AGENT SKILL

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.

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
vivi-methodology
metadata
{ "methodology": "Vivi" }
description
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.

03 SKILL.md 正文

Vivi Methodology

Feature implementation through evidence-grounded planning, test-anchored development, and structured self-improvement in brownfield codebases.


A — ANALYZE Phase

Step 1: Memory Recall

CRYSTALIUM path (when mcp__crystalium__* available):

mcp__crystalium__recall(
  scope  = { project: <cwd-project>, agent_class_visibility: "vivi" },
  query  = <task goal + domain + module area>,
  k      = 8,
  layers = ["semantic", "episodic", "procedural"]
)

episodic surfaces past task outcomes; semantic surfaces failure root causes and architectural decisions; procedural surfaces verified reusable skills. Fold relevant hits into mission context (≤ 1-2K tokens summarized).

Standalone path (when CRYSTALIUM absent): query agents/memories/ for past tasks in the same module, known reusable assets, and previous failure patterns.

Score matches by: path proximity → recency → outcome quality. Budget: ≤ 20 entries.

See agents/skills/memory-management/SKILL.md for the full routing decision and protocol.

Step 2: Repo Map Generation

Before reading any file in detail, generate a structural overview:

1. List directory tree for the target domain (2-3 levels deep)
2. Identify key files by convention:
   - Models/entities: app/models/DOMAIN/
   - Controllers/handlers: app/controllers/ or equivalent
   - Tests: spec/ or test/ mirroring source structure
   - Configuration: config/, .env patterns
3. For each key file, extract:
   - Public interface (method signatures, exported functions)
   - Dependencies (imports, includes, requires)
   - Test coverage existence (yes/no)
4. Rank by reference frequency: files imported by many others = high leverage

Output: Compressed structural summary (~50-100 lines). This is your navigation map.

Step 3: Requirements Decomposition

State explicitly:

  • Goal: What problem does this solve? One sentence.
  • Scope IN: Files, modules, features being changed
  • Scope OUT: What is explicitly NOT being changed (boundaries)
  • Acceptance criteria: Observable, testable conditions for "done"

Step 4: Asset Discovery (MANDATORY)

Search the codebase BEFORE planning. Discover what already exists.

Asset Type Search Pattern Purpose
Domain models app/models/DOMAIN/ Business logic, state machines
Repositories app/models/DOMAIN/repository.* Data access patterns
Services app/services/DOMAIN/ Orchestration, business rules
View components app/components/DOMAIN/ UI building blocks
Query objects app/models/DOMAIN/queries/ Complex data retrieval
Workers/Jobs app/jobs/, app/workers/ Async processing
Serializers app/serializers/ API response shaping
Shared utilities lib/, app/lib/ Cross-cutting helpers
Config/Constants config/, app/constants/ Feature flags, settings
Test factories spec/factories/, test/factories/ Test data patterns

For EACH discovered asset, record:

Field Values
Location file:line
Purpose One-line description
Relevance HIGH / MED / LOW
Quality Has tests? Recent changes? Known issues?
Verdict USE / EXTEND / WRAP / AVOID
Rationale Why this verdict (one sentence)

Step 5: Collision Mapping

Identify risk zones:

  • Files to modify (existing, may break things)
  • Files to create (new, may collide with in-flight work)
  • High-risk zones: Low test coverage, heavily imported, recently changed by others
  • Integration points: Where new code touches existing code

Phase output: Discovery Report (use agents/templates/discovery-report.md)


P — PLAN Phase

Step 1: Test Anchor Generation

BEFORE designing any solution, write the test expectations:

For each acceptance criterion:
  1. Describe the test case in plain language
  2. Specify: input state → action → expected outcome
  3. Identify what existing test patterns to follow
  4. Note which test helpers/factories already exist

These test anchors become the source of truth for implementation correctness. Implementation is done when these tests pass.

ANTI-OVERFIT rule (mandatory). Test anchors derive from the acceptance criteria + EXISTING test patterns in the codebase — never reverse-engineered from a candidate implementation. Writing tests after (or to fit) the code you intend to write reproduces the field's named failure: agents over-fit implementations to tests they authored in hindsight (DESIGN-RATIONALE Decision 3; SWE-bench-style contamination/overfitting). If you cannot state an anchor from the acceptance criteria alone, the criterion is under-specified — clarify it, do not invent an implementation-shaped test.

CAPTURE-LIVE-FIRST gate (mandatory). When the task parses external CLI stdout/stderr, or consumes a serde-renamed / IPC payload, stage the verbatim live capture as the fixture BEFORE writing the parser. Fabricated fixtures pass 10/10 tests vacuously while every assumption about the real output is wrong. Capture one real sample (or a one-line dump of the live payload) first; that capture is the anchor.

RED-GATE rule (mandatory — Stage 2). For a fix/repair task, the reproduction test anchor MUST be verified to FAIL on the unmodified base tree before any implementation begins. The substrate enforces this mechanically: pass --require-red to eidolons sandbox loop — a reproduction that already passes is VACUOUS (final=vacuous-reproduction) and cannot anchor a fix. The evidence is decisive: patching is near-solved when the repro test is right; repro-test validity — not patch generation — is the bottleneck (TDFlow: 88.8% vs 49% with verified repro tests). On a vacuous verdict, return to THIS step and re-derive the anchor from the acceptance criteria — do NOT burn loop attempts against a test that cannot fail.

Step 2: Strategy Generation (Tree-of-Thoughts)

Generate 3-5 genuinely different strategies. Requirements:

  • At least ONE strategy maximizes use of discovered internal assets
  • At least ONE strategy is the conservative/minimal-change approach
  • NO strawmen — every strategy must be plausibly the best choice
  • Each strategy must differ in at least one of: architecture, coupling, scope, or risk profile

For each strategy, document:

  • Approach: 2-3 sentence description
  • Files touched: List with change type (modify/create)
  • Assets used: Which discovered assets and how
  • Test impact: New tests needed, existing tests affected
  • Risk profile: What could go wrong

Step 3: Strategy Scoring

Score each strategy on four dimensions (1-3 scale):

Dimension 1 (Poor) 2 (Acceptable) 3 (Good)
Risk High blast radius, low coverage in affected areas Moderate, some coverage gaps Low blast radius, good coverage
Effort 3+ days, multi-team coordination 1-2 days, single team < 1 day, contained changes
Alignment Ignores internal assets, creates parallel paths Partial internal asset reuse Full Internal First compliance
Maintainability Adds technical debt, unclear ownership Neutral to codebase health Improves patterns, reduces duplication

Total: 4-12 (higher = better)

Step 4: Deep Evaluation of Top 2

Expand the top 2 scoring strategies with:

  • Detailed step-by-step implementation plan
  • Specific file changes with pseudocode
  • Dependency chain (what must be done first)
  • Abort conditions (what would make this strategy fail)
  • Blockers and unknowns

Re-score after deep evaluation. Hidden issues often emerge here.

Step 5: Selection with Justification

Document:

  • Selected strategy: Name + final score + one-paragraph justification
  • Runner-up: Name + score + why it was rejected
  • Confidence level: HIGH / MED / LOW
  • Abort conditions: Specific signals that mean "stop and re-plan"
  • Boundaries: Files/systems explicitly out of scope for this implementation

Phase output: Execution Plan (use agents/templates/execution-plan.md)

Memory: Plan Checkpoint (CRYSTALIUM)

After the Execution Plan is produced, call (if CRYSTALIUM available):

mcp__crystalium__plan_checkpoint(
  plan_id  = <task-slug + date>,
  state    = <full execution plan snapshot>,
  step     = "initial",
  metadata = { author_agent: "vivi", task_title: <title> }
)

Store the returned checkpoint_id in working context. If the plan is revised mid-cycle (the abort rules below fire), call plan_replan before re-entering Plan:

mcp__crystalium__plan_replan(
  plan_id            = <plan_id>,
  from_checkpoint_id = <checkpoint_id>,
  new_plan           = { diff: <what changed and why>, supersedes_id: <checkpoint_id> }
)

Graceful skip: if CRYSTALIUM unavailable, skip these calls silently.

