Release 1.0.21 — Workflows exports, Console model governance, agent activity timelines

A large release across Workflows, Console, Cockpit, and the Copilot chat runtime. New model-governance controls, deeper per-account admin surfaces, and a big set of reliability improvements to how chat, tools, and integrations behave under load.

Highlights

• Workflows — Export any workflow to JSON, PNG, or PDF. Corrected loop_v2 node with proper scoping, item binding, break, and working while conditions. A Stop button to cancel in-flight runs.
• Console — Model governance: pin specific models, per-model context windows, cascading summarization, and image inlining. Better verb-aware ranking for Composio and OpenAPI action discovery.
• Copilot — Per-turn agent activity timelines with full tool request/response detail. Mid-turn stream drops are now detected, resumed in place, and offered as a retry.
• Cockpit — Deactivate Copilot features per account. Feature ceiling visualized as a hierarchy. Cockpit audits now include Compute runtime activity.

What’s new

• Deactivate Copilot features per account. Turn specific Copilot features on or off for individual accounts, not just at the org level.
• Copilot feature ceiling as a hierarchy. A read-only tree view of which features roll up to which policies, so you can see the shape of your governance at a glance.
• Audit scope now includes Compute runtime activity. Sandbox runs are recorded alongside the rest of the audit trail.

• Model governance. Pin specific models, set per-model context windows, enable cascading summarization for long contexts, and inline images inside conversations that support them.
• Provider-aware function model contracts. Function-call contracts now match each provider’s specific format so calls behave consistently across providers.
• MCP OAuth flexibility. Tolerant server discovery, manual endpoint override, and static client import — so more MCP servers work without hand-holding.
• Composio + OpenAPI ranking. Action discovery ranks by verb intent, surfacing the right action first.

• Per-turn agent activity timelines. See every step an agent took inside a turn — tool calls with full request and response detail — durably recorded so you can review or share.
• Chat stream resilience. If a chat stream drops mid-turn, the runtime detects it, resumes in place, and offers a retry so nothing is silently lost.

• Export to JSON, PNG, or PDF. Ship your workflow definitions out of the product in three formats.
• Loop node, corrected. New loop_v2 with scoped body, item binding, break, and a properly working while.
• Stop button. Cancel in-flight runs from the Runs tab.
• Settings dialog. Rename, edit descriptions, and set per-workflow run limits via a gear icon.
• Configurable default run limits. Admins can set defaults through environment config — no DB migration needed.
• Runs tab polish. New Run button in the tab header, colored Step Results badges, and a fix that stops rename-to-empty.

• List layout with file stats and type filter. Browse the knowledge base as a list, see per-file stats, and filter by type.

• Governed extraction, tool-output budgets, and LLM proxy at the platform layer, so tool-call cost and output size are governed centrally.
• Co-browsing and seamless tier transitions inside chat runs.

Improvements

• Chat is more resilient end-to-end — durable terminal state, a tool-output store, and hardened chat recovery + provider error handling.
• Copilot context management is smarter — cascading summarization, BM25 search across tool outputs, and accurate trigger counts.
• Console streaming and traces hardened — fewer edge cases, cleaner spans.
• Governed chat sessions and durable knowledge hardened — sessions and stored knowledge survive more failure modes.
• AI context fullness scoped per user — the fullness signal reflects your own context, not shared state.
• Copilot integrations navigation restored with canonical branding.
• Thread context recovery and reasoning transport unified — one path for both, fewer inconsistencies.
• MCP credential ownership preserved — explicit credential owners survive session boundaries.

Fixes

• Fixed: user breakdown reports return data again (phantom columns removed).
• Fixed: cross-app Composio actions no longer blocked by a mis-scoped session.
• Fixed: Anthropic model calls no longer fail with a 400 on max_turns (stripped from pass-through kwargs).
• Fixed: Composio managed-upload actions correctly stage files into Composio storage (s3key).

• Fixed: publish-blocking save warnings now surface as a toast plus a badge and dialog — no more silent blockers.
• Fixed: canvas no longer crashes (React #185) when nodes are moved quickly.
• Fixed: mangled Composio action slugs resolved, and node maxTurns is forwarded correctly.
• Fixed: workflow-export PDF step table cleaned up — opaque IDs removed, proper columns.

• Fixed: agent dialog footer dead-space and marketplace “by Unknown” author.
• Fixed: guardrail-linked agent resources no longer dropped during batch category fetch.

• Fixed: legacy QUICK_APP node removed from the access-policy tree.

Rollout

All Marketplace subscribers were updated to 1.0.21 during the maintenance window on 20 July 2026. No customer action required.

Questions? Email support@alchemistudio.ai or reach us on LinkedIn .