Visible-Agent-Loop-Protocol

Failure Gallery

VALP is easiest to understand through failures.

This gallery lists failure shapes that normal chat transcripts often hide and the VALP evidence that catches them. The examples are intentionally small so a reader can map each failure to receipts, expected evidence, review gates, or audit output.

Failure 1: Runtime Says Completed, Evidence Is Missing

Symptom:

runtime status = completed
expected evidence file does not exist

Why it matters:

A runtime lifecycle event proves only that the runtime reached its internal state. It does not prove the assigned work produced the expected artifacts.

VALP catch:

• dispatch_completed must cite expected_refs.
• valp audit fails when expected evidence paths are missing.
• final-synthesis.md must reference concrete evidence instead of repeating “done”.

Try it:

• Run minimal-audit-demo.md.

Failure 2: Text Was Inserted But Not Submitted

Symptom:

dispatch text appears in an agent input box
no submit action is proven
no worker output or expected evidence appears

Why it matters:

Pane automation can make a task look delegated while the agent never actually received it.

VALP catch:

• dispatch_inserted is not delivery.
• Full Mode needs dispatch_submitted with adapter proof.
• Completion still needs expected evidence after submission.

Relevant docs:

• Dispatch receipts
• Runtime adapters

Failure 3: Worker Recommendation Is Ignored

Symptom:

reviewer or worker reports a next step
coordinator finalizes without accepting, rejecting, merging, or bounding it

Why it matters:

Multi-agent work becomes cosmetic if the coordinator silently discards the useful parts of worker output.

VALP catch:

• Recommendations from Leader-declared Agents must be resolved in agent-recommendations.json.
• The coordinator controls scope: adoption means explicit disposition, not unlimited expansion.
• valp audit fails non-trivial tasks when required recommendation resolution is missing or pending.

Relevant docs:

• Squad routing
• CLI audit

Failure 4: Long Context Pollutes Worker Dispatch

Symptom:

coordinator pastes the full conversation or long skill-router output into every worker prompt
worker spends context budget on stale or irrelevant material

Why it matters:

Large dispatch payloads increase drift risk and make the worker responsible for the leader’s context hygiene.

VALP catch:

• Dispatches should use a concise task brief plus task-local references.
• Long context belongs in task.md, routing.json, and skill-recommendations.json.
• The leader owns dispatch precision.

Relevant docs:

• Context compression
• Skill recommendation

Failure 5: Manual Work Is Marketed As Full Mode

Symptom:

human copied prompts and results
repo copy claims automated runtime dispatch proof

Why it matters:

Manual Mode can be useful evidence, but it is not runtime-backed submission or status proof.

VALP catch:

• Manual Mode uses manual receipt labels.
• Full Mode requires adapter submission proof and runtime preflight.
• Project status must distinguish repository proof from live-runtime proof.

Relevant docs:

• Manual Mode
• Project status

Add A Failure Case

Good failure reports include:

• what claimed completion;
• which receipt or status was present;
• which expected evidence was missing or invalid;
• what valp audit did;
• whether the fix needed routing, adapter, docs, schema, or workflow changes.

Use GitHub Discussions for open-ended failure stories and GitHub Issues for reproducible bugs.

This site is open source. Improve this page.