VerifiedPlugin
soloboard
SoloBoard - Autonomous development orchestrator for Claude Code.
1 skill3 commands1 MCP+2+2+224
@egorfedorovDevelopment
Plugin
claude-context-optimizer
Claude Code plugin that tracks token usage, identifies wasted context, and saves 30-50% on API costs. Heatmaps, ROI reports, budget alerts, efficiency scores, git-aware suggestions — all local, zero config.
Plugin Signal82/100Strong
maintained97
capable71
documented100
adopted44
Install on Claude Code
> /plugin marketplace add egorfedorov/claude-context-optimizer> /plugin install claude-context-optimizer@cco
Stats
111
Stars
10
Forks
Active
Maintenance
JavaScript
Language
MIT
License
6d ago
Last commit
6mo ago
Created
Repo: egorfedorov/claude-context-optimizer
What's inside
Skill(22 skills)
/cco-anatomyGenerate a compact project map so Claude understands the codebase without opening every file/cco-budgetConfigure token budget limits, auto-compact settings, and view current budget status (model-aware — Claude 5 lineup, Opus 5 default fallback, full 1M context at standard price)/cco-claudemdAnalyze CLAUDE.md files for token bloat and suggest optimizations
Hook(6 hooks)
README
The Problem
The average Claude Code session wastes 30-50% of tokens on files that are read but never actually used. Every Read call consumes context — whether the file was relevant or not.
- A 200-line config file? 800 tokens gone.
- A README you glanced at once? 2,400 tokens burned.
- That
package.jsonClaude reads "just in case"? 120 tokens, every time.
At $5/M input tokens (Opus 5) — $10/M on Fable 5.1 — a developer spending $100/month is lighting $30-50 on fire on irrelevant context.
The Solution
context-optimizer silently tracks every file read, edit, and search. It learns which files are actually useful and which are waste. Over time, it builds a profile of your coding patterns and tells you exactly where your tokens go — and how to stop wasting them.
What's new in v4.10 — Claude 5 prices, the 1-hour cache, and task state that survives /compact
The price table caught up. Sessions now report claude-fable-5-1 and
claude-opus-5; neither had a row. Fable was priced at the Opus tier — it costs
twice that ($10/$50) — and Sonnet 5 is cheaper ($2/$10) than the Sonnet 4.6 row it
was using. Fable 5.1 also bills cache reads at 0.025× instead of 0.1×, so it
carries its own rate; every dollar figure reads the per-model rates.
The cache math caught up. Claude Code keeps the prompt cache alive for an
hour on most sessions today (ephemeral_1h_input_tokens in the transcript).
CCO priced every write at the 5-minute 1.25× rate (real: 2×) and warned that the
cache "went cold" after a 5-minute pause when it had 55 minutes left. Now the
TTL is read per session: writes are billed at their real rate, and the
cache-break guard waits the full hour on 1h sessions.
Task state that survives /compact — the honest slice of Google/Purdue's
SKILL.state. That architecture replaces an agent's growing history with
instructions + a small structured state + the latest observation, so the prompt
stays O(1) in the step count. A plugin can't replace Claude Code's history (and
with caching, replayed history is cheap anyway) — but it can own the moment
history gets replaced. /compact writes a lossy free-form summary; CCO now keeps
an exact, bounded state per task and re-injects it right after compaction:
/cco-task patch '{"goal": "ship 1h-cache pricing", "done": ["parser", "tests"], "next": "README", "scratch": null}'
✓ State of task #7 updated (3 keys, 118/4000 chars)
… /compact …
[cco-task] Active task #7: ship 1h-cache pricing
Execution state (authoritative — trust this over any summary of earlier turns):
{"goal":"ship 1h-cache pricing","done":["parser","tests"],"next":"README"}
Patch semantics are SKILL.state's: set keys, null deletes. The cap (~1K tokens)
is the point — the state never grows with the task, and the patch is rejected
until you prune. Full notes in the CHANGELOG.
What's new in v4.9 — the optimizer learns, and the team shares
Your tools now price themselves. MCP and Agent costs were constants someone
picked once — mcp__* ≈ 200 tokens in, Agent ≈ 500. Real results vary by
orders of magnitude: a "list all issues" query and a one-row lookup are the same
tool name and nowhere near the same cost. On MCP-heavy sessions the budget meter
was guessing at its own biggest line item.
CCO already measured every tool result; now it remembers. After three calls a tool is budgeted from what it actually costs you:
/cco-tools
tool calls avg max total
mcp__linear__list_issues 41 38.2K 112.0K 1.5M ← where your budget really goes
Bash 380 133 342 50.5K
· mcp__github__get_pr 2 4.1K 6.0K 8.2K ← still on the constant
Pair it with /cco-overhead mcp: that finds servers you never call, this prices
the ones you do. (#38)
Patterns travel with the repo. What CCO learns — which files are usually waste, usually useful, edited together — was locked to one machine. A teammate's fresh clone started blind and re-learned the same lessons by wasting the same tokens.
/cco-patterns export # → .cco/patterns.digest.json — commit it
/cco-patterns import # teammate, day one on a fresh clone
The digest carries relative paths and counts only — no file contents, no absolute paths, no home directory. Anything that can't be expressed relative to the project root is dropped, and the file is audited before writing and on import; a digest with absolute paths is rejected, not sanitized.
Imported data is stored as a separate prior and never merged into your own
counts. That's deliberate: confidence means "how many sessions I observed
this in", and folding someone else's sessions into it would make your own
numbers a lie. The import fills in only where you have no evidence of your own.
