v0.9.8.6 release notes

v0.9.8.6 sharpens leadership and simplifies setup, driven by live test runs and cleared by a four-reviewer code cadre — the security pass blocked three rounds running until the Leader truly ran one model everywhere. This page is the delta on top of v0.9.8.5.

One model for the Leader — across every lane

The Leader now runs a single model for everything it does — decomposing objectives, talking to you, and verifying the team’s work — resolved from its roster agent (by tier, so a renamed Leader still resolves). This closes a real split where the Leader could decompose on one model and verify on another (or fall back to a producer/QC model) without anyone noticing.

The roster is now the single source of every seat’s model: set a seat’s model once and both the conversational Leader and the orchestration Leader use it. An engine guard fails loud if any code path ever re-introduces a divergent binding. The frozen default_models snapshot is no longer read to build any runner — the roster is authoritative for the TUI, daemon, ACP, and CLI alike.

Models and agents live in the Config tab now

The setup wizard no longer configures models or agents. It just gets the install bootable (system tools, vault, budget, first project, embedder) and launches the TUI. You build your team — Leader, QC, producers — and add models in the running TUI’s Config tab, an editable surface that beats a one-shot terminal flow. Edits are live: rename an agent, swap its model, add a producer, all without re-running setup.

A kickoff requires the full triad

A real run now refuses unless the roster has a Leader, a QC, and at least one producer, each with a model — failing fast with a clear “configure the team in the Config tab” message instead of running a hobbled team. A fresh install (empty roster) refuses until you set the team up; typing to the Leader before a model is configured nudges you to the Config tab.

Continuous-pull is the executor

The old wave-barrier dispatch is gone. A freed producer immediately pulls the next ready task instead of waiting for a whole wave to finish — the team stays busy, and producer concurrency is bounded by a global cap rather than wave size.

Task allocation routes by capability + load, never by skill

Producers are selected by capability (an image task won’t go to a text-only model) and load-balanced across qualifiers so nobody starves — skills never route a task. Every producer JIT-loads whatever skill a task needs from the shared library; the Leader and QC route by role.

A hung call can’t freeze the run

A single hung model call no longer stalls the whole team — the pull loop wakes on a tick to keep dispatching ready work to free producers while one call hangs. The Codex (GPT-5.5) streaming path gets a wall-clock deadline on top of the transport read timeout, on both the single-shot and tool-loop lanes, so a stream that trickles forever can’t wedge a seat. And a wedged call now shows an honest “call timed out” on its lane in the TUI instead of a phantom “working…” spinner.

Also in this release

Empty Leader reply on the converse path renders a clear “try again” fallback, not a silent bubble.
Codify (post-run learning) is time-bounded and env-tunable (MODULATIO_CODIFICATION_TIMEOUT_S), with a cancellation boundary so a slow learning step can’t block your deliverables or end report.
A heavy sourcing task that trips the compression-churn cap on its first pass is decomposed and recovered rather than left blocked.
Reload-services button on the Config → Agent tab.
The orphaned wizard model/agent modules + dead runner-builder code were deleted.

4872 tests pass. See the CHANGELOG for the full delta.