feat: Perfect Mode — goal-driven autonomous execution with relentless iteration #13

andeerc commented 2026年07月10日 14:16:44 +02:00

Summary

Add a Perfect Mode that, when activated, instructs Nibble to pursue a clearly stated goal relentlessly — iterating, testing, fixing, and validating until the goal is fully achieved, without stopping to ask for confirmation along the way.

Motivation

Today Nibble has YOLO Mode, which auto-approves tool calls without prompting the user for permission. Perfect Mode goes a step further: it is not just about skipping confirmations — it is about not stopping until the job is done right.

This is useful when the user wants a high-quality, complete result and is willing to trade token budget and time for it. Think of it as: "I don't care how many steps it takes, just get it right."

Concept

When Perfect Mode is activated, the user must provide a clear, explicit goal statement. Nibble will:

1. Acknowledge the goal and restate it to confirm understanding
2. Warn the user that this mode will consume significantly more tokens and time than a normal session
3. Request explicit confirmation before starting (e.g. "This will run autonomously until the goal is met. Proceed? [y/N]")
4. Execute autonomously — running tools, writing code, running tests, fixing errors — in a loop until the goal criteria are satisfied
5. Only stop when:
   • The goal is demonstrably achieved (tests pass, output matches spec, etc.)
   • A hard token/step budget is exhausted (with a clear warning to the user)
   • A blocking error occurs that cannot be resolved autonomously

Differences from YOLO Mode

YOLO Mode	Perfect Mode
Auto-approves tools	✅	✅
Requires a goal statement	❌	✅
Iterates until goal is met	❌	✅
Runs tests / validates output	❌	✅
Warns about cost upfront	❌	✅
Has a step/token budget guard	❌	✅

Proposed UX

`
/perfect

Enter your goal:
Implement the user authentication flow with full test coverage and passing CI.

⚠ Perfect Mode will run autonomously and iterate until the goal is achieved.
This may consume a large number of tokens and take several minutes.
Proceed? [y/N]: y

✦ Goal locked in: "Implement the user authentication flow with full test coverage and passing CI."
Starting Perfect Mode...
`

Proposed Behavior

• Goal locking: The goal statement is embedded into the system prompt for the session, so every agent decision is evaluated against it
• Self-evaluation loop: After each major step, the agent assesses whether the goal has been met (e.g. by running tests, checking output, reviewing diffs)
• No mid-task interruptions: Tools are auto-approved (inherits YOLO Mode behavior)
• Budget guard: A configurable maximum number of steps or tokens (e.g. perfectMode.maxSteps in nibble.json) prevents runaway sessions; the user is notified if the budget is hit before the goal is achieved
• Completion report: When the goal is reached, Nibble produces a summary of what was done, how many steps it took, and an estimated token cost

Configuration (nibble.json)

json { "perfectMode": { "maxSteps": 50, "maxTokens": 200000, "requireConfirmation": true } }

Notes

• Perfect Mode implies YOLO Mode — all tool calls are auto-approved for the duration of the session
• The goal statement should be stored in the session so it survives context reloads
• Consider adding a visual indicator in the TUI status bar when Perfect Mode is active (e.g. ✦ PERFECT)
• This mode is intentionally opt-in and gated behind an explicit confirmation to prevent accidental runaway sessions

## Summary Add a **Perfect Mode** that, when activated, instructs Nibble to pursue a clearly stated goal relentlessly — iterating, testing, fixing, and validating until the goal is fully achieved, without stopping to ask for confirmation along the way. ## Motivation Today Nibble has **YOLO Mode**, which auto-approves tool calls without prompting the user for permission. Perfect Mode goes a step further: it is not just about skipping confirmations — it is about **not stopping until the job is done right**. This is useful when the user wants a high-quality, complete result and is willing to trade token budget and time for it. Think of it as: *"I don't care how many steps it takes, just get it right."* ## Concept When Perfect Mode is activated, the user must provide a **clear, explicit goal statement**. Nibble will: 1. Acknowledge the goal and restate it to confirm understanding 2. Warn the user that this mode will consume **significantly more tokens and time** than a normal session 3. Request explicit confirmation before starting (e.g. "This will run autonomously until the goal is met. Proceed? [y/N]") 4. Execute autonomously — running tools, writing code, running tests, fixing errors — in a loop until the goal criteria are satisfied 5. Only stop when: - The goal is demonstrably achieved (tests pass, output matches spec, etc.) - A hard token/step budget is exhausted (with a clear warning to the user) - A blocking error occurs that cannot be resolved autonomously ## Differences from YOLO Mode | | YOLO Mode | Perfect Mode | |---|---|---| | Auto-approves tools | ✅ | ✅ | | Requires a goal statement | ❌ | ✅ | | Iterates until goal is met | ❌ | ✅ | | Runs tests / validates output | ❌ | ✅ | | Warns about cost upfront | ❌ | ✅ | | Has a step/token budget guard | ❌ | ✅ | ## Proposed UX ` /perfect > Enter your goal: > Implement the user authentication flow with full test coverage and passing CI. ⚠ Perfect Mode will run autonomously and iterate until the goal is achieved. This may consume a large number of tokens and take several minutes. Proceed? [y/N]: y ✦ Goal locked in: "Implement the user authentication flow with full test coverage and passing CI." Starting Perfect Mode... ` ## Proposed Behavior - **Goal locking**: The goal statement is embedded into the system prompt for the session, so every agent decision is evaluated against it - **Self-evaluation loop**: After each major step, the agent assesses whether the goal has been met (e.g. by running tests, checking output, reviewing diffs) - **No mid-task interruptions**: Tools are auto-approved (inherits YOLO Mode behavior) - **Budget guard**: A configurable maximum number of steps or tokens (e.g. `perfectMode.maxSteps` in `nibble.json`) prevents runaway sessions; the user is notified if the budget is hit before the goal is achieved - **Completion report**: When the goal is reached, Nibble produces a summary of what was done, how many steps it took, and an estimated token cost ## Configuration (nibble.json) `json { "perfectMode": { "maxSteps": 50, "maxTokens": 200000, "requireConfirmation": true } } ` ## Notes - Perfect Mode implies YOLO Mode — all tool calls are auto-approved for the duration of the session - The goal statement should be stored in the session so it survives context reloads - Consider adding a visual indicator in the TUI status bar when Perfect Mode is active (e.g. `✦ PERFECT`) - This mode is intentionally opt-in and gated behind an explicit confirmation to prevent accidental runaway sessions