HTI Fallback & Failure Handling

This Skill defines how to handle errors and "fallback" behavior in the hti-zen-harness project.

Core Principles

When working in this repo:

1. No fake success.

   • Do not hide failures behind:
     • Silent except Exception: pass
     • Returning empty objects ({}, []) or dummy values as if things worked
     • Swallowing errors and logging nothing
   • If something really cannot be implemented yet, mark it explicitly as such.

2. Prefer explicit failure over magical fallback.

   • It is better to:
     • Raise a clear, specific exception, or
     • Return an explicit error result / Result-style object
   • ...than to pretend everything is fine.

3. Be narrow and honest with error handling.

   • Catch specific exceptions (e.g., requests.Timeout, ValueError), not blanket Exception, unless there is a very strong reason.
   • When catching an exception, do at least one of:
     • Log it with clear context, OR
     • Re-raise it, OR
     • Convert it into a clearly named error type / return value.

4. No "mystery fallback" branches.

   • Avoid patterns like:
     pythonCopy

     def get_config():
         try:
             return load_from_disk()
         except Exception:
             return {}
     

   • Instead, make the failure visible:
     pythonCopy

     def get_config():
         try:
             return load_from_disk()
         except FileNotFoundError as e:
             raise ConfigMissingError("Config file not found") from e
     

Approved Alternatives to Bad Fallbacks

When you feel tempted to add "just in case" fallback logic, use one of these patterns instead:

1. Explicit TODO with loud marker

If something truly can't be implemented yet:

def fetch_remote_status():
    # HTI-TODO: Implement real remote status fetching.
    raise NotImplementedError("fetch_remote_status is not yet implemented for HTI harness")

Guidelines:

• Use HTI-TODO in comments to make search/refactoring easier.
• Do not silently return a fake value.

2. Logged error + re-raise

When you want extra context but don't want to hide the failure:

def run_harness_step():
    try:
        return do_step()
    except HarnessStepError as e:
        logger.error("Harness step failed: %s", e)
        raise

3. Structured, explicit error result

If the calling code already uses structured results:

@dataclass
class StepResult:
    ok: bool
    error: str | None = None
    data: dict[str, Any] | None = None

def run_harness_step() -> StepResult:
    try:
        data = do_step()
        return StepResult(ok=True, data=data)
    except HarnessStepError as e:
        return StepResult(ok=False, error=f"harness step failed: {e}")

Rules:

• Never set ok=True when something actually failed.
• Call sites should check ok and handle errors explicitly.

When You Think You "Need" a Fallback

Before adding fallback code, follow this checklist:

1. Ask: what is the correct behavior if this fails?

   • Crash fast and loudly?
   • Retry?
   • Skip this step but continue others?

2. If there is genuine choice, do one of:

   • Briefly outline 2–3 options in your response and pick the safest one by default.
   • Only ask the user when the choice meaningfully affects behavior or safety.

3. Document the behavior.

   • If you add a fallback, document:
     • When it triggers
     • What it does
     • Why this is acceptable in HTI/Zen context

Example:

def run_analysis():
    """
    Run analysis for the HTI Zen harness.

    Failure behavior:
      - If the upstream model call fails, we raise a clear error.
      - We do NOT silently fallback to dummy data.
    """
    ...

Patterns to Avoid in This Repo

Claude should avoid generating code like:

# BAD: silent swallowing
try:
    result = call_model(prompt)
except Exception:
    result = {"status": "ok"}  # Pretend success

# BAD: vague TODO fallback in production path
if not result:
    # TODO: handle better
    return {}

Or:

# BAD: doing something random "just to return something"
def pick_model():
    try:
        return pick_best_model()
    except Exception:
        return "gpt-5"  # arbitrary default with no explanation

Interaction with Testing

When writing or updating tests in this repo:

• Prefer tests that prove we're not silently falling back.

Example patterns:

• Assert that an invalid config raises a specific error.
• Assert that missing dependencies cause clear failures, not silent no-ops.
• If a fallback exists, add a test that triggers it and confirms:
  • It's documented
  • It's safe
  • It's clearly distinguishable from the normal path.

How to Use This Skill

Claude:

When editing code in hti-zen-harness, load and follow these guidelines whenever:

• You add error handling
• You touch network / API / MCP calls
• You consider "best-effort" or "fallback" behavior

If a user asks you to "just make it work even if X fails,":

1. Explain briefly what trade-off you're making
2. Choose explicit, debuggable behavior over silent fallbacks
3. Use HTI-TODO markers when something is intentionally left incomplete