# context-doctor 🩺

[![CI](https://github.com/KushalP1/context-doctor/actions/workflows/ci.yml/badge.svg)](https://github.com/KushalP1/context-doctor/actions) [![npm](https://img.shields.io/npm/v/context-doctor)](https://www.npmjs.com/package/context-doctor)

**See what's eating your LLM context window — and fix it.**

Every long-running LLM conversation slowly fills up with junk: duplicated documents, 10k-token tool outputs nobody reads again, base64 blobs, stale history. You pay for those tokens on **every single call**, and model quality drops as the window fills.

`context-doctor` is a zero-config profiler + optimizer for LLM contexts. It works with **Claude, GPT, Gemini** message formats, and plugs into **Claude Desktop, ChatGPT (developer mode), Cursor, Claude Code** — any MCP-capable app — or runs standalone from the terminal.

Built and maintained by [gAI Ventures](https://gai.ventures).

```
Where the tokens go
────────────────────────────────────────────────────────
Tool results      ████████████████░░░░░░░░░░░░  57%  ~41k
System prompt     ██████░░░░░░░░░░░░░░░░░░░░░░  21%  ~15k
Assistant replies ████░░░░░░░░░░░░░░░░░░░░░░░░  13%  ~9.4k
User messages     ██░░░░░░░░░░░░░░░░░░░░░░░░░░   9%  ~6.5k

Findings (4)
────────────────────────────────────────────────────────
✖ Message #12 contains a base64/binary blob (~8.2k tokens). [save ~7.4k]
   → Never put base64 in text content — use the provider's file/image APIs.
▲ Tool result at message #7 (web_search) is ~6.1k tokens. [save ~4.9k]
   → Truncate or summarize large tool outputs before they enter history.
```

## Quick start (30 seconds)

**One command sets up everything** — detects Claude Desktop, Claude Code, and Cursor on your machine, wires in the MCP server, installs the Agent Skill, and registers the Claude Code every-prompt hook:

```bash
npx context-doctor install
```

`install` configures every app it detects and does not stop at the first problem: a corrupt Claude Desktop config still gets you Claude Code and Cursor. It does not pretend either. Any target that failed is named with a ✗ line, the summary reads "Done with N problem(s)" instead of "Done.", and the **exit code is 1**, so dotfiles and onboarding scripts can react. A broken config file is never overwritten; fix it and re-run.

That single command is also all it takes to **set up context-doctor on anyone else's machine**. Prefer a global install, or want the unreleased `main`? Both work (Node 20+):

```bash
npm install -g context-doctor && context-doctor install
```

Restart your apps, then just ask Claude: *"what's eating my context?"* (`npx context-doctor uninstall` reverses it.)

**No API keys, ever.** Everything is deterministic local code; when an LLM is needed (summarizing pruned history), the model already running in your app does it. The proxy forwards *your app's* credentials untouched — context-doctor itself holds nothing.

## What `install` actually does — and what happens in every session after

One run of `npx context-doctor install` writes five things (each config edit makes a `.backup` first; `uninstall` reverses all of it):

1. **Claude Desktop config** (`claude_desktop_config.json`) — registers the MCP server
2. **Claude Code config** (`~/.claude.json`) — registers the MCP server
3. **Cursor config** (`~/.cursor/mcp.json`) — registers the MCP server
4. **Agent Skill** → `~/.claude/skills/context-doctor/` — context-hygiene playbook for Claude Code
5. **Every-prompt hook** → `~/.claude/settings.json` — the per-query context check for Claude Code

**In every chat afterward (Claude Desktop, Cursor):** when the conversation starts, the app launches the MCP server, which hands the model standing instructions that stay in force for the whole chat:

- summarize large pastes and tool results instead of carrying them verbatim,
- reference earlier content instead of re-quoting it, never inline base64,
- once the chat passes ~30 turns or accumulates big pastes, *proactively offer to profile it*,
- answer any "what's eating my context / cost / latency" question by calling `profile_context`, not by guessing.

**In every Claude Code / Cowork session afterward:** all of the above via MCP, plus two more layers:

- the **skill** loads whenever context work is relevant, and
- the **hook runs on every single prompt you send**: lean sessions cost a ~1ms file-size check; once a session is heavy it profiles on growth events and injects a note the model sees with your message — actual token count, cost per message, the single largest recoverable waste — with instructions to work leaner and offer you compaction. It re-fires only after ~40% further growth, can never break a prompt (any failure exits silently), and logs each deep check to a small local ledger that feeds `context-doctor report`.

**What it never does:** delete or rewrite your history without asking (pruning is consent-only, and the model writes the replacement summary so nothing is lost silently), send data anywhere (everything runs on your machine), or touch an API key.

