rdr-prelock

name	rdr-prelock
metadata	{"argument-hint":"<NNNN> <lens> [run] [--commit | --no-commit] # lens ∈ {grounding, 3amigo, critique, repeatability, cove}; run ∈ {1,2,3,diff}"}
description	Use to run one pre-lock lens and resolve its findings in the same pass. Lenses: grounding, 3amigo, critique, repeatability, cove. Trigger for pre-lock review, $rdr-prelock, or /rdr-prelock.

rdr-prelock — Stage 5+6 (Pre-Lock Review and Resolve, per lens)

Catch the defect classes a template slot can't absorb — PM/UX gaps, time-shifted failures, under-specified signatures, internal silences — and fix them in the same invocation. One lens = one cycle: run the lens, ground each finding, edit the draft, loop to convergence (or flapping). Dispatches into the lens prompts in pre-lock/ for the review half (doesn't redefine them) and runs 05-prelock-resolve.prompt.md for the fix half.

Usage

Codex: $rdr-prelock <NNNN> <lens> [run]        # lens in {grounding, 3amigo, critique, repeatability, cove}
Claude: /rdr-prelock <NNNN> <lens> [run]       # lens in {grounding, 3amigo, critique, repeatability, cove}

run (1 | 2 | 3 | diff) applies to repeatability only — 1–3 generates one run. Variant follows the profile, not the run arg (§repeatability-variant): mid/large = lite (run-1 only, then diff); foundational/escalation = full (run-1/2/3 then diff). Reject an unknown lens with stopped:bad-lens:<value>.

Pick the lens by profile ($RDR_HOME/stages/README.md matrix): mid → grounding 3amigo; large → grounding 3amigo critique; foundational → cove 3amigo critique repeatability (cove embeds the grounding sweep as its Step 0, so it leads and grounding is not run separately). A small RDR runs no lens (skip to /rdr-reconcile). Run lenses in that order — grounding/cove first (ground the frame before the personas debate it). mid/large additionally run repeatability-lite when the Stage 5 Determinacy trigger fires (algorithmic contract — see $RDR_HOME/stages/05-prelock.md); foundational runs the full ×3 repeatability lens.

Precondition. Stage 4 must have verified the assumptions — running a lens on unverified claims wastes it. If Critical Assumptions are still Pending without a plan, stop with stopped:assumptions-unverified and point at /rdr-resolve NNNN.

The cycle (3amigo · critique · cove)

One invocation runs the full loop for one lens:

1. Read rdr-common.md; run §seam-bind + §rdr-resolve. Bind {EVIDENCE_DIR} = <RDR_EVIDENCE>/<RDR_SLUG>/evidence/<lens>/ (§evidence). Re-entry is self-detected: a Status: Draft [revised from Final …; re-verify <IDs>] qualifier → write to iter-N/ (N = 1 + highest existing; loose files = iter-1) and delta-scope to <IDs>.
2. Run the lens prompt (pre-lock/0-grounding.md · 1-3amigo.md · 2-critique.md · 4-cove.md) → it writes element files to {EVIDENCE_DIR}. Heavy/dual-model → sub-agent returns the findings list, not a re-dump. This first run's findings are the origin ledger for the loop.
3. Review gate (below) — a bad pass is re-run on another model, not resolved.
4. Resolve — run the sibling 05-prelock-resolve.prompt.md; it owns the mechanics (grounding gate · origin anchor · durable disposition · tiebreaker-reduction · flag-as-you-go). Don't restate them here.
5. Loop or converge. Substantial fix (a rewrite — can open gaps) → re-run the lens delta-scoped to open ledger entries under iter-N/, then resolve again; a small fix needn't. Converged = no open entries (all fixed / dismissed-with-cite / charted). Cap = 3: still surfacing net-new findings against a barely-changed draft after three iterations is the plank problem — stop with stopped:verdict-flapping:<lens>:<NNNN> (name the churning entries), surface once; the cure is a human look or model switch, not a fourth pass.

Repeatability (the multi-session exception)

Not an in-skill loop — independence comes from separate sessions between runs, not from the invoking session being RDR-naive (the generation prompt reads the RDR anyway). Commit cadence is the §commit exception: each run session commits only its own run-<N>.md (chore(rdr): cli/NNNN repeatability run-N); the doc commit defers to the diff session — see rdr-common §commit.

