Skip to content
Chapter 7 · Decisions 7-48

No Continue gate; a deny settles the call, not the run

A run used to park in waiting_user for an explicit Continue when it hit a recoverable stop: a repetition that outlived its repair notes, a full context after compaction, a failed required checkpoint, a budget reservation the run could not get, or (before this decision, already only in documentation) an exhausted provider. Continue reopened a fresh budget window on the same pinned run. A user's approval deny, on the other hand, failed the whole run.

The owner asked for the opposite shape: a run that cannot go on should say why and stop, and anything the model can fix itself should be handed back to the model.

Decision

There is no Continue gate. Every former park ends the run with the code it used to park on (repeated_tool_call, context_limit_reached, checkpoint_*, subagent_family_budget_exhausted, tool_call_budget_exhausted, the provider's own code). HarnessRunTransition.ExhaustBudget and Continue, POST /agent/v1/runs/{runId}/continue, RPC run.continue, the SDK's continue() and ru agent continue are removed; the CLI command answers continue_removed. A new message in the session starts a fresh run that carries the bounded Session Run Handoff of the stopped one (ADR 0023), which is how work goes on.

Where the model can repair the cause, it is asked to before anything fails. The repetition guard keeps its two [REPETITION GUARD] notes. A tool batch larger than one budget window, which could never be reserved, now draws a model_tool_batch_too_large repair note (sharing the runner-rejection note budget) asking the model to split it.

A user's deny settles the asked call, not the run. The approval route commits, with the move back to Running, an approval_denied receipt for that call (ToolExecutionOutcome.Failed, no side effects, never dispatched), exactly as the ask_user answer route completes its call. The rest of the batch resumes; the retained-approval rehydrator and the coordinator's in-memory stash both skip a call that already has its receipt, so the same holds after a Host restart between the deny and the resume. The receipt tells the model not to repeat the call unchanged. Because a deny is the user's decision rather than a tool fault, it never draws unverified_after_tool_failure notes: the model may change course, ask, or simply finish. Stop remains how the user ends a run.

Approvals, ask_user questions and workspace trust are the only states that wait on the user. Startup recovery fails a run an older build parked for Continue with its park code (continue_removed when none was recorded); an ask_user wait, which shares WaitingUser, is told apart by its newest event and left alone. continue.requested stays readable for old databases and history replay.

Consequences

Work after a stop continues through the handoff, which is bounded (16,000 characters from at most four earlier runs) rather than the full native history Continue kept. ADR 0049 later sized the handoff by the model's window and carries the stopped run natively when the next run uses the same model. A provider that is still failing after the runner's retries (about twelve minutes; a Retry-After over five minutes is not waited out) ends the run, so an unattended run no longer waits for the user across an outage. Execution authority, approval tokens, policy and provider pins are unchanged; a deny never grants anything, and the denied call is never dispatched.

The terminal's typed Continue remains, but only to reopen a tracked run that is still running or waiting on an approval or a question in that chat; it no longer resumes anything. The desktop app drops its typed Continue entirely: the word is an ordinary message there.

Alternatives

Keeping Continue for provider outages only was rejected for one consistent rule: a stop is reported and ended. Making a deny end the run was the previous behavior; it threw away the rest of a batch and the model's chance to propose something acceptable.