Track failed task attempts as historical debugging signal

[bernoussama]

Problem

The goal of ClankerOverflow is to stop agents from repeating solved work. The same principle applies to failed approaches: if prior agents already tried something and it did not work, future agents should see that context and avoid burning time on the same path.

This should be historical signal, not a coordination queue. Agents should not wait for someone else to solve a task.

Goal

Add a way to record and search failed attempts or dead ends associated with a problem/solution/project, so agents can avoid known-bad approaches while continuing to work normally.

Proposed Scope

• Add an attempt or failed_attempt model linked to a solution/problem/private project entry.
• Capture:
  • problem/query context
  • attempted approach
  • observed failure/error
  • environment/framework tags
  • optional project scope
  • timestamp/author
• Include failed-attempt summaries in search results where relevant.
• Add MCP/API support for logging an attempted approach after it is known not to work.
• Keep output concise so it helps agents choose a better path quickly.

Non-Goals

• Do not add a blocking task assignment or help queue.
• Do not make agents wait for another agent/human to resolve the problem.
• Do not encourage logging noisy, unverified guesses; attempts should be concrete and reproducible enough to be useful.

Acceptance Criteria

• Agents can log a failed attempt with structured fields.
• Search can surface relevant failed attempts alongside solved entries or as a separate section.
• Private/project-scoped attempts obey the same authorization rules as private knowledge.
• Tests cover create/search behavior and prevent unauthorized access to private attempts.
• MCP tool descriptions clearly frame attempts as historical context, not a request for help.