AGENT SKILL

ramza-parallel-spec

TRANCE-only G3 evaluator-optimizer mode: fan out 2-4 perspective-diverse clean-context candidate specs, score with a bias-hardened evaluator anchored on ramza-score, judge-merge into ONE spec, terminate at confidence ≥85% or hard cap 3 iterations. Read-only; never the default — runs ONLY when the cortex authorizes TRANCE.

License 许可
Apache-2.0
所属仓库
Rynaro/magicite
最近更新
2026-10-14
Stars
★ 2
Forks
0

01 前言 Front Matter

name
ramza-parallel-spec
metadata
{ "phase": "TRANCE", "methodology": "RAMZA" }
description
TRANCE-only G3 evaluator-optimizer mode: fan out 2-4 perspective-diverse clean-context candidate specs, score with a bias-hardened evaluator anchored on ramza-score, judge-merge into ONE spec, terminate at confidence ≥85% or hard cap 3 iterations. Read-only; never the default — runs ONLY when the cortex authorizes TRANCE.

03 SKILL.md 正文

RAMZA — Parallel Spec Mode (G3 evaluator-optimizer)

This mode wraps the standard RS→S→P→E→C→T→R→A cycle — it does NOT replace it. It operationalizes the G3 TRANCE form named in the nexus trance-matrix ("Generator + evaluator + termination gate; cap 3 iterations").

Activation — TRANCE-GATED (never default)

This mode is TRANCE-gated. Do NOT fan out candidate specs at standard tier; run the single-pass RS→S→P→E→C→T→R→A cycle instead.

Activate ONLY when the cortex authorizes TRANCE — i.e. BOTH complexity AND stakes flags hold: complexity 10–12 (per ramza-score --rubric complexity) STRATEGIC/CHANGE specs, multi-service architecture, or high-rework-risk system design. At standard tier, fall back to the normal cycle.

Read-vs-write safety

RAMZA is READ-ONLY in every phase (DESIGN-RATIONALE.md D2), so the parallel generator branches are the explicitly-SAFE parallel-READ case — NO worktree isolation is required (R1-01). This is distinct from Vivi/APIVR-Δ's parallel-WRITE, which does need worktree isolation.

The cycle: GENERATE → EVALUATE → JUDGE-MERGE → TERMINATE

1. GENERATE (perspective-diverse fan-out)

Fan out 2-4 clean-context generator branches (default 3, hard cap 4). Each branch receives the SAME Scope + Pattern context but a DIFFERENT assigned perspective, producing a full candidate draft spec:

  • one conservative (low-risk, proven)
  • one pattern-leveraging (maximize reuse of existing conventions)
  • one innovative (novel approach)
  • optionally one risk-minimizing (blast-radius-first)

Perspective diversity is deliberate, not naive N-identical sampling — quality dominates raw diversity, so cap at 3-4 high-quality branches, never spawn N clones (R3-04, R3-06; cost-ceiling C1 max 5 → capped at 4 here because spec generation is expensive). Use clean-context subagents so branches cannot self-condition on each other's trajectory (R1-03).

2. EVALUATE (bias-hardened scoring)

A single evaluator scores each candidate through the tool — never hand-computed:

echo '{"alignment":9,"correctness":8,"maintainability":7,"performance":7,"simplicity":6,"risk":7,"innovation":5}' \
  | ramza-score --rubric explore --state <state> --label "<candidate-perspective>"

The 7-dimension weighted rubric (Alignment 25% + Correctness 20% + Maintainability 15% + Performance 15% + Simplicity 10% + Risk 10% + Innovation 5%, see docs/methodology/scoring.md) is arithmetic done by ramza-score, not the evaluator — the evaluator's job is producing honest per-dimension inputs while applying EXPLICIT LLM-as-judge bias mitigations (R3-09):

  • Strip authoring-branch identity/label from each candidate before scoring (counters self-preference / identity bias).
  • Randomize / rotate candidate presentation order across the comparison (counters position bias).
  • Length-normalize — judge on rubric-dimension content, not verbosity (counters verbosity bias).
  • Anchor on deterministic checks — prefer ramza-lint / ramza-ears-lint and the dependency/constraint layers already in the Test (T) phase as the trust anchor over pure LLM judgment.

Record the per-candidate ramza-score results (each call appends to the state file's gates[] and the calibration log) and the bias mitigations applied, so the evaluation is auditable in the final spec.