## Do you need MCP? Only sometimes — all the ways to use context-doctor

MCP is just one of six delivery mechanisms. It's only required when you want the AI **inside a chat app** to run the tools itself. Everything else works without it:

| How you use it | MCP needed? | What it requires |
|---|---|---|
| **CLI** — `analyze`, `optimize`, `session` on files/transcripts | ❌ No | Nothing but `npx` — works in any terminal, scripts, CI |
| **Proxy** — always-on optimization of your API apps | ❌ No | `context-doctor proxy` + one env var in your app |
| **Claude Code every-prompt hook** | ❌ No | Written by `install`; Claude Code invokes it directly |
| **Agent Skill** — hygiene behavior in Claude Code / claude.ai | ❌ No | A markdown file; `install` places it (or upload to claude.ai) |
| **Library** — `import { profileConversation } from "context-doctor"` | ❌ No | `npm install context-doctor` in your project |
| **In-chat tools** — Claude Desktop, ChatGPT desktop, Cursor chat | ✅ Yes | This is the only MCP piece — so the model itself can call `profile_context` / `optimize_context` mid-conversation |

Practical upshot: a developer who only wants cheaper, faster API calls never touches MCP (proxy + CLI). A Claude Code user gets the hook and skill without MCP either — the MCP server just adds in-chat tools on top. `install` sets up all of it at once precisely so you don't have to think about which mechanism is which.

## All commands at a glance

| Command | What it does |
|---|---|
| `context-doctor install` / `uninstall` | Wire (or remove) everything: MCP for Claude Desktop/Code/Cursor/Codex, the Agent Skill, the every-prompt hook |
| `context-doctor instructions [--copy]` | The ~90-token standing rules for claude.ai / ChatGPT preferences, for web and phones where no server runs |
| `context-doctor analyze <file>` | Profile a conversation: token breakdown, findings, cost + latency estimates. `--fail-over-budget` exits 1 on a breach, for CI |
| `context-doctor optimize <file>` | Apply the safe fixes; add `--strategy trim-tool-calls` for big inline file writes, `--strategy prune-history` for consented lossy compaction |
| `context-doctor session [file]` | Profile a Claude Code session: live context, findings, **measured tokens and prompt-cache economics**, **where the wall clock went** per tool, and **what its subagents cost** (their own windows, your bill; never in the parent's profile). Also reads ChatGPT data exports (`conversations.json`) |
| `context-doctor init [preset]` | Write a `.contextdoctorrc` from a preset (`chat`, `agent`, `batch`) — a budget you can adopt in one command and tune later |
| `context-doctor experiment --task "…"` | Run one task twice from the same commit, in a fresh session and forked from an `--existing` one, same model and tools; compare bill, cache split, wall clock, and whether `--check` passed. The only command here that spends money, so it caps spend per arm and refuses a dirty tree |
| `context-doctor diff <before> <after>` | Compare two profiles: what moved by category, which findings were resolved or introduced, and what it saves in money and latency |
| `context-doctor accuracy` | How much of what you are billed for is visible in your transcript — the fixed harness baseline and the per-turn injected content neither you nor the profiler can see |
| `context-doctor cursor [--list]` | Profile a chat from Cursor's local history (both storage formats) |
| `context-doctor report` | Machine-wide impact report (proxy savings persist across restarts): exact proxy savings, hook activity, recoverable waste in recent sessions |
| `context-doctor proxy` | Always-on local proxy that optimizes every Anthropic/OpenAI API request in flight (`/stats` for cumulative savings) |
| `context-doctor watch [file]` | Live monitor of a growing session/agent trace: token/cost line per change, findings as they appear |
| `context-doctor doctor` | Self-check the whole installation — one pasteable ✓/✗ diagnosis with fixes |
| `context-doctor dashboard` | Local savings dashboard on 127.0.0.1: tokens saved per day, sessions by context in use vs recoverable, budget status |
| `context-doctor statusline` | Claude Code status bar: live context vs window, cache share, cost. Wired by `install --statusline`; never overwrites a statusLine you already have |
| `context-doctor hook` | The every-prompt Claude Code hook (registered by `install`; you never run this yourself). Warning threshold tunable via `CONTEXT_DOCTOR_WARN_TOKENS` (default 80000) |
| `context-doctor-mcp` | The MCP server itself — stdio by default (what the installer wires); `--http [--port 8808] [--host H]` serves streamable HTTP at `/mcp` for URL-based clients like ChatGPT developer-mode connectors |

## What "always-on" means, per surface

| Where you run LLMs | Mechanism | Guarantee |
|---|---|---|
| Your own apps/agents (API) | `context-doctor proxy` rewrites every request in flight | **Every call, automatic** |
| Claude Code / Cowork sessions | `install` registers a **UserPromptSubmit hook**: every query measures the session; heavy sessions get injected hygiene guidance (silent when lean, rate-limited, never blocks a prompt) | **Every query checked** |
| Claude Desktop chat / Cursor | **MCP server instructions** — standing hygiene directives injected into every conversation where the server is enabled, plus prescriptive tool triggers | **Every conversation carries the rules** |
| claude.ai (web) / ChatGPT app | Upload `skills/context-doctor/SKILL.md` in the app's Skills settings for the same standing behavior | Manual one-time upload |

Nothing runs in the background for the Claude apps — the hook, skill, MCP server, and its instructions are all delivered by the app itself at the right moment. The proxy is the only long-running piece, and only your API-calling apps need it.

