Code

Shared-vendor OAuth scope ceiling is store-wide, not caller-scoped

ci: scan changed files for malicious code and disguised content

Agent ignores follow-up question after failed run

Multi-tool meeting research fails after retrieving calendar data

Agent exposes internal implementation details in user-facing response

Telegram messages render raw Markdown instead of formatted text

Hosted MCP: omitted destructiveHint is read as false, but the MCP spec defaults it to true

[QA] Platform-owned WIT and extension package changes fail the Reborn scope classifier

Configure all tools, channels, and extensions from AI chat as an Admin

Telegram voice notes and stickers fail the entire update parse (InvalidExternalRef)

Design System — Phase 2: DESIGN.md governance & guidelines

[QA] WASM guest diagnostics can expose detectable secrets through runtime, model causes, and tracing

Model budget enforcement is not wired in production — daily USD caps unenforced since #6174

[QA] Failed lazy delivery recovery is not retried within a coordinator lifetime

[QA] Report host-mediated egress ignoring ambient proxy variables in operator diagnostics

[QA] Concurrent coordinators can both send the same durable delivery attempt

[QA] Interrupted-delivery recovery can overwrite a concurrent Delivered status

[QA] Ambient proxy env vars bypass DNS-rebinding protection in ReqwestNetworkTransport

Time awareness without prompt-cache churn: append-only rollover context and duration evidence

extension_manager: five pre-existing findings surfaced by the WS2.4 split (false WriteFilesystem effect, untested lock predicate, two missing dispatch tests, six dropped causes)

WS5: close the attachments row — re-word §6.4.9 to exclude the read adapter, or re-home LoopAttachmentReadPort first

Split the product_wire DTO family in ironclaw_product_contracts (large_file exemption owner)

Changed-coverage gate: steering-queue slice error paths are crate-tier-only (no integration fault injection)

Monthly issue metrics report

Compaction: summarization inference must not pollute prompt cache or session affinity

Token accounting: hybrid provider-usage + tail estimates; fix ModelWorkRequest estimating from the content reference string

Compaction: derive context budget from the actual model window instead of hardcoded 128k

Cache: regression test pinning byte-identical prompt prefix across turns

Cache: keep the advertised tool array byte-identical — defer_loading/tool_reference instead of mid-run promotion

Add `hub` as alias for the `ironhub` CLI subcommand (IronHub dashboard compatibility)