Stop bad output early

TTSR (tool-text stream rules) watches the model’s output as it streams. If forbidden text appears inside a tool call (for example as any inside an edit payload), tsforge stops generation immediately, sends short correction text, and asks the model to retry.

The bad pattern never lands in your files.

Always on.

Built-in rules

Rule	Blocks	Tells the model
no-as-any	as any	use proper types
no-ts-suppression	@ts-ignore	fix the real error
no-empty-catch	empty catch {}	log or handle errors
no-console-log	console.log	remove debug logging

These apply to tool argument text under src/**/*.ts(x). Each rule has a cooldown so one mistake does not loop forever.

Custom rules

Add your own stream rules in .tsforge/rules.json at the project root. They load on top of the built-ins; a custom rule whose name matches a built-in is ignored, so the built-in safety rules always win.

[
  {
    "name": "no-todo-comment",
    "condition": ["//\\s*TODO", "//\\s*FIXME"],
    "guidance": "Do not leave TODO/FIXME markers; finish the work or open an issue.",
    "scope": "tool-args",
    "repeatMode": "cooldown",
    "repeatGap": 5,
    "fileGlobs": ["src/**/*.ts", "src/**/*.tsx"]
  }
]

Field	Required	Notes
name	yes	Unique; duplicates of a built-in are skipped
condition	yes	Array of regex sources; matches if any one matches
guidance	yes	Correction text sent to the model on a hit
scope	no	tool-args, content, or both (default content)
repeatMode	no	once (default) or cooldown
repeatGap	no	Turns between refires when cooldown
fileGlobs	no	Restrict to matching files; omit to apply everywhere

Invalid JSON or invalid regex is skipped silently. On load, tsforge reports how many custom rules were registered.

Learned rules

tsforge also writes its own TTSR rules: failures it made and then fixed in a repo are mined into .tsforge/learned-rules.json and loaded alongside the built-ins and your custom rules. So a mistake the model keeps making gets caught mid-stream next time. Automatically. See Learning from past runs.

Limits

• 3 interrupts per task, then TTSR turns off for the rest of that task

→ Learning from past runs · Big picture · Environment variables