| name | scout |
| description | Research & analysis workflow. Investigates questions, gathers findings, optionally validates with a fresh subagent. Produces a report — no code, no PRs, no commits. Use standalone or as a fellowship teammate. |
Scout — Research & Analysis
Overview
Investigates questions and analyzes codebases without producing code changes. Runs autonomously through Investigate → (Validate) → Deliver phases. When used as a fellowship teammate, the scout agent definition enforces read-only tool access. Also works standalone.
When to Use
- Research questions about a codebase ("how does X work?")
- Deep analysis ("what are all the entry points for auth?")
- Data collection ("list all API endpoints and their middleware chains")
Phase Flow
Investigate ──→ (Validate) ──→ Deliver
Process
Investigate
Goal: Gather thorough, factual findings about the question.
Actions:
- Invoke
/council to load task-relevant context
- Use Explore agents (Agent tool, subagent_type=Explore) to scan relevant code paths
- Read key files, trace call chains, understand behavior
- Document findings with specific file paths and line references
Investigate must produce:
High confidence = verified by reading actual code at specific lines.
Medium confidence = inferred from patterns, naming, or partial evidence.
Low confidence = assumption based on conventions or incomplete information.
If findings are incomplete, keep investigating. Don't move to validation or delivery with gaps.
Validate (conditional)
Goal: Adversarially verify findings using a fresh subagent with no context pollution.
When to validate:
- Deep analysis involving multiple systems or complex interactions
- Findings that will inform architectural decisions or code changes
- Questions where being wrong would waste significant downstream effort
- Any time there are Medium or Low confidence findings that matter
When to skip validation:
- Simple lookups ("where is X defined?")
- Straightforward questions with all High confidence answers
- Time-sensitive quick questions
Validation procedure:
- Write your findings to a working file (e.g.,
docs/research/<topic>.md) — do NOT commit
- Spawn a validator subagent via the Agent tool with subagent_type "general-purpose":
You are a research validator. Your job is adversarial: challenge
assumptions, verify factual claims, and flag anything wrong or unsupported.
FINDINGS TO VALIDATE:
<paste findings here, including file paths and line references>
INSTRUCTIONS:
1. For each factual claim, read the referenced file and line range.
Does the code actually do what the finding says?
2. For each Medium/Low confidence finding, investigate independently.
Can you confirm or refute it?
3. Produce a validation report:
- CONFIRMED: claims you verified are correct
- CONTESTED: claims that are wrong or misleading, with evidence
- UNVERIFIED: claims you couldn't confirm or deny
BOUNDARIES:
- Read any file. Do NOT modify any files or run commands.
- Be adversarial — your value is in catching errors, not agreeing.
- Review the validation report
- For contested findings: re-investigate, correct or remove the finding
- For unverified findings: downgrade confidence or investigate further
Deliver
Goal: Send findings to the requester in a structured format.
Fellowship teammate: Send via SendMessage to the lead (Gandalf).
Standalone: Present findings directly to the user.
Write findings file: In both modes, write the full report to .fellowship/scout-findings-{scout_name}.md (using the configured dataDir if overridden). This file enables scout-to-quest promotion — without it, findings can't be promoted into a quest. The {scout_name} comes from the scout's name (e.g., scout-auth-analysis). Do NOT commit this file.
Report format:
## Scout Report: <question>
### Findings
- <finding 1> (`path/to/file.ts:42-58`)
- <finding 2> (`path/to/other.ts:10-25`)
- ...
### Confidence
- High: <list>
- Medium: <list>
- Low: <list>
### Validation
- Confirmed: <validated claims>
- Contested: <claims corrected after validation>
- Skipped: <if validation was not performed, say why>
### Files Written
- <list any research files written, noting they are not committed>
Routing (fellowship only): If the spawn prompt includes a routing target (e.g., "→ send to quest-auth-bug"), also send findings to that teammate via SendMessage in addition to the lead.
Fellowship Integration
When running as a fellowship teammate, the scout agent definition restricts your tools to read-only source access plus Write for research notes (Read, Glob, Grep, Agent, Skill, TaskUpdate, SendMessage, Write). Phase metadata names for task updates: "Investigating", "Validating", "Done". See agents/scout.md for full details.
Key Principles
- Evidence over opinion. Every finding needs a file path and line reference.
- Confidence is honest. Don't mark something High if you inferred it.
- Validation catches errors. When in doubt, validate. Fresh eyes find what you missed.
- No side effects. In fellowship, tool restrictions enforce this. Standalone, respect it by convention.