Rolling Context for Claude Code

A transparent proxy that gives Claude Code rolling context compression — old messages get automatically summarized while recent messages stay fully verbatim. You never hit the context wall, and you never lose important details.

Zero config. Uses your existing Claude Code auth. No API key needed. Just install and forget.

Claude Code's built-in /compact replaces your entire conversation with a lossy summary. After a few compactions, you're summarizing a summary of a summary. This plugin only compresses old messages — recent context stays untouched.

It's also a cost story: every token you carry in context gets re-billed on every turn (at cache-read rates), so an unmanaged session's input cost grows with the square of its length. Capping the prefix makes it linear — the math works in relative units and holds for every model, and it matters more the bigger the context window, not less.

/compact vs Rolling Context

The economics: why capping the prefix matters

No price cards needed — the argument works in relative units that hold for every Claude model. Take the model's fresh-input-token rate as 1×. The API bills:

Claude Code re-sends the entire conversation on every turn. Even with caching working perfectly, each turn costs prefix size × 0.1. So a session's total input cost is the sum of the prefix over all turns:

• Unmanaged context (big window, auto-compact only near the limit): the prefix grows every turn, so total cost grows with the square of session length. A 1M window doesn't fix this — it just lets the prefix run to 900K+ before anything stops it, and every one of those tokens costs 0.1× again on every single turn.
• Rolling context: the prefix is capped between TARGET and TRIGGER, so total cost grows linearly.

The cache-miss blast radius matters even more in interactive use. The prompt cache has a TTL. Read a diff, think for a while, get coffee — and the next turn re-writes the whole prefix at 1.25×. At a 900K prefix, one cold turn bills the equivalent of ~1.1M fresh input tokens. With the prefix capped at ~100K, the identical cold turn is ~9× cheaper. Compression doesn't just shrink the average turn — it caps the worst one.

What compression itself costs: each cycle re-writes the new (much smaller) prefix once, and in native mode the summarization request is itself a cache read — a few hundred fresh tokens, measured (see below). Ballpark: sessions that accumulate past ~100K of context — a couple of hours of real work — come out ahead, and the gap compounds from there. Short sessions are a wash; don't expect magic on a 20-minute task.

On Pro/Max subscriptions none of this is dollars, but the same math applies in a different currency: rate-limit accounting weights cache reads far below fresh input, so the identical curves decide how fast you burn your 5-hour window.

Honest note: if cost were the only goal, lowering Claude Code's auto-compact threshold (CLAUDE_CODE_AUTO_COMPACT_WINDOW) buys a similar spend curve for free. What it can't buy is quality under repetition: built-in compaction replaces the whole conversation with a lossy summary every time it fires — at a low threshold it fires often, and you're soon working from a summary of a summary of a summary. The rolling design exists so aggressive compression doesn't cost you the session: recent work stays verbatim, and old work lives in one continuously-merged timeline instead of N generations of loss.

How It Works

Claude Code  ──HTTPS_PROXY──►  Rolling Context MITM (:5590)  ──►  core (:5588)  ──►  Anthropic API
                         │
                         ├─ context < 100K tokens? pass through unchanged
                         │
                         └─ context > 100K tokens?
                              1. summarize old messages in the background
                                 (native mode: your session's own model,
                                  served almost entirely from prompt cache)
                              2. keep ~40K tokens of recent messages verbatim
                              3. inject compressed context on next request
                              4. never blocks, never adds latency

How it attaches: the plugin sets HTTPS_PROXY (and NODE_EXTRA_CA_CERTS), not ANTHROPIC_BASE_URL. A non-anthropic ANTHROPIC_BASE_URL trips Claude Code's Remote Control / GrowthBook gate; routing through HTTPS_PROXY keeps the destination api.anthropic.com, so the gate stays untouched. A small MITM front-end terminates TLS with a locally generated CA (trusted via NODE_EXTRA_CA_CERTS) so bodies stay visible for compression; everything that is not api.anthropic.com (or your configured endpoint) is blind-tunnelled untouched.

