Cascade Dedup

Cascade dedup

Contract

Inputs

• The canonical baseline file system-prompt-baseline.md, must exist and be the agreed source of truth.
• The six output-style embeds: plugins/odin-core/output-styles/{axiom-mode,builder,duet,linus,odin,benchmark}.md.
• The two external harness embeds: ~/.omp/agent/AGENTS.md and ~/.codex/AGENTS.md.
• Explicit user authorization, recorded in the run, before any benchmark.md tail repair or authorized external embed edit.

Procedure

1. Map the cascade family and zones. The canonical baseline system-prompt-baseline.md is the whole-file source of truth and wins every conflict. For each output-style embed (axiom-mode, builder, duet, linus, odin, benchmark), the persona prefix above the charter <role> (the second <role> in each file) is the strip zone. The tail from the charter <role> to EOF is the byte-identity invariant zone and is never a dedup target. Never touch the benchmark.md auto-generated margin-runner preamble. The external harness embeds ~/.omp/agent/AGENTS.md and ~/.codex/AGENTS.md are editable in place but never committable; harness-adapted tool sections are legitimate divergence. Done when: the cascade family is mapped with strip zones, invariant zones, and external embeds identified.
2. Verify the invariant before dedup. Copy the canonical baseline to a temporary file. For each output style, diff the canonical against the file's tail of equal byte length:

   diff -q /tmp/canon.md <(tail -c "$(wc -c < /tmp/canon.md)" plugins/odin-core/output-styles/X.md)
   

   Drift in the invariant zone is a sync bug, not duplication. Repair by replacing the file's entire tail (from its charter <role> to EOF) with the canonical content as one block; dedup edits inside the invariant zone stay forbidden. For benchmark.md, obtain explicit user authorization before repairing. Record each diff result. Done when: all six tails are verified byte-identical (after repair where needed), or benchmark.md repair is deferred for lack of authorization.
3. Strip-zone scan (persona prefixes). benchmark.md has no eligible strip zone (its persona prefix sits inside the auto-generated block), so scan the five hand-authored styles and skip benchmark.md. Classify every sentence of each persona prefix as voice, duplicate, conflict, or unique-directive:
   • Normalize (lowercase tokens, stopwords removed). Jaccard >= 0.65 against a baseline rule -> duplicate: strip it; the baseline copy stands.
   • Jaccard >= 0.45 plus opposing modal verbs (must/never, always/must-not) -> conflict: the baseline wins; delete or rewrite the prefix line.
   • Persona voice (identity, tone, register) is not a directive: keep.
   • Unique directives (persona-specific rules with no baseline pair): keep. Done when: every prefix sentence is classified and every strip cites its baseline pair.
4. Harness embeds. Per external file (~/.omp/agent/AGENTS.md, ~/.codex/AGENTS.md):
   • Internal duplication within the file, same thresholds as step 3.
   • Directive-level comparison against the baseline. Classify each divergence: harness-adaptation (names that harness's tools or commands: keep), accidental drift (same rule, mutated wording: align to baseline wording), conflict (baseline wins, unless the divergence fits harness-specific tooling). Done when: a divergence ledger with exactly one classification per divergence is produced, none unclassified.
5. Apply, verify, then commit.
   • Apply all repo edits: strips, conflict resolutions, tail repairs, and the releaseVersion bump if strips or conflict resolutions occurred. Let the package-surfaces generator (scripts/render-plugin-surfaces.mjs) re-render every package surface (plugins/<id>/plugin.json and .claude-plugin/plugin.json) from that one literal. A run that only re-syncs an embedded baseline with zero strips is a pure sync change -> no bump. Any benchmark.md change (tail repair included) ships only with explicit user authorization recorded in this run.
   • Run verification before committing: re-run the step 2 invariant diffs (all six tails must be byte-identical 6/6) and run prek run --all-files. If either fails, do not commit; report the failing diff or check and the file responsible.
   • Only after verification passes, commit all repo files in one atomic commit (edits plus regenerated surfaces).
   • External files: edit in place, never commit; end the report with an explicit warning listing every externally edited path for user review.
   • Output styles load at session start: smoke-test doctrine effects in a fresh session. Done when: the final report contains strips per file with citations, conflicts resolved with the winner named, kept harness-adaptations, externally edited paths, the prek result, and the invariant re-verified 6/6, with the atomic commit created only after verification passed.

Failure and recovery

• Invariant-zone drift detected: a sync bug, not duplication. Repair by replacing the entire tail as one block; never edit inside the invariant zone. If the canonical itself is suspect, stop and surface the discrepancy rather than rewriting the canonical.
• benchmark.md repair without authorization: do not touch benchmark.md (tail or preamble) until explicit user authorization is recorded in the run. Without it, leave the file unchanged and report the needed authorization.
• Ambiguous classification (Jaccard in the 0.45-0.65 band without opposing modals): do not strip; classify as unique-directive and report it for human review.
• prek failure or invariant not 6/6 after edit: the done predicate does not hold. Do not commit; report the failing diff or check and the file responsible.
• Partial-result rule: a run that completes classification but fails final verification produces no commit. The divergence ledger and classification record are still reported as the partial result; the invariant must be re-verified before any commit.
• Rollback: verification runs before the atomic commit, so a failed verification produces no commit to revert. Uncommitted repo edits are discarded with git checkout -- <files>. External file edits are not VCS-tracked; record the original content before editing and restore from that record on failure.

Output

A final report containing the 6/6 invariant re-verification diff results, strips per file with baseline-pair citations, conflicts resolved with the winner named, kept harness-adaptations, the complete divergence ledger (one classification per divergence), every externally edited path flagged for user review, the prek run --all-files result, and the releaseVersion bump decision (bumped literal plus regenerated surfaces, or no-bump rationale); repo changes land as one atomic commit, external edits land in place and are never committed.