Optional belt-and-braces for any chat app: add one line to your profile preferences — *"Practice context hygiene: summarize large content instead of re-quoting it, and use context-doctor's tools when conversations get heavy."*

Or use the CLI directly, no install needed:

```bash
npx context-doctor analyze conversation.json --model claude-sonnet-5
```

```bash
npx context-doctor optimize conversation.json --out slimmed.json
```

Input is any of: OpenAI chat format, Anthropic messages format (with `system` and content blocks), or a bare `[{role, content}]` array. Use `-` to pipe from stdin.

Reports include **dollar and latency estimates**, not just tokens:

```
Cost:  ~$3.30 input per call · ~$3302 per 1k calls · ~13.2s of latency per call (estimates)
...
Potential recovery: ~5.9k tokens (~73% of context) ≈ $17.69 per 1k calls, 0.2s faster per call
```

## Profile your actual Claude Code sessions

```bash
npx context-doctor session            # profile your most recent session
npx context-doctor session --list    # browse sessions
```

Parses the transcripts Claude Code writes locally and answers "where did my tokens go today?" — it will happily tell you that one giant skill load is 67% of your context.

## Always-on: optimize every request automatically

Run the proxy and every Anthropic/OpenAI API call your apps make gets optimized in flight — no code changes:

```bash
npx context-doctor proxy
```

Then point your app or SDK at it:

```bash
export ANTHROPIC_BASE_URL=http://localhost:8787      # Anthropic SDKs / tools
export OPENAI_BASE_URL=http://localhost:8787/v1      # OpenAI SDKs / tools
```

The proxy dedupes repeated content, trims stale tool results, and strips base64 blobs from the message history of each request, then forwards it to the real API. Your API key passes through in headers untouched, streaming (SSE) works unchanged, and per-request savings are logged with upstream latency:

```
[context-doctor] POST /v1/messages → 200 in 842ms | optimized 7.3k → 518 tokens (2 changes) | session total: 6.9k tokens ≈ $0.021 saved
```

`GET http://localhost:8787/stats` returns cumulative savings (requests, tokens, estimated USD), **exact upstream usage** read from every response (JSON and SSE), and **prompt-cache advisories** — the proxy watches your real traffic and flags big stable prefixes missing `cache_control` or prefix churn that silently re-bills the cache. Per-model behavior via `--config`: The advisor now says *where*: for a large system/tools prefix it names the block to mark (the last tool definition, or the last system block), and when the older messages were byte-identical to the previous request it names the exact message to put `cache_control` on, with the token count that run is re-billing each turn. Anyone who already placed breakpoints is left alone.

```json
{ "routes": [{ "modelPrefix": "gpt", "strategies": ["strip-base64"], "keepRecent": 4 }] }
```

Because prompt caching matches byte-identical prefixes, deterministic strategies are chosen so repeated requests stay stable — but if you rely on aggressive cache prefixes, start with `--strategy strip-base64 --strategy dedupe` and add more as you verify.

> **Note on desktop chat apps:** Claude Desktop and the ChatGPT app talk to their own backends — no tool can sit in that path. For those, use the MCP integration below and add a line to your custom instructions like: *"When a conversation gets long or includes large pasted content, proactively use context-doctor's profile_context tool and tell me what to trim."* The model will then invoke it on its own.

## Use with the Claude & ChatGPT apps

`context-doctor` ships an MCP server, so the AI itself can profile and slim context on demand. Before the setup table, the honest question: **on which surfaces does it act by itself, and on which does it only nudge?** MCP gives a server no way to see the conversation or intercept a turn; only a hook or a place in the data path can do that.

