---
name: triage-backlog
description: Read-only LCM-X issue and pull-request triage for one report or duplicate cluster at a time. Use when assessing current-main applicability, P0-P4 priority, reproduction quality, duplicate relationships, linked PRs, ownership, Electric Sheep impact, or the next evidence gate without changing GitHub.
---

# Triage Backlog

Triage one issue, PR, or duplicate cluster into a bounded evidence packet. Treat GitHub as the
live source and make no writes unless a maintainer separately authorizes an exact mutation.

## Minimum Capability

Read-only triage requires live access to the repository, current refs, issues, pull requests,
linked commits, and authorization state. Stop if any required read is unavailable. A
maintainer-authorized mutation and its read-back additionally require the corresponding GitHub
write capability.

## 1. Establish Identity

Record:

- repository `electricsheephq/lcm-x`;
- exact current `main` SHA;
- item number, URL, state, author, timestamps, labels, and linked work;
- whether the item came from LCM-X, upstream `stephenschoettler/hermes-lcm`, or another source.

Stop on the wrong repository, unavailable live state, credentials in evidence, or a request that
would require customer/runtime mutation.

## 2. Search Before Classifying

Search open and closed LCM-X issues and PRs using the title, error text, component, invariant,
upstream number, and linked commits. Read plausible matches before naming a duplicate.

For a duplicate cluster:

- select the earliest or best-evidenced active item as canonical;
- preserve every original URL, author, and upstream link;
- distinguish the same root cause from merely similar symptoms;
- do not close or comment on any item.

## 3. Verify Current Applicability

For a reported bug, require either:

- reproduction on the exact current `main` SHA; or
- evidence that current code violates a named mandatory invariant in `AGENTS.md`.

Record the supported path, expected result, observed result, exact command or artifact, affected
SHA, frequency/impact, workaround, and whether data is missing, duplicated, reordered,
misattributed, or visible across sessions/profiles.

Use `needs-repro` when evidence is insufficient. Labels, CI, reviewer severity, an upstream claim,
or an unmerged PR are leads—not proof.

For a linked PR, inspect its complete diff, exact head/base, tests, checks, review history, issue
coverage, dependencies, overlap with other PRs, and any behavior it introduces. Do not call it
merge-ready. Hand readiness or an explicitly authorized merge decision to `land-pr`; that
handoff grants neither approval nor source-push authority.

## 4. Classify And Prioritize

Use one primary classification:

- `bug`, `security`, `data-integrity`, `performance`, `feature`, `docs`, `test`,
  `best-practice`, `nit`, `needs-repro`, `duplicate`, `superseded`, or `out-of-scope`.

Assign:

- `P0`: active catastrophic data loss, disclosure, or unusable core path;
- `P1`: verified severe integrity, security, crash, lockout, or release-blocking failure;
- `P2`: significant bounded correctness or resource failure;
- `P3`: limited bug, feature gap, or moderate maintainability/performance issue;
- `P4`: documentation, tooling, portability, testing, or cosmetic work.

Also record:

- confidence: `verified-current`, `likely`, `needs-repro`, `not-a-bug`, or `superseded`;
- LCM-X/Electric Sheep impact: `direct`, `possible`, or `none`;
- source lifecycle: `active-continuation`, `archive-record`, `local-only`, or `unknown`;
- continuation target: the live LCM-X or upstream issue/PR URL, or `none`;
- disposition: `migrate`, `reference-only`, `needs-repro`, `duplicate`, `superseded`,
  `defer`, or `do-not-migrate`;
- accountable owner and the next evidence or decision gate.

Nits default to P4. Best-practice work remains P3/P4 unless current evidence proves a violated
mandatory invariant.

Treat an open executable upstream item or continuing upstream discussion as an
`active-continuation`. Do not silently turn it into an archive-only record, mark it superseded,
recommend `do-not-migrate`, or recommend closing it. If current source state or its continuation
target is ambiguous, return `OWNER_GATE` and preserve the item unchanged.

## 5. Return One Bounded Packet

Return:

```text
Item:
Exact main SHA:
Classification / priority:
Confidence:
Current-main applicability:
Supported-path impact:
Evidence:
Canonical duplicate / related work:
Linked PR assessment:
Electric Sheep impact:
Source lifecycle:
Continuation target:
Owner:
Recommended disposition:
Exact next gate:
Proof boundary:
```

Deduplicate findings while retaining every source URL and contributor attribution. State clearly
whether a failure is reproduced, inferred, or still hypothetical.

## Write Boundary

Remain read-only by default. Do not comment, label, assign, close, reopen, create issues, edit
milestones, approve, push, or merge.

Invoking this skill never authorizes a mutation. Approval and source pushes are outside this
skill and require separately authorized workflows. Use `land-pr` only for a separately requested
readiness or explicitly authorized merge decision; neither invocation nor handoff creates
approval, source-push, or merge authority.

Routine reversible metadata means labels, assignees, and milestones. One exact current
maintainer authorization plus the live-state checks and read-back below is sufficient; a routine
metadata change does not require a second code-owner approval solely because the item is
classified as security or data-integrity work.

Public security or data-integrity disclosure and every close or reopen are sensitive or terminal
actions. Require current non-author human code-owner approval, confirm the exact public content or
lifecycle transition, and stop if an active continuation would be archived or closed. Route
private vulnerability handling outside this public triage workflow.

If a maintainer explicitly authorizes one of the remaining GitHub metadata mutations:

1. name the exact item and requested change;
2. re-fetch current item, repository, and authorization state;
3. apply the routine-metadata or sensitive/terminal gate above and record any required
   non-author human code-owner approval;
4. stop on any other drift or ambiguity;
5. perform only the named write;
6. read back the result.

AI output is proposal and evidence. Deterministic live-state checks must guard every authorized
write. Never process the whole backlog, manufacture issues from hypotheses, or post one comment
per item when a bounded summary will do.