Instead of replacing everything, this plugin:

1. Keeps recent messages untouched — recent context stays verbatim
2. Only compresses when needed — triggers at 100K (real API token count), compresses old messages, grows naturally until next trigger
3. Merges summaries — each compression cycle merges with the previous summary, building a rolling timeline
4. Never blocks — compression runs in the background, applied on the next request
5. Full transcripts preserved — Claude Code still saves everything to JSONL in ~/.claude/projects/

Install

Option 1: Claude Code Plugin (recommended)

Run these two commands inside Claude Code:

/plugin marketplace add https://github.com/NodeNestor/nestor-plugins
/plugin install rolling-context

Restart your terminal and start a new Claude Code session. On the first start, the plugin configures HTTPS_PROXY + NODE_EXTRA_CA_CERTS and starts the proxy. Since those env vars only take effect on the next terminal, restart your terminal once more — after that, everything works automatically. No pip install needed — pure Python stdlib (a local CA is generated with cryptography, which Claude Code already ships).

Option 2: Manual install

Linux / macOS:

git clone https://github.com/NodeNestor/claude-rolling-context.git ~/claude-rolling-context
cd ~/claude-rolling-context
bash install.sh

Windows (PowerShell):

git clone https://github.com/NodeNestor/claude-rolling-context.git $HOME\claude-rolling-context
cd $HOME\claude-rolling-context
powershell -ExecutionPolicy Bypass -File install.ps1

The installer configures HTTPS_PROXY + NODE_EXTRA_CA_CERTS and registers the plugin. Restart your terminal and you're done. Requires Python 3.7+ (no pip install needed — pure stdlib).

How Compression Works

When the message array exceeds the trigger threshold:

BEFORE (hit 100K trigger):
  [msg1] [msg2] [msg3] ... [msg60] [msg61] ... [msg100]
  |<——————————————— ~105K tokens ——————————————>|

AFTER (compressed):
  [rolling summary] [ack] [msg61] ... [msg100]
  |<— ~5K summary —>|    |<—— verbatim ————————>|

NEXT CYCLE (grows back to 100K, triggers again):
  [rolling summary] [ack] [msg61] ... [msg140]
  |<——————————————— ~105K tokens ——————————————>|
  → new summary merges old summary + msg61-msg100
  → keeps msg101-msg140 verbatim

The summary preserves a structured record of everything that happened:

• Active Goal — what the user is currently asking for, constraints, do/don't rules
• Previous Goals — completed or shifted-away-from goals (kept brief)
• Timeline — chronological numbered steps: every file change, decision, error, and user instruction
• Current State — what's done, in progress, and next
• Key Details — file paths, configs, decisions that must survive compression

Goals evolve naturally across rolling compressions — the latest request stays prominent while completed goals move to the previous section. User instructions are never lost.

How the summary is generated (native mode)

By default the proxy doesn't build a separate summarization request. It clones the exact request Claude Code just sent — same model, system prompt, and tools, with the conversation truncated at the cut point — and appends one user message asking for the summary (the same way Claude Code's own /compact works). Two big wins:

• It's a prompt-cache read. The cloned prefix was just sent by the chat request, so the API serves it from cache. Measured in practice: a ~72K-token compression request cost ~400 fresh input tokens.
• It's genuine Claude Code session traffic. Pro/Max subscription OAuth tokens are classified server-side — standalone requests that don't look like Claude Code get routed to the overage lane and rejected with 429. The cloned request passes because it is the session's own request shape.

Setting ROLLING_CONTEXT_MODEL pins a different summarizer model (the request shape stays native, but a different model means no prompt-cache reuse). If the session's model answers the compaction request with a 200 and no text — on claude-fable-5 a safety classifier can return stop_reason: refusal with no content, deterministically for that conversation — the proxy logs the stop reason and retries once on ROLLING_CONTEXT_FALLBACK_MODEL instead of cooling the conversation down for nothing (#11). Configuring any ROLLING_CONTEXT_SUMMARIZER_* variable switches to a standalone flattened request instead — see below.

