SNIPER

Appearance

Learning System ​

SNIPER captures learnings from protocol executions -- CI failures, review findings, retrospective insights -- and codifies them into project memory. This creates a learning loop where agents improve across runs.

How Learning Works ​

Protocol execution → retrospective → learnings captured → stored in memory → injected into next agent spawn

The learning system has three components:

  1. Self-healing hooks -- PostToolUse hooks that detect failures in Bash output and instruct agents to fix them
  2. Retrospectives -- Post-protocol analysis that extracts patterns and codifies them
  3. Memory injection -- Agents receive relevant learnings in their spawn prompts

The /sniper-learn Command ​

Use /sniper-learn to manually submit, review, or deprecate project learnings:

  • Submit a new convention, anti-pattern, or decision
  • Review candidate learnings and promote or reject them
  • Deprecate outdated learnings that no longer apply

Self-Healing CI ​

Beyond signal capture, SNIPER includes a self-healing pattern (inspired by Aider). The settings-hooks.json PostToolUse hook detects test/lint failures in real-time and instructs the active agent to fix them before proceeding:

11 lines
json
{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Bash",
        "description": "Self-healing CI: detect test/lint failures and instruct agent to fix",
        "command": "if echo \"$CLAUDE_TOOL_OUTPUT\" | grep -qiE '(FAIL|FAILED|ERROR|error TS|ESLint)'; then echo 'WARN: Tests or linting failed. Fix the issues before continuing.'; fi"
      }
    ]
  }
}

This creates a tight feedback loop:

Agent writes code → runs tests → tests fail → hook fires → agent fixes → runs tests again

The agent does not mark its task complete until tests pass. No human intervention needed.

Learning Injection ​

When agents are spawned, relevant learnings are injected into their context:

  1. The lead-orchestrator reads .sniper/memory/ (conventions, anti-patterns, decisions)
  2. Learnings matching the current phase and agent role are selected
  3. Selected learnings are appended to the agent's spawn prompt under a ## Project Memory section

This means if a test failed because of a date mocking issue last week and it was captured as a convention, the developer agent spawned this week receives that learning automatically.

Retrospectives ​

After protocols with auto_retro: true complete, the retro-analyst:

  1. Analyzes gate results (what passed, what failed, what was borderline)
  2. Reviews signal history for the protocol run
  3. Generates recommendations

Retrospective output is stored in .sniper/memory/retros/.

Configuration ​

Enable the learning system in .sniper/config.yaml:

2 lines
yaml
visibility:
  auto_retro: true        # Run retrospectives after protocol completion

Next Steps ​