/tres-request-skill-update

TRES Plugin Feedback

Guide the user through submitting clear, actionable feedback about the TRES Claude plugin and save it via the save_ai_conversation_feedback TRES MCP tool. No extra configuration required — works for any user with a TRES API token.

This skill operates in two modes:

1. Explicit feedback — the user asks to submit feedback, report a bug, etc. Run the full guided flow starting at Step 1.
2. End-of-conversation prompt — the conversation is wrapping up. Offer a quick, non-intrusive nudge (see Step 0) and only proceed to the full flow if the user opts in.

Step 0 — End-of-Conversation Prompt

When the conversation appears to be ending (the user says "thanks", "that's all", "looks good", etc.) and this skill is triggered proactively:

Ask casually:

Before you go — any feedback on how this session went? Bugs, ideas, or things that could be better? Happy to log it for the team in 30 seconds.

• If the user declines ("no", "I'm good", "nah") — say "No worries, have a great day!" and stop. Do NOT push further.
• If the user shares something — transition into Step 1 and run the full guided flow, but keep it lightweight. Skip questions they've already answered in the conversation and aim for a quick turnaround.
• Keep the tone light and optional — this should never feel like a mandatory survey.

Step 1 — Identify Submitter

1. Call get_viewer (TRES MCP, no arguments) to identify the submitter.

   • Extract the organization name from the response.
   • If get_viewer fails, use "Unknown org" as the submitter and continue.

2. Get the user's git identity for the feedback record:

   git config user.name && git config user.email

   Use line 1 as agent_name and line 2 as agent_email. If git is not available, use the system username and leave email empty.

Tell the user: "You're connected as {orgName}. What feedback would you like to share?"

Step 2 — Understand the Feedback

Let the user describe their feedback in their own words first. Then classify it into one of these categories based on what they said:

Confirm the category with the user: "It sounds like this is a {category} — is that right?"

If the feedback is about a specific skill or MCP tool, identify which one. If it's general or about the plugin overall, note that.

Known skills:

Maintainer note: update this table when new skills are added to the plugin.

Step 3 — Dig Deeper

Based on the category, ask targeted follow-up questions to make the feedback actionable. Ask one or two questions at a time — keep it conversational, not interrogative.

For bug reports:

• What exactly happened? What did you see?
• What did you expect to happen instead?
• What were you doing when it happened? (steps to reproduce)
• Did you see an error message? If so, what did it say?
• Can you share a specific example? (tx hash, wallet address, input you used)

For feature requests & improvements:

• What problem would this solve for you?
• How do you handle this today without the feature?
• Can you describe what the ideal experience would look like?
• How often do you run into this need?

For new skill ideas:

• What workflow or task would this skill automate?
• Who on your team would use it, and how often?
• Can you walk through a concrete example of how you'd use it?
• What data source would it need? (TRES MCP, external API, local files)

For MCP / data issues:

• Which MCP tool or query was involved?
• What data did you get back, and what was wrong about it?
• What did you expect the data to look like?
• Can you share the specific query or identifiers you used?

For positive feedback:

• What specifically worked well?
• Was there anything that surprised you (in a good way)?
• Is there a particular workflow or use case where it really shined?
• Would you change anything to make it even better?

For general feedback:

• What part of the experience are you reacting to?
• Was anything confusing or unclear?
• What would have made it better?

Adapt based on what the user has already told you — skip questions they've already answered. The goal is to get enough detail that someone reading the feedback can understand the context and take action without needing to ask follow-up questions.

Step 4 — Preview the Feedback

Compose the feedback and show it to the user for review. Format it clearly:

Headline: {concise summary — max 80 chars}

Category: {Bug report | Feature request | Improvement | New skill idea | MCP issue | Positive feedback | General}
Area: {skill name, "MCP", or "General"}
Org: {orgName}

---

{Well-structured description that includes:
 - What the feedback is about (context)
 - The core issue, idea, or praise (substance)
 - Supporting details — steps to reproduce, examples, expected behavior, etc.
 - Impact — how often this comes up, how many people it affects, how it blocks work}

Tags: {comma-separated list}

Writing the description:

• Synthesize the user's answers into a clear, readable narrative — don't just dump Q&A pairs.
• Lead with the most important point.
• Include concrete details (tx hashes, error messages, specific steps) — these are what make feedback actionable.
• For positive feedback, be specific about what worked and why it mattered.

Also show the user the conversation excerpt that will be submitted (see Step 5 format). This gives them informed visibility into what leaves their machine.

Ask: "Here's what I'll submit — does this capture everything? Want to change anything?"

Iterate if the user wants edits. Only proceed on explicit confirmation.

Step 5 — Submit Feedback

Call save_ai_conversation_feedback with these arguments:

Conversation format — include only the last 30 exchanges (user+assistant pairs) from your context window, not the entire session history. Apply a redaction pass before including any message:

• Replace Bearer [A-Za-z0-9._\-]{20,} → Bearer [REDACTED]
• Replace standalone 64-character hex strings → [REDACTED_HEX]
• Replace file paths matching /Users/<name>/ or /home/<name>/ → /Users/[REDACTED]/
• Remove any apparent BIP-39 mnemonic phrases (12–24 dictionary words)

Format:

USER: <text, redacted>

ASSISTANT: <text, redacted>

TOOL_CALL: <tool_name>(arg1=value1, arg2=value2)
TOOL_RESULT: <one-line summary of the result>

... continue for the last 30 exchanges only ...

Step 6 — Report Result

On success: Tell the user:

Your feedback has been submitted — thank you! The team will review it.

If the feedback was a bug or blocker, add:

If this is urgent, reach out to the team directly as well.

On failure: Surface the error from the MCP tool and suggest:

• Check that the TRES API token is valid
• Try again — it may be a transient issue

Error Handling