Using any API or a local model for compression

Summarization can run on a completely separate endpoint — any Anthropic-format API, or any OpenAI-compatible one (Ollama, LM Studio, vLLM, OpenRouter, DeepSeek, Groq, ...):

# Separate Anthropic API key (billed there instead of your subscription)
export ROLLING_CONTEXT_SUMMARIZER_KEY=sk-ant-api03-...

# Local model via Ollama / LM Studio / vLLM (OpenAI-compatible)
export ROLLING_CONTEXT_SUMMARIZER_URL=http://127.0.0.1:11434
export ROLLING_CONTEXT_SUMMARIZER_FORMAT=openai
export ROLLING_CONTEXT_MODEL=qwen3:8b   # required for openai format

# OpenRouter (or any hosted OpenAI-compatible API)
export ROLLING_CONTEXT_SUMMARIZER_URL=https://openrouter.ai/api
export ROLLING_CONTEXT_SUMMARIZER_FORMAT=openai
export ROLLING_CONTEXT_SUMMARIZER_KEY=sk-or-...
export ROLLING_CONTEXT_MODEL=deepseek/deepseek-chat

Non-Anthropic endpoints (GLM/Z.ai, DeepSeek, OpenRouter, a local gateway)

Point Claude Code at your endpoint the usual way and install as normal. The hook chains it — ANTHROPIC_BASE_URL stays your endpoint, the MITM front-end intercepts that host, and your endpoint is stored as ROLLING_CONTEXT_UPSTREAM — so both the chat traffic and the background compaction go there. No extra configuration.

Third-party endpoints implement the /v1/messages shape to varying depth. Native mode clones the request Claude Code just sent, which can include cache_control breakpoints and tool_choice; if your endpoint rejects one of those, the proxy retries the compaction as a plain flattened summary using the session's own model rather than giving up for the cooldown period. Both paths are covered by tests/ , which runs the real proxy against a mock endpoint — including a deliberately strict one that rejects those fields.

Check it landed — the startup banner names the compaction endpoint separately from the chat endpoint, and they should match:

Forwarding to: https://api.z.ai/api/anthropic
Compacting via: https://api.z.ai/api/anthropic (third-party — flattened fallback armed)

On auto-compact. Rolling Context doesn't disable Claude Code's auto-compact; it keeps the reported context low enough that auto-compact never gets near its threshold. If auto-compact still fires on a custom endpoint, compaction is failing — read ~/.claude/rolling-context-debug.log rather than raising the threshold. Claude Code sizes the window from the model name and falls back to 200K for names it doesn't know; for a non-claude-* model you can state the real window with CLAUDE_CODE_MAX_CONTEXT_TOKENS, which is cleaner than encoding it in the model name.

Architecture

The proxy is fully stateless — no sessions, no databases, no tracking. It works by hashing message content:

1. When a response comes back from the API with a high token count, the proxy compresses the messages and stores the result keyed by content hashes
2. On the next request, it hashes the incoming messages and checks for a matching compression
3. If found, it swaps in the compressed version transparently

This means:

• Multiple conversations work automatically — each conversation has unique content, unique hashes, no collision
• Subagents and branches just work — the proxy doesn't care about sessions, only content
• No state to corrupt — restart the proxy anytime, worst case is one extra compression cycle
• Claude Code sees nothing different — the proxy is invisible, JSONL transcripts are unmodified

Turning it off

Some sessions want the whole repo in view — deep planning, refactor design, an optimization pass where the model should see everything. Compression is not what you want there.

/rolling-context:off      # this conversation only, other sessions unaffected
/rolling-context:on       # back on

/rolling-context:off --global   # every session on this machine
/rolling-context:on --global

/rolling-context:status   # what is the machine-wide setting?