3. JUDGE-MERGE (mandatory aggregation)

Synthesize ONE spec by taking the highest-scoring approach per dimension:

  • Record which candidate won each dimension and why, via [DECISION] markers (per-dimension provenance table).
  • Carry every rejected candidate's rationale into the Rejected Alternatives section (a full-tier-required section, see docs/methodology/tiers.md) — never silently discard a losing branch.

The result is one synthesized dual-format spec (Markdown + YAML/JSON) that still passes through the normal Assemble exit gates — ramza-freeze, then ramza-verify-emit before any envelope is emitted. Downstream sees exactly one spec + one validated ECL envelope, identical to the standard cycle. The parallel mode is invisible at the hand-off boundary.

4. TERMINATE (bounded)

Stop when the merged-spec confidence ≥85% (ramza-score --rubric confidence at the Assemble gate) OR the iteration count reaches the hard cap of 3 — whichever comes first. The cap is inviolable (inherited unchanged from SPECTRA, DESIGN-RATIONALE.md D2; the nexus trance-matrix R4 refusal gate). On non-convergence, emit a [GAP]/[BLOCKED] gap report and escalate to the human — do not loop past the cap.

Hard constraints (P0)

  1. TRANCE-only — never the default; standard tier runs the single-pass cycle.
  2. READ-ONLY; no worktree isolation needed (read-only ⇒ safe parallel).
  3. Cap branches at 4 (default 3); cap iterations at 3.
  4. Evaluator MUST apply identity-strip + order-rotate + length-normalize + deterministic-anchor and record them auditably; scoring arithmetic always runs through ramza-score, never hand-computed.
  5. JUDGE-MERGE produces ONE spec with per-dimension [DECISION] provenance, and still passes through ramza-freeze + ramza-verify-emit before emission.
  6. Non-convergence → [GAP] escalation, never an unbounded loop.

See SPEC.md "## Parallel spec mode (TRANCE-gated)" for the full methodology section and DESIGN-RATIONALE.md for the succession's research basis.


RAMZA — Parallel Spec Mode (TRANCE / G3)

04 同仓库更多 Skills

Rynaro
Rynaro/magicite

wine-dxvk-cache-clear

Clear a stale DXVK shader cache that is causing graphical corruption in a Wine or Proton game. Use when: a game shows corrupted textures or shader flicker that started after a driver or DXVK version change.

★ 2
Rynaro
Rynaro/magicite

steam-download-region-fix

Fix a Steam client stuck at very low download speed by changing its download region. Use when: Steam downloads are far below the local internet speed for every game, not just one.

★ 2
Rynaro
Rynaro/magicite

proton-battleye-eac-toggle

Toggle BattlEye or Easy Anti-Cheat runtime support for a Steam Proton game. Use when: a multiplayer game refuses to launch with an anti-cheat error under Proton.

★ 2
Rynaro
Rynaro/magicite

vivi-verify-incoming

Load when reading any upstream artefact handed off to Vivi that carries a sibling .envelope.json. BLOCKING per ECL §6.2.2: the orchestrator MUST have verified the envelope's SHA-256 (eidolons verify-envelope --block / eidolons run --verify) and recorded a verify_pass before dispatch. If no verify_pass exists for the message_id, or a verify_fail is present, REFUSE to process the payload and hand back to the orchestrator. Symmetric receiver gate — every Eidolon enforces it identically.

★ 2
Rynaro
Rynaro/magicite

vivi-parallel-tracks

Load ONLY under TRANCE authorization when the Plan phase yields N independent, non-overlapping implementation tracks (disjoint file sets). Operationalizes the TRANCE G4 form — parallel multi-track implementation in mandatory git-worktree isolation, max 5 clean-context tracks, a per-track verifier cascade, a non-fungible per-track ≤3 reflection budget, explicit stop conditions, and a single-threaded merge/aggregation step. Absent TRANCE gating, Vivi runs exactly as today: single-track A→P→I→V→Δ/R.

★ 2
Rynaro
Rynaro/magicite

vivi-methodology

Full Vivi cycle reference — A(nalyze) → P(lan) → I(mplement) → V(erify) → Δ(Delta)/R(eflect). Use when the task is a non-trivial feature implementation in a brownfield codebase and you need the complete methodology reference: complexity routing, evidence-grounded planning rules, test-anchoring requirements, and failure-escalation thresholds.

★ 2