| Surface | Runs by itself | What that means |
|---|---|---|
| **Claude Code** | Yes: hook on every prompt, status line on every refresh | Past ~80k tokens the model receives hygiene guidance naming the largest waste; compaction is offered. Measured: 115 automatic checks, 48 warnings, across 32 sessions on one machine |
| **Cursor** | **Yes**, since 0.15: Cursor loads Claude Code's hook config (`~/.claude/settings.json`) and runs the same hook on every agent prompt, passing its own transcript. Output is accepted through Cursor's Claude-compat layer | Same guidance as Claude Code, inside Cursor's agent, for everyone who ran `install`. Before 0.15 the hook fired but could not read Cursor's transcript format, so it said nothing |
| **API traffic through the proxy** | Yes: every request rewritten in flight | Fewer tokens, guaranteed, model not consulted |
| **Claude Desktop** | The standing instruction in every chat (we confirmed in the app bundle that Desktop's `LocalMcpServerManager` reads it), a one-click `context_checkup` prompt, and since 0.17 a `profile_context` the model can actually afford to call from chat | Until 0.17 the tool wanted the whole conversation as its argument, so calling it from chat meant re-typing 50k tokens; nobody did, and Desktop's log showed zero calls in a month. Now the model passes a ~100-token **sketch** (turn count, the large or repeated blocks) and gets a sized estimate, findings and the fix to apply. Still a nudge, not a hook: Desktop chat has no hook API and no transcript on disk |
| **Codex (OpenAI): ChatGPT.app's Codex tab, the Codex IDE extension, the `codex` CLI** | **Yes**, since 0.16: `install` writes the hook to `~/.codex/hooks.json`, the MCP server to `~/.codex/config.toml`, and the skill to `~/.codex/skills/`. Codex uses Claude Code's hook contract almost verbatim and passes its own rollout transcript, which carries the API's real usage figures | Same guidance as Claude Code, from measured tokens. One extra step, Codex's rule not ours: a new hook runs only after you trust it once (type `/hooks` in Codex). `session` and `session --list` read Codex rollouts too |
| **ChatGPT chat UI** | No | No MCP, no hooks, no data path in the chat product itself. Use Codex, or a developer-mode connector at a URL you host |

So "every chat inherently better" is true for Claude Code, Cursor, Codex and the proxy; for Claude Desktop it is "the rules ride in every chat and the checkup is one cheap tool call away"; and not a claim we make for the ChatGPT chat UI.

**Where there is no hook and no MCP at all** (claude.ai on the web, the Claude and ChatGPT phone apps, plain ChatGPT): the app's per-account preferences are read on every turn, which is the closest those surfaces have to a hook. `context-doctor instructions --copy` puts the ~90-token rules on your clipboard and tells you where to paste them (claude.ai Settings > Profile; ChatGPT Settings > Personalization > Custom instructions).

**Do you need to configure anything by hand? Usually no:**

| App | Setup |
|---|---|
| Claude Desktop | `npx context-doctor install` writes the config — just restart the app. Or one click: download `context-doctor-<version>.mcpb` from the [latest release](https://github.com/KushalP1/context-doctor/releases) and open it (Settings > Extensions). The bundle runs on Desktop's own Node, no npm needed |
| Claude Code | Same command — MCP + skill + every-prompt hook, all automatic |
| Cursor | Same command — writes `~/.cursor/mcp.json`; the every-prompt hook is picked up from Claude Code's config, which Cursor reads |
| Codex (OpenAI) | Same command — `~/.codex/config.toml`, `~/.codex/hooks.json`, `~/.codex/skills/`. Then, once, `/hooks` in Codex to trust the hook |
| ChatGPT (developer mode) | **Manual + a reachable URL** — ChatGPT connects to servers over the internet, never local commands. Run `context-doctor-mcp --http` on a host/tunnel, then add the URL as a connector. Normal ChatGPT (no dev mode) has no MCP — use the CLI |

For any other MCP client, the server entry is:

```json
{
  "mcpServers": {
    "context-doctor": {
      "command": "npx",
      "args": ["-y", "context-doctor-mcp"]
    }
  }
}
```

### How it works in Claude Desktop, step by step

1. Run `npx context-doctor install` (writes the config above for you) and restart Claude Desktop. Or open the `.mcpb` from the latest release: same server, no npm, installs as an Extension.
2. From then on, **every conversation carries context-doctor's standing instructions**. The MCP server hands them to Desktop on connect and Desktop puts them in front of Claude: summarize big pastes instead of re-quoting them, refer to earlier content by name, never inline base64, and past ~30 turns or on any question about tokens, cost, speed or limits, call `profile_context` before answering.
3. That call is cheap on purpose. Claude cannot export a Desktop chat, so it passes a **sketch**: how many turns, which blocks are large, repeated, stale or images, with one size hint each (~100 tokens). The server sizes it (±30%, and it says so), prices the per-turn re-read (on a subscription that is what spends your usage limit), and returns ranked findings with the action for each: "summarize *the nginx config* into the points still needed", "refer to *test output* by name", "offer a 300-token handoff summary for a fresh chat". The reply ends with an instruction to apply the top one, not just suggest it.
4. One click instead of asking: the `context_checkup` prompt in the **+** menu sends that request for you.
5. Say *"optimize it"* on an exported conversation and Claude applies the safe fixes; if you agree to pruning old history, **Claude itself writes the replacement summary** (that's the no-API-key summarization).
6. For the same rules on your phone and on claude.ai, where no MCP server runs: `context-doctor instructions --copy`, then paste into Settings > Profile > personal preferences.

What this does not do: read the chat behind Claude's back or trim it for you. Desktop chat has no hook API and no transcript on disk (checked in the app bundle, v2.2553). The model does the trimming, when the rules and the checkup tell it to.

