Propose Skills from Wiki Traces with wikiskill-proposer

Where this fits

Picture this: your agent has failed the same type of task three times this week. You have execution traces piling up and a wiki that documents known patterns, but turning those failures into actionable skill updates is a manual, error-prone chore. wikiskill-proposer automates that diagnostic loop. It reads your wiki and traces, identifies root causes, and writes a structured skill proposal (create, patch, or no_action) that your harness can apply and validate.

Why agents benefit

• Root-cause diagnosis from evidence: Instead of guessing, the agent cross-references wiki/index.md, rejected proposals in skill-impact.md, and at least four execution traces per iteration.
• Avoids repeating past mistakes: It reads the full content of previously rejected proposals before deciding, so new suggestions actually converge.
• Structured output for automation: The proposal JSON goes to runs/proposals/iter-NN.json, ready for a harness to apply, validate, and close the loop.
• Balances exploration with intent: The workflow targets trace exploration using the training summary table, making it cheaper than letting the agent randomly read every log.
• Separates decision from action: The skill proposes; your system disposes. This keeps humans in control while letting agents do the analysis.

Practical scenarios

Recurring API timeout failures

Your agent repeatedly fails on a third-party API call. wikiskill-proposer reads the wiki pattern page for timeout handling, scans recent traces, and proposes a patch that adds a retry-with-backoff step to the existing skill—not a brand-new skill that duplicates logic.

New domain where no skill exists

You start making agents handle PDF invoice extraction, but no wiki pattern covers it. After reading several failed traces, the skill proposes creating a new invoice-parser skill with the exact error-prone steps identified in the logs.

Avoiding rejected solution repetition

A previous skill proposal was rejected because it used regex instead of a proper parser. The proposer reads skill-impact.md first, sees that rejection, and instead suggests a patch using the already-endorsed document parsing library.

Add it to your agent workflow

Install it with:

npx skills add ashutoshsinghpr7/wikiskill --skill wikiskill-proposer

Then give your agent a task prompt that includes the wiki paths, trace paths, and output path. The skill handles the rest. A typical proposal output looks like:

{
  "action": "patch",
  "skill": "api-client",
  "change": "add retry with exponential backoff on 429 responses"
}

Compare before adopting

Before you commit, compare it against alternatives like a hand-written runbook or a generic prompt that says “analyze these logs.” Look at the quality of the proposal schema, how fresh the wiki content is, whether the skill is actively maintained, and whether your trace format is compatible. A skill that produces clean, validated JSON is more valuable than one that writes long natural-language explanations.

Why it is worth tracking

With 40 stars and a focused design around trace-driven skill proposals, wikiskill-proposer is an early but promising signal that agent workflow tooling is moving toward self-improvement loops. Track it if you’re building autonomous agents that need to learn from failures without human intervention. For ongoing details and updates, see the skill page on OpenAgentSkill and evaluate it again once more community usage and validation results appear.