[tzhouam]

Copilot and ReviewBot currently duplicate backend invocation, knowledge validation and review orchestration. This refactor gives Copilot one implementation for those capabilities and leaves scheduling, GitHub business state, publication and delivery receipts in ReviewBot.

• Route SDK/CLI/MCP through the application workflows and durable ReviewRuntime; retain delegating Direct/Strict compatibility APIs.
• Consolidate provider transports, structured decoding, accounting and process supervision. Compose read-only agent sessions through the same local/SSH Docker abstraction.
• Share knowledge parsing, evidence normalization, validators, bounded curation and quality execution while retaining each init stage's acceptance policy and append-only/lifecycle rendering contracts.
• Share review checks and publication eligibility, including stale-head refusal; inject domain policy into the generic execution kernel and remove superseded implementations.

This is the framework delivery PR. Real model quality, remaining native vendors, deep production recovery and release/rollout acceptance remain separate. The user explicitly deferred the proposed additional 182 model calls; no new model dispatch is authorized by this PR.

Paired consumer

ReviewBot companion PR: https://github.com/JiusiServe/omni-reviewbot/pull/249. Merge/release coordination is required; this PR does not update production deployment pins.

Validation

Historical full-suite production snapshot: ec81d9278f36151d3f5cd195ae9558df26fe9cc4; these full-suite results predate current-main integration.

• Full Copilot suite: 6023 passed, 4 environment skips, no failures/errors. Corresponding installed-resource/real Linux namespace complements: 5 passed; original skips retained.
• Paired ReviewBot suite with this installed SDK: 2925 passed.
• Fresh isolated wheel environment: 114 selected mainflow tests, public SDK release smoke and pip check passed. A separate SHA256-locked offline installation also passed.
• Actual Direct application local/SSH Docker configuration and scoped bridge controls stopped before model dispatch. Local control's erroneous directory postcondition is retained and verified by read-only inspection of its original canceled attempt; it was not rerun.
• Production Python sources and installed payloads match the validated wheels. Mock/model-boundary tests do not establish real model quality or production deployment readiness.

Review entry points: architecture/ownership and deletion handoff, full paired integration, complete plan, source/AST changes. Historical reports retain their original artifact scope.

Current-main integration

Resolved conflicts with upstream 6a6d87fb0bab609552e6907c1f747f764a0211ff and pushed 00c65b0f61761c6115814fda408d9eac691a02a0. Preserved the shared ZCode supervisor, upstream typed errors, lazy SDK exports and both resource sets. RFC CLI/MCP now consume the shared application client; the public SDK re-exports it.

Merged-tree checks: 162 affected ZCode/SDK/MCP tests passed; RFC/process/boundary checks had 365 passed and 1 Playwright collection skip; 145 paired ReviewBot checks passed against the newly installed SDK. Wheel installation and merged-resource probes passed. These are affected-scope checks, not a new full-suite or real-model-quality claim.

Conflict resolution and validation record. Live CI is reported separately by GitHub. This remains a framework draft; production pins and deployment are unchanged.

Documentation CI repair

Converted 237 host-local evidence links into labeled inline paths and rendered the quoted invalid model-generated link as text with its original target. Artifact identities and historical results are unchanged; no private/local artifacts were published. The link checker remains unchanged. python tools/check_doc_links.py passes for all 529 documents, both in the working tree and a separate tracked-files-only snapshot without local output artifacts. Production code is unchanged by this documentation repair. GitHub CI has been retriggered.