Skip to content

fix(crashlytics-autotriage): pre-install Firebase CLI so the MCP server connects#89

Merged
stozo04 merged 1 commit into
mainfrom
feature/crashlytics-autotriage-mcp-startup
Jun 23, 2026
Merged

fix(crashlytics-autotriage): pre-install Firebase CLI so the MCP server connects#89
stozo04 merged 1 commit into
mainfrom
feature/crashlytics-autotriage-mcp-startup

Conversation

@stozo04

@stozo04 stozo04 commented Jun 23, 2026

Copy link
Copy Markdown
Owner

Description

After #88, the workflow runs once and the Crashlytics tool group is activated — but Claude still reported the Firebase MCP server "failed to connect" (3 runs on #87, no stacktrace).

The run log shows the action itself is healthy (OIDC ✓, Claude installed ✓, 12 turns, success) and that the MCP server startup happens inside the SDK's hidden output. The most likely cause: the MCP config launches npx -y firebase-tools@latest mcp, which cold-downloads the entire firebase-tools package at MCP-startup time, exceeding the SDK's connection window on a fresh runner.

Changes

  1. Pre-install firebase-tools in a step before the Claude step, so the MCP server launches instantly.
  2. .github/firebase-mcp.json now calls firebase-tools (no @latest), so npx resolves the pre-installed global copy instead of re-downloading.
  3. Temporarily set show_full_output: "true" on the action so the MCP startup logs are visible if anything still fails. Revert this once the run is green.

Type of Change

  • 🐛 Bug Fix (CI / MCP startup)

How Has This Been Tested?

  • Reasoned from the [Crashlytics][TEST] Phase 0 retest — real ID #87 run log (action healthy; MCP startup hidden; npx @latest cold-download is the prime suspect).
  • Re-test after merge: re-label an issue crashlytics-auto; with full output on, confirm the firebase MCP server connects and crashlytics_get_issue returns data.

Follow-up

  • Once green, open a tiny PR to drop show_full_output again.
  • If it still fails with full output, the logs will show whether it's auth (roles/firebasecrashlytics.viewer / Crashlytics API not enabled) vs. another startup error.

Checklist

  • 🔍 Self-review performed.
  • 📖 Workflow YAML + JSON validated.

…er connects

The MCP config ran `npx -y firebase-tools@latest mcp`, which cold-downloads the
whole firebase-tools package at MCP-startup time — exceeding the SDK's connection
window on a fresh runner, so the firebase server "failed to connect" and the
crashlytics_* tools never loaded (issue #87, 3 runs).

- Add a step to `npm install -g firebase-tools@latest` before the Claude step.
- Point firebase-mcp.json at the pre-installed copy (drop `@latest` force-fetch)
  so the MCP `npx` resolves instantly instead of re-downloading.
- Temporarily enable `show_full_output: true` so the MCP startup is visible if
  anything still fails (revert once green).
@stozo04 stozo04 merged commit bb82c41 into main Jun 23, 2026
1 check passed
@stozo04 stozo04 deleted the feature/crashlytics-autotriage-mcp-startup branch June 23, 2026 20:14
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant