ry-repair

name: ry-repair description: "Нормализация репозитория: source-of-truth scan, semantic entropy cleanup, stale docs/memory repair, validators, docs/memory sync. Используй для: ry-repair, почини систему, нормализуй репозиторий, убери противоречия. EN triggers: repository repair, semantic entropy cleanup, stale AI-tool context, contract normalization."

Purpose

Normalize a repository so OpenCode, Codex, and Claude Code can work from the same verified facts with minimal semantic entropy. This is a technical repair flow, not permission to change business logic silently.

For rldyour AI CLI configuration repositories, /ry-repair also verifies deterministic install/update/sync convergence through the root config/ry-repair-sync-contract.json and scripts/ry_repair_sync.py contract when that control plane is present.

Mode Selection

Work project: run the full authoring repair workflow below.
rldyour control plane or adapter config repo, goal = converge the installed system: run consumer mode instead - update the checkout FROM GitHub (git pull --ff-only, git submodule update --init --recursive), then python3 scripts/ry_repair_sync.py --check and --apply --install-system. In consumer mode do NOT author changes into the repository: no commits, no doc/contract/memory edits, no fullrepo publication. If fast-forward fails or validators reveal repository drift, report it and switch to authoring repair only on explicit owner instruction.

Workflow

Detect repository type, branch/worktree, submodules, CI surface, deploy surface, and effective project policy from RLDYOUR_PROJECT_POLICY, .rldyour/project-policy.local.json, .rldyour/project-policy.json, or built-in defaults.
Read native instruction/config surfaces: AGENTS.md, .claude/CLAUDE.md, opencode.json, .opencode/commands, .opencode/skills, .opencode/agents, .opencode/plugins, and repository contracts.
Inspect Serena memories, plans, and research archives for stale facts, unsupported claims, missing taxonomy, duplicated rules, or contradictions with current code.
Inspect GitHub issues, pull requests, and recent history through MCP or CLI when available. Verify every issue against current code before treating it as a fact.
Inspect MCP/LSP/tooling config, plugin hook lifecycles, commands/skills/agents, CI gates, release manifests, dependency baselines, and docs source-of-truth declarations.
When the root control plane is present, run python3 scripts/ry_repair_sync.py --plan --target "$PWD" and use --check before claiming local repo, system AI CLI config, Serena memory, or GitHub/fullrepo sync. For OS/mode work, pass explicit flags: --os macos|linux|wsl|windows, --mode standard|orchestrator, and --cmux only for macOS orchestrator mode.
Treat approved active inventories as system-wide: installed Claude/Codex/OpenCode configs, active agent/tool surfaces, CI workflows, runtime pins, docs, and release gates must match current MCP/provider policies. Do not keep permanent tool-specific absence gates for removed components.
Detect semantic entropy: duplicated docs, stale pins, conflicting instructions, dead config, unclear source-of-truth, missing ADR/CONTEXT/FUTURE facts, broken validators, adapter parity drift, and natural-language policy overrides that are not materialized in project policy JSON.
Produce a repair plan that separates:
- technical repairs the agent may apply;
- business, functional, security-posture, deployment-target, data-model, or ADR decisions that require the owner.
Ask the owner in Russian before changing any decision-class item. Present concise options with a recommendation, reason, and impact.
Apply technical-only repairs using existing project patterns and native OpenCode surfaces. If the user explicitly states repository policy and no policy file exists, write .rldyour/project-policy.json or .rldyour/project-policy.local.json only from that explicit instruction; do not infer ownership/destructive permissions from chat tone.
Run matching validators, tests, schema checks, plugin/hook checks, release/archive checks, installed-config checks, and instruction/memory freshness checks.
Synchronize durable docs and Serena memories from verified code/config state, then finish through flow-post-task-sync when durable artifacts changed.

Non-Negotiables

Current code, config, runtime checks, and verified GitHub state are the source of truth. Memories and docs are derived evidence.
Do not edit ADR meaning, business logic, functional behavior, pricing, deployment targets, security posture, or data contracts without owner approval.
Hooks and plugins stay bounded and deterministic. They may mark state; ry-repair performs the repair.
Do not hide unresolved drift behind green summaries. Every blocked check names the blocker and next proof command.
Do not report local repo, system config, Serena memory, or GitHub parity without evidence from ry_repair_sync.py, installed-config validators, current git state, and fullrepo/GitHub checks where applicable.

Output

Report in Russian:

Scope and source-of-truth map.
Confirmed drift, grouped by severity.
Technical repairs applied.
Decision-class items left for owner approval.
Exact validation commands and results.
Docs/memory/fullrepo/git synchronization status.