# Triage Findings > Use when an agent needs to decide what to do with a static review's findings before any fix is planned; groups findings into shared causes, lists the repairs each cause admits and enumerates what each one touches, and records a per-finding disposition that carries into the next review round, rather than sequencing the fixes, editing code, or re-reviewing it. - Skill: `tacticaldoll/triage-findings` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add tacticaldoll/triage-findings` - Raw SKILL.md: https://api.skillmd.com/api/skills/tacticaldoll/triage-findings/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: tacticaldoll (https://skillmd.com/u/tacticaldoll) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/tacticaldoll/triage-findings --- # Triage review findings Use this skill to turn a list of review findings into decisions: what the causes are, which repairs each cause admits, and which findings are accepted, declined, or deferred. Governing intuition: **a finding names a place; a cause names the thing to change.** A review reports locations, so a repair that touches only the location leaves the cause standing and the same findings come back next round. Name the cause, then list the repairs it admits and **enumerate** what each one touches — a repair confined to the finding's own place is then visible as exactly that, with nothing to classify and no bucket to mis-assign. **Input**: a review findings list — a `static-review` Review Record, findings pasted inline, or a path to either — plus the prior round's Disposition Record when one exists, and the codebase the findings concern; if none is given, ask which findings to triage. **Boundary**: decides causes, repairs, and dispositions — produces a Disposition Record; does not sequence or apply the repairs, review the code or add findings of its own, or persist the record itself. ## Workflow ### Phase 0: Resolve the findings and the prior round Normalize the input into one flat list of findings, whatever shape it arrived in — a Review Record's gate tables, its non-gated tracks, a pasted list, or a file holding either. A section the producer declares carries no findings is not flattened into that list, and its rows are not counted against the record's stated total — a Review Record's `Findings` field names its own exclusions. Such a section is still evidence: read it for a cause, and record a mismatch between it and the findings as Record integrity rather than as a finding of its own. When a section's finding status is not declared at all, that is `not claimed` — not a claim that its rows are findings. Neither flatten nor count them, and record the absent declaration for Record integrity. This state is the concession to records this repository did not produce: `static-review` requires the declaration from its own producer, so a Review Record it wrote never reaches it. A Structural Causes row naming no finding is itself one. It states what is wrong and the unit that carries it, which is all Phase 1 needs to key it; the review reported it as a cause because the gate that would have carried it never opened, not because it is a lesser kind of thing. A row that does name a finding is that finding's stated cause, and Phase 2 verifies it like any other. Audit the Review Record before reading the code: reconcile its Verdict with the Gate Index, its stated finding count with the finding rows it actually contains, and its Coverage claim with the units and gates it enumerates. Preserve each mismatch for Record integrity; do not silently repair the input by reinterpreting a count, scope, or status. Then load the prior round: - When a prior Disposition Record exists, every finding it declined or deferred **keeps that disposition and its recorded reason**. Reopen one only on new evidence, and name the evidence. - Record **this round's scope** — the files, diff, or component the input covers. Membership is decidable **only when the input enumerates what it covered**. When it states a scope without enumerating it, a prior disposition's membership is `undetermined` — record it as that rather than deciding it. Undetermined is neither carried nor closed. Re-evaluate it from the next round's input: a matching re-report enters the ordinary finding flow; newly enumerated scope decides inside versus outside; another unenumerated scope leaves it undetermined. - Against an **enumerated** coverage, compare each prior disposition to it. One that lies **inside** it and that the input did not re-report is a closure candidate, not automatically carried or closed. One that lies **outside** it is `out of scope this round`: this round looked elsewhere and says nothing about it. These slots are in the record (Phase 5); a scope change is not a closure. - When none exists, record `first round`. Do not infer a prior decision from the code. ### Phase 1: Give each finding a stable identity Key each finding by **what it violates plus the unit that carries it** — not by `file:line`. The unit is the smallest named code or document element that owns the violated contract (function, type, module, manifest, section); use the file only when no smaller stable owner exists. Preserve the logical unit when code moves or is renamed and the same contract remains. A line number drifts with the next edit, so a line-keyed finding