rdr-doctor

rdr-doctor — verify the seam + engine

Read-only health check. Runs the same resolution every /rdr-* skill's §seam-bind does, but checks every invariant (never stops at the first failure) and reports each with the one command that fixes it. The answer to "a skill said stopped:… — what's wrong?" It writes nothing — sources the marker, stats paths. Safe anytime, from any repo or worktree.

Run it — one Bash call

All checks live in a script file shipped with the engine (doctor.sh, beside this SKILL.md). Run that file — do not paste its body into a shell (pasting a long quoted block is what made earlier runs drop their labels). Resolve the engine the same two ways §seam-bind does (the doctor can't use the seam to find itself), then run it:

GC=$(git rev-parse --git-common-dir 2>/dev/null) && GC=$(cd "$GC" && pwd -P)
WS=$(dirname "$(dirname "$GC")")
for H in "$CLAUDE_PLUGIN_ROOT" "$CODEX_PLUGIN_ROOT" "$WS/rdr"; do [ -f "$H/skills/rdr-doctor/doctor.sh" ] && { sh "$H/skills/rdr-doctor/doctor.sh"; break; }; done

Under Codex, if $CODEX_PLUGIN_ROOT is unset, use this skill's filesystem-backed source locator instead: run the doctor.sh that sits beside this SKILL.md.

That is the whole invocation — one sh <file>, no embedded quoting to mistranscribe. It prints the full report (pure ASCII, one line per check + a Verdict). POSIX sh, so identical under bash (Linux) and zsh (macOS).

Print the script's stdout verbatim inside a fenced block — it is the report (pure ASCII, one line per check + a Verdict). Do not reformat it into a table, re-run it to "confirm", or paraphrase the lines — that's wasted turns. The first run is the answer; if a line says FAIL, surface its fix as-is, don't improvise a repair.

Corpora (optional, after a clean structural run). If arc is present, list the named corpora in $RDR_RESOURCES and note any that don't resolve as [WARN] (degraded mode — research stages run hollow until built). Never a FAIL; skip if no arc.

Domain priors (optional, after a clean structural run). Read $RDR_RESOURCES and check its Domain priors section: if it is absent, empty, or names no competitor/prior-art/standard for the project's problem space, report [WARN] domain priors thin — Propose enumerates approaches from the model prior, not prior art. This is the breadth gap: with no priors to reach for, Propose invents candidate approaches from training memory and Resolve can only verify the winner of a race the strongest alternatives may have sat out. Never a FAIL (the priors are the user's to populate); the one fix is "populate Domain priors in $RDR_RESOURCES with the competitors/prior art for this domain."

Review gate

• Read-only? No file/symlink written, nothing installed — the doctor diagnoses; /rdr-init (or the user) repairs.
• Complete? Every check ran; one FAIL never skips the rest.
• Each FAIL names one fix? No "might be" — the command.

Next step (rdr-common §next-step)

• All PASS → Next: /rdr-status, or /rdr-seed <idea>.
• Any FAIL → run the named fix (usually /rdr-init), then re-run /rdr-doctor.