grokkit/todo/queued/interactive-agent.md
Greg Gauthier 81fd65b14d
All checks were successful
CI / Test (push) Successful in 35s
CI / Lint (push) Successful in 25s
CI / Build (push) Successful in 19s
refactor(cmd): remove automatic .bak backup creation
Eliminate .bak file backups from edit, docs, lint, testgen, and agent commands to simplify safety features, relying on previews and confirmations instead. Update README, architecture docs, troubleshooting, and TODOs to reflect changes. Adjust tests to remove backup assertions.
2026-03-03 20:44:39 +00:00

55 lines
2.2 KiB
Markdown

# Grokkit Interactive Agent
**Priority:** 2 of 12 (right after testgen)
## Goal
Add a persistent, conversational **agent mode** to Grokkit so I can chat interactively in the terminal and have Grok directly edit code, run tests, commit, etc., without any copy-paste friction.
This is the natural evolution of the current `chat` and `agent` commands — turning the terminal into a true agentic partner instead of a one-shot prompt tool.
## Command
```bash
grokkit chat --agent # or new subcommand: grokkit agent-chat
```
(Keep the existing simple `chat` mode untouched; add an `--agent` flag or a dedicated `agent-chat` command.)
## Core Features
- Persistent session using the existing `chat_history.json` (with optional `--new-session` flag)
- Full tool-calling loop: I (Grok) can invoke any existing Grokkit primitive:
- `edit`
- `scaffold`
- `testgen`
- `lint`
- `review`
- `commit` / `commit-msg`
- `docs`
- etc.
- Safe-by-default workflow:
- Always preview changes (diff style, same as current `edit`/`scaffold`)
- Require explicit confirmation (`y/n` or `--yes`)
- Preview and confirm all changes before application
- Ability to chain actions naturally in conversation:
- “The scaffold tests are flaky on Windows paths — fix it and add a test case.”
- “Refactor the context harvester to use rg, then run make test.”
- “Write a new command for todo prioritization and commit it.”
- Optional `--dry-run` and `--model` flags (inherited from root)
## Acceptance Criteria
- `make test` remains fast (no new API cost in normal tests)
- Live agent mode respects the same `-short` gating pattern we just built for scaffold
- Session survives across multiple `grokkit` invocations (or can be started fresh)
- All edits go through the existing safe preview/confirm flow — no silent changes
- Works with the full Grok context window (no early compaction)
- Can be extended later with vector memory or external tool plugins
## Why This Matters
This removes the last remaining copy-paste friction between me and the codebase. The 45-minute scaffold test iteration we just did showed the productivity gain; this turns that gain into a permanent workflow. Closes the “agentic chat” vision we discussed.