cannot be matched across rounds, and a finding that cannot be matched cannot be closed. One defect often arrives more than once: as a gate finding, again as a contract row, again as a refuted claim. Findings that key alike are **one** finding — record every source id against it rather than triaging the same defect several times. Match every finding against the prior record before treating it as new. Findings in the same unit that violate different contracts remain different identities. When one reappears, compare the recorded repair and its Reach with what actually changed: - **No recorded Reach changed** — the accepted repair never landed; a repair nobody attempted is no evidence about anything. Keep the **disposition and its reason**, and re-derive the repairs in Phase 3. A disposition is a decision, so freezing it is what stops re-litigation; a repair is a reading of the cause, so freezing it would preserve a misreading the code never justified. - **Only part of the recorded Reach changed** — the repair did not fully land. Keep the disposition and reason, record the changed and untouched locations, and do not call it recurring. - **The complete recorded Reach changed** — the repair landed and the finding survived. Record it as recurring, verify whether the original cause still holds, and re-derive the repairs in Phase 3; either the Reach was incomplete or the stated cause was wrong. For each closure candidate — a prior finding inside an enumerated scope that the input did not re-report — verify all of the following before closing it: 1. The input `gate-reviewed` the unit, or lists it as `partially-gate-reviewed` with the relevant gate among those opened, or explicitly verifies the repair. A `triage-only`, `unread`, or partial unit whose opened gates omit the relevant gate cannot close a finding by silence. 2. The code at the recorded cause shows that the cause no longer holds. 3. The input does not contradict the closure elsewhere in its record. When all three hold, record the finding as `closed` with the code and input evidence. When the cause still holds, or any closure evidence is unavailable, carry the prior disposition and original reason forward; say which closure condition was not established. Review silence alone is never closure. Condition 1 has three arms and the third is not a lesser one. A repair a round can re-run — revert the unit the finding names, watch a named test fail, restore — is stronger evidence than a reading of that unit, and where a repository keeps such probes as a ledger, running one satisfies condition 1 outright. What silence never does is stand in for any of the three. **A finding whose unit is prose closes on the reading that settled it.** A docstring's wording, a registry entry, a record's own cell: no test holds them and no gate opens over them, so requiring a gate-reviewed unit would carry them forever, which reports the same rows every round and retires nothing. For these, condition 1 is met by the recorded reading — the Disposition Record's own statement of what was read and what it said — and conditions 2 and 3 are unchanged. Carrying a prose finding is for when the reading has not been done, not for when no gate can be opened. ### Phase 2: Group findings into causes For each finding ask: **what one change would remove it?** Findings that answer the same way share a cause. - A cause is not a category. "Naming" is a gate; "the module exposes two responsibilities under one name" is a cause. - A cause is stated as the thing to change, not as the symptom that revealed it. - **Verify the cause against the code.** A review states a cause as well as a location, and the stated cause can be wrong while the finding is real — a wrong cause points at a wrong repair. Check it, and when it does not hold, record the cause the code supports and say what the review claimed. - A finding whose cause cannot be named from the code, or from the documents the code answers to, stays **ungrouped**. Do not invent a cause to make the table tidy — an invented cause produces a repair that fixes nothing. ### Phase 3: List the repairs each cause admits A cause may admit more than one repair, and they are not interchangeable — one may correct prose where another changes behaviour. **List every repair the cause admits**, not the one you would pick: a cause recorded with a single repair hides the choice from whoever plans the work. Each repair carries two things. **Reach** — every location the repair touches, enumerated as a file plus the unit inside it: the name of the function, type, or section that owns the change, or a quoted phrase where no name owns the text. Enumerate; do not classify. Whether a repair is confined to the finding's own place, spans several, or dissolves them is then read off the list itself. Not `file:line`, for the reason Phase 1 gives about keying a finding: a line number drifts with the next edit to that file, so a Reach keyed to one cannot be checked once the record is a round old, and a record's Reach is read by every round after it. A symbol survives an unrelated edit and can be verified against the file; a line number can only be maintained by hand, and by the time it is wrong nothing says so. **Kind** — what the repair does: | Kind | The repair | |---|---| | `guard` | add the missing check where it is missing | | `derive` | replace a written-down set with one computed from the source it must cover | | `index` | replace a hand-written scan or matcher with lookup through an existing key or ordering capability | | `converge` | delete one complete implementation of an operation and call the other implementation that already performs it | | `distinguish` | replace an undifferentiated error or result boundary with variants a caller can tell apart | | `restate` | correct prose, a spec, or a document | | `declare` | record a limit or a count as governed policy | | `forbid` | refuse an input that was admitted | These eight are the values this skill's own use has produced, from one repository. The list is **not closed**: a repair none of them fits is named in its own words and reported as a gap, never forced into the nearest value. A cause whose repair is a decomposition of a unit, or a re-drawn seam between units, is outside this vocabulary — do not invent a Kind for it; hand the cause on (Rules). A cause whose repairs cannot be named from what the review reached has **none listed** — say so, and say what would produce one. A cause the review reached through a blocked or unopened gate is the common case. When one repair's reason disappears once another lands, note which voids which. A repair that a sibling makes pointless is not a second task. ### Phase 4: Disposition each finding Decide per finding, not per gate, and never accept a batch because a review produced it. - `accept` — the cause is real and at least one of its repairs is worth doing. - `decline` — the finding does not hold for this codebase, or every repair costs more than the defect. - `defer` — real, but not now; name what would make it now. Every `decline` and `defer` carries a reason. The reason is the only thing that stops the finding returning next round, so a disposition without one has not actually been made. ### Phase 5: Produce the Disposition Record ```markdown ## Disposition Record **Source**: [the review record or list triaged] **Scope**: [the files, diff, or component this round's input covers] **Prior round**: [the record carried forward | first round] ### Record integrity | Check | Input claim | Reconciled evidence | Result | |---|---|---|---| | Verdict / Gate Index | the gate the verdict names | that gate's recorded status | pass \| mismatch \| not claimed | | Calibration / Gate Index | the gates the calibration declares | the gates the index records as opened | pass \| mismatch \| not claimed | | Finding count | stated count | current Review Record finding rows | pass \| mismatch \| not claimed | | Coverage | stated scope and coverage | four enumerated unit sets and the gates opened for each partial unit | pass \| mismatch \| not claimed | | Non-finding sections | the sections the `Findings` field excludes from the count | the row-bearing sections the record contains, whether each declares its finding status, and the section's rows against the gate findings — a row stating what a gate declares a finding, or a finding naming a unit the section omits, surfaces here | pass \| mismatch \| not claimed | [`not claimed` when the input carries no such claim at all — a record predating a field, or one from another producer, has nothing to reconcile and is neither consistent nor contradictory about it. Do not read a missing claim as a passing one. Code defects the review missed do not belong here — send those back for review. A Result may carry a **qualifier**: one of the three values, then `, `, then what the qualifier says — `mismatch, stated by the input`. The three values are the domain and the qualifier says why; there is no fourth value, and a Result that merely begins with one of the three is not one of them.] ### Prior scope resolution | Prior finding | Scope evidence stated by this round's input | Membership | This round's slot | |---|---|---|---| | id | four enumerated unit sets \| scope stated but units not enumerated | inside \| outside \| undetermined | current finding → Dispositions \| closure candidate → Closed or Carried forward \| out of scope → Out of scope this round \| undetermined → Undetermined | [Use only scope evidence the input states. Finding locations and code inspected during triage do not enumerate the Review Record's coverage.] ### Causes and candidate repairs | # | Cause (the thing to change) | Findings | Repair | Kind | Reach (every location it touches) | Route | |---|---|---|---|---|---|---| | 1 | stated as the change, not the symptom | every source id | 1a | listed Kind \| ` (gap)` | `file` + `symbol`, … | where it goes | | 1 | " | " | 1b | the alternative, when the cause admits one | `file` + `symbol`, … | " | ### Pattern [Two or more causes sharing one shape: the shape, which causes hold it, and a pattern-level repair that catches the next instance — repairing the instances does not remove the shape. `none` when no shape recurs.] ### Coupling [Repairs whose reason disappears once another lands, as `1b voided by 3a`. `none` when independent.] ### Dispositions [Every finding reported by this round, including a stable id matched to prior history. A prior finding not re-reported belongs in exactly one exclusive lifecycle home below — Carried forward, Closed, Out of scope this round, or Undetermined — and is not duplicated here. Each slot from Prior scope resolution has exactly