token-diet

token-diet

🌐 English · Français

Aggressive token reduction for OMP dev teams. Bundles three best-in-class external token savers and wires them into OMP the native way — on top of what OMP already does (compaction, astGrep/summarizeCode, provider prompt caching), not in place of it.

Layer	What	Win	Upstream
ctx-wire	transparent CLI proxy that filters command output + scrubs secrets (full logs kept on disk); EN+FR filter overrides for git status + dotnet build/test (VSTest & MTP)/restore/run/tool	big cuts on git/build/test/lint noise	pivanov/ctx-wire
context-mode	native OMP plugin that sandboxes tool output and indexes it (FTS5/BM25, language-agnostic) — keeps raw payloads out of context + survives compaction	~98% on giant/unstructured output; any locale (incl. ro)	mksglu/context-mode
CodeGraph	MCP symbol/call graph — query instead of grep+read	~96% on "who calls X / impact / architecture"	colbymchenry/codegraph
context7	MCP library docs lookup — up-to-date API docs on demand	eliminates stale-knowledge hallucinations on library APIs	upstash/context7
caveman	terse, fragment-style output (on demand)	~65% output tokens	JuliusBrussee/caveman
yagni	write less code — YAGNI / laziest-senior-dev (on demand)	~80–94% less code; fewer tokens now + every future turn	DietrichGebert/ponytail
read-dedup + context-dedup	re-reads of unchanged files + byte-identical repeated blocks inflate input	LOSSLESS, on: re-read → stub; identical blocks collapsed before each call	caveman-code's "Read Dedup", reimplemented on OMP's tool_call/context hooks
context-compress	old prose context stays verbose every turn	protect-masked prose shrink of old messages — code/paths/numbers byte-identical (safe on by default; lite/full opt-in)	quality-preserving take on caveman-code's LLMLingua/Provence
Provider isolation	excludes all foreign-tool user configs from OMP context	eliminates agent noise from Claude Code / Codex / Gemini / Cursor / Windsurf / Copilot / OpenCode plugin registries	built-in OMP settings
Lean tool surface	tools.discoveryMode: all — hides non-essential tool schemas behind OMP's on-demand discovery tool, keeping only the hot path loaded	startup "System tools" ~18K → ~10K (full dev-team startup ~29K → ~20K), no capability lost	built-in OMP settings
C# LSP	csharp-ls wired as OMP-native language server	go-to-definition, references, diagnostics on .cs/.csx	razzmatazz/csharp-language-server

Install

omp plugin install token-diet@omp-dev-team
bash plugins/token-diet/install.sh   # installs ctx-wire + codegraph, indexes your repos,
                                     # and turns everything on. Restart omp afterwards.

Active by default after install.sh — no manual flags: ctx-wire shims compress command output, the CodeGraph MCP server is enabled, and the skills are turned on. install.sh prompts for the root of your sources (a directory of many repos) and indexes each git repo it finds (--sources-root=PATH, --depth=N).

