Skip to main content
Ask at synapse.synheart.ai lists what Synapse told your coding agents to do, newest first. Each time your agent uses Synapse over MCP, or the guard checks an action on your machine, Synapse records a short decision: what it advised and why. Ask shows your own decisions only. There’s no organization-wide list and no way for a manager to open someone else’s. Leads and above see team totals on the Agents page, never individual rows (see Team totals).
Ask used to be the name of the dashboard’s search page. That page is now Search, and old /ask?q=… links go there.

​
What happened

Filter the list by what happened, by period (7, 30 or 90 days) and by task. All checks shows the first three. Proceeded has its own filter. Synapse itself never blocks. Recall and the intent check can at most tell the agent to check with you. Only your organization’s own guard rules block an action or need approval, and the CLI on your machine enforces them.

​
Reasons and evidence

Each row shows:
  • What happened, and what the agent was told, in one sentence.
  • Reasons, in plain words. For example: a current decision governs this code, a high risk is present, nothing current was found, or a team rule stopped the action.
  • Evidence, as identifiers: the decisions Synapse cited (each links to Search), the kind of risk, or the rule id.
  • The coding agent (Claude Code, Cursor, Droid and so on), the time, and the outcome.
Open a row to see the full explanation. An agent can get the same explanation with the MCP tool explain_decision, for your own decisions only.

​
Tasks

Decisions are grouped by task so you can follow one piece of work across sessions and agents. Synapse works out the task from, in order: an issue or PR number, an issue number in the branch name (feat/123-title, issue-123), the branch itself, or the agent’s session. Synapse keeps a keyed digest of a branch, not its name. Default branches such as main are never a task.

​
Accept or reject

In the decision’s detail panel you can record whether you accepted the agent’s edit, with an optional reason. A verdict you give in the dashboard counts as your own, exactly like one from the terminal. A decision with no task can’t take a verdict. From the terminal, with CLI 0.23.0 or later:
The task comes from the current branch unless you pass --issue, --pr or --task. Reasons are wrong-approach, incorrect, incomplete, out-of-scope, style, violates-decision, breaks-checks, security, looks-good, accepted-with-changes and other. synheart feedback runs only for a person at an interactive terminal. It refuses in an agent’s shell, so an agent can’t accept its own work.

​
Relayed feedback

If you tell your agent what you think of its edit, the agent can pass that on with the MCP tool synapse_feedback. Synapse stores it as relayed: the row says Relayed by the agent, and relayed verdicts are counted separately from your own. An agent’s credentials can’t record a verdict as yours. Your own verdict from the dashboard or the CLI takes precedence.

​
Summary

The cards at the top count, for the period:
  • Asked you and Blocked.
  • Correction cycles: on each task, how many times you rejected an edit before you accepted one, averaged over the tasks you accepted.
  • First-pass acceptance: the share of accepted tasks where you accepted without rejecting first.
Both measures are rough guides to how well agents land work with you. They aren’t for ranking people or judging performance.

​
Getting decisions here

Decisions appear once either of these is set up:
  • Your agent uses Synapse over MCP. Recall and the intent check (related_context) each record a decision. See Connect your agent.
  • The guard shares its decisions. Each guard check can be sent as a decision record. This is off by default. Turn it on, from your own terminal:
It sends the rule id, what the rule did, the coding agent, the time and the agent’s session id. It never sends the command, a file path, a prompt, the repository or the branch. synheart guard team decisions off stops it. Set up the guard first: see Developer setup.

​
What’s stored

Decision records hold no content. Each record keeps:
  • the mode and reason codes;
  • evidence identifiers;
  • the task id and the coding agent;
  • the time.
Synapse never stores the question or instruction your agent sent, or any prompt, code, diff, excerpt, secret, branch name or session id in clear. The instruction text of an intent check is used for the search and then discarded. Records hold no human state either. Turning off team human-state views doesn’t affect Ask or the team decision totals, and neither of them ever shows human state. Retention. Decision records and your accept/reject verdicts are kept for 90 days and deleted daily after that. Erasing your Synapse data deletes them straight away, as does erasing the organization’s data.

​
Team totals

Leads and above (Synapse Leads, Synapse Admins, and the organization owner with a seat) see totals for the whole team under Agents → Team:
  • decisions by what happened;
  • the most common reasons;
  • the decisions cited most;
  • team correction cycles and first-pass acceptance;
  • what the intent check found.
These are totals only, with no names, tasks or per-person rows. They’re withheld until enough developers contribute, at your organization’s cohort floor.