Both scopes apply from the next request — no Claude Code restart, no proxy restart. Machine-wide off wins over a session that opted in.

Turning it off never throws work away. Compressions already computed stay in the proxy's memory; they just stop being injected, and no new compression is triggered. Turn it back on and the session resumes with everything it had, so a conversation that ran hot before the toggle never pays to compress that history twice.

Changing the default

A fresh install compresses out of the box. If you would rather opt in per session, flip the default in ~/.claude-rolling-context/config.json:

{ "enabled": false }

Sessions that have not run /rolling-context:off or :on follow this; an explicit toggle in a session always wins over it. A missing or malformed config falls back to on, so a typo can never silently disable your install. /rolling-context:status prints the current default and the config path.

Subagents follow the session that spawned them

/rolling-context:off covers the subagents your session launches, so a planning session where agents explore the repo keeps full context everywhere — not just in the main thread.

The command prints a marker into the transcript, and the proxy latches it against the X-Claude-Code-Session-Id header Claude Code sends on every request. A subagent gets its own transcript, so it never carries the marker, but it inherits its parent's session id — which is what carries the setting in. Other sessions on the same machine are unaffected either way.

The latch also outlives the marker: if a conversation is /compacted and the marker is summarized away, the session keeps its setting.

One caveat: the latch lives in the proxy's memory, bounded to the 512 most recent sessions. Restart the proxy (or blow past 512 sessions) and it forgets, so a session falls back to the machine-wide setting until its next marker is seen. That costs savings for a request, never correctness — and re-running /rolling-context:off restores it.

Configuration

All settings via environment variables (all optional — defaults work great):

Is it actually firing?

Every 25 requests at or over the trigger, the proxy logs how often it injected:

[STATS] prefix injected on 61/72 requests at or over the trigger (85%)

A conversation's first turns are always misses — nothing has been compressed yet, and compression happens in the background after a response, so there is a short warmup before the first summary exists. After that, every turn should carry it. A rate that stays low once sessions are warm means something is wrong; before v1.11.4 it sat near 21% because one conversation compacting blocked all the others (issue #8).

Logs

Two files live in ~/.claude:

• rolling-context-debug.log — the proxy's own log, rotated and capped by the settings above (30 MB total by default).
• rolling-context-proxy.log — the proxy's stdout, captured by the start hook. It carries WARNING and above only, so it stays small; it exists to catch crashes that never reach the logger.

Before v1.11.3 the proxy logged at DEBUG unconditionally into an unrotated file and duplicated every line into the second one. A long-lived proxy could reach several gigabytes (issue #7). If you are upgrading, delete the old files once — nothing prunes them retroactively.

Stop the proxy first. A running proxy holds both files open, so deleting them frees no disk at all until the process exits — the space only comes back on restart. Reported by @drewdrewthis after clearing a 6.86 GB log and finding the disk unchanged:

# stop the proxy (it restarts automatically on the next Claude Code session)
kill "$(cat ~/.claude/rolling-context-proxy.pid)" 2>/dev/null
rm -f ~/.claude/rolling-context-proxy.log ~/.claude/rolling-context-debug.log

Proxy Chaining

Already using another proxy (model router, API gateway, etc.)? Rolling Context auto-detects this and chains through it:

Claude Code  ──HTTPS_PROXY──►  Rolling Context MITM (:5590)  ──►  core (:5588)  ──►  Your Proxy  ──►  Anthropic API

If ANTHROPIC_BASE_URL is already set to a custom endpoint when you install, the plugin keeps it as the host Claude Code talks to, saves it as ROLLING_CONTEXT_UPSTREAM, and has the MITM front-end intercept that host too — so your endpoint is still reached and bodies are still rewritten. No manual config needed.

You can also set it explicitly:

export ROLLING_CONTEXT_UPSTREAM=http://localhost:8080  # your existing proxy

Health Check

curl http://127.0.0.1:5588/health

Returns who is on the port (service, version, pid) and compression stats: how many compressions, tokens saved, etc.