one of those sections as its home; a `closure candidate` resolves into Closed or Carried forward by the Phase 1 test. A finding re-reported after it was Closed returns here as `carried`; reappearance alone does not make it Recurring.] | Finding | Cause | Carried | Disposition | Reason (REQUIRED for decline and defer) | |---|---|---|---|---| | id | # \| ungrouped | new \| carried | accept \| decline \| defer | why — not "as reviewed" | The `Finding` cell opens with the stable id and nothing else: capitals, digits and hyphens. What may follow it is one of two forms and no third — ` — ` and a description, or `, alternative ` and the repair the row is about. The id is what a later round matches on, so a cell that runs the id into anything else has no id, and a reader taking the text before a separator it guessed at would make two rows keyed alike look distinct. ### Carried forward [Prior dispositions lying inside this round's scope that the input did not re-report: id, disposition, the original reason unchanged, and which closure condition was not established. A disposition leaves this section only when the closure test in Phase 1 proves its cause no longer holds, never because the review stopped reporting it. `first round` when there is no prior record.] ### Closed [Prior dispositions whose closure test passed: id, prior disposition, code evidence that the cause no longer holds, and the input evidence that the relevant unit was reviewed without re-reporting the finding. Preserve the stable finding id so a later re-report can match the closed history. `none` when no closure is established.] ### Out of scope this round [Prior dispositions the input's **enumerated** coverage decidably excludes: this round looked elsewhere and says nothing about them. Neither carried nor closed. Their ids and count.] ### Undetermined [Prior dispositions whose membership this round cannot decide, because the input states a scope without enumerating what it covered. Not out of scope — out of scope is a decision, and this is the absence of one. Neither carried nor closed. Their ids, count, and the enumeration that would settle each. Re-evaluated from the next round's input (Phase 0). This is a **compatibility state**, not a resting place: a `static-review` Review Record enumerates its coverage, so this section is `none` for one. It exists for a record from another producer, or from before that field. A round that keeps filling it is reading records that cannot support cross-round state at all — say so.] ### Recurring [Findings that came back after every location in a recorded repair's Reach changed: the stable id, the prior repair and Reach, whether its cause still holds, and the newly derived repairs. This is additional analysis for a current Dispositions row, not a mutually exclusive lifecycle slot. A finding that was previously Closed is not Recurring merely because it appears in a later round.] ### Ungrouped [findings whose cause could not be named, and what would settle each] ### Self-check | Check | This record's answer | |---|---| | Every prior id sits in exactly one exclusive lifecycle home | pass \| the ids in more than one, or in none | | Every accepted cause carries at least one repair with an enumerated Reach | pass \| the causes missing one | [This is about the record you are producing; Record integrity above is about the input. Completed last, because the sections it checks are written before it.] ``` ## Rules - Decide only. Do not edit code, sequence the work, or put the record anywhere but your reply — no file, no published page, no external destination, however durable the caller needs it to be. When the prior record arrived as a path, that file is read-only input. Losing the record is a real cost and not yours to solve: say where it should be kept, and let the caller keep it. - Triage only what the input reports. A defect in the code that the review missed is not yours to add — say so plainly and send the code back for review. A defect in the **record** is different: it is the input's own fitness, and you are its consumer, so it goes in Record integrity (Phase 5). Keep that separate from Self-check, which is the same question asked of the record you are writing. - List every repair a cause admits. Recording one where the cause admits two makes a choice that belongs to whoever plans the work, and hides it as if there were nothing to choose. - Stay in lane; hand off at the boundary. Route each accepted repair by its Kind: a `restate` has no code handoff — name the document and who owns it, and stop there; for every other Kind, hand off to `plan-implementation`. Most repairs go there, and saying so is more honest than a routing table that reads richer than the work. Route a named Kind gap by the action it actually performs: a code repair goes to `plan-implementation` unless one of the structural cases below applies. When a cause's repair decomposes one unit, hand off to `plan-split`; when it redraws the seam between units, hand off to `design-boundaries`. When a cause cannot be named without tracing the fault, hand off to `diagnose-issue`; when settling it needs runtime evidence rather than static reading, hand off to `plan-testing`. For a fresh review of the code itself, route to `static-review`. Name the handoff rather than half-doing the other workflow's job.