Clear your findings backlog — without handing the AI the pen.
Dethereal Triage is a Claude Code plugin that turns the findings the platform computed into a clean set of decisions. It groups identical findings, has an AI analyst propose a grounded verdict for each, and walks you through a fast review loop. Nothing reaches your model until you approve the batch at one gate.
> /dethereal-triage:triage ./payments-api 34 findings · 32 pending → 9 decision units · evidence: full # exposure (group) verdict conf. 1 Unencrypted data flow keep high 2 Missing rate limit set aside medium 3 Weak auth on admin path supersede medium 4 Overbroad IAM role needs you — → nothing is written until you approve at the gate
A triage run never changes anything on its own. It groups your findings, proposes a grounded verdict for each, then stops and waits. Every verb you type edits a plan in memory; the model is untouched until you approve the whole batch at a single gate. Walk away mid-run and nothing has happened.
Accept a proposal and the analyst’s reasoning is recorded under your approval; override it and your verdict and your words go in instead. Either way the decision is stamped, so months later a reviewer can see the human behind every muted finding.
A whole backlog,
one reviewable loop.
One command runs the whole thing from your terminal. It reuses your Dethereal sign-in, so there’s no separate login — just point it at a model and work the proposals at your own pace.
- 01
Run it at your model
/dethereal-triage:triagePoint it at a model folder — no model ID to look up. The run reads the findings the platform already computed, prints a header stating how much evidence it can see, and a lifecycle summary of every finding by state. Only pending, platform-generated findings enter the loop; your own authored findings are left alone.
claude — dethereal-triage> /dethereal-triage:triage ./payments-api Evidence mode: full · knowledge graph: available ✓ 34 findings · 32 pending to triage → grouping into decision units…
- 02
Decide once for identical findings
grouping + evidenceA model can raise the same gap on a row of similar services. Rather than making you answer it over and over, the run groups identical findings into one decision unit, gathers the evidence behind each, and — if members turn out to differ — splits the group and tells you it did. One group, one decision, applied to every member.
claude — dethereal-triage⠋ gathering evidence per group… ✓ 32 findings → 9 decision units → group 3 split: one element differs (3#1, 3#2)
- 03
Read the proposals, make the calls
accept · dispose · supersede · skipAn AI analyst proposes a verdict per group — keep it, set it aside with a reason, or replace it with a corrected version — each with a confidence and one-line rationale you can expand. You work the list one verb at a time: accept the proposal, override it with your own verdict, split it, or skip it. Every verb edits a plan in memory; nothing touches your model yet.
claude — dethereal-triage> accept 1 Group 1 accepted — dispose / NOT_APPLICABLE > dispose 4 COMPENSATING_CONTROL Group 4 — control recorded: mesh sidecar TLS > skip 7 Group 7 skipped (stays pending)
- 04
Approve once — the only write
done → Apply N dispositions?Type done and the run compiles your decisions into a concrete list of operations, itemized by type. That list is the gate: one question, Apply N dispositions?, and only yes writes anything. The summary reports every outcome — applied, invalid, failed — and every disposition is stamped with who decided it.
claude — dethereal-triage> done Compiled 18 ops · Apply 18 dispositions? [yes/cancel] > yes ✓ applied 15 · 1 invalid · 2 failed (nothing partial) → 9 affirmed · 6 disposed · 3 superseded, each stamped
Built so a serious mute
can never slip through.
The speed is in the routine decisions. The guardrails make sure the decisions that matter never ride in on a shortcut.
Skip is the default
Anything you don’t act on stays pending. Silence is never consent — the run never mutes a finding because you missed it.
One write, one gate
Every verb edits an in-memory plan. Nothing reaches your model until you answer “Apply N dispositions?” with yes.
High-severity mutes held out
A proposal to set aside a High or Critical finding is pulled into its own section and locked out of every bulk action — decide each one yourself.
Needs-human rows wait
When evidence is thin or two reviewers conflict, the AI proposes nothing — it hands you both views and waits for your explicit call.
Confidence with a ceiling
Run without your local model files and every proposal is capped at medium — no high-confidence call ever rests on evidence the run couldn’t verify.
An audit marker on every decision
Each applied disposition is stamped operator-approved or operator-edited, so a later reviewer sees whether a human ratified the AI or made the call.
Load the plugin. Work your findings.
Dethereal Triage is a paid Claude Code plugin that loads alongside Dethereal and reuses its sign-in — no separate login. Point it at a model and start a run just to see what it finds; nothing writes until you say so.