florianbuetow/codebasescout

Spawned by the scout-router to define the scoring formula for the run. Sets Impact x Opportunity as the score, the 1-5 rubric for each axis for this objective, the discard rule, and the priority matrix. Writes formula.md.

Spawned by scout-router as the final step. Tags each task in the map with a recommended model and a one-line rationale, matching task difficulty and scope to model capability -- cheap models for mechanical/localized tasks, premium models for cross-cutting or architectural ones. Produces the deliverable tasks.md.

Spawned by scout-router. Does the meta-level pattern analysis -- finds when findings share a deeper underlying issue that produces the same class of problem across the codebase -- and plans the ARCHITECTURAL fix that closes the whole class. Designs and plans only; it never edits or implements code. Writes groups.md.

Spawned by the scout-router to turn the abstract Impact and Opportunity axes into concrete, measurable signals for the chosen objective, and to name the data sources to pull. Ships presets for tech debt, dead code, conversion, bugs, security, and performance. Writes profile.md.

Spawned by the scout-router after all scorers finish. Merges every score-*.md slice, applies the Impact x Opportunity priority matrix, sorts the survivors, and classifies each finding's likely type so the task-writer can turn them into tasks. Writes ranked.md -- this is not the final deliverable; it feeds the task map.

The cheap-model sweep worker. Spawned in parallel by the scout-router, one instance per slice of files, to rate each file's Impact (1-5) and Opportunity (1-5) independently per the run's formula and profile, multiply, and discard 1s and 2s. This is the ready-to-use scoring prompt operationalized. Writes score-<slice>.md.

Scout a codebase and produce a ranked map from codebase to tasks, each tagged with a recommended model. Identifies work; does not fix it.

Use to scan a codebase and produce a ranked map from codebase to tasks, each tagged with a recommended model -- planning the fixes but never implementing them. Lead orchestrator for codebasescout. Guards the target, gates scan-vs-skip, runs the identify-and-score sweep, runs meta-level pattern analysis, plans the point and architectural fixes, maps findings into tasks with model recommendations, and assembles the deliverable tasks.md. Examples include "scout this repo for tech debt", "map this codebase into tasks with model recommendations", "where is the high-value work and which model should do each piece".

Spawned by the scout-router after the target is set. Decides whether scoring is worth it at all. Recommends skipping the sweep for very small repos or when the available signal is garbage (no git history, unreliable metrics), so the model is pointed directly instead. Writes scope.md.

Spawned first by the scout-router. Guards against aiming a model with no target ("fix everything"). Forces a single concrete objective, an explicit in-scope file set, and success criteria before any scoring happens. Writes target.md.

Spawned by scout-router. For each ranked finding, plans the POINT fix -- the localized approach that resolves the issue at that exact spot -- and flags findings that are really one instance of a class (whose architectural fix the pattern-grouper plans). Plans and points out the fix; it never edits or implements code. Writes tasks-draft.md.