ss-troubleshooting-workflow

End-to-end defect fix — takes an issue link, alert ID, or problem description and orchestrates tier-adaptive root-cause investigation, human confirmation, branching, a fix plan, multi-agent coding wit

SKILL.md

GitHub

Troubleshooting Delivery Workflow

Read workflow-common.md, workflow-checkpoint.md, agent-discipline.md, and delta rules.

Context

Input is issue/alert/link/text or a completed matching investigation. Main only coordinates; the same logical owner investigates, plans, codes and repairs at this boundary. Preserve stable Problem ID and sanitized rawInput, normalizedFlags, gateCompletion, authorizedScope, workspace/not-yet-created and currentFingerprint in the one checkpoint through every delegation, multi-repo handoff and recovery. Read full original artifacts on demand; recency and a terse summary are not proof.

Ordered chain

  1. Owner calls report-only ss-inspect, or reuses a complete verified report for this Problem ID. Require independent corroborating evidence and understanding of relevant design/implementation, causal flow and affected paths before repair. Keep Violated Invariant:, Owning Boundary:, Patch-Style Alternative Rejected: and Repositories Requiring Fix:.
  2. Before root-cause confirmation, validate applicable delta when semantics change. A restoration of an existing promise uses no-delta only with the specific authoritative promise, unchanged semantics rationale and evidence. Use protected external staged validation mirror when checkout must not yet be mutated. New semantics are a scope decision, not an inferred bug-fix permission.
  3. Main presents report + repair contract/delta and records real human root-cause confirmation, or explicit eligible --auto policy completion. --decide alone is not approval. Gate binds current report/proposal/spec bytes and evidence; auto is never fabricated human confirmation and grants no new permissions.
  4. Confirmed multi-repo scope follows detection, carrying flags/gates/checkpoint/Problem ID and stopping this single-repo chain.
  5. Create authorized full workspace with ss-create-branch, or verify lite workspace. Materialize identical gated artifacts. Owner reuses sufficient repair breakdown or calls ss-plan for a short interface/invariant/test-intent map; no repeated RCA.
  6. Call ss-coding for real focused regression RED -> minimum GREEN -> refactor; archive/materialize/rebase and complete required commits before review, only with existing Git authorization. Freeze the committed identity for PR delivery or preserve honest local delivery; then fresh independent review. Group accepted fixes with the same owner and obtain fresh impact review when substantive.
  7. One retained verifier runs final full validation AFTER APPROVED review on the same currentFingerprint. Preserve failures -> targeted owner fix -> fresh impact review -> retained current validation. Never parallelize review and full tests.
  8. ss-create-pr reports an actual PR or honest local delivery; never merge/deploy. Cleanup needs separate authorization. Full and lite share all quality gates.

Incomplete RCA/review, unknown writer, input drift, cancellation or host failure preserves the checkpoint and original results. Supported safe continuation resumes the same logical owner within existing authority; no CLI/model switch or duplicate writer. Report gateCompletion, currentFingerprint, evidence, delivery and blockers.