threnody-fast-review

Threnody fast review

Purpose

Use this when a review spans many files and should not collapse into a tiny number of agents. This mode trades the deeper file × dimension review swarm for faster file-level parallelism:

wave 1: one read-only review agent per file
wave 2: one synthesis agent

This is the default for broad swarm review. Use threnody-swarm-review only when the user explicitly asks for deep review, security-critical audit, threat-modeling, or a named specialist dimension.

Tiering is medium by default. Ordinary risk words such as auth, token, or secret add security attention inside the file reviewer but do not automatically force high tier. High tier is reserved for explicit deep/security-critical wording, concrete exploit primitives, or large/dense files.

The global swarm.max_agents cap still applies. If the requested count is clamped, the largest and riskiest files are kept (ranked by exploit signals, then risk signals, then LOC x density; listed order is only the last tiebreak), and the rest are dropped. Report the requested_vs_effective_agent_count field and tell the operator which files were not reviewed.

Workflow

0. Resolve files

Use file paths from the user's message. If none are provided, resolve changed source files:

git diff --name-only HEAD
git diff --name-only --cached

Filter out generated files, lock files, binary assets, Markdown docs, and large schemas unless the user explicitly asks to review them.

1. Run cheap deterministic checks

Run relevant linters or type checks first when available. Keep failures as context for synthesis; do not stop the swarm solely because a linter found findings.

2. Call execute_swarm

Use the FAST_REVIEW: sentinel and request one agent per file plus one synthesis agent:

execute_swarm(
  task="FAST_REVIEW: <space-separated file paths>",
  topology="dag",
  max_agents=<number_of_files + 1>,
)

The sentinel activates fast file-level review in the heuristic planner.