[IanYet]

HOLD: unpublished review candidate. This Draft targets the existing unpublished Doctor feature branch behind #27, not main. Do not merge it into a formal Git marketplace source, install it for users, package it or release Doctor. #27 and its existing branch history remain unchanged.

Hosted Doctor follow-ups need a stable original report and saved observations of the actual native tool result. Use the explicit original report and its SHA as the baseline; append each real supplied tool observation in a separate file without changing the original. Keep the actual host/scope/service/project and exposed call identifiers where available, and preserve unknowns. A supplied result is not independent proof of the original session, production account authentication or recovery.

Provide optional offline capture and feedback from existing reports. Reuse one canonical collector with generated self-contained copies, aligned with the collector 1.2.6 candidate in #29. Existing report files and sidecars remain intact; the original report is explicitly referenced for each follow-up.

Saved-report delivery has a single owner and a separate SHA-bound receipt. An attempted POST without a valid complete ACK, including HTTP500/503 or a malformed/interrupted response, remains delivery-uncertain and blocks a blind resend. No-attempt validation failures retain local evidence. ACK still does not prove complete storage. Content-only native MCP results can be saved without inventing projects[] or structuredContent; explicit failed results/configuration/tool discovery/independent probes cannot be recorded as a successful native recheck.

Validation on Windows:

• Implementer compatible source suite: 156/156 passed; 4 known media fixtures containing '?' in filenames are excluded and their earlier full-suite failures remain disclosed.
• Root independently reran 49 selected source tests covering capture, report identity, follow-ups, upload ownership/ambiguity and source-input resilience: passed.
• Generated helper consistency and git diff --check: passed. Plugin manifests/versions are unchanged.
• Tests use controlled local files, mocked HTTP and owned loopback services. No production report upload, installed plugin change, model call, login or deployment.

Still required before any release decision: actual natural-language triggering in target Codex and Claude sessions, justified repair of a real evidenced failure, an actual read-only tool call in the original host/scope/service/project, saved linked recheck results, and appropriate production storage verification. This Draft must not be described as real-user recovery or completed end-to-end acceptance.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}