### How it works in ChatGPT, step by step (honest version)

ChatGPT's MCP support differs fundamentally from Claude Desktop's: **it never spawns local processes**. Its custom connectors (developer mode) have OpenAI's servers connect to a **URL** — so the MCP server must be reachable from the internet.

1. **Normal ChatGPT (no developer mode): no MCP at all.** context-doctor still helps via the CLI: export the conversation and run `npx context-doctor analyze chat.json --model gpt-5` / `optimize` — no account settings required.
2. **ChatGPT developer mode**: run our HTTP transport somewhere reachable — `context-doctor-mcp --http --port 8808` on a small host (bind `--host 0.0.0.0` there), or expose your machine temporarily with a tunnel (`ngrok http 8808`). Then Settings → Connectors → Advanced → Developer mode → add connector with URL `https://<your-host>/mcp`.
3. Once connected, GPT gets the same three tools with the same trigger guidance: ask *"what's eating my context?"* → it calls `profile_context`; *"optimize it"* works the same, including GPT writing the pruning summary itself.

Security note for step 2: the HTTP endpoint is unauthenticated — put it behind your tunnel's auth or a reverse proxy if it stays up long-term.

### claude.ai on the web and the phone apps

Your local MCP server can't reach the website, but the behavior can. Two options: `context-doctor instructions --copy` and paste into Settings > Profile > personal preferences (applies everywhere you are signed in, phone included), or upload `skills/context-doctor/SKILL.md` under Settings → Capabilities → Skills. Either way web and mobile conversations gain the same standing habits: summarize, don't re-quote, offer a handoff when heavy.

### MCP tools

| Tool | What it does |
|---|---|
| `profile_context` | Token breakdown by category, largest messages, findings with estimated savings. Takes either `conversation` (full JSON or text) or `sketch` (turns + large/repeated blocks, for chat apps) |
| `optimize_context` | Rewrites the conversation: dedupe, trim stale tool results, strip base64, optional history pruning |
| `context_best_practices` | Curated checklist, optionally specialized for Anthropic / OpenAI |

## Use it as a library

```ts
import { parseConversation, profileConversation, optimizeConversation } from "context-doctor";

const profile = profileConversation(parseConversation(chatJson), "claude-sonnet-5");
console.log(profile.totalTokens, profile.findings);

const { conversation, tokensBefore, tokensAfter } = optimizeConversation(chatJson, {
  strategies: ["dedupe", "trim-tool-results", "strip-base64"],
});
```

## Subagents: their own windows, your bill

A subagent has its own context window, so its tokens are correctly absent from the parent's profile. They are not absent from the bill. Claude Code writes each one to `<session>/subagents/agent-<id>.jsonl`, and `session` now reads them:

```
Subagents (their own windows, your bill)
────────────────────────────────────────────────────────
49 subagent(s) made 4560 API calls: 626.0M input billed, 1.0M output, ~$608.51.
That is 10% on top of the parent session's own input cost ($5846.45), and none of it appears in the profile above.
    $42.52   51 calls  ctx   319k    6m  You are auditing part of a FastAPI backend at /Users/kp/tech
    $37.48  107 calls  ctx   229k    6m  You are auditing the Turtle AI backend (FastAPI, Python) at
  … and 44 more
18 subagent(s) ended above 200k tokens of context. A subagent that big is doing a main session's job; give it a narrower brief, or split the task.
```

Per subagent: what it was asked, how many calls it made, the context it ended with, how long it ran, and its cost at list price with cache reads and writes priced correctly. Models without a price on file are counted but marked unpriced rather than costed at zero. On this machine that was 195 subagents across 19 sessions and about $1,844 at list price that no profile had ever shown.

## The same number inside VS Code and Cursor

An extension in [`vscode/`](vscode/) puts context health in the editor's own status bar:

```
⌁ ctx 848k · 85% · cache 100% ⚠
```

It reads the newest Claude Code transcript for the open workspace folder, shows live context, share of window and cache share, turns to the warning colour past 70% (configurable), and opens a terminal running `context-doctor session` when clicked. Nothing leaves the machine; it only reads files Claude Code already writes. Until it is on the marketplace, build and install it locally:

```bash
cd vscode && npm ci && npm run package
code --install-extension context-doctor-vscode-0.1.0.vsix     # or: cursor --install-extension …
```

## Context health in Claude Code's status bar

```bash
context-doctor install --statusline
```

Claude Code shows the first line a `statusLine` command prints, on every refresh, while you type. With this on, that line is the number that matters:

```
ctx 801k/1.0M ▮▮▮▮▮▮▮▮░░ 80% ⚠ · cache 100% · $12.34
```

Live context against the model's window, a warning mark from 70%, the share served from cache, and the session's cost. It reads the size from the status payload when Claude Code provides it, and otherwise from the last 256KB of the transcript (about 1ms on a 20MB file; 80ms end to end including Node startup). It is opt-in and polite: there is only one status line, so it never overwrites one you already have, and `uninstall` removes only its own. Any failure prints nothing rather than an error.

