Agent setup

Settings

The sandbox's own configuration, plus what it has saved you and which rules fired

On this page(5 sections)

How this sandbox is configured
Change the sandbox settings
What the token-saving measures were worth
Read a built-in system prompt
When each rule last did something

Read and write the settings that govern how agents behave here. The other three routes are read-only reports on their effects: what the token-saving measures were actually worth, the text behind a built-in prompt, and when each rule last did something.

GET/settingsHow this sandbox is configured

Every setting that governs how agents behave here, with the defaults filled in for anything nobody has chosen.

What you send

Nothing. Call it as it is.

What comes back

Field	Type
`stableSystemPrompt`Keep the instructions identical between turns…	boolean
`skills`Which skills are switched on	string[]
`hashlineEdits`Have the agent edit files by…	boolean
`terseOutput`Ask the agent to say less	boolean
`terseHoldout`What share of turns to run…	number
`systemPromptMode`Which instructions the agent starts from:…	"intentic" \| "claude" \| "custom"
`systemPrompt`Your own instructions, used only when…	string
`iqSearch`Teach the agent how to use…	boolean
`iqSearchHoldout`What share of conversations to run…	number
`workspaceMap`Open every conversation with a map…	boolean
`outputCleaners`Which command outputs to trim before…	string
`outputHoldout`What share of commands to leave…	number
`quickModel`Which models do the small automatic…	string[]
`changelogRepos`Which repositories keep a changelog, and…	string[]
`agentRunModels`Which models run the work a…	string[]
`agentRunEffort`How hard those runs should think	string
`autoTier`Whether an easy-looking turn may run…	"off" \| "shadow" \| "on"
`autoTierEagerness`How readily a turn counts as…	"cautious" \| "balanced" \| "eager"
`autoFastModels`Which cheaper model a downgraded turn…	string[]
`agentRetentionDays`How many days a finished conversation…	number
`resumeAfterOutage`Whether a turn killed by the…	boolean
`autoResumeOnRestart`Whether a turn killed by the…	boolean
`rules`Standing instructions you give the sandbox…	object[]
`id`	string
`label`	string
`moment`	"turn.ending" \| "push.starting" \| "agent.finished"
`when`	object
`repo`	string
`paths`	string[]
`outcome`	"clean" \| "error" \| "conflict"[]
`action`	object
`when kind is "command"`	shape
`command`	string
`timeoutMs`	number
`when kind is "instruct"`	shape
`text`	string
`when kind is "verdict"`	shape
`verdict`	"allow" \| "hold"
`when kind is "builtin"`	shape
`name`	"verify-edits"
`enabled`	boolean
`automationFailureLimit`How many failures in a row…	number
`admission`Whether work started from outside may…	object
`schedule`	"allow" \| "hold" \| "deny"
`event`	"allow" \| "hold" \| "deny"
`listener`	"allow" \| "hold" \| "deny"
`webchat`	"allow" \| "hold" \| "deny"
`workspace`	"allow" \| "hold" \| "deny"
`workflow`	"allow" \| "deny"
`actionRules`What an agent may do out…	object
`commandRules`What an agent may run inside…	object
`subagentsAtOnce`How many helper agents may work…	number
`subagentsPerTurn`How many a single turn may…	number
`subagentDepth`How many levels deep the delegation…	number

Try itanswered in this tab

curl

curl "$SANDBOX/settings" \
  -H "x-intentic-control: $INTENTIC_TOKEN"

TypeScript

import { sandbox } from "@intentic/sandbox-client";

const result = await sandbox.settings.get();

POST/settingsChange the sandbox settings

Writes the settings whole, so send the complete object rather than the fields you changed.

What you send

Field	Type	Where
`stableSystemPrompt`Keep the instructions identical between turns…	boolean	body
`skills`Which skills are switched on	string[]	body
`hashlineEdits`Have the agent edit files by…	boolean	body
`terseOutput`Ask the agent to say less	boolean	body
`terseHoldout`What share of turns to run…	number	body
`systemPromptMode`Which instructions the agent starts from:…	"intentic" \| "claude" \| "custom"	body
`systemPrompt`Your own instructions, used only when…	string	body
`iqSearch`Teach the agent how to use…	boolean	body
`iqSearchHoldout`What share of conversations to run…	number	body
`workspaceMap`Open every conversation with a map…	boolean	body
`outputCleaners`Which command outputs to trim before…	string	body
`outputHoldout`What share of commands to leave…	number	body
`quickModel`Which models do the small automatic…	string[]	body
`changelogRepos`Which repositories keep a changelog, and…	string[]	body
`agentRunModels`Which models run the work a…	string[]	body
`agentRunEffort`How hard those runs should think	string	body
`autoTier`Whether an easy-looking turn may run…	"off" \| "shadow" \| "on"	body
`autoTierEagerness`How readily a turn counts as…	"cautious" \| "balanced" \| "eager"	body
`autoFastModels`Which cheaper model a downgraded turn…	string[]	body
`agentRetentionDays`How many days a finished conversation…	number	body
`resumeAfterOutage`Whether a turn killed by the…	boolean	body
`autoResumeOnRestart`Whether a turn killed by the…	boolean	body
`rules`Standing instructions you give the sandbox…	object[]	body
`id`required	string	body
`label`required	string	body
`moment`required	"turn.ending" \| "push.starting" \| "agent.finished"	body
`when`	object	body
`repo`	string	body
`paths`	string[]	body
`outcome`	"clean" \| "error" \| "conflict"[]	body
`action`required	object	body
`when kind is "command"`	shape	body
`command`required	string	body
`timeoutMs`	number	body
`when kind is "instruct"`	shape	body
`text`required	string	body
`when kind is "verdict"`	shape	body
`verdict`required	"allow" \| "hold"	body
`when kind is "builtin"`	shape	body
`name`required	"verify-edits"	body
`enabled`	boolean	body
`automationFailureLimit`How many failures in a row…	number	body
`admission`Whether work started from outside may…	object	body
`schedule`	"allow" \| "hold" \| "deny"	body
`event`	"allow" \| "hold" \| "deny"	body
`listener`	"allow" \| "hold" \| "deny"	body
`webchat`	"allow" \| "hold" \| "deny"	body
`workspace`	"allow" \| "hold" \| "deny"	body
`workflow`	"allow" \| "deny"	body
`actionRules`What an agent may do out…	object	body
`commandRules`What an agent may run inside…	object	body
`subagentsAtOnce`How many helper agents may work…	number	body
`subagentsPerTurn`How many a single turn may…	number	body
`subagentDepth`How many levels deep the delegation…	number	body

What comes back

Field	Type
`ok`Always true	true

Try itanswered in this tab

curl

curl -X POST "$SANDBOX/settings" \
  -H "x-intentic-control: $INTENTIC_TOKEN" \
  -H "content-type: application/json" \
  -d '{"stableSystemPrompt":false,"skills":["…","…"],"hashlineEdits":false,"terseOutput":false,"terseHoldout":0,"systemPromptMode":"intentic","systemPrompt":"","iqSearch":false,"iqSearchHoldout":0,"workspaceMap":false,"outputCleaners":"off","outputHoldout":0,"quickModel":["…","…"],"changelogRepos":["…","…"],"agentRunModels":["…","…"],"agentRunEffort":"","autoTier":"off","autoTierEagerness":"cautious","autoFastModels":["…","…"],"agentRetentionDays":3,"resumeAfterOutage":false,"autoResumeOnRestart":false,"rules":[{"id":"a1b2c3d4","label":"Nightly changelog","moment":"turn.ending","when":{"repo":"root","paths":["src/app.ts","README.md"],"outcome":["clean","error"]},"action":{"kind":"command","command":"pnpm dev","timeoutMs":900000},"enabled":true},{"id":"e5f6a7b8","label":"Release notes","moment":"push.starting","when":{"repo":"site","paths":["src/app.ts","README.md"],"outcome":["clean","error"]},"action":{"kind":"command","command":"pnpm build","timeoutMs":900000},"enabled":true}],"automationFailureLimit":0,"admission":{"schedule":"allow","event":"allow","listener":"allow","webchat":"allow","workspace":"allow","workflow":"allow"},"actionRules":{"src/app.ts":"allow","README.md":"hold"},"commandRules":{"src/app.ts":"allow","README.md":"hold"},"subagentsAtOnce":20,"subagentsPerTurn":200,"subagentDepth":3}'

TypeScript

import { sandbox } from "@intentic/sandbox-client";

const result = await sandbox.settings.set({
  "stableSystemPrompt": false,
  "skills": [
    "…",
    "…"
  ],
  "hashlineEdits": false,
  "terseOutput": false,
  "terseHoldout": 0,
  "systemPromptMode": "intentic",
  "systemPrompt": "",
  "iqSearch": false,
  "iqSearchHoldout": 0,
  "workspaceMap": false,
  "outputCleaners": "off",
  "outputHoldout": 0,
  "quickModel": [
    "…",
    "…"
  ],
  "changelogRepos": [
    "…",
    "…"
  ],
  "agentRunModels": [
    "…",
    "…"
  ],
  "agentRunEffort": "",
  "autoTier": "off",
  "autoTierEagerness": "cautious",
  "autoFastModels": [
    "…",
    "…"
  ],
  "agentRetentionDays": 3,
  "resumeAfterOutage": false,
  "autoResumeOnRestart": false,
  "rules": [
    {
      "id": "a1b2c3d4",
      "label": "Nightly changelog",
      "moment": "turn.ending",
      "when": {
        "repo": "root",
        "paths": [
          "src/app.ts",
          "README.md"
        ],
        "outcome": [
          "clean",
          "error"
        ]
      },
      "action": {
        "kind": "command",
        "command": "pnpm dev",
        "timeoutMs": 900000
      },
      "enabled": true
    },
    {
      "id": "e5f6a7b8",
      "label": "Release notes",
      "moment": "push.starting",
      "when": {
        "repo": "site",
        "paths": [
          "src/app.ts",
          "README.md"
        ],
        "outcome": [
          "clean",
          "error"
        ]
      },
      "action": {
        "kind": "command",
        "command": "pnpm build",
        "timeoutMs": 900000
      },
      "enabled": true
    }
  ],
  "automationFailureLimit": 0,
  "admission": {
    "schedule": "allow",
    "event": "allow",
    "listener": "allow",
    "webchat": "allow",
    "workspace": "allow",
    "workflow": "allow"
  },
  "actionRules": {
    "src/app.ts": "allow",
    "README.md": "hold"
  },
  "commandRules": {
    "src/app.ts": "allow",
    "README.md": "hold"
  },
  "subagentsAtOnce": 20,
  "subagentsPerTurn": 200,
  "subagentDepth": 3
});

GET/settings/savingsWhat the token-saving measures were worth

Measured rather than estimated: what each mechanism actually saved over a range of days. The same day range the spending ledger takes, so one calendar filters both.

What you send

Field	Type	Where
`from`First day to include, as YYYY-MM-DD…	string	query
`to`Last day to include, as YYYY-MM-DD…	string	query

What comes back

Field	Type
`input`	object
`updatedAt`	number
`commands`	number
`rawTokens`	number
`emittedTokens`	number
`savedPct`	number
`perCleaner`	object[]
`id`	string
`commands`	number
`savedTokens`	number
`holdout`	object
`cleaned`	number
`heldOut`	number
`measuredSavedPct`	number
`gaps`	object[]
`command`	string
`commands`	number
`tokens`	number
`output`	object
`metrics`	object[]
`metric`	"proseChars" \| "searchCalls" \| "openingSearches"
`on`	object
`turns`	number
`mean`	number
`off`	object
`turns`	number
`mean`	number
`controlTurnsNeeded`	number
`marginPct`	number
`deltaPct`	number
`saved`	number
`minTurns`	number
`sampleUnit`	"turns" \| "conversations"
`cohort`	string
`search`	object
`metrics`	object[]
`metric`	"proseChars" \| "searchCalls" \| "openingSearches"
`on`	object
`turns`	number
`mean`	number
`off`	object
`turns`	number
`mean`	number
`controlTurnsNeeded`	number
`marginPct`	number
`deltaPct`	number
`saved`	number
`minTurns`	number
`sampleUnit`	"turns" \| "conversations"
`cohort`	string
`tier`	object
`judged`	number
`fast`	number
`atStakeUsd`	number
`routed`	number
`routedUsd`	number
`escalated`	number
`denied`	number

Try itanswered in this tab

curl

curl "$SANDBOX/settings/savings" \
  -H "x-intentic-control: $INTENTIC_TOKEN"

TypeScript

import { sandbox } from "@intentic/sandbox-client";

const result = await sandbox.settings.savings();

GET/settings/system-prompt/{base}Read a built-in system prompt

The actual text behind one of the built-in modes, so a settings screen can show the prompt instead of asking anyone to trust a description of it, and so either can be forked into a custom one.

What you send

Field	Type	Where
`base`required	"intentic" \| "claude"	address

What comes back

Field	Type
`text`	string
`version`	string

Try itanswered in this tab

curl

curl "$SANDBOX/settings/system-prompt/intentic" \
  -H "x-intentic-control: $INTENTIC_TOKEN"

TypeScript

import { sandbox } from "@intentic/sandbox-client";

const result = await sandbox.settings.builtinPrompt({
  "base": "intentic"
});

GET/settings/rule-firingsWhen each rule last did something

A separate read rather than a field on the settings, because a rule firing is not somebody editing anything: folding it in would turn every firing into a settings write and put a self-changing value inside the object a screen edits.

What you send

Nothing. Call it as it is.

What comes back

A plain value rather than an object. The example below is the whole of it.

Try itanswered in this tab

curl

curl "$SANDBOX/settings/rule-firings" \
  -H "x-intentic-control: $INTENTIC_TOKEN"

TypeScript

import { sandbox } from "@intentic/sandbox-client";

const result = await sandbox.settings.firings();