Skip to content

armgit / Ix

Public Imported from GitHub
0

Understand any codebase instantly. System intelligence for codebases, built for humans and AI.

3.7 KB
Raw Blame

Reset outcome reconciliation

An asynchronous reset can change graph or ingestion state before the client loses its connection or status entry. Repeating it can delete work created since the original reset. A missing status does not establish that nothing ran.

The CLI now requires a 202 start acknowledgment with a UUID and a status body identifying that same operation with running, done or failed. It stops with explicit reconciliation guidance after missing, denied, invalid or unavailable status, a partial failure, or an observation deadline. It reports the operation ID when known and never submits another reset on these paths.

An administrator should inspect the original operation, graph effects and pipeline claim before deciding how to recover. Do not delete a durable claim or create a new reset solely because a status is absent. The status ledger is process-local and bounded: restarts, replica selection and eviction can each make an operation unavailable while the reset itself completed.

Authorization for the status route is whatever the transport in front of it enforces; the route itself resolves an operation ID against the ledger.

Treat the blast radius of a repeated reset as wider than the caller expects. That is the reason this client stops rather than retrying, and why recovery is an operator decision made against the deployment's own records rather than something the CLI can infer.

Local synchronous reset and the existing synchronous fallback for an absent async start route remain. A 404 from status never invokes that fallback.

The client never follows redirects for async start, status polling, local sync reset or legacy sync fallback. A redirected start may already have run; following 307/308 can replay its POST, while following a redirect to a 404 can incorrectly trigger sync fallback. A refused redirect target also says nothing about whether the initial reset ran. Redirect responses therefore require reconciliation, with the original operation ID retained when known. Configure the direct backend endpoint before any operator-approved recovery; do not repeat an uncertain reset. Because the backend itself never redirects, a redirected start or sync request also names the redirect target's origin (never its path or query) and points at IX_ENDPOINT as the likely fix; a redirected status poll keeps the operation ID instead, since its start was already accepted.

Authentication errors at start remain errors. This patch does not add tenant headers, credentials or an authorization bypass; the remote transport must establish the expected verified caller identity.

ix-cli/src/client/__tests__/reset-outcome.test.ts executes the real client with synthetic HTTP responses and asserts the exact request sequence. It covers success for both reset kinds, invalid/foreign operation IDs, partial failure, transport loss, deadlines and local/older-server compatibility. ix-cli/src/client/__tests__/reset-redirect.test.ts drives the same client against real loopback HTTP servers and asserts that no reset request is replayed or reinterpreted after a 3xx on any of the four request paths. Live authenticated remote transport and supported-release integration still need verification before claiming deployment acceptance.

Pro is an optional, separately installed private plugin; the OSS package does not fetch it as a dependency. Its explicit runtime resolution is therefore listed in Knip's dependency exceptions. The loader returns normally only when Pro is absent or fully initialized. Missing peer modules, invalid registration exports, and credential-guard failures in an installed Pro abort command startup instead of silently falling back to unguarded OSS execution.