This is the "editor status bar" roadmap item, delivered for the editor most users of this tool are actually in. A VS Code / Cursor extension for the same number remains future work.

## Does a smaller context actually help? Measure it

Everything else in this tool measures what is *in* the context. None of it can tell you whether the task succeeded, so a smaller transcript can be a cheaper failure. `experiment` is the honest test:

```bash
context-doctor experiment \
  --task "Add a null check to parseHeader in src/parse.ts and make the tests pass" \
  --check "npm test" \
  --existing 9c6f9dc9-457b-4d09-bf6d-a499c2f2f919 \
  --model sonnet --budget 1
```

It runs the task in a **fresh** session, resets the tree to the starting commit, then runs it again **forked from your existing session** (`--resume … --fork-session`, so your real session is never touched), same model, same tools. For each arm it records what you were billed (input, cache read, cache write, output), cost, wall clock, turns, and whether your `--check` command passed, then puts them side by side:

```
                               fresh      existing
──────────────────────────────────────────────────
billed input                     20k           65k
  of which cache read              0           60k
  of which cache write          8.0k          4.0k
cost                          $0.110        $0.420
wall clock                        4s            9s
check                           PASS      FAIL (1)

Verdict: fresh was cheaper AND passed; existing failed the check. Clear win for fresh.
```

The verdict line is the point: cheaper only counts if it also passed. Because this spends your Claude budget it caps spend per arm (`--budget`, default $1), refuses to start on a dirty tree (both arms must begin from one commit, and the tree is reset between them), and refuses to run from inside a Claude Code session, where the CLI cannot start. `--dry-run` shows the exact commands first.

## Exact counts, and what they teach the estimator

The default token count is a chars-per-token heuristic so everything runs with no key and no tokenizer. Its error is content-dependent, and there is no honest way to fix that from transcripts alone (the billed number includes content the transcript never sees). `analyze --exact` fetches a true count for the exact bytes just estimated (Anthropic's count-tokens API with `ANTHROPIC_API_KEY`; tiktoken for GPT if installed) and prints the drift.

Since 0.13.9 it also **remembers the comparison**, per model family, on this machine, and later estimates for that family are scaled by it. Nothing about this is silent: the profile header says `estimates calibrated +12% from 3 exact count(s) you ran on this machine`. No exact count ever run means no calibration and unchanged numbers; out-of-range samples are ignored; `CONTEXT_DOCTOR_NO_CALIBRATION=1` returns to the raw heuristic.

## What it detects

- **Oversized tool results** — the #1 context killer in agent loops
- **Oversized tool calls** — a `Write` or a `cat > file <<EOF` puts the whole file in context permanently. In file-heavy agent work these outweigh every tool result combined, and `--strategy trim-tool-calls` reclaims them
- **Duplicate content** — the same doc/result pasted twice
- **Near-duplicates** — the same doc re-pasted with different surrounding words (shingle similarity, ≥60%)
- **Repeated file reads** — the same file pulled in three or more times, every copy still in context. Counts shell reads too (`cat`, `head`, `tail`, `less`), which is where most of them hide in agent sessions
- **Retained error output** — stack traces and failed commands kept verbatim long after the fix landed
- **Repeated identical tool calls**, split into the two things they can mean: a **retry** (the same call after a failure, where the fix is in the error text, and three or more is a loop) and a **re-read** (the same call after a success, where the model forgot it already had the answer). Across 42 local sessions that was 15 retries against 151 re-reads, so the old combined advice was wrong for most of them
- **Base64 / binary blobs** in text content — checked by character distribution, not just alphabet, so hex digests and long identifiers are not mistaken for encoded binary
- **Long history** past the point where models track the middle
- **Cache-hostile ordering** — volatile content before stable content breaks prompt caching (Anthropic `cache_control`, OpenAI automatic prefix caching)
- **Window pressure** — usage % against the target model's real context window

## What it fixes (deterministically — no LLM calls, no API keys)

| Strategy | Lossy? | Default |
|---|---|---|
| `dedupe` — replace repeated content with a reference | No | ✅ |
| `trim-tool-results` — truncate stale tool outputs | Mostly no | ✅ |
| `strip-base64` — remove inline binary blobs | No (for the model) | ✅ |
| `trim-tool-calls` — shrink the arguments of calls that already ran | Mostly no | opt-in |
| `prune-history` — collapse old turns into a stub for summarization | Yes | opt-in |

`trim-tool-calls` is the big one for agent sessions. Writing a file through a tool call puts the entire file in context permanently, so in file-heavy work the calls outweigh every tool result combined — on a real 278k-token session, the default set reached 248k and adding `trim-tool-calls` reached 102k. It is opt-in because it edits what the model itself wrote, and that is measured, not cautious: across 42 real sessions, of 73 large writes 18 were later edited and **16 of those edits had no read in between**, meaning the model built the edit from its own earlier `Write` input, at a median distance of 43 messages. Trimming those would turn each into a failed edit plus a recovery read. In the offline optimizer the rest of the conversation is known, so exactly those writes are left intact and the 63% never touched again are trimmed; in the live proxy the future is not known, which is why it stays off there unless you turn it on.