How it's wired into OMP

• ctx-wire → installed by install.sh, then ctx-wire shims install drops transparent wrappers into ~/.local/bin (first on PATH, inherited by OMP's bash tool) so the agent runs commands normally — no prefix — and their output is filtered + secret-scrubbed before it hits context (full logs kept on disk). (ctx-wire init claude only wires Claude Code, not OMP, so we use shims.) ctx-wire gain shows savings; ctx-wire doctor verifies; ctx-wire mcp-wrap can also compress MCP-server output. Replaces the earlier RTK integration (RTK is also English-only, so it offered no localization advantage). Multilingual filters → install.sh merges EN+FR overrides (ctx-wire/filters.d/) for git status + dotnet build/test (VSTest and Microsoft.Testing.Platform)/restore/run/tool into ~/.config/ctx-wire/filters.toml, so the same compaction fires in fr_* locales (FR strings taken verbatim from git/MSBuild/VSTest localization). Only git+dotnet are localized: every other ctx-wire filter is either structural (grep, git-log, ls) or wraps an English-only toolchain (npm/cargo/go/…). No Romanian — git and .NET ship no ro translation, so they emit English in a ro_RO locale; Romanian only appears in data, handled by context-mode. See ctx-wire/README.md.
• context-mode → omp plugin install context-mode (run by install.sh, --no-context-mode to skip). A native OMP plugin on the tool_call/tool_result/session_start/session_before_compact hooks that sandboxes tool output and indexes it with language-agnostic FTS5/BM25 — the locale-agnostic safety net for any-language output (incl. Romanian) and for session continuity across compaction. Layers on top of ctx-wire's deterministic collapses, not in place of them. It also compresses MCP tool output (it hooks tool_result), so the verbose Atlassian/Miro/GitHub MCP JSON is reduced too — the ctx-wire shims only see shell commands, not MCP. For self-defined MCP servers you can additionally use ctx-wire mcp-wrap --compress; see ctx-wire/README.md.
• acli → the official Atlassian CLI (Jira/Confluence/Bitbucket), installed to ~/.local/bin by install.sh (--no-acli to skip; re-run to update — versions are supported ~6 months). acli is our go-to for Atlassian — for both reads and writes, instead of the Atlassian MCP. install.sh also offers to run acli jira auth login when interactive. Its output is English/structural — the bundled ctx-wire/filters.d/acli.toml compacts it and redacts bare ATATT… API tokens (ctx-wire already scrubs GitHub/ADO/Atlassian tokens in header/URL/key=value form).
• CodeGraph → an MCP server (.mcp.json, codegraph serve --mcp, enabled by default) exposing codegraph_search/node/callers/callees/explore/impact/files/status. See skill://codegraph. Auto-syncs on file changes.
• skills → install.sh appends config.snippet.yml to ~/.omp/agent/config.yml enabling skill commands and applying provider isolation. --no-config to skip.
• context7 → CLI mode (ctx7 library / ctx7 docs via bash — no MCP process). install.sh installs the ctx7 CLI globally. The bundled context7 skill (skill://context7) instructs the agent to fetch current docs automatically whenever a library, framework, or API is involved.
• Lean tool surface → config.snippet.yml sets tools.discoveryMode: all with an essentialOverride hot path (read, bash, edit, write, find, search, task, todo). OMP otherwise inlines every tool's JSON schema into the system prompt on every request (~18K with dev-team); discovery mode hides the non-essential ones behind the built-in search_tool_bm25 discovery tool, so they cost nothing until used. Drops startup "System tools" to ~10K. Widen the always-loaded set (e.g. add ast_grep) via tools.essentialOverride; set discoveryMode: auto to only hide MCP tools past 40, or off to disable.
• Provider isolation → config.snippet.yml sets disabledProviders + enableClaudeUser/Project/CodexUser: false so OMP only loads its own plugins and project-level AGENTS.md/CLAUDE.md. Excluded: ~/.claude/plugins, ~/.codex, ~/.gemini, ~/.cursor, ~/.codeium/windsurf, ~/.copilot, ~/.config/opencode, .clinerules. Existing users who re-run install.sh get a disabledProviders block appended without touching other settings.
• C# LSP → install.sh installs csharp-ls via dotnet tool install -g csharp-ls (if .NET SDK present) and writes ~/.omp/agent/lsp.json to register it for .cs/ .csx files. OMP auto-activates it when a .sln/.slnx/.csproj is detected.
• caveman → a native OMP skill (/caveman, levels lite/full/ultra) rather than the upstream installer, so it's first-class in OMP. See skill://caveman.
• yagni → a native OMP skill (/yagni, levels lite/full/ultra/off) porting the ponytail "laziest senior dev" YAGNI discipline: a do-I-even-need-this ladder + review/audit/debt modes. Lazy ≠ negligent — security/validation/data- loss/a11y/tests are never cut (and it won't edit .feature specs to dodge work). See skill://yagni.
• read-dedup / context-dedup / context-compress → native OMP extensions, mirrored into ~/.omp/agent/extensions/token-diet by install.sh (OMP doesn't load extensions from marketplace cache installs). The two dedups are lossless and on by default: read-dedup gates the read tool (tool_call) so a byte-identical re-read of an unchanged file returns a stub instead of the bytes (compaction-aware), and context-dedup uses the context hook to collapse byte-identical large blocks repeated across tool/assistant messages (keeping the newest verbatim) — catching duplicates from any source, including bash/cat and MCP. context-compress runs at safe by default (near-lossless: strips ANSI + collapses whitespace only, never drops words) — the quality-preserving realization of caveman-code's LLMLingua/Provence context transform. A protect mask keeps code, paths, numbers and identifiers byte-identical (the same set you'd hand real LLMLingua-2 as force_tokens), only prose is touched, and the recency window + every user/system message are left untouched. Go further (lossy — drops filler/articles) or disable with TOKEN_DIET_CONTEXT_COMPRESS=lite|full|off. Pure logic in extensions/lib, unit-tested by bun scripts/extensions.test.ts. Full analysis + the heavier real-LLMLingua escalation path: research/caveman-code.md.

What OMP already does (so you don't double up)

Compaction/handoffs (history summarization), native AST tools (astGrep, astEdit, summarizeCode, blockRangeAt), and provider prompt/context caching. This plugin fills the remaining gaps: raw command output, persistent cross-file symbol graph, and verbose model output. See skill://token-diet for the full decision guide.

Notes

• Pairs naturally with copilot-preset (cheap per-token models) — fewer tokens × cheaper tokens.
• The CodeGraph entry previously sitting (disabled) in dev-team/.mcp.json is now centralized here with the correct serve --mcp invocation.
• Independent of the other plugins; install only what you want.