool, GlobTool, EditTool, and WriteTool.
  • Consider whether EditTool should explicitly reject overlapping/nested edits instead of relying mainly on exact-match uniqueness.

Models

  • Add backend smoke tests for:
    • OpenAICompModel
    • AnthropicModel
    • ArgoModel
    • TFModel
  • Decide whether supports_native_tools=True is the right name for backends that emulate tool calls through strict JSON prompting.
  • Document the practical support matrix for reasoning/token accounting across providers.

Docs

  • Keep docs focused on code-backed behavior and remove speculative framework comparisons unless they are needed and maintained.
  • Reconcile README.rst examples and defaults with the current CLI and loop implementation.