v0.9.6 release notes

v0.9.6 is a reliability release. The headline: the team always finishes the job. A producer’s attempts are now budgeted per task, and when that budget is spent the quality check finishes the work itself — so a run lands a real deliverable instead of getting stuck. Around that: a leaner team (the Leader is the only required role), lifecycle tooling to keep an install tidy, a more conversational Leader, and fail-closed confinement for subscription seats.

Still Beta — the Beta calibration still applies

This release hardens how a run recovers and tightens security; the engine’s honest limits on the Beta calibration page still hold. This page is the delta on top of v0.9.5.

---

The team always finishes the job

The purpose: a multi-agent run should never wedge. Before, a producer could keep retrying the same task forever, or — on a subscription seat — quietly re-launch the model to work around its retry budget, and a stubborn task could end blocked with nothing to show.

What changed: a producer’s attempts are now tied to the task, not the call. The budget counts across every retry, hand-off, re-assignment, and re-run, and it never resets — so no model earns a fresh budget by being re-dispatched. When the budget is spent, QC steps in and finishes the work itself: it patches the existing draft in place, or writes the artifact from the task’s brief when there’s nothing to patch. Either way the task completes, the producer is freed, and the run moves on. There’s nothing to configure — it’s the default. In practice you’ll see a task that a producer couldn’t land come back completed, authored by QC, with the run continuing past it.

Clearer signals when something’s off

See that QC checked the work. Each QC review now appears in the activity feed against the task it reviewed, so you can tell at a glance that a producer’s output was actually verified — not just that the producer “wrapped up.”

A plain message when the Leader’s model is down. If the model you picked for the Leader is unavailable when you start a job (an overloaded provider, say), Modulatio now tells you plainly to switch the Leader’s primary model and try again — instead of failing with a stack trace.

Producers stay on contract

The producer brief now directs each model to do exactly what the task asks, and stop — no re-planning the problem, no over-gathering, no padding. This keeps reasoning-heavy models from drifting off the brief and inflating the work, which in turn keeps them inside their budget and on-spec for the quality check.

A leaner team

The Leader is the only required role. Team formation now treats QC and producers as optional: stand up a solo Leader, add a QC verifier if you want one, and add producers (1–10) as the work needs them — instead of a forced minimum team.

Lifecycle tooling

• modulatio uninstall — remove the install with tiered, clearly-named choices (settings, project folders, finished deliverables, pandoc). Every removal of your own data is backed up first; a --pristine flag does a full never-installed reset; a standalone uninstall.sh can clean up even when the package itself is too broken to import. A vault Modulatio didn’t create is never auto-deleted.
• modulatio repair — fix broken model presets and agents, recreate a missing vault or default project, and clear configuration in tiers (each gated behind its own confirmation, backed up first). Setup now opens with an Install / Repair choice when it detects an existing configuration.

Talk to the Leader

/models opens the model picker, /new archives the current conversation aside (kept, never deleted) and starts fresh, and /editor composes a message in your $EDITOR. You can stop the Leader mid-thought with ESC in the conversational / solo-coding lane — cooperative, so an in-flight call finishes first and the stopped turn is recorded as a first-class interrupt.

Fail-closed confinement for Clay seats

A producer / QC / planning seat run through your Claude Code subscription (claude -p) is now restricted to a fixed set of non-process built-in tools, runs with customizations disabled (no project/user CLAUDE.md, skills, plugins, hooks, or MCP servers), and explicitly bars the shell and the sub-agent spawners. A confined seat can no longer spawn a hidden crew or re-launch the CLI to work around its retry budget. The interactive Leader (converse / solo-coding) lane is unaffected and keeps its full tool loadout.

Reviewed, end to end

Every change cleared a four-lens cadre review before it shipped: a security pass (confinement + the budget state machine, verified at the command-line level), a hull pass (terminal-state correctness and concurrency), a code-quality pass, and a coherence pass. Findings were remediated and re-verified by a reviewer who ran the code, not just read it.