| name | silk-triage |
| description | Validate, deduplicate, specify, size, and prioritize Silk maintenance issues from the Linear Backlog. Use only when Julia explicitly asks to triage Silk issues or invokes $silk-triage. |
Triage Silk issues
Read ../../silk-manager/WORKFLOW.md, ../../silk-manager/LINEAR.md, and
../../silk-manager/TRIAGE.md completely before acting. Follow TRIAGE.md's required investigator
and skeptic fan-out.
Input may be explicit Linear issue identifiers or a batch size. Default to the five strongest
Backlog issues, selected by likely impact and evidence rather than identifier order. Unless Julia
says otherwise, choose stabilization leads before feature leads for the default batch.
Procedure
- Fetch the canonical Linear project by ID from
LINEAR.md and verify its team. Never resolve or
recreate it by display name. Read every Backlog issue, the current Todo queue including feature
issues, and plausible duplicates in every state. The existing Todo queue is required context for
enforcing stabilization-first priority. If the canonical project is unavailable, stop before
writing.
- Launch the parallel investigation waves from
TRIAGE.md. Every selected issue must receive a
subagent investigation. Every proposed Todo, Duplicate, or Canceled verdict must receive an
independent skeptical pass from a different subagent. The coordinator adjudicates and writes;
it does not substitute for either pass.
- Consolidate fragments that describe one coherent change. Preserve the best issue, add missing
evidence there, and mark the others Duplicate with a native relation. Split an overly broad lead
when independent changes have separate value. Leave uncertain groupings separate.
- For each selected issue, verify the five triage claims in
WORKFLOW.md. Use the current
baseline, codebase graph, relevant docs, tests, package exports, open and archived OpenSpec
changes, and GitHub PRs. Treat the issue's proposed solution as a hypothesis.
- Resolve investigator and skeptic disagreements against repository evidence. When evidence is
still insufficient, leave the issue in Backlog and add a focused comment stating what remains to
investigate.
- If a claim fails, move the issue to Canceled and comment with the failed claim, evidence, and a
concrete reopen condition. If another issue owns the work, use Duplicate instead.
- If all claims hold, rewrite the description into the queue-ready issue shape from
WORKFLOW.md.
Preserve the discovery rationale when verified or replace it with the stronger rationale from
investigation. Include a concrete ## Why this matters section that names the present cost or
risk, affected workflow or system property, causal chain, and consequence of doing nothing.
Never leave Todo with only an outcome, scope, and acceptance list. Make every acceptance item
observable. Assign priority, estimate, and a broad existing label. Move it to Todo.
- When completion is genuinely blocked, keep it Todo, add the smallest accurate
## Gate, and
apply the Blocked label. Use native issue relations for issue-to-issue blockers. Split an
independently useful unblocked slice only when it can ship on its own.
- Unless Julia says otherwise, place validated stabilization issues from any of the five
maintenance themes above feature issues. Encode this in Linear using the priority floor and cap
from
WORKFLOW.md. Reconcile existing Todo feature priorities only when needed to preserve this
default, and report every such change. Do not reprioritize In Progress or In Review work.
- The coordinator writes Linear sequentially, reads each changed issue back, and reports the
verdict, priority, estimate, investigator/skeptic coverage, and one-line reason. Also name any
untriaged issue that plausibly outranks the new queue head.
Language and standard-library changes must have acceptance that includes the OpenSpec artifacts
required by AGENTS.md. Do not preserve compatibility paths merely because the current code has
them; this repository is green-field.