**Optimization is cache-aware.** Prompt caches match a byte-identical prefix, so editing a message in the middle invalidates everything after it — and the naive "trim everything older than the last N messages" boundary moves every single turn. On a 25-turn agent conversation that invalidated the cached prefix on 22 of 24 turns, paying the 1.25x cache-write price on the whole prefix to save a few hundred tokens. The trim boundary is quantized so it holds still between steps (8 of 24 on the same fixture), while still reaching 15 of 20 tool results.

The step size is a trade-off, not a formula, and it is yours to set: `"trimBoundaryStep": 20` in `.contextdoctorrc` (default 10). Measured on a growing agent session at 400 turns: a step of 10 invalidated the cache on 21% of turns with ~2 stale results waiting on average; 20 gave 12% and ~4; 40 gave 10% and ~9. Adaptive steps were worse everywhere, because a step that changes size moves the boundary by itself. Heavy API users who lean on caching want a bigger step; interactive users who want stale output gone promptly want a smaller one.

Everything the optimizer does is inspectable: it prints exactly which messages changed and how many tokens each change saved.

**Summarization without an API key:** when `prune-history` runs through the MCP tools, context-doctor hands a digest of the pruned turns back to the model that called it (the Claude/GPT already running in your app) and asks *it* to write the replacement summary — LLM-quality compaction, zero extra cost, no keys.

## The Agent Skill

`skills/context-doctor/SKILL.md` (installed by `npx context-doctor install`) teaches Claude to practice context hygiene proactively: summarize big tool results after consuming them, never re-paste duplicated content, keep stable content cache-friendly, and offer compaction when a session gets heavy. In Claude Code and Cursor the every-prompt hook enforces the heavy-session part; the skill covers the habits in between.

## Measuring the impact: `context-doctor report`

```bash
npx context-doctor report
```

One report for your whole machine, led by a headline of **tokens context-doctor saved**, built only from measured sources:

- **exact** proxy savings (real before/after on every request),
- **exact** savings from every optimization applied via the CLI or the in-chat tools — split by model family (Claude vs GPT), with dollar estimates,
- **observed per-session shrinkage**: real context reductions recorded between the hook's deep checks after hygiene warnings — shown per session in the table alongside remaining waste.

Honest measurement note: proxy numbers are exact. Session numbers are measured-now. What no tool can report is the counterfactual — tokens Claude *avoided* adding because of the hygiene guidance — since the same session can't be re-run without it. The report says so instead of inventing a number.

## Context budgets (`.contextdoctorrc`)

Drop a `.contextdoctorrc` in a project (or your home directory) and context-doctor enforces your limits instead of its defaults:

```json
{
  "budget": { "maxTokens": 120000, "maxCostPerMessageUsd": 0.5, "maxWindowPct": 60 },
  "strategies": ["dedupe", "trim-tool-results"],
  "keepRecent": 6
}
```

The nearest file wins (walking up from the working directory, then `~`). `analyze` and `session` print a budget verdict, the every-prompt hook uses `maxTokens` as its warning threshold and names the breach to the model, and `optimize`/`proxy` pick up the defaults when you do not pass flags.

## Prompt-cache economics (Claude Code sessions)

Caching is the largest lever on LLM cost, and transcripts record exactly how it went — so `session` reports it as fact rather than estimate:

```
Prompt cache: 95.6% of input served from cache across 1117 requests
  read 558.6M · written 25.5M · uncached 2k
  input cost $438.88 — caching saved $2481.89 against $2920.76 uncached (list prices)
```

A cache read bills at ~10% of input while a write bills at ~125%, so a session that keeps invalidating its prefix can cost *more* than one with no caching at all. context-doctor warns on the two failure modes: a **low hit rate** (something early in the prompt changes every request) and **cache churn** (writes rivalling reads).

## Enforce a budget in CI

```bash
npx context-doctor analyze conversation.json --fail-over-budget
```

Exits 1 when the `.contextdoctorrc` budget is breached, so a pull request can be gated on context size the same way it is gated on tests.

## Sharing a profile safely

A profile quotes message previews and file paths, so pasting one into an issue pastes fragments of real work. `--redact` keeps every number and the finding structure but replaces content with `[redacted]` and masks paths:

```bash
npx context-doctor session --redact
```

`context-doctor doctor` is safe to paste as-is: it reports integration status, never conversation content.

