v0.9.8.5 release notes

v0.9.8.5 is a reliability release, driven by live test runs and cleared by a four-reviewer code cadre. No interface redesign — the engine getting steadier and the Leader getting sharper. This page is the delta on top of v0.9.8.

Still Beta — the Beta calibration still applies

The engine’s honest limits on the Beta calibration page still hold.

---

Producers run thinking-off; judgment seats reason

Producers (drafters, research) now run with their inner monologue disabled by default — /no_think for reasoning-toggle models, reasoning_effort="disable" where the provider honours it — while the judgment seats (Leader, planner, QC) keep reasoning on. A per-agent disable_thinking override in the agent file wins either way. An always-on producer runbook (the bar-commit working spine) rides every producer prompt, so a thinking-off producer stays rigorous rather than sloppy. The reasoning belongs at the judgment layers; the producers act.

Task count follows the work, not a fixed cap

The old six-task-per-goal limit is gone. The planner now sizes each task to fit a producer’s context budget — small enough to finish below the compression trigger, with headroom — and fans as wide as the work actually needs (one task or a hundred). The runtime compression/churn cap bounds an oversized task, and the engine-enforced no-standalone-verification-goal invariant still blocks decompose-storms. The count cap was a proxy; the context budget is the real measure.

The dispatch floor also changed from a gate to an ordering: an idle under-floor producer takes work when the qualifiers are busy, so nobody starves and the team load-balances honestly.

The Leader sees the team’s deliverables

A Clay (claude -p) Leader is now granted the run directory read-only so its own file tools can open and judge the produced files — in both the goal-verify and the conversational paths. It can inspect, never mutate. (A metered-API Leader reaches the same files through the in-process helpers; the destination is identical, the path differs.)

Clay as the Leader — the honest edges

Clay functions as a full Leader, but running as a sandboxed claude -p subprocess has caveats: it inspects deliverables with its own read-only tools (not the in-process helpers); its goal-verdict can lean on task summaries (a conservative “on the fence” with a reservation for your spot-check rather than a deep read of every file); a claude -p turn can rarely come back empty (retry / model-fallback recovers it); and reads widen while writes stay gated by the same operator-widen permission gate as any seat. For fully unattended, deep-verifying runs, a metered-API Leader (api_key) has no subprocess caveats. See Providers.

The sign-off shows the real verdict

The end-of-run sign-off now surfaces the Leader’s actual verdict (satisfied / on the fence / disappointed) plus a Product Quality Report digest in the conversation — not just a stats line. And the verdict no longer false-fails on a long report: the human-facing report rides as a Markdown section outside the verdict JSON, so prose with quotes and newlines can’t break the parse.

Security — verified by the cadre

A round-one cadre BLOCK caught a real hole: a Clay Leader’s deliverable-visibility grant was mounted read-write, so a Clay seat handed a run directory to inspect could have modified the deliverables it was meant to review. Fixed — the visibility grant is now mounted read-only (--ro-bind), and the reviewer re-verified the sandbox mount before signing off.

Also in this release

• Model-picker capability letters (reasoning / vision / tools).
• Bug-report opens the issue tracker — no maintainer token needed.
• A recovered task’s failure ticket auto-resolves at run-end.
• Delivery + the end report happen before best-effort codification, so a slow learning step never delays your result.
• The send-log modal never requires a token and is always exitable.
• Downloadable offline docs from the DOCS tab.

---

4906 tests pass. See the CHANGELOG for the full delta.