NeuBird Ops Agent

Production ops superpowers, powered by NeuBird — the AI SRE that lives in your terminal.

When to Use

✅ USE this skill when the user asks about anything in production:

❌ DON'T use this skill when:

• neubird desktop binary is not installed — direct user to neubird.ai
• The question is about code review, writing code, or pre-deploy checks
• The user wants a dashboard — open the observability platform directly

CLI Interface

# List available projects
neubird projects

# Run a named capability
neubird run <capability> --project <project-name> --session /tmp/

# Free-form investigation (alias for 'run investigate')
neubird investigate "<prompt>" --project <project-name> --session /tmp/

# Follow-up question (project inherited from session)
neubird run <capability> --session /tmp/nb-<timestamp>.json

# Clean up session when done
neubird run --cleanup --session /tmp/nb-<timestamp>.json

All 9 Capabilities

Session Behavior

• --session /tmp/ → auto-generates /tmp/nb-.json, prints path to stderr
• --session /tmp/nb-123.json → creates on first call, resumes on follow-ups
• --project required on first call; inherited from session on follow-ups
• Use --cleanup when done to remove the session file

Exit Codes

Output Format

Investigations stream over 60–180s. Output has two layers:

• Spinner on stderr — ⠋ thinking, ⠙ exploring, etc. — ignore
• Findings on stdout — narrative markdown, ends with Completed in XmYs

Use --verbose to see tool calls and MCP server logs during debugging.

Agent Workflow

1. Understand the request — identify which capability fits (see table above); for ambiguous requests default to investigate
2. Determine the project — infer from context ("prod" → prod_cop, "staging" → staging_auto); if ambiguous run neubird projects and ask
3. Set expectations — tell the user: "Running [capability] against [project] — this takes 1–3 minutes..."
4. Start the run — session path is printed to stderr as Session: /tmp/nb-.json:

For a named capability:

```

neubird run --project --session /tmp/

```

For a free-form investigation:

```

neubird investigate "" --project --session /tmp/

```

1. Narrate findings — lead with the bottom line, don't dump raw output:
• State the headline conclusion first
• Summarize key findings with supporting evidence
• Give a concrete recommended action when warranted
• Offer to drill deeper or follow up

1. Follow-up if needed — reference the session path, no --project required:

```

neubird investigate "" --session /tmp/nb-.json

```

1. Clean up when done:

```

neubird run --cleanup --session /tmp/nb-.json

```

Project Names

Common project slugs: prod_cop, staging_auto, dev_cop, prod_cop_sev2.

Run neubird projects to list all available projects with their IDs.

References

Load these when relevant to the findings:

Constraints

MUST DO

• Lead every response with the headline conclusion
• State blast radius / scope before recommending action
• Give a concrete next step, not just analysis
• Offer to drill deeper after every finding
• Clean up session files when done

MUST NOT DO

• Dump raw neubird output without narration
• Fabricate findings if the command fails — report the error clearly
• Skip scope/blast radius — "unknown" is valid but must be stated
• Recommend rollback without checking if a recent deploy is in scope