Cascade dedup

Cascade dedup

The cascade family replicates one canonical prompt across three harnesses. Dedup means: verify the intentional replication byte-for-byte, strip the accidental repetition, and resolve conflicts with the canonical as winner.

File	Role	Zones
/home/alpha/.claude/claude/system-prompt-baseline.md	Canonical; wins every conflict	Whole file is the source of truth
output-styles/{axiom-mode,builder,duet,linus,odin,benchmark}.md	Style embeds	Persona prefix above the charter <role> (the second <role> per file) = strip zone. Tail from the charter <role> = byte-identity invariant zone, never a dedup target
output-styles/benchmark.md auto-gen preamble (margin-runner header)	Generated	Never touch. Repairs to benchmark.md's cascade tail require explicit user authorization first
/home/alpha/.omp/agent/AGENTS.md, /home/alpha/.codex/AGENTS.md	Harness embeds	Harness-adapted tool sections are legitimate divergence; targets are internal duplication and accidental drift. Outside the repo: editable, never committable

Step 1 — Verify the invariant before dedup

Extract the canonical: cp system-prompt-baseline.md /tmp/canon.md (its charter <role> is line 1, so the canonical is the whole file). Per output-style:

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

Drift here is a sync bug, not duplication. Repair means 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, get explicit user authorization before repairing.

Completion criterion: 6/6 byte-identical (after repair where needed), each diff result recorded.

Step 2 — Strip-zone scan (persona prefixes)

benchmark.md has no eligible strip zone — its entire persona prefix sits inside the auto-generated margin-runner block — so Step 2 covers the five hand-authored styles and skips benchmark.md completely. Classify every sentence of each persona prefix as voice | duplicate | conflict | 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.

Completion criterion: every prefix sentence classified; every strip cites its baseline pair.

Step 3 — Harness embeds

Per external file (~/.omp/agent/AGENTS.md, ~/.codex/AGENTS.md):

1. Internal duplication within the file, same thresholds as Step 2.
2. 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 the baseline wording.
   • conflict — the baseline wins, unless the divergence exists to fit harness-specific tooling.

Completion criterion: a divergence ledger with exactly one classification per divergence — none unclassified.

Step 4 — Apply and verify

1. Repo files: one atomic commit. Any benchmark.md change (tail repair included) ships only with explicit user authorization recorded in this run. Re-run the Step 1 diffs after editing; run prek run --all-files.
2. Strips and conflict resolutions are doctrine changes → patch-bump all three manifest version fields (plugin.json .version, marketplace.json .version and .plugins[0].version) from the origin/main base in the same commit. A run that only re-syncs an embedded baseline, zero strips, is a pure sync change → no bump.
3. External files: edit in place, never commit; end the report with an explicit warning listing every externally edited path for user review.
4. Output styles load at session start — smoke-test doctrine effects in a fresh session.

Completion criterion: 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.