@sailfish-ai/autofix

Standalone installer for Sailfish Autofix — enables AI-powered bug fixing across AI coding agents without the desktop app.

Quick Start

npx @sailfish-ai/autofix setup

This single command will:

1. Open your browser to authenticate with Sailfish
2. Detect which AI coding agents are installed on your machine
3. Install the Autofix skill and MCP server for each detected agent

Once setup is complete, open any supported agent and use /sailfish-autofix <issue-id> to investigate and fix issues.

Prerequisites

• Node.js >= 18.0.0
• A Sailfish account (sign up at app.sailfish.ai)
• At least one supported AI coding agent installed

Installation

Option A: One-command setup (recommended)

No global install needed — npx runs it directly:

npx @sailfish-ai/autofix setup

Option B: Install globally

npm install -g @sailfish-ai/autofix
sailfish-autofix setup

Option C: From the Sailfish web app

Click "Fix with CLI" on any issue page to get a one-liner with an embedded auth token:

npx @sailfish-ai/autofix setup --token <token>

Option D: Via Agent Skills

If your tool supports the Agent Skills standard:

npx skills add @sailfish-ai/autofix

Note: This installs only the skill. Run npx @sailfish-ai/autofix setup separately to configure authentication and the MCP server.

Supported Agents

| Agent | Skill Location | MCP Config Location | |-------|---------------|---------------------| | Claude Code | ~/.claude/skills/ | ~/.claude.json | | Cursor | ~/.cursor/skills/ | ~/.cursor/mcp.json | | Codex (OpenAI) | ~/.codex/skills/ | ~/.codex/config.toml | | GitHub Copilot | ~/.agents/skills/ | ~/.vscode/mcp.json | | OpenCode | ~/.agents/skills/ | ~/.config/opencode/config.json | | Windsurf | ~/.windsurf/skills/ | ~/.windsurf/mcp.json | | Gemini CLI | ~/.gemini/skills/ | ~/.gemini/settings.json | | Goose | ~/.agents/skills/ | ~/.config/goose/config.yaml | | Generic (Agent Skills) | ~/.agents/skills/ | Manual MCP setup |

All agents also get a copy at ~/.agents/skills/sailfish-autofix/ (cross-client convention).

Commands

setup — All-in-one install

npx @sailfish-ai/autofix setup [options]

Authenticates, detects agents, and installs the skill + MCP server configuration.

| Option | Description | |--------|-------------| | --token <token> | Use a pre-authenticated token from the Sailfish web app | | --global | Install skill globally for all projects (default: true) | | --agents <list> | Comma-separated list of agents to configure (e.g. claude-code,cursor) | | -y, --yes | Skip confirmation prompts |

uninstall — Clean removal

npx @sailfish-ai/autofix uninstall [options]

Removes all installed skills, MCP configurations, and optionally credentials.

| Option | Description | |--------|-------------| | --keep-credentials | Keep authentication credentials (only remove skill + MCP config) | | -y, --yes | Skip confirmation prompts |

login — Authenticate

npx @sailfish-ai/autofix login [options]

Opens your browser to sign in via OAuth. Credentials are stored at ~/.sailfish/credentials.json.

| Option | Description | |--------|-------------| | --token <token> | Exchange a short-lived token instead of browser flow |

logout — Remove credentials

npx @sailfish-ai/autofix logout

status — Show current state

npx @sailfish-ai/autofix status

Displays authentication status and which agents have the skill and MCP server installed.

doctor — Diagnose issues

npx @sailfish-ai/autofix doctor

Checks credentials, backend connectivity, and agent installations. Suggests fixes for any issues found. If credentials are expired, doctor will attempt to re-authenticate automatically.

serve — Run MCP server

npx @sailfish-ai/autofix serve

Starts the MCP server in STDIO mode. This command is used internally by agent configurations — you don't need to run it manually.

fork — Resume an autofix session in another agent

npx @sailfish-ai/autofix fork <issue-id> [--agent <target>]

Downloads the conversation from the latest Autofix run for an issue and resumes it inside a coding agent of your choice — so you can pick up where the automated run left off, in the tool you prefer. It reconstructs the session (task, action log, tool outputs, files touched, decisions) and opens the agent on it, then the agent greets you with a short branch/worktree orientation that asks how you want to proceed before it touches any code.

