From request to finished work
Follow a task through model responses, tools, approvals and its final result.
In this topic
- 01Message and context→
- 02Model proposes an action→
- 03Permission and tool result→
- 04Continue or finish
Some turns need clarification or stop with an error. Inspect the final state rather than assuming a tool card means completion.
A Mellow task is a conversation that uses tools to produce an outcome. The model proposes steps, the runtime enforces available capabilities and permissions, and tool results return to the conversation. The final message should explain what happened; the files and action results let you verify it.
Describe a finish line
Give the agent a bounded goal, the relevant inputs, and the evidence you expect back. For example:
Read the CSV files in this folder. Create a monthly totals spreadsheet
and a chart. Keep the original files unchanged. Report rows you could
not parse, and include the output filenames and validation totals.
This allows the agent to distinguish a complete result from a partial one. For ambiguous tasks, it may ask a clarification. For longer work, a task list can show progress.
Choose the execution location
| Location | Appropriate work | Important check |
|---|---|---|
| Working Folder | Read or edit a real host project | Confirm the selected directory |
| Sandbox | Run software in the supported isolated environment | Confirm runtime and permissions |
| Paired Mac | Execute on the selected remote host | Confirm host identity and available inputs |
| Cloud workspace | Assign supported work to the connected service | Confirm account scope and published target |
Do not treat these locations as aliases. Selecting a local path does not copy it to another Mac or a cloud agent. Picking a Working Folder can disable Sandbox mode; review the resulting state.
The loop
The agent can publish a plan through todo, ask a necessary question through clarify, call permitted tools, then end through complete. These loop controls organize a run; they do not prove the external operation succeeded.
An approval pauses a proposed action for review. A denied or failed operation must remain visible in the outcome. If the run stops, inspect the last tool result before deciding whether to retry, revise the task, or change configuration.
Working Folder details
Mellow persists a selected folder with the conversation. A fresh chat can receive an agent default, then a project default where applicable. A folder chosen explicitly for that chat takes precedence.
For an attended custom-agent conversation without a folder or Sandbox, the agent can request a folder selection when the task requires one. Automated or external execution cannot rely on an interactive folder picker appearing; configure its location in advance.
File operations
| Tool family | Purpose |
|---|---|
file_read, file_search | Inspect content, locate files, and narrow a large input |
file_write, file_edit, file_copy | Create, modify, or duplicate permitted files |
file_operation_history, file_undo | Inspect and reverse recorded supported changes |
detect_pii, redact_file | Find or replace recognized personal data |
shell_run | Execute a command through the configured permission path |
git_status, git_diff, git_commit | Inspect repository changes and create a commit where available |
File tools support bounded reads. Large documents may require a page, section, line range, or search rather than repeatedly requesting the entire file. Office and PDF content extraction is not identical to viewing the original layout, so inspect rendered output when formatting matters.
Review and undo file changes
Use the File Changes inspector to compare recorded operations and final file content. A revert depends on retained snapshots and the file's subsequent state. Keep version control or backups for valuable work; no chat history should be your only recovery plan.
For a command-driven change, inspect the diff or files after execution. A zero exit code establishes that the process reported success, not that it met every user requirement.
Bulk edits and on-device redaction
Use targeted or batched edits for repetitive changes instead of regenerating a whole file unnecessarily. Atomic edit batches validate their matches before applying the change. If a match fails, narrow the target and reread the current content.
Personal-data detection depends on the configured local detector. Pattern-only fallback can recognize less than a richer model-assisted detector. Review representative output before treating redaction as complete, especially for domain-specific identifiers.
Sharing artifacts
A generated file becomes easy to access when the agent shares it as an artifact. Writing a file and exposing its card in the conversation are separate operations. Ask for the output path and a shareable card when the deliverable is not visible.
Files returned by a delegated run can be surfaced in the parent conversation. Review the artifact itself, especially when multiple specialists produced related versions.
Check completion against the request
For code, look for the relevant check or test. For a report, open the document and compare its claims with the inputs. For an external operation, inspect the returned identifier or state. A task can be partly complete even when the chat has stopped generating.
If the agent lacks a required capability, choose an authorized alternative or stop at a clearly identified partial result. See Subagents for bounded delegation and Tool contract for operation schemas.
Continue exploring · Your daily workspaceFiles and knowledge collections →Give agents a reference library, manage indexing and understand what a collection provides.