thearchitectit/aiggp-agentic-framework/three-strikes

Three Strikes Rule

Track your attempts on each task. Never continue beyond 3 failures.

Why Three Strikes?

Continuing beyond 3 attempts:

• Wastes tokens and compute
• Contaminates context with failed attempts
• Frustrates the user
• Rarely succeeds (debugging by random mutation)

Strike Tracking

Maintain mental (or explicit) state:

Task: <description>
Strike 1: <what was attempted, why it failed>
Strike 2: <alternative attempted, why it failed>
Strike 3: <HALT — escalate to user>

What Counts as a Strike?

A strike is counted when:

• A targeted fix or approach is attempted and fails
• The same error recurs after adjustment
• A different error surfaces that indicates the approach is wrong

A strike is NOT counted when:

• A syntax fix is needed as part of the same attempt
• A dependency is missing and must be installed first
• The test environment needs setup

After the Third Strike

1. STOP immediately — Do not attempt a 4th fix
2. Summarize attempts — List what was tried and what failed
3. Describe current state — What is broken, what works, what is uncertain
4. Ask user for guidance — Present options or ask for direction
5. Wait for response — Do not proceed until user clarifies

User Message Template (Third Strike)

HALT: Three Strikes
 
I've attempted this task 3 times without success:
 
1. <First attempt>: <what was tried> → <result>
2. <Second attempt>: <what was tried> → <result>
3. <Third attempt>: <what was tried> → <result>
 
Current state:
- <What is working>
- <What is broken>
- <What I am uncertain about>
 
I need guidance on how to proceed. Possible paths:
A) <Option A>
B) <Option B>

Exceptions

The Three Strikes Rule can be overridden ONLY by explicit user instruction:

• "Keep trying"
• "Try approach X"
• "It's okay, keep going"

Without explicit override, HALT at 3 strikes every time.

Task

Track failure attempts on the current task. If this is the first or second failure, adjust and retry. If this is the third strike, halt immediately and escalate to the user with a summary of attempts and current state.

References

• skills/halt-conditions/SKILL.md — Full halt conditions checklist
• AGENTS.md — Core safety protocols