Treatment
Yuj calls its default group of long-task and recovery settings the treatment base. The --treatment option selects this base.
Activate the environment that contains Yuj before you run a command on this page. Otherwise, replace yuj with that environment’s bin/yuj path.
Use the default base
Run:
yuj "Fix the issue and run the relevant tests."
Yuj applies four groups of rules:
| Group | What Yuj does |
|---|---|
halflife context | Keep recent tool results in the form first sent to the model. Shorten older tool results when the model nears its input limit. |
| Command handling | Correct known shell and task-format problems before Yuj sends a command. |
| Output handling | Limit large or repeated command output before it fills the model input. |
| Recovery | Record completed observations and apply the guards enabled by the selected settings. |
The assistant CLI applies the treatment base file for command handling, output handling, and recovery. It selects the halflife context mode separately.
Yuj may filter or shorten command output before the model sees it. The trace saves the result that Yuj kept, not necessarily every byte from the command. Yuj may put a large kept result in .tool_output/. Read Saved files for the exact rules.
CLI and paper settings differ
The CLI treatment base is not a complete paper setting.
Each paper comparison also fixes the model, input limit, runtime settings, and detector limits. The paper configuration guide lists the exact files and their order.
Use the plain base
Run:
yuj --no-treatment "Fix the issue and run the relevant tests."
--no-treatment selects the plain base file and the full context mode.
The paper starts its control settings from the same plain base. A complete paper control also fixes its model, input limit, runtime settings, and detector limits.
Use the plain base when you want to compare the two bases. Do not treat this single CLI option as a complete paper reproduction.
Change one value
Put your change in a small TOML file:
# my-yuj.toml
[loop]
max_turns = 160
[llm_hurdle_detector]
cadence_turns = 4
[llm_hurdle_detector.trace_nets]
reread_min_gap = 9
Apply the file after the base:
yuj --config my-yuj.toml "Fix the issue."
Repeat --config to apply more files. Yuj applies them from left to right. A later value replaces an earlier value.
Your changed values make a new setting. Do not call it the shipped default or an exact paper setting.
Read Configuration for the full setting order.
Repeated source inspections
The current trace_nets backend records observations. It does not choose a response from the medicine ladder. To enable the repetition policy, add these values to your settings file:
[loop]
duplicate_guard_enabled = true
duplicate_warn_count = 2
duplicate_abort = 0
The second identical completed call gets one warning. From the third call, eligible unchanged reads and source searches return a reference to the earlier answer. A changed file, changed query or range, pending background work, or an unknown search input requires fresh execution. Compaction and rewind clear the reference. Arbitrary shell commands continue to execute. Repetition never ends a session; the overall turn, time and token limits still apply.
The policy uses file digests for reads and native entry metadata for searches over explicit task subtrees. These checks do not form an atomic filesystem snapshot. The trace records when an answer was reused without executing its tool again.
Change only the context mode
Keep the other treatment settings. Choose another context mode:
yuj --context full "Fix the issue."
Read the CLI reference for every command-line option. Read Saved files to see which records remain complete when a context mode shortens the model’s view.