Skip to main content
Retain Session, Run, operation IDs, and the last cursor before investigating. Before retrying, determine whether an operation may already have taken effect. Do not attribute every failure to user credentials or recreate work by default.

Example: no new events

Query the Run. For waiting_approval, inspect the approval and Tool call, then submit an authorized explicit decision. For completed, investigate cursor handling and read output. For unreachable Runtime, retain the binding/stage and observe recovery without launching another Harness. All three can look like “no new text,” but need different actions. Recovered connectivity does not verify an external write; unknown Tool outcomes need reconciliation.

Start with a bounded read-only check

When you have the original Session / Run IDs, use the repository diagnostic program. Install the appropriate SDK using the example setup, then choose one client from the repository root:
Replace the placeholders with the original IDs. These CLIs discover a local workspace; they are not hosted Token configuration examples. The program checks Run/Session association, then reads state, bounded event pages and artifact metadata. It does not submit, cancel, approve or close work. Output still contains private resource IDs and belongs in local diagnostic records.
  • run_state and error_code help select a branch; an absent error code does not establish healthy dependencies.
  • event_type_counts covers only this read window, not necessarily the full history.
  • reached_empty_event_page does not mean task completion; last_observed_cursor is not a durable application-consumer checkpoint.
  • Reads are not an atomic snapshot. State can change between queries; investigate using the same identifiers rather than recreating work because observations differ.

Recorded real case: recovering a stopped credential ingress

The 2026-11-18 automatic credential-ingress acceptance records a real cloud scenario: the control plane remained running while the ingress for protected model requests was deliberately stopped. API reachability and the Sandbox’s ability to make protected upstream requests are separate paths. The record also preserves an implementation defect: an overly broad recovery guard treated idle Runtimes as active work that must block origin replacement. It was narrowed to persisted nonterminal invocations without deleting original task resources to bypass the guard. Distinguish an incorrect guard from genuinely active work; terminating original tasks is not evidence that recovery is correct. For credential_ingress_recovery_blocked, inspect whether a persisted active invocation still needs the original origin. Explicitly disabled or external/hosted ingress retains its deployment policy; the platform does not silently open a development tunnel. Follow the networking guide for that deployment, and confirm the original task state, actual request outcome and cleanup.

Locate the responsible layer

Report useful evidence

Include code/SDK version, Provider/Harness version, model protocol, stage, sanitized error, reproduction, expected/actual results, evidence location, and cleanup status. Keep real resource IDs privately for cleanup; use redacted references publicly. Exclude Tokens, database passwords, preview tickets, raw headers, and upstream bodies. Record failed and later successful attempts separately. Distinguish simulation from real cloud reproduction. Do not erase failures to make the history appear uniformly successful.

Relevant procedures

Use original states and sanitized diagnostics to select the procedure. Completion requires observed recovery and confirmed resource outcomes, not merely disappearance of an error message.