triage

name	triage
user-invocable	false
description	Internal drydry classifier that buckets duplication findings by convergence cost.
allowed-tools	["Read","Grep","Glob","Bash(rg *)","Bash(grep *)","Bash(find *)"]
effort	medium

Triage

The three-bucket triage pass behind the drydry audit pipeline. Takes a findings list (output of drydry:sweep) and classifies each finding into cheap-and-safe / partial / needs-design with a one-line convergence cost. Not user-invocable.

Input contract

Caller supplies through args:

• findings: the markdown structure returned by drydry:sweep, either inline or as a path to a markdown file.
• scope: the directory or package the findings were collected over. Used to estimate convergence cost (count of call-sites, presence of public exports).

Output contract

Return the findings list, with one triage: and one cost: line added per finding:

## Pattern: <pattern_id>

- Finding 1
  - file_a: `<path>:<line>`
  - file_b: `<path>:<line>`
  - drift_hypothesis: <as-is from sweep>
  - verifier_command: <as-is from sweep>
  - verifier_expected: <as-is from sweep>
  - triage: cheap-and-safe | partial | needs-design   (the by-design case is a sub-class of needs-design; the bucket label stays needs-design)
  - cost: <one-line convergence cost: file count touched, public-API surface affected, migration complexity>

When a finding is classified as needs-design, also include a design_question: line capturing the open decision (which side wins, what to name the unified abstraction, how callers migrate). When partial, include a next_step: line naming the smallest action that would unblock convergence.

The three buckets (Chapter 5)

cheap-and-safe

There is an obvious convergence direction. The change is local. No public contract moves. Both call-sites are caller-owned; no external consumers.

Signals:

• Both files are within the same package or module
• The unified shape is obvious (factory, helper, partial, shared component)
• Touched file count is small (typically <=5 within the audit scope)
• No public method or exported symbol changes name or signature

Convergence path: apply the unified abstraction in the same audit. Both paths migrate to the new shape; the diff is local; reviewers see the convergence at a glance.

partial

Convergence requires design judgement, but the judgement is small and the operator can resolve it inside the audit if they make one decision. Typical pattern: the unification is clear except for one open question.

Signals:

• The two paths agree on what they do but disagree on a name, a parameter order, or which side owns the shared abstraction
• A public symbol moves but the rename is mechanical (one find-and-replace, no client-side reasoning)
• Touched file count is moderate (5 to 20)
• The operator can answer the design question in one short conversation

Convergence path: surface the design question in the findings artefact under design_question:. When the operator answers, the finding promotes to cheap-and-safe.

needs-design

The finding looks like a duplicate from outside, but a real constraint or genuine domain difference makes convergence either non-trivial or wrong. Document why so future audits do not re-flag it (Chapter 7's contrarian second-pass gates this verdict). When the blocking constraint is a framework or external API ("by-design" in informal speech), the verdict and the artefact entry are still needs-design; the constraint becomes the body of the design_question: line.

Signals:

• A framework constraint forces the parallelism (e.g., AppIntent's perform() signature requires concrete types; type erasure costs more than the duplication)
• The two paths share surface but encode genuinely different domain rules (a "user" in onboarding and a "user" in billing may be the same word and different concepts)
• The unification would break a public contract with downstream consumers outside the scope
• The unification would require a multi-quarter migration the operator has no appetite for

Convergence path: document the verdict and the constraint in the findings artefact. The orchestrator's audit-mode step 5 runs an independent contrarian pass to verify the verdict is not hollow (Chapter 7).

Workflow

1. Read the findings. Parse the markdown structure from args.findings.
2. Read the scope context. For each finding, briefly inspect file_a and file_b to count: (a) call-sites of the duplicated function or component, (b) whether either path crosses a package or module boundary, (c) whether either path is exported (public API surface).
3. Classify. Apply the bucket signals above. When multiple signals fire across buckets, the most conservative wins: cheap-and-safe requires every signal to point to local; partial requires the design question to be small and isolatable; needs-design otherwise.
4. Estimate cost. Write one line per finding: <N> files touched, <M> public symbols moved, <complexity tag>. Complexity tags: mechanical (rename only), local-refactor (extract method or component), interface-migration (public contract changes), framework-bound (constrained by an external API shape).
5. Add design questions and next-steps. For partial, name the next-step in one sentence. For needs-design, name the design question.
6. Return. Hand back the annotated findings markdown to the caller.

Rules

• The bucket is the verdict; the cost is the evidence for the verdict. A cheap-and-safe classification with a cost of "30 files touched, interface migration" is contradictory; promote it to partial or needs-design.
• No fourth bucket. Chapter 5 is explicit: three buckets, no later, no backlog, no polish next sprint. Either fix it (cheap-and-safe), park the design question (partial), or document the by-design constraint (needs-design).
• Triage is not approval. This skill does not decide whether the operator will fix the finding; it decides how much room the convergence needs. The caller acts on the triage.