RECOVERY INSTRUCTIONS

# Recovery Instructions

This file is the active working-memory surface for non-canonical continuity.

Keep here only the smallest recovery payload that would be costly to reconstruct after compact, reset, handoff, or surface drift.

Use it only for signal that has already passed the feedback loop strongly enough for repo carriage but is not yet ready for a clearer long-term home.

Do not treat this file as canon, a backlog, or a long lessons log.

If the signal has not yet passed that loop strongly enough, keep it in the local temporary restore overlay instead.

## Current State

Status:
- Provisional and derived from the current feedback-loop and tooling-stability thread.
- Current Anchor role-development continuity is under feedback-loop audit before any further mutation.
- For the current Anchor role-development thread, broader multi-role feedback-loop coverage is the default expectation rather than a single human-observer lane alone.
- For the current Anchor role-development thread, no next slice begins without explicit human approval; human roles in the chat may intervene before progression.
- The current next-stream prep assumes a no-sound stream and must remain readable to a cold viewer without spoken narration.
- Current candidate status should remain provisional: bounded AI-side holds do not count as human approval or as a `working candidate` label.

Hard:
- Treat `flooding -> compact -> inference drift -> worse decisions -> more drift` as an active stability chain, not a cosmetic quality issue.
- Keep the current error classes separate: `routing error`, `recoverability error`, and `propagation error`.
- `routing error` was intentionally handled first in `ai-vscode-tools`; repo-level routing guards now exist there.
- The next repo-level slice is `recoverability error` in this AI repo, not broader role propagation.
- For the current Anchor candidate, accepted content remains disposable; if process drift or chat-carried inference threatens recoverability, discard and re-ground from the handshake and referenced files.
- The latest process miss was underuse of the maintained feedback loop on the latest candidate change, not a defect in the maintained loop itself.
- For the current Anchor role-development thread, do not let `Source role` collapse into a default `Human observer` lane when broader role coverage was expected for the exercised loop.
- If the current prep chat is deleted to create an honest cold start, only the smallest externalized carry-surface should survive: process philosophy in maintained transition artifacts and the exact next-stream pack in the YouTube repo.
- For the current role-development stream prep, the next investigation target is Kodax as a genuine failed slice, but the stream must not stop at failure display alone; the target arc must also show positive demonstrated ability after the boundary repair.

Soft:
- Open questions should be answered before action when the question itself is still unresolved.
- Tooling-first evidence lanes should outrank raw transcript or raw JSON reading when they can answer the same question with bounded output.
- Recovery capture should happen earlier than perfection would normally tempt.
- Human acceptance of the latest candidate slice does not by itself imply retention or permission to progress.
- Stream maintenance through `youtube/tools/stream-state.mjs` is part of the continuity handshake and should stay aligned with the current validated phase.
- Human approval to continue may be broader than one person; human roles present in the chat may intervene in the loop before any next-slice progression.
- For a no-sound stream, explicit state transitions and failure snapshots are part of inference transfer, not cosmetic presentation.

Next:
- Keep this file as the shortest durable carry-surface for active non-canonical inference.
- Move stable parts of this state into the smallest correct canon-bearing files once their destination is clear.
- Only then lock broader deterministic role-propagation changes.
- Verify correct feedback-loop use on the current Anchor candidate slice before any further mutation.
- Keep the next conversation grounded through the primary handshake, referenced files, and current stream state rather than this chat.
- Preserve the broader multi-role default and the explicit human approval gate when the next conversation resumes this Anchor thread.
- If the next conversation resumes no-sound stream preparation or a live role-development stream, read `youtube/assets/posts/technical-feedback-loop.md` and `youtube/assets/posts/next-stream-kodax-pack.md` before proposing new framing.
- Preserve the distinction between `candidate held on bounded AI-side testing` and `human-approved`; do not collapse provisional AI-side progress into approval language.

Avoid:
- Treating this file as canon.
- Letting this file become a running archive of every discussion.
- Re-normalizing raw session JSON as the default evidence lane when tooling can provide bounded inspection.
- Starting broad role rewrites before recoverability is stable enough to protect the result.
- Treating a smoke check or subagent reading as completion of the feedback loop.
- Letting accepted candidate text, chat summaries, or chat approval become the continuity surface.
- Treating one human-observer lane as the full intended loop when broader multi-role coverage was part of the current thread's expectation.
- Treating silence after a slice as permission to begin the next slice without explicit human approval.
- Treating a cold-start cleanup as permission to rely on deleted chat reasoning instead of the externalized pack and maintained process artifacts.