Skip to content
Browse documentation

Agents

Execution limits and recovery

Understand inactivity deadlines, refusals, structured results, dependencies and saved work.

Choose what a timeout measures

Execution > Timeout measures defaults to inactivity: progress restarts the deadline, so productive stages can run longer than the configured duration. Elapsed mode limits total active execution time. Pauses and permission waits are excluded in both modes. The agent's Timeout (milliseconds) overrides the Settings > Advanced default; an empty value inherits it. A command/operation also has its own timeout and can be silent long enough for the stage to expire.

Max Vopit turns applies to the Vopit loop used by API, ChatGPT and Ollama connections. Native CLIs own their internal turns and context. Speed mode also changes the models, effort and correction rounds of each stage; see Speed modes and parallel requests. Increasing Vopit turns will not fix a native inactivity timeout, a denied dependency, or an invalid final result. Inspect the error mode and last activity first.

A refused optional read is an observation

Protected-path refusals and declined infrastructure proposals are returned as access-denied observations so the agent can continue permitted independent work. The editable Access denied instructions explain this recovery. No bypass is authorized. If the assigned deliverable truly depends on the missing access, the agent must report that concrete blocker. A provider failure, expired deadline or invalid structured result can still fail a stage.

A public answer is not always a valid stage result

The selected output contract must parse and satisfy the configured evidence rules. A plan should complete its own preparation instead of demanding a downstream deployment. A DevOps verdict must describe actual deployment and health evidence, not merely an upload. Codex completed without a valid structured stage result means the expected JSON contract was not accepted; the native session is retained for Resume. Inspect files and evidence first because the stage may already have made changes.

Read dependency and correction failures

A success edge requires a successful predecessor, a failure edge requires a correction verdict, and always accepts either outcome. A dependency without a direct conditional edge requires success. Correction loops are bounded by the configured review/visit budget. An unresolved-dependency diagnostic identifies the required outcome, actual outcome and predecessor evidence, or an unscheduled branch. Changes requested without a correction branch requires fixing the blocker or workflow; it is not a successful review.

Partial results and open points

A request can end partial instead of failed, with its work kept:

  • Review budget used up: the request shows Implementation saved · review remaining. The latest reviewer note is preserved; Resume request continues the pending corrections with a fresh review budget.
  • Open points: the workflow ran to its end, but blockers remain that no later stage resolved. The request shows Almost there · open points remain, with up to 40 open points listed with the result. After-work stages such as a Git Manager commit still run.
  • A later stage could not complete after earlier stages succeeded, and no recovery was planned: the request also ends with open points rather than failing.
  • A linked folder that could not be committed, or a request worktree kept with uncommitted changes, is added as an open point.

For a request with open points, Fix open points creates a new request asking the team to resolve only the listed points, with the original request as context. Existing files stay in place; if the original request had its own worktree, the new request continues in that worktree and branch.

Changed sources and interrupted native stages

Before a stage builds on upstream results, Vopit checks the files those results relied on. When they changed outside this request, unaffected work is kept and the stage is asked to requalify what depended on them; the activity shows Requalification required. A native stage that stopped with the app leaves its pre-stage snapshot on disk, so a later stage of the same request attributes the files it wrote to the request instead of treating them as someone else's work.

Resume the saved request deliberately

  1. Inspect the failed ticket, error, modified files and recorded validation before changing settings. Export evidence if needed.
  2. Resolve the concrete missing access, model, tooling, configuration or output-contract issue. Stop active work before editing a team/project/account.
  3. Use Resume request for a failed, stopped or review-remaining request. Saved completed stages and native thread continuity are reused where available; interrupted work uses the current connection. Old requests without checkpoints can repeat planning or the graph.
  4. Restart or Run again creates a new request. Neither Resume, Restart, Stop nor history deletion undoes files, commits, deployments or database changes. Deleting a request removes its worktree only when nothing is left uncommitted in it.
  5. Verify the remaining deliverable and current files. Git resume recognizes a recorded request commit rather than making a duplicate; external changes can still block recovery.

Edit a request or rerun a board card

Editing a request that is still queued and has not started replaces its text in place. Editing a completed, partial, failed or stopped request starts a new request with the new text; the original stays in History. Stop a running request before editing it. A board card linked to the request takes the new text. Running a failed, stopped or partial card again resumes its request, unless the card text changed since that request started: then the card starts a new request. See Plan work on the Board.

All documentation