By akii09
Token-saving Claude Code plugin: routes work across Haiku/Sonnet/Opus per pipeline stage. Adds session-start brief, opt-in local telemetry with savings dashboard, and advisory routing hints. Composes with built-in CLAUDE.md + auto memory and with claude-mem / code-review-graph.
Locates the minimum set of files relevant to a task. Returns a precise file list with one-line summaries. Does NOT read whole files, does NOT edit anything, does NOT plan or implement. Invoke before planning or implementing whenever the target files are not already known.
Applies a single phase of an approved plan to a specified set of files. Edits only — does not rewrite, does not explore. Receives an exact file list, a phase description, and a verification command. Produces a focused diff and runs verification.
Produces a phased implementation plan from a task description and a file list. Reads files, identifies risks, breaks work into small verifiable phases. Does NOT write code. Returns a plan the user can review before implementation begins.
Reviews a recently-applied diff for bugs, regressions, missing tests, and quality issues. Reads the diff first, files only as needed to judge it. Returns a prioritized issue list and a ship verdict.
Runs the full Yukti pipeline (explore → plan → user-confirm → implement → verify → review) for a task. Delegates each step to the right specialist subagent. Does not do the work itself — only coordinates.
[Deprecated as standalone — internal pipeline stage] Locate the minimum set of files relevant to a task. Returns a precise file list. Runs on Haiku. Use /yukti:smart for code-change tasks; the orchestrator invokes the explorer internally. Direct invocation is preserved for power users but may be removed in v0.3.
[Deprecated as standalone — internal pipeline stage] Apply ONE phase of an approved plan to specified files. Edits only, no rewrites, no exploration. Runs on Sonnet. Use /yukti:smart for the full code-change pipeline; the orchestrator invokes the implementer per phase. Direct invocation is preserved for power users but may be removed in v0.3.
Produce a phased implementation plan for a task. Reads files, identifies risks, breaks work into small verifiable phases. Runs on Opus. Does NOT write code — produces a plan for review before implementation.
Review a recently-applied diff for bugs, regressions, missing tests, and quality issues. Reads the diff first, files only as needed. Runs on Opus. Returns a prioritized P0/P1/P2/P3 issue list and a SHIP/FIX-FIRST/NEEDS-REWORK verdict.
Smart entry point for Yukti. Code changes (add/fix/refactor) route through the explore→plan→confirm→implement→review pipeline that saves ~50-60% cost. Analysis, questions, and explanations are answered directly in the current session — no extra waiting. No fork, no rules, just type what you want.
Uses power tools
Uses Bash, Write, or Edit tools
Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimBased on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
A lightweight Claude Code plugin that routes each pipeline stage to the right model — Haiku to explore, Opus to plan + review, Sonnet to implement. Claude Code stays the same to use; you just type /yukti:smart <task> instead of letting one model do everything.
Target savings: ~50% cost reduction vs always-Opus on routine code work, with quality on par for that workload. Numbers below are illustrative — real measurement lands in v0.2.0. The architecture is designed for it; the receipts are still being collected.
If you've seen plugins claim "80% token savings," they're either measuring cost (not tokens) or compromising on quality. This README won't make that mistake. We will publish measured numbers from real-world usage when we have them, not before.
Composes with the ecosystem. Yukti is the routing layer. For cross-session memory install claude-mem. For structural code review with blast-radius analysis install code-review-graph. We don't try to be everything.
It splits a coding task across four specialist subagents, each running on the cheapest model that can do that step well:
| Step | Subagent | Model | Why this model |
|---|---|---|---|
| 1. Locate files | explorer | Haiku | Fast, cheap, good enough for grep + glob |
| 2. Plan phases | planner | Opus | Long-horizon reasoning matters here |
| 3. User confirms | (you) | — | The quality firewall |
| 4. Implement each phase | implementer | Sonnet | Near-Opus on focused tasks, much cheaper |
| 5. Verify | hook + Bash | — | Typecheck/test gate |
| 6. Review | reviewer | Opus | Catches what Sonnet missed |
The orchestrator (smart-orchestrator, also Opus) runs on a deliberately tiny tool set — only Agent, Read, Bash — so it physically cannot do the work itself. It can only delegate.
This plugin is the codification of practices that Claude itself recommended for making Sonnet punch above its weight:
PreToolUse hook injects limit: 500 on any Read call without an explicit limit. Prevents accidental 5000-line reads.Run inside any Claude Code session:
/plugin marketplace add akii09/yukti
/plugin install yukti@akii09-yukti
Yukti is installed once at the user level and available across every project you open with Claude Code. Skills are namespaced (/yukti:smart, /yukti:status, /yukti:plan, /yukti:review) so they never collide with other plugins or future built-in commands.
To upgrade later:
/plugin upgrade yukti
| Your situation | Use this |
|---|---|
| Normal user, want savings on every project | Marketplace install above ✅ |
| Locked-down env (corporate proxy, no marketplace, CI runner) | Fallback installer (below) |
| You want to enforce Yukti for every contributor of one specific repo | Fallback installer, commit .claude/ to that repo |
| Plugin development / quick eval against a fork | claude --plugin-dir ./yukti (below) |
Drops the agents, skills, hooks, and bin scripts into the current project's .claude/ directory. Skills are not namespaced in this mode — they're invoked as /smart, /plan, /explore, /implement, /review. That means autocomplete won't surface them under /yukti, and they may collide with anything else in your project that defines those names.
curl -sSL https://raw.githubusercontent.com/akii09/yukti/main/install.sh | bash
If jq is available it's used for clean JSON merging; otherwise the script falls back to sed.
Clone and load the plugin directly without a marketplace:
git clone https://github.com/akii09/yukti.git
claude --plugin-dir ./yukti
Skills appear as /yukti:smart etc. Useful for evaluating a fork or developing changes.
The main command:
/yukti:smart Add a dark mode toggle to the settings page
npx claudepluginhub akii09/yukti --plugin yuktiUltra-compressed communication mode. Cuts ~75% of tokens while keeping full technical accuracy by speaking like a caveman.
Memory compression system for Claude Code - persist context across sessions
Multi-model consensus engine integrating OpenAI Codex CLI, Gemini CLI, and Claude CLI for collaborative code review and problem-solving.
Curate auto-memory, promote learnings to CLAUDE.md and rules, extract proven patterns into reusable skills.
Comprehensive UI/UX design plugin for mobile (iOS, Android, React Native) and web applications with design systems, accessibility, and modern patterns
Editorial "Web Designer" bundle for Claude Code from Antigravity Awesome Skills.