Roam — Codebase Comprehension Skill

Repository: <https://github.com/Cranot/roam-code>;

Roam pre-indexes codebases into a semantic graph (symbols, dependencies, call graphs, architecture layers, git history) stored in a local SQLite DB. Query it via CLI instead of repeatedly grepping files and guessing structure.

Setup

Ensure roam-code is installed and the project is indexed:

pip install roam-code   # or: pipx install roam-code
cd <project-root>
roam init               # indexes codebase, creates .roam/index.db

After git pull or major changes, run roam index to refresh (incremental, near-instant if few files changed). After large refactors: roam index --force.

Command Decision Table

Use this table to pick the right command for the situation:

Core Workflow

1. Orientation (first time in a repo)

roam understand        # tech stack, architecture, health, conventions
roam tour              # onboarding: key symbols, reading order, entry points
roam map               # project skeleton with top symbols by PageRank

2. Before Making Changes

Always run roam preflight <symbol> before modifying code. It combines blast radius + affected tests + complexity + coupling + fitness into one check:

roam preflight MyClass
# Output: blast radius, affected tests, complexity, coupling, fitness verdict

If you only need files to read:

roam context MyClass
# Output: definition file + callers + callees with exact line ranges

3. After Making Changes

roam diff              # blast radius of uncommitted changes
roam diff --staged     # blast radius of staged changes
roam pr-risk           # risk score (0-100) + suggested reviewers

4. Debugging

roam diagnose <symbol>  # root cause ranking by z-score risk
roam trace <A> <B>      # dependency path between two symbols
roam effects <symbol>   # DB writes, network I/O, filesystem, global mutation

Output Modes

• Default: Human-readable text (also optimized for LLM consumption)
• roam --json <cmd>: Structured JSON with consistent envelope
• roam --budget N <cmd>: Token-capped output (N = max tokens)
• roam complexity --select '.symbols[:5]': Project one bounded JSON field without shell parsing (--select is repeatable)
• roam grep <pattern> -C 5: Search source and return bounded live code packets in one call
• roam grep <pattern> --whole-symbol --max-packets 3: Deduplicate hits into enclosing functions/classes; stale spans fall back safely
• roam --sarif <cmd>: SARIF 2.1.0 for CI integration

Prefer --json when you need to parse output programmatically. Prefer --budget 2000 when context window is tight. Prefer --select '.summary' or a bounded list slice when only one envelope section is needed.

Key Commands Reference

roam search <pattern>

Find symbols by name (regex). Results ranked by PageRank.

roam search "Auth.*Service"
roam search "handle_request" --kind fn

roam context <symbol>

AI-optimized file list with line ranges for reading. Supports --task modify|debug|review for context tuning.

roam context Flask
roam context myfile:MyFunction    # disambiguate with file prefix

roam preflight <symbol|file>

Compound pre-change check. Run this before every modification.

roam preflight UserController
roam preflight src/auth/login.py

roam health

Composite score (0-100). Use --gate for CI (reads .roam-gates.yml).

roam health
roam health --gate               # exit 5 on failure

roam diff

Blast radius of uncommitted or committed changes.

roam diff                       # uncommitted
roam diff --staged              # staged only
roam diff HEAD~3..HEAD          # commit range

roam algo

Detect algorithm anti-patterns (23 patterns: O(n^2) loops, N+1 queries, quadratic string building, etc.) with confidence levels and fix suggestions.

roam algo
roam algo --confidence high     # high-confidence only
roam algo --task nested-lookup  # specific pattern

roam impact <symbol>

Full blast radius using Personalized PageRank.

roam impact Flask

roam symbol <name>

Symbol definition + callers + callees + metrics.

roam symbol open_db
roam symbol --full open_db      # include source code

roam affected-tests <symbol|file>

Trace reverse call graph to find covering tests.

roam affected-tests UserService

roam agent-export --write

Auto-generate agent instructions for the project. Detects CLAUDE.md, AGENTS.md, .cursor/rules, etc.

roam agent-export --write
roam agent-export --brief           # compact top-level summary

roam minimap --update

Inject/refresh annotated codebase snapshot in CLAUDE.md.

roam minimap --update           # update sentinel block in CLAUDE.md

Discovering More Commands

This skill covers the most common commands, but roam has 287 commands. To explore what's available:

roam --help                 # list all available commands
roam <command> --help       # detailed usage for a specific command

For full documentation, examples, and the latest features, see the roam-code repository.

Tips

• One roam command replaces 5-10 grep/read cycles. Always try roam first.
• Use roam search instead of grep/glob for finding symbols — it understands

definitions vs. usage and ranks by importance.

• roam context gives exact line ranges — more precise than reading whole files.
• After git pull, run roam index to keep the graph fresh.
• For disambiguation, use file:symbol syntax: roam symbol myfile:MyClass.

"Find every reference to X" — prefer roam refs over multi-shape grep

A common agent pattern is:

Grep(pattern: "->ekfpa|\.ekfpa\b|'ekfpa'|\"ekfpa\"")

The intent is "find all references to symbol ekfpa". The multi-shape regex tries to catch function calls (->ekfpa), attribute access (.ekfpa), and string literal mentions ('ekfpa', "ekfpa"). It works, but:

• False positives. Matches comments, docstrings, and unrelated string

literals — the agent must filter those out manually.

• Unstructured. Returns raw line text; the agent has to parse to learn

which symbol owns each match.

• Misses one-off shapes. Ruby obj.ekfpa!, Python f-string f"{ekfpa}",

decorator @ekfpa, etc. each need another regex shape.

roam refs <symbol> (alias for roam uses) walks the indexed call/import/ inherit graph and returns only real references — every result is a named symbol with kind/file/line, grouped by edge type:

$ roam refs find_symbol
VERDICT: 'find_symbol': 50 production consumers, 7 test consumers in 27 files

-- Called by (30) --
fn  affected_tests   src/roam/commands/cmd_affected_tests.py:230  production
fn  annotate         src/roam/commands/cmd_annotate.py:21         production
…
-- Imported by (20) --
…

Latency on this repo: ~700ms via subprocess, sub-100ms via the MCP server (`roam_uses` tool, which agents in MCP-enabled clients should prefer because the import-cost is paid once at server start). 2-5× slower than ripgrep in raw wall-time, but the result is already usable — there's no follow-up "now read the matching files and figure out the structure" step.

Use roam refs when:

• Finding every caller / importer / inheritor of a symbol.
• Planning an API change ("what would break if I rename X").
• Sweeping for usages before a deletion.

Stay with grep / Grep when:

• The target is a literal string, not a symbol (e.g. an error message,

a translation key, a SQL fragment).

• The codebase isn't indexed yet (roam init first if you'll be doing

many of these queries).