You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Minor Changes
512a432: assistants now self-heal when the inference provider rejects a chat as malformed: the runtime trims history to the last 5 user messages, prepends a recovery notice that nudges the agent to recover lost context via its tools, and retries — instead of leaving the thread stuck.
6cf658b: Every assistant now exposes a platform toolset to its runtime alongside its user-attached toolsets, with no user-facing toolset row and no setup required. Removes the assistant_memory product feature flag in the process: GET /rpc/productFeatures.get no longer returns assistant_memory_enabled, and POST /rpc/productFeatures.set no longer accepts "assistant_memory" as a feature_name — the assistant memory tools are always-on.
707bc98: Outbound Slack messages can now render rich Block Kit content. chat.postMessage and chat.postEphemeral accept an optional typed Blocks field (section, actions+button, context, divider) alongside the existing text fallback. Button clicks come back as block_actions interactions on the existing Slack trigger webhook, are correlated to the originating thread, and reach the assistant as a new turn carrying action_id, action_value, and block_id — so assistants can present options and receive the user's choice in the same conversation.
fa5ef43: Add Codex (OpenAI) hooks support. A new /rpc/hooks.codex endpoint accepts all six Codex hook events (SessionStart, PreToolUse, PermissionRequest, PostToolUse, UserPromptSubmit, Stop), enforces org-level risk policies on blocking events, and records telemetry to ClickHouse. The plugin generator now produces a downloadable Codex observability plugin (ZIP and install script) that registers the hooks with a Gram marketplace entry in ~/.codex/config.toml. The install instructions dialog gains a Codex tab alongside Claude Code and Cursor.
eb65287: Remove the legacy Speakeasy IDP authentication layer and migrate to WorkOS-native auth. Authorization, token exchange, and session management now go directly through the WorkOS SDK instead of the intermediate Speakeasy IDP proxy. Deterministic UUIDv5 user/org IDs bridge cross-system identity without runtime lookups. Adds OAuth CSRF nonce validation and browser-binding cookie to the login flow.
bbfecc5: Allow adding multiple GitHub collaborators when publishing plugins to a marketplace. The publish dialog accepts a list of usernames as chips, and the publishPlugins API now takes github_usernames (array) instead of github_username (string).
1057ea9: Add OTEL forwarding: customers can configure a URL and headers on the Org Logs page, and a body-tee middleware mirrors every payload received on /rpc/hooks.otel/v1/* to that endpoint. Forwarding is org-wide, async (bounded worker pool, fire-and-forget on failure), capped at 4 MiB per request, and gated behind org:admin for writes / org:read for reads. Header values are encrypted at rest and never returned by the API.
a5e0990: Added support for configuring webhooks to deliver audit log events to external destinations.
Patch Changes
491f3b8: add an opt-in L1 ML prompt-injection classifier (deberta-v3) that runs alongside the heuristic baseline. enable the new "ML classifier (deberta-v3)" rule under the Prompt Injection category in the policy editor to layer the classifier on top of L0 heuristics. detection runs in a sidecar service; configure with PI_CLASSIFIER_URL and PI_CLASSIFIER_THRESHOLD (default 0.9)
7290607: Removed the 1-public-MCP-server cap on accounts without an active subscription. Users can now enable as many public MCP servers as they want on any plan.
ad3c963: /rpc/tools.list now accepts a tool_types filter and can return direct external MCP tools, unblocking the toolset editor's "Add Tools" picker for tools from already-attached external MCP servers.
ec37cf7: quiet false-positive Temporal workflow failure alerts: benign ContinueAsNewError and CanceledError log at Info, and VerifyCustomDomain is non-retryable on NXDOMAIN.
6305bd6: harden AnalyzeBatch against Presidio degradation
44ccc02: The assistant runtime now spills oversized MCP tool results to a file inside the assistant workdir instead of letting them 413 the provider. The in-band tool result is replaced with a pointer ({ truncated, saved_to, original_bytes }) so the model can read or grep the full output via the filesystem tools — no information loss, no provider error.
f872cc2: Drop trigger dispatches whose target assistant has been deleted instead of failing the activity; retrying can't recover a missing row.
44be24a: Fix plugin re-publish so Claude Code, Cursor, and Codex marketplace clients refresh installed copies. Every plugin manifest now ships with a per-publish version (0.1.<unix_ts>) instead of a hardcoded 0.1.0, so platform clients see a newer version on republish and pull the updated content.