## Performance: what context-doctor itself costs

A tool that promises speed must be near-free. Measured overhead per touchpoint:

| Touchpoint | When it runs | Overhead |
|---|---|---|
| Every-prompt hook (Claude Code) | Every prompt | **~80ms** (Node startup; logic ~1ms). Lean sessions exit on a single `stat()` — the transcript is never read. Full profiling (~200ms on a 4MB session) happens only when the transcript has grown ~40% since last checked |
| MCP server | Spawned once per app session | Tools run only when called; standing instructions cost **~110 tokens per conversation** — deliberately terse |
| Proxy | Per API request | ~1–3ms of CPU (parse → optimize → re-serialize) against typical model latencies of hundreds of ms; responses stream through chunk-by-chunk, never buffered |
| Skill | Loads only when relevant | ~1k tokens while active; its always-present description is ~60 tokens |
| Profiling a session | On demand, and on hook growth events | ~160ms for an 8.5MB / 1,855-message transcript (near-duplicate pairs that cannot clear the similarity bar are skipped without comparison) |
| CLI / library | Only when you run it | Not in any hot path |

Net effect is strongly negative overhead: the tokens these touchpoints save on every subsequent call dwarf what they cost.

## Why token counts are "~" (and where they are exact)

Counting exactly needs each provider's tokenizer, so the default is a calibrated chars-per-token heuristic (denser for code and JSON). It is good enough to rank what is heavy and to measure the effect of a fix, and it keeps the tool offline and zero-config.

Two ways to get real numbers instead:

- **`analyze --exact`** uses the Anthropic count-tokens API for Claude models (set `ANTHROPIC_API_KEY`; opt-in network call, key never stored) or tiktoken for GPT models (install it alongside), and reports how far the heuristic drifted.
- **Sessions report measured tokens automatically.** Claude Code transcripts record what the API actually charged, so `session`, the hook and the reports use that figure when it is present — no key, no estimate.

One honest caveat worth knowing: a transcript stores the conversation, **not** the harness's system prompt, tool schemas or skills. Measured against the API's own numbers here, a message-only estimate undercounts the true context by roughly 60%. That is why sessions prefer the reported figure, and why the message breakdown is labelled as covering messages only.



Exact counts require each provider's private tokenizer. `context-doctor` uses a calibrated chars-per-token heuristic (denser for code/JSON) that lands within ~10% — plenty accurate for finding what's heavy and measuring savings, and it keeps the tool fully offline with zero configuration.

## Roadmap

See [ROADMAP.md](./ROADMAP.md) for the full plan with rationale. Headlines: **v0.5** trust & automation (tag-based publishing, `doctor` self-check, live `watch`), **v0.6** accuracy (exact tokenizers, semantic dedupe, more session formats), **v0.7** proxy pro (response accounting, prompt-cache advisor), **v1.0** budgets + local dashboard. Non-goals, permanently: cloud services, telemetry, silent history rewriting, mandatory API keys.

Contributions welcome — this project is small on purpose. Open an issue before a big PR.

## Releasing (maintainers)

```bash
npm version patch        # or minor/major — bumps package.json + git tag
npm test                 # 131 tests must pass; CI runs the same on 3 OSes x Node 20/22/24
npm publish              # prompts for the npm 2FA code
git push --follow-tags
npm run build:mcpb       # context-doctor-<version>.mcpb for Claude Desktop; attach it to the GitHub release
```

**What the npm download number measures.** `install` writes `npx -y context-doctor-mcp` into MCP configs, and npx re-fetches the tarball whenever a new version exists. So every release is downloaded once by every active install within about a day, and the daily count is almost entirely those refreshes: on this package, release days run ~170 downloads and non-release days ~27. Read it as "size of the active installed base × number of releases", not as new users — a quiet week with no releases will look like a decline while nothing has changed. Two corollaries: the release-day figure is a live count of machines running context-doctor, and a broken release reaches all of them automatically, which is why `prepublishOnly` runs the full test suite. npm's stats also lag by several days and occasionally record a day as zero; a zero on a release day is a gap in their pipeline, not in usage.

Known gotcha: if `npm publish` fails with **`404 Not Found - PUT …/context-doctor`** on a package that clearly exists, the real cause is an **expired npm login token** — npm reports unauthenticated publishes as a 404, not a 401. Check with `npm whoami`; if that errors, run `npm login` and publish again.

Also keep the MCP server version in `src/mcp.ts` in sync with `package.json`, and remember `dist/` is committed — run `npm run build` before committing so the CI dist-sync check passes.

## Contributing

Issues and PRs welcome — see [CONTRIBUTING.md](./CONTRIBUTING.md) for the six rules that keep this tool trustworthy (no API keys, nothing leaves the machine, no silent data loss, measurements not guesses, the hot path stays cheap, tests with every change) and a list of good first issues. What is planned next lives in [ROADMAP.md](./ROADMAP.md).

## License

MIT © [gAI Ventures](https://gai.ventures)