| name | rdr-resolve |
| metadata | {"argument-hint":"<NNNN> [--commit | --no-commit]"} |
| description | Use to verify critical RDR assumptions against source, spikes, prior art, and reuse audits. Runs Stage 4 before pre-lock review. Trigger for resolve assumptions, $rdr-resolve, or /rdr-resolve. |
rdr-resolve — Stage 4 (Resolve Assumptions)
Verify every Critical Assumption against source, live spikes, prior art, and
standards so the pre-lock rounds critique a true document. The load-bearing
front-half stage. Running a pre-lock round before this one wastes the round.
Usage
Codex: $rdr-resolve <NNNN>
Claude: /rdr-resolve <NNNN>
- Read
rdr-common.md; run §seam-bind + §rdr-resolve
to bind $RDR_ENV, $RDR_RESOURCES, RDR_PATH, RDR_SLUG, {SPIKE_DIR}.
- Run the prompt
04-resolve.prompt.md
with those bound. Have a live spike target reachable. The prompt is read-heavy —
delegate corpus searches / reuse audits / dependency spelunks to a sub-agent that
returns verdict + evidence pointer (§return-packet), not raw hits (rdr-common §delegation; spawn
with the Task tool).
- Fan-out heuristic (mined): group the Source-Search assumptions into one
sub-agent (multi-pass, one verdict each) and give each Spike assumption its
own sub-agent (it needs the live target). Don't spawn one agent per source-search
claim — they share the same corpus/source reads. Hold only the verdicts in the
main context; author the
Status: Verified edits here, not in the sub-agent.
- Self-detected re-entry (no flag). The prompt reads the RDR
Status: line: a
Draft [revised from Final …; re-verify <IDs>] qualifier scopes the run to the
listed assumption IDs (+ anchors the demotion touched), carrying the rest forward
as Verified; a bare Draft verifies every assumption cold. Do not pass a resume
flag — the RDR's state drives it.
Review gate (what the human checks — Stage 04-resolve-assumptions.md)
- Every load-bearing assumption actually
Verified, Evidence not self-reference and
not bare Docs Only.
- Spikes actually ran (command + captured output under
{SPIKE_DIR}).
- The reuse audit ran against the
$RDR_ENV reuse-audit paths.
- No research finding contradicts the approach — if one does, that's not a
citation fix.
Next step (rdr-common §next-step)
- If autocommit is on, run §commit for
resolve first (+ a separate spike-evidence commit if a spike wrote; see the rdr-common table).
- Research refuted the approach or surfaced existing capability → back to
/rdr-propose NNNN (or /rdr-refine NNNN); rework, then re-run this.
- Assumptions verified → forward. This stage sets the
Profile Metadata
field (the routing latch) from the contract count; the next pointer reads
that field — it does not re-derive size ($RDR_HOME/stages/README.md matrix):
Profile: small → skip Pre-Lock: Next: /rdr-reconcile NNNN.Profile: mid+ → the first lens of the profile's set (grounding runs
first whenever present): mid/large → Next: /rdr-prelock NNNN grounding;
foundational → Next: /rdr-prelock NNNN cove (cove subsumes grounding as
its Step 0). Then the remaining lenses in matrix order
($RDR_HOME/stages/README.md).
/rdr-status NNNN to re-orient.
