v0.9.5 release notes

v0.9.5 lets you bring subscriptions you already pay for to the team. Two new seat types reach a model through its vendor’s own harness instead of a metered API key: Clay runs any seat on your Claude Code subscription, and GPT-5.5 runs on your OpenAI Codex subscription. And each seat can now carry fallback models, so a rate-limited or down provider no longer stalls a run.

Still Beta — the Beta calibration still applies

This release adds new model sources and per-seat resilience; the engine’s honest limits on the Beta calibration page still hold. This page is the delta on top of v0.9.4.

Patched in v0.9.5.1 (2026-06-20)

Two fresh-install fixes. Setup now offers Clay to put a Claude subscription on the team — the old “Anthropic OAuth” quick-add pointed at the metered api.anthropic.com and 401’d a subscription token, and never surfaced Clay; that path is gone, replaced by the Clay quick-add (and the Codex quick-add now targets the ChatGPT subscription backend, not the metered OpenAI API). And the boot splash now appears on launch — it had been enabled only on the modulatio-tui entry point, so the plain modulatio launch skipped it. Already have a stale OAuth preset from 0.9.5? Clear it with modulatio models remove <key> then modulatio auth clear-all.

---

Clay — a Claude avatar seat

Clay puts Claude on your team through the subscription you already have. It’s a model seat backed by claude -p — the Claude Code CLI — running headless: assign Clay as your Leader, your QC, or any producer, hand it a task, and it works the task with Claude’s own hands (read, edit, run) and hands back the result.

The point is the path it takes to Claude. Clay reaches the model through the official Claude Code harness — your logged-in subscription — never a metered API key and never api.anthropic.com. So a Clay seat spends the Claude subscription you’re already paying for, not per-token billing.

Clay is treated like any other agent in the role it’s given. It isn’t a privileged back door and it isn’t a special case: it inherits the same confinement and the same rules as every other seat. By default it works confined to its own folder (a working sandbox is required), and when a task needs a real project path it raises the same operator approval as the rest of the team. The Claude binary’s own sign-in is the only thing involved — Modulatio never reads or stores your token.

Set it up in Config → Models: add the “Clay — Claude avatar” provider, install Claude Code and run claude to sign in, then pick a model. It’s purely additive — your existing Anthropic API-key path is untouched, so you can run both.

GPT-5.5 through your Codex subscription

The same idea for OpenAI. A model on your Codex (ChatGPT) subscription now works as any seat, tool use included — reached through the ChatGPT backend’s Responses API where the subscription is valid, instead of the metered api.openai.com (which rejects a subscription token outright). Add the “OpenAI Codex (subscription)” provider, sign in with codex login, and pick gpt-5.5.

Per-seat model fallbacks

Every seat — Leader, QC, each producer — can carry its own ordered list of backup models. If a seat’s model is unavailable for a provider reason (a rate limit, an auth failure, a 5xx), the engine warns you and restarts the whole task on the next backup — never a mid-task model switch that would degrade the result, and the primary is tried first on every new task so it auto-resumes the moment it recovers. A down provider stops stalling the team. Configure it in Config → Agents: pick a seat, open Fallbacks, and list the backups in priority order. Direct-subscription models (Grok via xAI, GPT-5.5 via Codex) can never be given an OpenRouter fallback — enforced in the engine.

Quiet hours for seat alerts

Config → Agents gains a Quiet hours window: during the hours you set, a seat’s model-fallback notices — and the optional Telegram auth pings — are held back and rolled into a single digest at the end of the window, so an overnight run that trips a transient 429 doesn’t buzz you at 3am. Off by default, set per project.

Also fixed

• The boot splash holds long enough to read. The Feng-Tui boot frame could be skipped in a blink — the Enter that launches modulatio leaked into the new screen and dismissed it before the tagline registered. It now holds for a readable beat (up to ten seconds, or any key), with a short opening guard so a stray launch keystroke can’t skip it.
• Widened-exec sandbox hardening (security). When the solo Leader is granted exec in a real folder, a command that named a file there (rather than cd-ing into it) could, under the global dev/test sandbox bypass, run unsandboxed with the parent environment. Widening is now derived from the whole command, and the global bypass never reaches a widened run.

Reviewed, end to end

Clay cleared full code cadre review — coherence, hull/wiring, bypass-surface, and contract — with its one real finding (a sandbox bind that could have re-exposed your home directory) caught, remediated, and re-verified on a live claude -p round-trip before merge. The whole release is green at 4611 tests. Clay’s confinement is enforced in the engine and proven at the real subprocess path, not asserted in a prompt.

---

See the CHANGELOG for the full delta, and the roadmap for what’s next.