When to Use

Whenever you are about to act on a belief about the state that you did not just verify. After a compaction. After a handoff. Returning to a session left open. Before trusting a summary.

The split this skill enforces

Observing is mechanical and the model must not do it. Orienting is judgement and a script must not do it. Every token spent having a model rediscover which branch a tree is on is a token not spent on the only part that needs a model, which is noticing that two records disagree.

That split is the cost argument. An unstructured reorientation runs eight to twelve exploratory commands and reads all their output. This runs one.

Phase 1: Observe, with the tool that already exists

scripts/reorient          # from the workspace root, or name it: scripts/reorient ROOT
scripts/reorient --fetch  # when remote state matters

One command. Read its output. Do not go looking for more. It ends with a COVERAGE block naming what it read and what it deliberately did not; if something you need is outside that block, the fix is to extend the tool so the next run has it too, not to run an ad hoc command that leaves no trace.

Check for this tool before writing another one. This skill was first written around a second, near-identical tool built without looking, on a tree that already had this one and a flagged thread describing it. The duplicate was deleted the same session. Prior art in this tree is usually an entry in scripts/ or bin/, and, where the tree has a justfile, a just recipe that just --summary lists.

Phase 2: Orient

Four questions, answered against that output. Skip any with nothing to say.

1. What moved that I did not move? A branch tip, a capture, mail, a dirty file, a standing file older than the thing it describes. This one is always worth asking.
2. Where does the record disagree with the tree? A summary, a capture, a queue or a register saying one thing while the tool says another. Name the disagreement rather than silently taking the tool's side, because which one is wrong is itself the finding.
3. What did I carry in that does not survive this? State the assumption you arrived with and whether it held. A carried assumption that is never stated cannot be falsified.
4. Does the next concrete action change? Either name the new one or say the old one stands.

Cost discipline

The report is a short message, not a document. Ten lines is normal and one line is fine: "Nothing moved since the capture; the next action stands." A long orient report means orienting turned into working, which is the failure this skill is shaped to avoid.

Write nothing to disk. Anything worth keeping belongs to session-log or session-close, and a finding recorded in two places goes stale in one of them.

Why this one may be model-invoked

daily-dashboard sets disable-model-invocation because it chooses a day's priority, and that is a person's call. This skill chooses nothing. It reads and reports, and the compaction boundary that most needs it is an event only the agent notices. So it is available to the agent, and it still ends without starting any work.

Scope Pointer

• daily-dashboard: the full start-of-day procedure. It reads the standing files, verifies carry-over, ranks todos and ends in questions. This is the cheap subset answering only "where are we", and running it does not discharge that one.
• session-log: the write half of the same boundary. This reads state, that records decisions and findings, and a compaction wants both.
• session-close: end of session. Nothing here substitutes for any part of it.