Maple Harness home

Maple Harness 2.6.2 released! What's new in 2.6.2

Single userTeams & hubs

One person, one machine: the agent, the proxy and your own usage.

Several people, from one shared proxy up to a hub of hubs.

Settings by goal #

On this page
  1. In plain words
  2. I want to spend less on long runs
  3. I want the work to stay as good
  4. I want to stop runaways
  5. I want results I can rely on

Maple Harness aims at three things at once: to be efficient (fewer tokens for the same work), efficacious (the work stays as good, and defaults are safe) and effective (whole tasks succeed and waste stops early). This page maps a goal to the settings and commands that serve it. It makes no promise about your bill: results vary from run to run and task to task, so measure your own, and turn off anything that does not help on your work. Every setting listed has an off switch.

In plain words #

  • What it is. A map from what you want (spend less, keep quality, stop runaways, get reliable results) to the setting that serves it.
  • What to do. Find your goal below, change the setting where it says (most are on the setup page's Models & routing tab, or on Advanced, which Find a setting reaches by name), and measure the result on your own work.

I want to spend less on long runs #

Do thisWhereTrade-off
Keep one model for a long task instead of switching/model <name>, or /model auto (it keeps your model when it fits)A fixed model may be more than a small job needs
See where the money goes/cost, the setup page's Cache card, maple-proxy cacheNone
Keep the provider's cache workingproviders[].prompt_cache (on by default), prefix_stabilizer (on by default)None for the defaults
Trim old tool output in very long conversationscontext_trim on the Guards card (off by default)The model may ask for trimmed output again
Keep noisy searches out of the main conversation@name <task> or /agent, with a subagentA subagent starts with only the task you give it
Stop a stuck run earlyloop_guard.mode (on by default)It may end a run you would have let continue
Shorten long failing logstool_output.failed_max_lines (on by default)The model sees less of a long log
Cap what a provider can cost per daybudgets[] on the Budgets cardRequests are refused, or fall back, when reached
Start fresh tasks in fresh sessions/clear, or a new sessionYou lose the earlier context

I want the work to stay as good #

Do thisWhereTrade-off
Let the compiler check the model's editslsp (on by default), /lsp to see what is runningThe first edit starts the server in the background
Keep the safe defaultspermission.mode: ask, sandbox.mode, file_policy.*More questions, fewer surprises
Prefer the tools that show a diffbash_file_writes: warn (default) or denydeny refuses shell writes
Plan before changing anything/planAn extra step
Use a capable model for hard work/model, auto_model_mapCosts more per turn
Keep automatic choice away from models that decline long tasksauto_avoidA pattern may hide a model you would have liked
Check your setupmaple-harness doctorNone

I want to stop runaways #

Do thisWhereTrade-off
End a turn that keeps failing the same way, with a summaryloop_guard.mode: stop, loop_guard.stop_afterSee above
Cut a reply that repeats itselfdegenerate_guard (the repeat guard) on the proxy's Guards cardA repetitive answer could be cut short
Refuse or trim a request bigger than the model's windowcontext_guardtruncate forgets old turns
Cap the turns a subagent may takesubagents.max_turnsA long job may need more
Limit how fast a client may call the proxyserver.rate_limit.*Bursts are slowed
Move to another model when one runs outmodel_fallbacksThe fallback may be different or weaker

I want results I can rely on #

Do thisWhereTrade-off
Get told when a session needs younotifications.*, maple-harness notify testNone
Resume where you left off--continue, --resume, /sessionsOld sessions take disk; cleanup removes them
Keep a checklist for a big requestledger: auto, /ledgerOne extra model call
Make the model carry on when it stops earlyon by default; disable_auto_continue turns it offNone for the default
Check a change against your own tasksmaple-harness --prompt with --json-summary, repeated with a setting on and offIt costs the runs you make

Read next: Proxy settings, Commands and keys and Harness settings.

Made with ❤️ by Kinn Coelho Juliao in Ottawa, eh — fuelled by double-doubles, bud. 🍁