(#37)
This repo ships its own
.cco/patterns.digest.json— clone it and/cco-patterns importto see the feature work on real data.
What's new in v4.8 — the metric gets honest, the knobs come out
- Token estimates are per-language now. The headline "tokens saved" number
was computed from one constant — 35 chars per line, for every file type. So I
measured 6,344 real source files instead of guessing.
.svgaverages 100.5 chars/line (nearly 3x under-counted before),.css25.3 and.txt27.8 (over-counted),.ts44. Notably the measurement contradicted my own priors:.jsonis 38.8, not the ~20 I expected. (#35) /cco-config— 11 hardcoded knobs, now tunable. Re-read warning points, Read-Cache staleness, prompt-coach length bands,/cco-packbudget cap. Every value range-validated: a typo is ignored in favour of the default and flagged, never silently applied. (#39)- The decision logic is under test. The formatters had coverage; the code that decides what actually reaches your context didn't. Budget gating, cache-break detection, Read-Cache staleness and session waste classification are now pure, exported, and covered — 238 tests, up from 183. (#36)
- Version drift can't ship again.
marketplace.jsonquietly sat two releases behind; CI now fails on any mismatch.
What's new in v4.7 — Windows works
CCO quietly assumed POSIX paths. On Windows that broke four things at once, and none of them looked broken — they just reported less:
/cco-overheadsaid "no session transcripts found" even with transcripts right there. The cwd → folder encoding dropped\,:and spaces, so it looked in~\.claude\projects\C:\Program Files\Gitinstead of…\projects\C--Program-Files-Git. Fixed — which also restores the CLAUDE.md/memory itemization that had been collapsing into "system prompt & tools (unattributed)". (#46).contextignorewas inert on Windows. CRLF files left a\ron every pattern, so not one of them matched. Every lockfile you thought you'd excluded was still being counted. (#33)- Globs now understand
\—C:\proj\dist\a.jsmatchesdist/**. - Project-root detection and path display are separator-aware, so reports
stop leaking raw
C:\…paths.
Plus a CONTRIBUTING.md with the platform rules, so this doesn't regress. 183 tests, green on Node 18/20/22.
Hooks still need a POSIX shell (Git Bash or WSL) — native Windows hooks are tracked on the roadmap.
What's new in v4.5 — model auto-detection & the dollar leaks
The session's REAL model now drives everything, and the three biggest avoidable dollar leaks get caught live:
- Model auto-detection. The budget hook reads the exact model id from the
session transcript — window size, pricing, the
/ccoboard, and Read-Cache staleness thresholds all follow the model the session actually runs on (Fable 5, Opus, Sonnet, Haiku). Switch/modelmid-day; nothing miscalibrates.config.modelis only the fallback. Claude 5 family added to the pricing table. - Bash & MCP output finally counted. Shell output was the biggest untracked
context consumer. PostToolUse now measures the actual
tool_responsesize for every tool instead of stat-based guessing — and any single result ≥10K tokens gets a one-line fix ("pipe through tail/grep", "read with offset/limit"). - Cache-break guard. Resume after a ≥5-min pause with a warm context and the
hook names what that break just cost (
re-warming 150K of context ≈ $0.86 extra) — the prompt cache TTL is 5 minutes, and a broken cache re-bills your whole context at 12.5× the cached rate. Batch pauses;/compactbefore stepping away. - Context-rot warning. On 1M-window models a one-shot notice fires at ~350K context: quality degrades in the ~300-400K "dumb zone" long before the window fills. Budget-% warnings can't catch this; a quality signal can.
- Observation → rule, automatically. Files read-but-unused in 3+ sessions now
surface in the hook path once per session (when the recurring waste is ≥30K
tokens) with the exact
/cco-shield applyfix — previously you had to run the CLI to see them. - CLAUDE.md size nudge. Project + user memory over 200 lines → a one-shot
pointer to
/cco-claudemd. Memory loads into every prompt; it's the most expensive place for bloat. - The headline. Session-end summary now leads with the number that matters:
★ CCO saved $654.36 this session — 86% of what it would have cost.
What's new in v4.3 — cache economics & a politer coach
- Real cache-aware pricing. Session cost is now computed at true prompt-cache rates (cache reads bill at 10%, writes at 125% of input) from exact transcript
File tree
More plugins by @egorfedorov.
View @egorfedorov’s profile@egorfedorovFAQ
claude-context-optimizer is a Claude Code plugin with 22 hand-picked skills for development work, indexed on Flowy. Install it with the command on its page. It includes cco-anatomy, cco-budget, cco-claudemd. Its skills do not fire on their own yet. Request auto-invocation to have Flowy route them as you prompt. Free and open source.
- What is claude-context-optimizer?
- Claude Code plugin that tracks token usage, identifies wasted context, and saves 30-50% on API costs. Heatmaps, ROI reports, budget alerts, efficiency scores, git-aware suggestions — all local, zero config.
- How do I install claude-context-optimizer?
- Run these in Claude Code: /plugin marketplace add egorfedorov/claude-context-optimizer then /plugin install claude-context-optimizer@cco. Then prompt normally.
- Does claude-context-optimizer auto-invoke its skills?
- Not yet. It is indexed on Flowy as a plain plugin. Request auto-invocation on its page and Flowy will route its skills for you as you prompt.
- Is claude-context-optimizer free?
- Yes. Flowy is free and open source, with nothing gated. You can read every skill in full before you install.
Is this plugin yours?
Claim it with GitHubSubmit a pluginPromote it