[ancongui]

GitHub issue conversations now have separate, finite assistant consent bound to the observed account, project, repository, issue and selected profile. Review shows each mapped user's current read/edit readiness and refuses insufficient membership before writing active consent. Model-capable turns are reserved before provider entry; duplicates, uncertain outcomes and later operator decisions retain their original custody without automatic retry.

Feedback carries the assistant's confirmed reply and questions, preserves slash-separated prose and line breaks, and distinguishes the latest attempt from an earlier completed response. Selected AI communication keeps the complete-candidate writer, separate semantic classifier and one final correction within the existing three-request ceiling. Changed facts or authority hold delivery without a substitute message. GitHub pagination accepts only the verified repository's canonical numeric alias on the same API origin, resource and exact bounded query.

A separate versioned source-only command requires native administrator consent tied to the same issue, repository, run, task, source candidate, profile and SCM provenance. It preserves the SCM caller's identity and the original run and command allowance. Production activation still requires a separately qualified matching capability; this change registers no production native profile.

Studio adds the functional-consent walkthrough and exact-receipt recovery guidance. Version files and release notes move to 0.2.36. The 301-resource catalog retains all 276 historical prompt resources byte for byte, and documentation now matches the 166 registered operator tools. The published read-only patch-index fix and two-tab reservation regression remain intact.

The final implementation also includes the public constants required by the isolated workspace exporter, preserves all historical MCP tool contracts alongside the four additive functional-consent tools, and checks the actual v5 reader/filesystem policy. Oversized assistant feedback uses a clearly labeled excerpt only after validating its complete immutable reference in the same repository; selected-AI failures retain an empty held body. The role-save browser test waits for each exact successful response before asserting the unchanged complete payload.

Validation:

• Exact-head Verify CI: all five jobs passed; Python 4,742 passed, 55 skipped and 4 warnings; Studio 571 unit tests and 782 browser cases passed. Omitted cases are not claimed as acceptance. Documentation CI passed; PR deployment was correctly skipped.
• The actual same-run distribution archive matched its official SHA-256. Configured payload and private-boundary checks verified all 2,142 source files, 619 wheel entries and 2,171 source-distribution entries against the exact commit. All 28 Studio files matched the independently qualified local build.
• The actual downloaded wheel passed 209 controlled SDK, checkpoint, consent and feedback tests using pyfly 26.10.7 and SDK 2.58.0, with package origins checked outside the source runtime. The original signed long-reference PostgreSQL regression passed in the controlled native fixture. Unpublished recovery changes are outside this PR.
• Local qualification retained 524 current canonical screenshots and 16 directly reviewed frames. An initial intermittent external-tools scroll assertion and bounded interrupted attempts are preserved; the assertion remained unchanged and subsequent scoped captures passed. No ordering cause is claimed.

Known nonblocking copy follow-up: two general functional-consent labels still say “CRM run”; they are deferred to the separate UI cut. This qualification does not establish live-model or generated-product acceptance.