ECL emit on FORGE consultation

If Plan-phase reasoning calls for a FORGE consultation (adversarial reasoning, trade-off arbitration), emit a reasoning-request.envelope.json next to the question artefact (template at templates/reasoning-request.envelope.json). Required: to.eidolon=forge, performative=REQUEST, artifact.kind=reasoning-request. Body validates against schemas/_base-profile.v1.json. ise.assertion_grade="self-attested" (ECL v2.0 §6.5.2; this envelope precedes any loop-native verification — see skills/loop-native/SKILL.md §6 for the grade rationale). Skip the envelope when ECL_VERSION is absent.


I — IMPLEMENT Phase

Memory: Skill Reuse (CRYSTALIUM)

Before building, if the A-ANALYZE recall surfaced a procedural entry for this task type, invoke it (if CRYSTALIUM available):

mcp__crystalium__skill_invoke(
  skill_id = <procedural entry id from recall>,
  context  = <current task context>
)

Use the result to short-circuit re-derivation. After verifying a new reusable pattern, commit it:

mcp__crystalium__commit(
  layer   = "procedural",   # "semantic" if unverified
  payload = <pattern: name, location, purpose, usage_hint, quality>,
  provenance = { author_agent: "vivi" }
)

Graceful skip: if CRYSTALIUM unavailable, skip these calls silently.

Execution Priority

Follow this order strictly:

  1. USE — Assets marked USE AS-IS. Wire them in directly.
  2. EXTEND — Assets marked EXTEND. Add methods/features to existing code.
  3. WRAP — Assets marked WRAP. Create adapter layer for legacy interfaces.
  4. CREATE — New code only when Discovery confirmed no suitable internal alternative.

Architect/Editor Separation

For Complex-tier tasks, separate reasoning from editing:

ARCHITECT PASS (reasoning):
  - Describe WHAT needs to change and WHY
  - Specify the interface contracts between components
  - Define the data flow through the change

EDITOR PASS (implementation):
  - Translate architect output into actual code edits
  - Follow existing code style and conventions exactly
  - Produce minimal, targeted diffs (not rewrites)

Implementation Rules

  • Write tests for new functionality FIRST (test-anchored from Plan phase)
  • One logical change per commit. Each commit should pass linter + existing tests.
  • If you discover an asset not found in Analyze, STOP and update the Discovery Report.
  • If implementation reveals the plan is wrong, STOP and return to Plan phase.
  • Track progress with structured task list:
## Task Progress
- [x] TASK-1: Create factory method for Widget — DONE
- [ ] TASK-2: Extend WidgetRepository with #find_active — IN PROGRESS
- [ ] TASK-3: Add WidgetComponent for list view — BLOCKED (needs TASK-2)
- [ ] TASK-4: Wire controller action — PENDING

Targeted Test Execution

Run tests incrementally, not all at once:

  1. Run the SINGLE most relevant test after each change
  2. Fix that failure before moving to the next change
  3. Run the broader test suite only after all individual tests pass
  4. This prevents the overcorrection cascade (fixing one thing, breaking another)

ECL emit on Implement-phase exit

On phase exit, emit vivi-completion-report.envelope.json next to the completion artefact (template at templates/vivi-completion-report.envelope.json). Required: to.eidolon=idg, performative=PROPOSE, artifact.kind=vivi-completion-report, integrity.method=sha256 matching the payload bytes. Profile schema: schemas/vivi-completion-report-profile.v1.json (required keys: files_changed_count, tests_run, tests_passed). ise.assertion_grade="validated" (ECL v2.0 §6.5.2) — justified because this envelope is only reachable after the V-phase closed loop passed pass^k against the regression + anchoring tests (skills/loop-native/SKILL.md §4, §6); that is the spec-mandated verification gate the validated grade requires, not a self-report. Skip when ECL_VERSION is absent.

After the envelope is produced and verified (V-VERIFY phase), ingest it into CRYSTALIUM (if available):

mcp__crystalium__ingest(
  envelope = <vivi-completion-report.envelope.json contents>,
  payload  = <completion report contents>
)

Graceful skip: if CRYSTALIUM unavailable, skip silently.


V — VERIFY Phase (loop-native)

Vivi does not "run the checks once and decide." It drives the closed loop eidolons sandbox loop (see skills/loop-native/SKILL.md) and acts as its --fix-hook: the loop runs the verification, and on failure re-invokes Vivi's repair step (R, below) in fresh context, bounded, until green or capped. This is the core of the loop-native cycle and the reason Vivi supersedes its predecessor — which ran V once and handed the running back.

