A saved tool runs its skills in order and returns the final step's output.
Before You Run
On the Tools page, select a deployment and review:
- Input Sources — primary files, folders, or documents.
- Reference — saved reference sources. In the current beta these are not read. The run emits a warning and ignores them.
- Output Folder — the destination given to file-writing skills. A folder produces a new timestamped file; a document is appended to instead. See Tool Output and Results.
- Model — the AI route assigned to this deployment. This appears only when the tool contains an AI skill.
- Settings — identity, skill-specific configuration, and applicable execution controls. AI instructions and guided mode appear only for AI tools. Python tools require an interpreter, script, arguments, input mode, timeout, and trust for the current script version.
- Triggers — optional schedules or new-file watches.
Press Run in the deployment inspector, or choose Open in Window for the separate runner.
The runner window and dashboard widgets expose their own output destination in the Context tab. A destination selected there replaces the deployment's configured output folder for that run.
Runtime Sequence
- The invoking surface builds the input.
- A local file or folder source under the projects root scopes SK Memory to that project.
- If the input is a pointer to context rather than content, ThinkForge expands and reads the supported sources and combines their text.
- ThinkForge verifies that every step names an available skill and, when an allow-list exists, that the skill is allowed.
- Steps execute in ascending order.
- The final result is returned.
Chaining and condition fields do not alter this sequence.
Data Passed to Each Step
Each step receives:
- The static arguments saved on that step
- The original resolved input text
- The immediately preceding step's output
- A transcript of completed steps, when available
- Tool instructions and skill-specific saved settings
- Source, project, and model-routing information
- Output information: the destination folder, plus the destination file when the caller chose a document rather than a folder
The runtime is a linear pipe: one step's structured result becomes the next step's data. A skill that needs a URL, path, or another specific value extracts it from that data itself. There are no branches.
Autonomous and Guided Modes
Autonomous mode proceeds without approval pauses.
Guided mode asks for approval after AI steps. If you reject the result, execution stops.
Interaction behavior depends on where the tool runs:
- The desktop Tools page can prompt you interactively.
- The Mobile Dashboard surfaces pending approvals to the mobile client.
- Power Strip, browser commands, and triggers run headless and auto-confirm.
Use autonomous tools for unattended triggers. A guided tool invoked headlessly will not pause for review.
Cancellation
ThinkForge checks for cancellation between ordinary skills. It cannot interrupt most skills that are already running, so Cancel normally takes effect after the current step returns.
RunPythonScript is an exception. While its process is active, Stop or timeout terminates the Python process and its child process tree, captures the available output, and prevents later steps from starting.
Failures
Execution stops when:
- A skill is missing or outside the tool's allowed capabilities
- A skill throws an error
- A skill returns a recognized error signal
- Python exits with a non-zero code, times out, is stopped, or fails executable-code trust
- Guided approval is rejected
- Cancellation is observed between steps
The result includes success, duration, error text, and an execution transcript. Some surfaces show these details; others expose only a status message.
Streaming
A tool can emit output while a run is active, and desktop and mobile surfaces may display that progress. The final result is still collected in full before it becomes input to the next step.