silk-discover

name	silk-discover
description	Verify the checkout is exactly current remote main, then run a deep parallel Silk sweep for stabilization, documentation drift, simplification, dead-code, and technical-debt leads and file every lead passing the Backlog intake bar in Linear. Use only when Julia explicitly asks for a Silk maintenance sweep or invokes $silk-discover.

Discover Silk maintenance work

Read these project-local references completely before acting:

• ../../silk-manager/WORKFLOW.md
• ../../silk-manager/LINEAR.md
• ../../silk-manager/DISCOVERY.md

This is an inspection and intake skill. It may create or enrich Linear issues. It must not edit repository source, change GitHub state, switch branches, fetch, or clean the working tree. It must not run pnpm check, the full test suite, native differential acceptance, or pnpm release:candidate unless Julia explicitly requests broader validation. Focused tests and minimal reproducers are allowed only to give one lead a concrete breadcrumb.

Follow DISCOVERY.md's required three-subagent fan-out, coordinator assignment, coverage ledger, cross-check, and gap-driven second pass. Respect a user-supplied scope or limit. Without one, do not cap the result or stop after the first strong finding.

Procedure

1. Run DISCOVERY.md's remote-main freshness gate before any Linear read, source inspection, or subagent spawn. Stop immediately unless local HEAD exactly equals the SHA returned by git ls-remote --exit-code origin refs/heads/main. Never trust a local tracking ref and never update the checkout on Julia's behalf.
2. Fetch the canonical Linear project by ID from LINEAR.md and verify its team. Never recreate it from its display name. If it cannot be resolved, stop before writing. Read all canonical-project issues in every status for deduplication.
3. Record the verified remote-main SHA and dirty paths. Review recent commits, active Codex tasks, open PRs, and active OpenSpec changes so work underway is not rediscovered.
4. Execute the parallel assignments in DISCOVERY.md. Use the codebase graph first for structural discovery. If an index lookup fails, select the index that corresponds to the current repository rather than abandoning that lens.
5. Wait for all scouts, then re-run the remote-main freshness gate before doing more expensive work. Stop without Linear writes if remote main moved. Build the coverage ledger, remove exact duplicates and clearly owned work, and spot-check lead provenance. Do not perform triage's skeptical review, debatable consolidation, worth judgment, sizing, or prioritization. Run the required second pass when the first pass yields fewer than five novel investigable leads or any material coverage area remains weak or unchecked.
6. Apply the Backlog intake bar from WORKFLOW.md. Immediately before the first Linear write, re-run the remote-main freshness gate once more and stop if it moved. File every distinct investigable lead with its honest confidence, a visible ## Why this matters section, and triage questions, including low-confidence leads with a specific starting point. The rationale must connect the observation to a concrete affected workflow or system property and a possible consequence if ignored; distinguish evidence from inference. Use no priority or estimate. Use Bug only for directly demonstrated defects; otherwise use Improvement or no broad label.
7. For an existing issue, add only genuinely new evidence as a dated comment. Do not reopen, reprioritize, or rewrite a triaged issue during discovery.
8. Read every changed issue back. Report new issues, enriched issues, and intake exclusions limited to exact duplicates, clearly owned active work, or observations too vague to investigate. Also report scout completion, the full coverage ledger, commands run, and material gaps.

Every new issue must meet the Backlog intake bar in WORKFLOW.md, include concrete breadcrumbs, an explicit causal justification, triage questions, and the source baseline. Draft acceptance is optional because triage owns the final specification. One excellent issue is not evidence of a thorough sweep; breadth is demonstrated by completed scout assignments, coverage, and the second-pass rule.