Flows
Runs and debugging
How to follow a LoCrew flow run, approve or reject it, inspect every step's input, output, events, tool calls and spend, compare attempts, replay a run, and recover from failures with retries, Cancel and Resume.
Every time a flow runs, its steps, outputs, approvals and spend are recorded and kept across restarts. Open Runs in the sidebar, or use Open run details in the editor or Inspect run on a run card in chat.
#Run states
| State | Meaning |
|---|---|
| Idle | Created, about to start. A run card in chat shows Queued. |
| Running | Steps are in progress. |
| Waiting for you | Nothing is running and an approval is pending. |
| Completed | The run reached its end. |
| Failed | A step could not be carried out, the run went over its time limit, or an internal error occurred. |
| Rejected | You rejected an approval that had no Rejected path. |
| Cancelled | You cancelled it. |
| Interrupted | LoCrew stopped mid-run. Nothing is retried until you Resume. |
Steps have the same states plus Skipped (Not run: this path was not taken. or Not run: an earlier step failed.). Every badge has an icon and words, never colour alone.
#Why a step failed
| Label | Meaning | Retried automatically |
|---|---|---|
| Failed | The runtime or server reported a failure. | Yes, if the step allows more attempts |
| Ran out of time | The step hit its time limit. | Yes, if allowed |
| Cannot succeed as it is | A missing agent or credential, a budget limit, a sign-in problem, or an earlier failure. | No |
| Stopped | The step was cancelled. | No |
| Outcome unknown | LoCrew stopped, or a request may have reached the server, so it was not repeated. HTTP error statuses that are not retried, and responses missing a kept value, are also labelled this way. | No |
A step that failed and then succeeded on retry shows Succeeded after N failed attempts on the canvas.
#The Runs view
- Show runs of filters by flow. Each row shows the flow, its state, when it ran, how long it took, and needs you when an approval is pending.
- The run header has Cancel run (while active), Resume (when interrupted), a link back to the conversation it was started from, and Open flow.
- Facts: started, ended, duration, started by, flow version and spend.
#Spend
The run's spend adds up only measured spend. Not measured means no step reported a price; $0.42+ (2 steps not measured) means some steps could not be priced. Model and external agents are never priced.
#Approving a run
When a run waits for you, a badge appears on Runs, and a run card in chat shows Review approval. Open the run:
- Read the approval card: its title, notes and the output of the steps it shows.
- Optionally write Feedback; later steps can use it.
- Press Approve or Reject.
Approving lets the flow continue. File changes still need their own approval. A pending approval survives a restart.
#The run canvas
The run shows the flow as it was when the run started. Click a step to inspect it, or an arrow to see what moved along it:
- Thick arrows carried data.
- Solid arrows were taken; dashed ones were not.
- Dotted red arrows send work back.
Clicking an arrow lists each value that moved (research.text → write (input)), or explains why nothing did.
Replay steps through the recorded events with a slider, redrawing each step's state as it was. It re-runs nothing.
Below the canvas, Steps lists every step in order with its attempts, and Run log lists every event (200 per page, Load more).
#The step inspector
Click a step to open it. With several attempts, choose Attempt N or Compare two side by side.
| Tab | Shows |
|---|---|
| Overview | Why it did not succeed or was skipped; agent, runtime, status, attempt, times and cost; What went in and Where it went. |
| Input | What the agent received (its full prompt), or the HTTP request as sent; the values from earlier steps; the configuration. |
| Output | The output and the steps it was passed on to. For agents, the final reply; changes made through tools are under Tools. |
| Events | The flow's and the agent's events, in order. Streamed text is folded. |
| Tools | Each tool call with arguments, result and duration, as the runtime reported it. External agents report none. |
| Metrics | Total duration, queue wait (for the agent, the concurrency limit or the directory lock), execution time, attempts, tokens, turns and cost. Copy metrics copies them. |
Each stored value is marked Captured, Captured, secrets masked, Truncated or Not available. Values are capped at 64 KB.
Note
What is kept depends on Settings → General → Flow step history. With Only which steps ran…, values are not stored. Stored inputs and outputs older than the retention period are cleared at startup. See Settings.
#Debugging a failed run
- Open the run and click the step marked failed. Overview says why, and the label tells you whether a retry could help.
- Check Input to see exactly what the step received: a missing value, a wrong field, an unexpected prompt.
- For HTTP steps, read the error message and compare the request as sent with what the API expects.
- For agent steps, check Tools and Events for what the agent did, and Output for its reply. A JSON reply that missed a declared field shows the problem.
- Back in the editor, failing steps show their reason on the canvas, and the step panel has In the last run with the full error and a link to the run.
- Optionally use Ask an agent why it failed at the bottom of the step panel. See Build a flow.
- Fix the flow, Save, and Run again. Earlier runs keep the version they used.
#Retries, time limits and budgets
- Retries: set per Agent and HTTP step with Attempts. Each retry is a new attempt you can inspect.
- Time limits: each step has its own; the agent's own limit and the global Max task duration still apply. The whole run can have a limit too; when it expires: The run went over its time limit of … and was stopped.
- Budget: before every agent attempt, LoCrew refuses to start if the run has started 200 agent steps, or has passed Max spend per chain, or a step's spend could not be measured while that limit is on. The step fails with Not started: …
#Cancel and resume
- Cancel run (or Cancel on a chat run card) stops live steps, cancels pending approvals and marks unfinished steps cancelled.
- If LoCrew stops mid-run, the run becomes Interrupted: LoCrew stopped while this run was in progress. Resume it to re-run the interrupted steps. Runs that were only waiting for an approval keep waiting.
- Resume re-runs only the steps that were cut off. Completed steps keep their results, and the run's time limit starts again in full.