Skip to content

analyze_issue_with_seer can't replace a completed Seer run #2273

Open

Description

@sentry-junior

Pattern: once an issue has a completed Seer run, analyze_issue_with_seer always returns that run. Callers can't start a fresh analysis through MCP, even when the stored root cause is stale (for example, it points at code that has since been fixed and shipped), and the instruction argument is dropped without notice.

Reports: 1 (user feedback, MCP server 0.42.0, analyze_issue_with_seer, impact: blocked)

Current behavior (packages/mcp-core/src/tools/catalog/analyze-issue-with-seer.ts @ 516cad4)

  • L149: the tool calls startAutofix only when there's no previous run or the previous run's status is error.
  • L209-225: any other terminal run (completed, waiting for input) is returned as-is with Found existing analysis (Run ID: …). params.instruction is only passed to startAutofix (L155), so on this path it is silently ignored.
  • The input schema (L83-92) has no way to ask for a new run, and the description says Results are cached - subsequent calls return instantly (L80).

Sentry API (getsentry/sentry@fe43ef0)

  • POST /issues/{id}/autofix/ starts a new run whenever neither run_id nor sentry_run_id is passed (src/sentry/seer/endpoints/group_ai_autofix.py L286-299). An existing completed run doesn't block it.
  • GET /issues/{id}/autofix/ returns the group's latest run (get_autofix_agent_state → fetch_latest_run_state, src/sentry/seer/autofix/autofix_agent.py L558-575). So the gap is on the MCP side only.

Evidence from the report

  • The sampled analyze_issue_with_seer call made GET …/issues/{id}/ and GET …/issues/{id}/autofix/ and no POST. It returned Found existing analysis (Run ID: …) / ## Analysis Complete with the old root cause. That sampled call had no instruction argument; the reporter says an earlier call did pass one. We sample 30% of traces, so that call may not have been recorded.

Reporter's proposal: a force/refresh parameter on analyze_issue_with_seer.

Related: #1270 adds a runId guard on the same code path that stops the tool from starting a new run. Any change here needs to fit with it. Starting a new run spends Seer credits, which is likely why the tool reuses the existing run today.

via automation evt_905b9c478746fd9deeb853f85ad7441b · Run ID dispatch_603ac936cce8aaa570e9b7eae0aa9519.

--

View Session [Sentry]

Activity

  1. linear-code commented on Nov 14, 2026

    @linear-code
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    MCPSentry MCP tools, transports, and documentationmcp-feedback-pattern

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions