Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS
JavaScript ai-agentsdashboarddeveloper-toolsdigital-twinquantified-selfsecond-brainself-hostedtailscale
/**
* Agent TUI spawn config
*
* The PURE launch-shape builder: provider + model + posture in, the argv /
* command-line quadruple out. No PTY, no session, no timers — nothing here
* observes or mutates a live run.
*
* Extracted from `agentTuiSpawning.js` so "what command does this provider
* launch?" is answerable without loading the live session state machine
* (`sessionController.js`). The two concerns had drifted into one closure's
* neighbourhood purely by history: a vendor-recipe change and a paste-retry
* change touched the same file for no shared reason.
*/
import { formatShellCommandLine } from '../../lib/shellCd.js';
import { resolveInteractiveShell } from '../../lib/interactiveShellResolver.js';
import { buildCodexAgentThreadArgs, isClaudeCommand, applyLeanClaudeArgs } from '../../lib/providerModels.js';
import { isCodexCommand } from '../../lib/codex.js';
import { applyCredentialBootstrap } from '../../lib/credentialBootstrap.js';
import { buildVendorSpawnConfig, injectTuiModelAndEffort, supportsTuiPublicReviewPosture } from '../../lib/providerVendors.js';
import { publicReviewPostureForProfile } from '../../lib/agentExecutionProfiles.js';
import { DEFAULT_TUI_PROMPT_DELAY_MS, inferTuiCommand, applyCommandDefaults } from '../../lib/tuiHandshake.js';
export function buildTuiSpawnConfig(provider, model, {
systemPromptFile = null,
effort = null,
maxConcurrentThreads = null,
safetyProfile = null,
shell = resolveInteractiveShell(),
} = {}) {
// A public-content stage's argv is the vendor's maintained recipe, never the
// generic assembly below — that path forwards `provider.args` and applies
// `applyCommandDefaults`, either of which can hand a contributor-controlled
// review a saved `--dangerously-skip-permissions`. `tui: true` tells the
// recipe to drop only the flags that require `--print` and keep every
// enforcement flag. Fail closed when the pairing declares no attachable recipe: the
// caller is supposed to have asked `supportsTuiPublicReviewPosture` first, so
// reaching this is a routing bug and must not silently open a PTY whose
// posture is decorative.
const posture = publicReviewPostureForProfile(safetyProfile);
if (posture) {
if (!supportsTuiPublicReviewPosture(provider, posture)) {
throw new Error(`Provider '${provider?.id || provider?.command || 'unknown'}' cannot run an attachable ${posture} session`);
}
const recipe = buildVendorSpawnConfig(provider, {
effectiveModel: model,
effort,
maxConcurrentThreads,
systemPromptFile,
safetyProfile,
tui: true,
});
// A `sandboxed-actions` posture never gets credential-bootstrap-wrapped —
// the enforced recipe above IS the sandbox there, and it is spelled entirely
// in argv. A `no-tool` posture IS wrapped, because spawned bare it would
// carry no credential at all (#7720). `applyCredentialBootstrap` owns that
// split (keyed on `safetyProfile`); either way this branch returns the same
// shape as every other, with `command`/`args` still naming the harness.
const { command: spawnCommand, args: spawnArgs } = applyCredentialBootstrap(provider, recipe.command, recipe.args, { safetyProfile });
return {
command: recipe.command,
args: recipe.args,
spawnCommand,
spawnArgs,
commandLine: formatShellCommandLine(spawnCommand, spawnArgs, shell),
promptDelayMs: provider?.tuiPromptDelayMs || DEFAULT_TUI_PROMPT_DELAY_MS,
};
}
const command = provider?.command || inferTuiCommand(provider?.id);
const baseArgs = applyCommandDefaults(command, [...(provider?.args || [])], provider);
// Model+effort injection (including the antigravity-validates-the-pair special
// case) is shared with tuiHandshake.js#buildTuiInvocation via
// providerVendors.js#injectTuiModelAndEffort, so the two spawn paths can't
// drift — they already had once, on cursor, before #3618.
let args = injectTuiModelAndEffort(command, baseArgs, provider, model, effort);
if (isCodexCommand(command)) {
args = [...args, ...buildCodexAgentThreadArgs(maxConcurrentThreads)];
}
// Lean mode for Ollama-backed claude sessions (no-op otherwise) — must come
// before the system-prompt flag so `--bare` is present when the contract
// file rides along.
args = applyLeanClaudeArgs(provider, args, command);
if (systemPromptFile && isClaudeCommand(command)) {
args = [...args, '--append-system-prompt-file', systemPromptFile];
}
// `command`/`args` keep naming the harness itself — every consumer of this
// config (ready-text detection, permission-dialog handling, error messages)
// keys off them by identity. `spawnCommand`/`spawnArgs` are what actually
// gets launched: the bootstrap CLI in front of the harness invocation for a
// credential-bootstrap-configured provider (see credentialBootstrap.js), or
// an identical copy otherwise. `commandLine` (used both to type the harness
// into a login shell and to display "what's running") is built from the
// SPAWNED pair so it always matches the real process.
const { command: spawnCommand, args: spawnArgs } = applyCredentialBootstrap(provider, command, args);
const commandLine = formatShellCommandLine(spawnCommand, spawnArgs, shell);
return {
command,
args,
spawnCommand,
spawnArgs,
commandLine,
promptDelayMs: provider?.tuiPromptDelayMs || DEFAULT_TUI_PROMPT_DELAY_MS
};
}