§repeatability-variant — resolve the variant from profile before writing any run

The intended variant lives on disk, not in session memory — pre-lock clears context and switches models between runs, so nothing you hold survives. run-1.md's variant: header line (written at generation, beside model:) is the durable record; if run-1.md doesn't exist yet, resolve from the RDR's Profile (foundational = full run-1/2/3; mid/large = lite run-1 only) and stamp it into that line. Never infer the variant from how many run-*.md files exist — that is what silently promotes a large RDR to full ×3. A run-2/run-3 request against a variant: lite header stops with stopped:repeatability-lite-no-run-2:<NNNN> — escalate to full only by rewriting that line to full (escalated: <reason>) first (3-repeatability.md Escalate), else diff on run-1.

• Run the generation prompt directly — bind {RDR_PATH}, {EVIDENCE_DIR}, <N> and execute 3-repeatability.md. Write run-<N>.md and stop. One session writes exactly one run; never write a second run in the same session.
• Next run needs a fresh session (full only) — after writing, stop with stopped:repeatability-needs-fresh-session:run-<N+1>; the next /rdr-prelock NNNN repeatability <N+1> invocation in a fresh session writes the next run, relaunched on the alt model for the cross-model draw. Lite stops after run-1 with next Next: /rdr-prelock NNNN repeatability diff (never a run-2 pointer); do not continue to reconcile.
• Diff only when complete + clean — completeness is judged against the profile's variant (§repeatability-variant), not the file count: lite needs run-1, full needs run-1/2/3. This session wrote none; else stopped:repeatability-incomplete:<missing>. ≥1 run on a different model.
• Resolve once diff.md lands — this same skill runs the resolve prompt on diff.md; its REPEATABILITY DIFF clause governs each divergence. If autocommit is on, the diff session does the doc commit (docs(rdr): prelock cli/NNNN — repeatability) and one evidence commit over the whole repeatability/ dir (chore(rdr): … repeatability evidence) — self-healing: it sweeps any run files an earlier session left uncommitted (no-op guard skips ones already in). See rdr-common §commit.

Review gate (Stage 05-prelock.md)

• Read findings against the lens's own Expected signal. Healthy = concrete, named passages. Generic advice / over-agreement / identical persona lists → switch model and re-run; don't proceed on a bad pass.
• After resolving: edits stayed brief (no change-history narration); fixes were grounded (a fix to satisfy a finding whose cited code doesn't exist on main is wrong); the needs-verification list is honest; net-new scope was charted, not folded in; tiebreakers surfaced only when evidence was indeterminate.

Next step (rdr-common §next-step)

• If autocommit is on: non-repeatability lenses run §commit for prelock <lens> here (doc + a separate <lens> evidence commit). repeatability follows its own cadence — run files commit per session, the doc commit + a whole-repeatability/-dir (self-healing) evidence commit land at the diff (see above / §commit).
• Before printing Next: re-read the RDR's current Profile field and bind the full lens-set from the matrix above. Then scan Normative Contracts for the Stage 5 Determinacy trigger; for mid/large, append repeatability-lite unless evidence/repeatability/ already contains run-1.md + diff.md or a one-line determinacy: n/a - <reason> disposition. Treat that as the required checklist, then subtract only completed lens evidence under <RDR_EVIDENCE>/<RDR_SLUG>/evidence/; for repeatability, a folder alone is not complete. This is mandatory after a reset, demotion, or escalation: an RDR that became foundational still owes cove 3amigo critique repeatability even if it previously ran the mid or large subset, and a completed critique is not the end of Stage 5 unless repeatability is also complete. critique itself isn't "complete" from a folder alone: read the evidence's Model: stamp (rdr-common §model-stamp) — a foundational RDR needs the dual-model diff (or recorded single-model fallback), and a re-entry under a different model is the second pass to run, not a no-op.
• 3amigo | critique | cove converged → next lens is the first missing item in the current checklist: Next: /rdr-prelock NNNN <next-lens>.
• repeatability — pick the variant from Profile (§repeatability-variant), never from files present. Lite (mid/large): missing run-1 → repeatability 1, then diff — never point at run-2. Full (foundational/escalation): missing run → Next: /rdr-prelock NNNN repeatability <N+1> until run-1/2/3, then diff. Diff runs in a fresh session; it also resolves diff.md.
• All profile lenses and Determinacy obligations done → Next: /rdr-reconcile NNNN (carry the needs-verification list).
• stopped:verdict-flapping → resume after a human look / model switch, or chart the churning entry.
• A finding refuted an assumption → it's on the needs-verification list (Stage 6); if it forces a redesign, /rdr-resolve NNNN / /rdr-propose NNNN now.
• /rdr-status NNNN to re-orient.