Prerequisites: you must be authenticated first (login or setup) — fork downloads the run from the Sailfish backend. For the auto-launching targets you also need that agent's CLI on your PATH. The <issue-id> is the numeric ID from the issue's page in the Sailfish web app.

| Option | Description | |--------|-------------| | --agent <target> | Which agent to resume in (default: claude-code). See targets below. | | --native | opencode/gemini only: use the native import/checkpoint path (higher fidelity, but pinned to the tool's version) instead of the default portable RESUME.md. | | --cwd <path> | Repo to resume in (default: current git root). The session's file paths are patched to this checkout. |

Targets (--agent <target>):

| Target | Aliases | Default | How it resumes | |--------|---------|---------|----------------| | claude-code (default) | cc, claude | native | writes the session into ~/.claude, claude --resume | | goose | — | native | goose session import (via goose's own CC importer) → resumes the named session | | codex | — | portable | .sailfish-resume/ package, auto-launches Codex pointed at it | | cursor | cursor-agent | portable | .sailfish-resume/ package, auto-launches Cursor (@.sailfish-resume/RESUME.md) | | opencode | oc | portable | .sailfish-resume/ package + manual steps. --native: opencode import (pinned to opencode's version) | | gemini | gemini-cli | portable | .sailfish-resume/ package + manual steps. --native: /chat resume checkpoint (pinned to gemini's version) | | windsurf | devin | portable | .sailfish-resume/ package + manual "@-mention in Devin Desktop" steps |

Resume modes. claude-code and goose resume natively — they ride on Claude Code's own transcript format, so it's robust. opencode and gemini can resume natively, but their native paths reverse-engineer those tools' private, version-pinned formats, so they default to the portable path (a RESUME.md the agent reads, which survives version bumps); pass --native for full-fidelity native resume. codex/cursor/windsurf are portable-only (no native path — --native errors).

codex/cursor auto-launch the portable package; opencode/gemini/windsurf print manual steps — open the agent, then paste the one-line resume prompt it prints in a framed block. That prompt @-mentions @.sailfish-resume/RESUME.md, so the agent pulls the file into context and follows it. Their TUIs/GUI can't be launched with an initial prompt.

The portable package lands in a dedicated .sailfish-resume/ subdir with a bundled .gitignore — so it never clobbers a file you own and never lands in a commit.

# Resume issue 3088309's autofix run in Cursor
npx @sailfish-ai/autofix fork 3088309 --agent cursor

When it works, the agent opens already loaded with the run's conversation and greets you with the branch/worktree orientation. No autofix run found for issue … means autofix hasn't run on that issue yet; No session log found for this run means the run is still in progress or too old to resume.

Note: Copilot and the Generic target (listed under Supported Agents above) are configured by setup but are not fork targets — use one of the agents in the table above.

How It Works

The package installs two things for each detected agent:

1. Agent Skill (SKILL.md) — A workflow file following the Agent Skills open standard. It tells the AI agent how to investigate and fix issues using Sailfish telemetry data.

2. MCP Server — A Model Context Protocol server that provides 12 tools for accessing telemetry:

   | Tool | Description | |------|-------------| | autofix_get_issue_overview | Fetch the issue overview: description, customer report, exception details, stack traces, session IDs | | autofix_get_backend_logs | Server-side logs for a recording session | | autofix_get_exceptions | All captured exceptions in a session/window (co-occurring or unattributed, not just the issue's) | | autofix_get_frontend_logs | Browser console logs | | autofix_get_network_request_logs | HTTP request/response logs | | autofix_get_network_hops | Service-to-service call chain (distributed tracing) | | autofix_get_print_statements | stdout/print output | | autofix_get_user_actions | User interaction events (clicks, navigation) | | autofix_get_frontend_display_data | Frontend display state snapshots | | autofix_update_issue_status | Report fix progress back to Sailfish | | autofix_create_pull_request | Create a PR via GitHub/Bitbucket integration | | autofix_clone_repos | Clone company repos locally |

Credentials & Authentication

Credentials are stored at ~/.sailfish/credentials.json with mode 0600 (owner read/write only).

The MCP server reads credentials via the SAILFISH_JWT_FILE environment variable, which is set automatically in agent configurations.

Coexistence with Desktop App

This package works alongside the Sailfish desktop app (Veritas). They use compatible credential formats but separate configuration paths:

• Desktop app: writes MCP config as sailfishTelemetry / sailfishEnterprise
• Standalone CLI: writes MCP config as sailfishAutofix
• Both can coexist in the same agent configuration

Troubleshooting

"No supported agents detected"

The installer checks for known config directories and binaries. If your agent isn't detected:

npx @sailfish-ai/autofix setup --agents claude-code

"Credentials expired"

npx @sailfish-ai/autofix login

"Cannot reach backend"

Run npx @sailfish-ai/autofix doctor to diagnose connectivity issues. Ensure you can reach https://ui-api.sailfish.ai from your network.

Verify installation

npx @sailfish-ai/autofix status

Shows which agents have the skill and MCP server installed. All entries should show skill and mcp.

License

MIT