Sasha MCP reference

suggestImprovement

Suggest an improvement

Suggest an improvement to Sasha. Feature ideas are wanted: if a tool would have made your task easier, say so. File one call per distinct issue, at the moment you hit it. Do not batch a session into one report. When the user gives feedback about Sasha itself (they say "feedback", "tell Sasha", "suggestion" or "Sasha should..." about its behaviour, skills, agents or docs), file it at once, without asking, with origin "human-reported"; do not offer to change a skill or write a spec instead. Do not use this for a knowledge-content problem that a write tool can fix. Before you file severity "blocking", corroborate the mechanism against document bytes or tool results, not memory. Put the user's own words in detail as User said: "<exact words>" (date), so the Sasha team reads them verbatim. When the user answers a question you asked about the issue, call again with followUpTo set to the reference you got and their answer in detail; it is added to the same report. A failed filing never blocks your main task; try once more near the end if it still matters. Returns an id and a short reference such as FEEDBACK-012; give the person the reference. Call checkSuggestion with either one in a later session to see what happened.

Arguments

ArgumentTypeRequiredDescription
kindenum (feature, bug, friction, docs)requiredfeature | bug | friction | docs
summarystringrequiredOne line, 10 to 300 characters
detailstringoptionalWhat you tried, what happened, what you expected. At most 2 KiB.
toolNamestringoptionalThe tool this is about, if any, named exactly as this server publishes it (for example app__time_sheet_entry__entries__create). Your client's own prefix is accepted.
projectstringoptionalThe project this is about, if any
severityenum (blocking, major, minor)optionalblocking | major | minor
originenum (agent, human-reported)optionalhuman-reported when the user told you about it; agent (the default) when you found it yourself
followUpTostringoptionalThe id or FEEDBACK reference of a report you filed earlier in this conversation; detail is added to it

Scopes

None. Any authenticated connection can call it.

Annotations

  • Read only: no
  • Destructive: no

When to use

Call suggestImprovement at the moment a tool, a limit or a document layout gets in your way. Feature ideas are wanted: if a tool would have made your task easier, say so.

It needs no scope. Any authenticated connection can call it, including a member's.

  • File one call for each distinct issue. Do not put a whole session into one report.
  • When the person gives feedback about Sasha itself (they say "feedback", "tell Sasha", "suggestion" or "Sasha should..." about its behaviour, skills, agents or docs), file it at once, without asking, with origin human-reported. Do not offer to change a skill or write a spec instead.
  • kind is feature, bug, friction or docs.
  • summary is one line of 10 to 300 characters.
  • detail is optional, at most 2 KiB: what you tried, what happened, what you expected.
  • toolName is optional. Name the tool as this server publishes it, for example readDoc. A name with your client's own prefix is also accepted. It must start with a letter, hold only letters, digits, _ and -, and have at most 128 characters.
  • project is optional, at most 200 characters. It is kept only when this person can see that project. If not, the suggestion is still recorded, without the project.
  • severity is optional: blocking, major or minor. Before you file blocking, check the cause against document text or tool results, not memory.
  • origin is optional: human-reported when the person you work for told you about the problem, agent (the default) when you found it yourself.
  • Put the person's own words in detail as User said: "<exact words>" (<date>), so the Sasha team reads them as they were said.
  • followUpTo is optional: the id or ref of a report this person filed earlier. When you asked them about an issue and they answered, call again with followUpTo and their answer in detail. The answer is added to that report, which becomes human-reported. summary and kind are still required but are not stored.
  • Do not use this tool for a problem in the knowledge content that a write tool can fix.
  • A failed filing never blocks your main task. If it still matters, try once more near the end of the session.

Example

{
  "kind": "feature",
  "summary": "searchKnowledge should return every matching line in a file, not only the first",
  "detail": "I searched northwind-advisory for 'discount'. pricing-policy.md mentions discounts on several lines, but only the first line came back, so I had to read the whole document to find the approval rule.",
  "toolName": "searchKnowledge",
  "project": "northwind-advisory",
  "severity": "minor",
  "origin": "agent"
}
{ "id": "sug_4f1c9a2e7b3d5a6c8e0f1b2d", "ref": "FEEDBACK-012", "status": "received" }

The status field tells you what happened:

  • received: the suggestion is recorded. ref is a short reference such as FEEDBACK-012; give it to the person you work for, who can quote it later. Call checkSuggestion with the id or the ref in a later session.
  • duplicate: this connection filed the same summary in the last hour. The result holds the id and ref of that earlier suggestion. Nothing new is recorded, and the earlier suggestion keeps its own origin.
  • rate-limited: this connection or this person filed 20 suggestions in the last hour. The result has retryable: true and retryAfterSeconds: 3600. Try again later.
  • updated: a followUpTo call added the answer to the earlier report. The result holds that report's id and ref.
  • not-found: the followUpTo report does not exist, or this person did not file it (for a member, this connection did not). The result has retryable: false. File a new suggestion instead.
  • not-recorded: the instance could not save it. The result has retryable: true. Try once more near the end of the session.
{ "status": "rate-limited", "retryable": true, "retryAfterSeconds": 3600 }

Refusals and what to do

  • A kind, severity or origin outside the lists, a summary shorter than 10 or longer than 300 characters, a detail longer than 2 KiB, a project longer than 200 characters, or a toolName that does not start with a letter, has characters other than letters, digits, _ and -, or is longer than 128 characters, or a followUpTo that is not a sug_ id or a FEEDBACK- reference, is refused as invalid arguments. Correct the value and call again.
  • rate-limited and not-recorded are answers, not errors. Continue your main task.

Related

  • checkSuggestion reads the status of a suggestion by its id.
  • getDocs has an improvements section that lists the dispositions.
Built with BernardCookie settings