Configure the loop

  • --tests, or split --regression/--reproduction: the success command(s). Success = the Plan-phase test-anchors plus the existing suite, with regression-first, then the new acceptance test (passing only the new test FAILS — anti-reward-hacking).
  • --protect <glob>: the anchoring tests; the fix-hook MUST NOT edit them (a mutation aborts → VIGIL).
  • --k <n>: pass^k — a green that is non-deterministic across k re-runs is flaky → BLOCKED, not accepted.
  • --max-attempts N + --via <isolation>: the bounded cap and the host sandbox (untrusted/LLM-authored code needs ≥ container isolation, R8-03).
  • --require-red: mechanical red gate — the reproduction anchor must FAIL on the base tree before any fix attempt (P-phase RED-GATE rule, enforced).
  • --fanout N: parallel-sample-and-select shape for standard-tier hosts (see Host-adaptive shape below) — N independent fresh-context candidates, externally selected; NO self-repair iteration.
  • --judge-hook <cmd>: external diff-review judge over a surviving candidate (layered hack detection — the sealed holdout alone is insufficient).

Verification checks (run by the loop each iteration)

Check Pass criteria
Regression (full suite) no new failures — runs FIRST
Test-anchors (new) all pass
Linter zero new violations
Coverage no decrease in affected files
Build / Type check clean
pass^k green stable across k re-runs (else flaky → BLOCKED)

Decision

  • Loop reaches green (regression + reproduction + pass^k) → it emits a candidate diff (diff-not-apply; the human applies) → proceed to Δ.
  • Loop caps / flaky / protected-tests-mutated → proceed to R (already bounded); on the 3-same-category / cap threshold, escalate to VIGIL.

Host-adaptive shape (Stage 2) — iterate vs fanout

The loop's gain belongs to an RL-trained / loop-competent host; on a weak host self-repair iteration is the wrong shape — feeding failures back injects new errors (RLEF; the self-repair literature). Vivi therefore picks the LOOP SHAPE by host tier instead of falling off the loop entirely:

Host tier Shape Invocation
thinking / loop-competent iterate --max-attempts 3 --k 2 --require-red
standard / weak fanout --fanout 3 --max-attempts 1 --k 2 --require-red

In fanout, each candidate is an INDEPENDENT fresh-context, single-shot fix generated from the same base tree + the same localized base-failure feedback; the substrate selects the survivor externally (tests + pass^k + holdout + judge). The weak-host model never judges its own retry — selection is external by construction (R2E-Gym hybrid-verifier evidence: +8pp over either filter alone; parallel sampling beats sequential self-repair on weak models).

Host-contingency / graceful degradation

If no adequate isolation is available, or the substrate (eidolons sandbox) is absent entirely, degrade: run the checks once, emit a diff for the human, and recommend the conservative fallback (eidolons add apivr). Never run untrusted code without --via or an explicit --allow-unsafe-host.


R — REFLECT Phase (Failure Only) — fresh-context repair

R runs as the loop's --fix-hook on each failing iteration. Load skill: skills/failure-recovery/SKILL.md.

Fresh-context per attempt (the decisive discipline)

Each repair attempt starts from a CLEAN context: the localized feedback ($EIDOLONS_SANDBOX_FEEDBACK — failing test ids, assertion, file:line loci, full log) + the original acceptance criteria + the current working tree — NOT the accumulated "prior error + new hypothesis" transcript. Carrying the prior trajectory re-creates self-conditioning (long-horizon execution failure is dominated by models conditioning on their own prior errors). This is the single biggest change from the predecessor's same-context retry, and it is why the loop helps rather than degrades.

Evidence Gate (MANDATORY)

STOP if you have no concrete artifact (the loop's localized feedback IS the artifact: assertion + file:line). No artifacts = ESCALATE. Never guess at fixes.

