remediate

vyuh-dxkit remediate

Close the debt the deterministic lanes cannot version-bump away: the grandfathered broken build, advisories with no direct upgrade path, the lint backlog, missing tests, missing docs. Each run plans that debt as finite work orders and works them in two tiers: deterministic recipes execute first at $0 (lockfile resync, override pins, dependency declarations, lint autofix), and only the remaining orders go to a scoped agent, one order per run under a derived budget and an enforced envelope. A class that keeps failing is paused by the circuit breaker instead of re-bought every firing. Everything runs inside a verified frame, and the agent's own claim of success is never trusted: an entry correctness-floor snapshot is captured on the pristine tree before any fix, then after the run a leftover sweep, the floor attributed against that entry (only net-new failures block; pre-existing debt is disclosed, never weaponized), and the guardrail as final arbiter. Work lands only as a PR carrying the verification ledger.

Usage

vyuh-dxkit remediate plan [path] [--json] [--with-floor]  # dry-run: no key, no spend
vyuh-dxkit remediate [path] --task <id> [--land pr] [--dispatch-override] [--json]
vyuh-dxkit remediate configured [path] [--land pr] [--json]

• plan shows, per enabled task, the task > tier > driver-native model resolution, the effective per-task budget, and the spend-ceiling-trimmed matrix the managed workflow reads. It also lists the planned WORK ORDERS (workOrders in --json): the finite units the run consumes (recipe-tier orders execute deterministically in the frame; the remaining agent-tier orders are dispatched one order per agent run, up to remediate.maxOrdersPerRun), built from the entry floor, deferred advisories inside their window (joined to the live dependency scan), and the lint backlog, each with its class, tier (recipe where a registered recipe matches, else agent), derived budget, and done criterion. Findings no class can take are listed under undispatchable with the reason. The floor is read from the baseline's recorded envelope (or the loop snapshot) so the plan stays cheap; --with-floor runs the live floor instead, and the output says which source it used (workOrderFloorSource). The scheduled matrix (matrixTasks) is derived from the OPEN orders in value order: a task with no open orders spawns no scheduled job, open-ended tasks (improve-tests, write-docs) appear only when policy lists them explicitly (disclosed as the legacy shape), and classes the circuit breaker paused are shown with the reason and the unpause conditions (pausedClasses). When the plan's evidence is degraded (an unreadable baseline, no floor evidence anywhere), the matrix falls back to the static policy task list with the reason named, so a fail-open zero can never silently turn the schedule off. A dispatch naming a task bypasses the matrix and runs exactly that task.
• --task <id> runs one task. With --land pr a verified outcome (or a budget-exhausted one under the draft-pr salvage policy) pushes the standing branch dxkit/remediate-<task> and opens or updates its PR.
• configured runs every policy-configured task through the same executor; this is the scheduled workflow's entry point. Tasks share one tree, so a task that leaves unlanded work stops the loop and the remaining tasks are named as skipped.