ForgeFlowForgeFlow / Reference
Review Routing
ForgeFlow classifies changes before spawning review agents. Start with /review in Claude Code or $forge-review in Codex. The route determines specialist coverage; every review does not need the full team.
Modes
| Mode | Use Case |
|---|---|
| skip-mode | No changed files, or documentation-only changes of at most 200 changed lines outside source paths. |
| thin-mode | Test-only changes, or at most two low-risk non-frontend files and 50 changed lines. |
| full-mode | Standard multi-agent review. |
| deep-mode | Auth, security, migrations, schemas, permissions, crypto, or broad high-risk changes. |
Route Helper
scripts/forgeflow/explain-review-route.js --json
With explicit files:
scripts/forgeflow/explain-review-route.js --json --files changed-files.txt --lines 120
When callers precompute line counts, they can also pass source detail:
scripts/forgeflow/explain-review-route.js --json --files changed-files.txt --lines 120 --tracked-lines 90 --untracked-lines 30
The JSON includes lines_changed, tracked_lines, and untracked_lines so routing decisions can explain whether new untracked files affected the selected mode.
With calibration:
scripts/forgeflow/explain-review-route.js --json --calibration .forgeflow/<project>/calibration-summary.json
Verifier Verification
High-risk or historically noisy findings can be sent through Verifier before they become blockers.
Verifier returns:
CONFIRMEDREJECTEDBLOCKED
The goal is not to suppress specialists. The goal is to separate useful attention from final judgment.
Local Context Before Review
Review mode can prepare bounded context before agents are spawned:
scripts/forgeflow/build-context-pack.js --json
scripts/forgeflow/check-context-budget.js --root .forgeflow --warn-only --json
scripts/forgeflow/advise-context.js --root .forgeflow --record --json
The generated reviewer packets give specialists a focused view of the change. The context advisor reports low-savings packets, budget violations, and trend deltas from previous runs so the workflow can trim context before spending tokens.
Interpret The Result
High-risk non-test paths take precedence over documentation and small-change rules. A large documentation change can receive a full review. Explicit mode overrides and CI policy can change the route; read the helper's reasons rather than assuming a file extension guarantees a skip.
A routing skip is not an approval from specialists. Review history records the final workflow verdict for shipping; dashboard verdict events and triaged review-outcomes.jsonl records serve different purposes. Save real review evidence before recording verdicts, and record confirmed/rejected outcomes only after triage. Empty charts on a fresh installation are expected. See Dashboard and Telemetry Readiness.