ECL emit on 3-failure escalation

When the 3-same-category threshold fires (the methodology counter is the authority; the loop's --max-attempts is the substrate ceiling — whichever trips first), wrap the escalation in a repair-failed-report.envelope.json (template at templates/repair-failed-report.envelope.json). Required: to.eidolon=vigil, performative=ESCALATE, trust_level=high, assumptions[0]="trigger: 3-failure-same-category". Profile schema: schemas/repair-failed-report-profile.v1.json (keys attempts>=3, failure_category, last_test_command). ise.assertion_grade="self-attested" (ECL v2.0 §6.5.2) — the escalation is Vivi's own bounded-failure judgement, not an externally-gated pass. Skip the sidecar when ECL_VERSION is absent (the escalation still happens).

Failure Protocol (fresh-context, localized)

See skills/failure-recovery/SKILL.md for the full taxonomy. Quick reference:

Iteration Condition Action
Each localized feedback present Targeted fix to the reported loci, from fresh context
2nd, same category — A genuinely different approach (the feedback drives it, not the prior narrative)
3rd, same category — ESCALATE to VIGIL (cap reached)
Any LOW confidence / no artifact ESCALATE immediately
Any fix would edit an anchoring test or exceed scope ABORT (anti-reward-hacking) → escalate

Escalation Format

When escalating, provide:

## Escalation: [task description]

### What was attempted
1. [Approach 1]: [what happened]
2. [Approach 2]: [what happened]

### Evidence collected
- [error output, test results, etc.]

### My assessment
- Root cause hypothesis: [best guess with confidence]
- What I need: [specific help required]

### Suggested next steps for human
1. [concrete suggestion]

Δ — DELTA Phase (Success Only)

After successful verification, evaluate the touched code for normalization opportunities.

Candidate Scoring

Priority = (Severity + Frequency + Velocity) - Cost

Each factor scored 1-3. Threshold: ≥ 3 to suggest.

Factor 1 2 3
Severity Cosmetic Moderate coupling Architectural debt
Frequency Seen once Seen 2-3 times Pattern across codebase
Velocity Stable area Moderate change rate Active development area
Cost Major refactor Moderate effort Quick improvement

Anti-Criteria (Reject if ANY match)

  • First occurrence only → premature abstraction
  • Dormant area (> 6 months since last meaningful change)
  • High cost but affects ≤ 2 files
  • "Might be useful someday" reasoning
  • Would require changes outside current domain

Output Format

## Delta Suggestions

### Δ-1: [Title]
- Pattern: [what was observed]
- Location: [file:line references]
- Score: Severity(X) + Frequency(X) + Velocity(X) - Cost(X) = [total]
- Suggestion: [specific improvement]
- Effort estimate: [hours/days]

Status: SUGGESTION ONLY — Do not implement

CRITICAL: Delta suggestions are OUTPUT ONLY. Never implement infrastructure suggestions. Log them to agents/memories/delta-history.md for future reference.


Post-Task: Memory Update (Δ/R phase)

After every task (success or failure), update memory via the active path. See agents/skills/memory-management/SKILL.md for the full protocol.

CRYSTALIUM path (when available):

# Task outcome
mcp__crystalium__commit(layer="episodic", payload={task_title, domain, outcome,
  summary, key_decisions, lesson}, provenance={author_agent:"vivi"})

# Failure root causes (if any)
mcp__crystalium__commit(layer="semantic", payload={failure_category, context,
  root_cause, prevention, domain}, provenance={author_agent:"vivi"})

# New patterns (if discovered and verified)
mcp__crystalium__commit(layer="procedural", payload={pattern_name, location,
  purpose, usage_hint, quality:"verified"}, provenance={author_agent:"vivi"})

# End session
mcp__crystalium__session_end()   # triggers Dream consolidation

Standalone path (when CRYSTALIUM absent): write to agents/memories/ files per skills/memory-management/SKILL.md §Standalone Fallback.


Vivi Methodology — Flow-engineered, test-anchored, context-aware

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-memory-management

Load at Vivi session start, session end, or when you notice a repeating pattern across tasks. CRYSTALIUM-primary memory protocol: when mcp__crystalium__* tools are available, route ALL memory through CRYSTALIUM (recall/commit/ingest/plan_checkpoint/session_end). The local agents/memories/*.md files are the standalone fallback for when CRYSTALIUM is absent — EIIS conformance. Never write both.

★ 2