This is also how the start hook decides whether to start anything. It does not trust the PID file: a crashed proxy leaves that file behind, and the OS is free to give its number to an unrelated process, so "the PID is alive" was answering yes while nothing was listening — the hook logged Proxy already running and every session failed with ConnectionRefused until someone deleted the file by hand (issue #9, reported by @drewdrewthis). Asking the port cannot go stale.

If sessions fail with API Error: Connection refused, read ~/.claude/rolling-context-hook.log — since v1.11.7 the hook waits for the proxy to answer before it exits, and says which of these happened:

Upgrades never cut a running stream

The proxy is shared by every Claude Code session on the machine. After a plugin auto-update, sessions started on the old version and sessions started on the new one each run the start hook from their own plugin cache directory. Up to v1.13.1 the hook restarted the proxy on any version difference, so the two cohorts took turns restarting it — and every restart dropped every session's in-flight /v1/messages stream as API error / Connection dropped (nestor-plugins issue #1, reported by @drewdrewthis: 188 such restarts in one log). Since v1.13.2:

• Never downgrade. A running proxy of the same or a newer version is left alone.
• Upgrade only when idle. /health reports active_requests; the hook waits for it to reach zero (up to ~6s), and otherwise defers the upgrade to a later session start rather than killing mid-stream. Proxies too old to report the field are judged by their established client sockets instead.
• Drain on SIGTERM. The proxy stops accepting, finishes what is in flight (up to ROLLING_CONTEXT_DRAIN_SECONDS, default 20), then exits.
• /rolling-context:restart forces a restart when you want the new version now; it still waits briefly for in-flight requests.

Sessions still running a pre-1.13.2 plugin keep the old "restart on any difference" hook, so the flapping stops fully once those sessions have ended.

A transport change never strands a running session

Claude Code live-reloads the env block of settings.json into already-running sessions. NODE_EXTRA_CA_CERTS, though, is read once at process launch: a session started before it was wired in can never be made to trust the MITM cert by a later rewrite. So flipping a running session from the old ANTHROPIC_BASE_URL transport onto HTTPS_PROXY + MITM used to break it — every request reset with Connection dropped (ECONNRESET) until the session was restarted, and the front-end logged nothing (issue #12, @drewdrewthis). Since v1.13.3:

• The flip waits for idle. If any session is established on the proxy ports when the start hook would flip the transport onto the MITM, the hook leaves settings.json unchanged and logs DEFERRED transport change. Running sessions keep working on the transport already in the file (the core port still listens); the flip happens at a later session start once they drain. ROLLING_CONTEXT_FORCE_WIRE=1 overrides it — then restart or --resume the old sessions.
• No connection is dropped silently. The MITM front-end now logs every connection it closes before a usable CONNECT — an empty connection, a client that spoke TLS straight at the proxy port, a non-CONNECT request — and a TLSV1_ALERT_UNKNOWN_CA handshake failure now says the client was started before NODE_EXTRA_CA_CERTS was wired in and should be restarted.

502 CERTIFICATE_VERIFY_FAILED is the proxy's own TLS, not an outage

The proxy verifies its outbound TLS to api.anthropic.com against the host Python's CA store. On some installs that store is empty — notably a python.org macOS build where Install Certificates.command was never run — so every forwarded request failed CERTIFICATE_VERIFY_FAILED, surfaced to Claude Code as an opaque 502 that reads exactly like an Anthropic outage (issue #13, @drewdrewthis). Since v1.13.4 the outbound context loads the bundled certifi roots in addition to the OS store (a union, so corporate roots stay trusted), so verification works regardless of whether the OS cert-setup step was ever run. If certifi is not importable and the OS store is also empty, the proxy logs an actionable warning and the 502 body now names the real cause (run pip install certifi or the Python installer's Install Certificates.command, then restart the proxy) instead of passing through a bare SSL error.

Debug

curl http://127.0.0.1:5588/debug/compressions

Returns the stored compression entries with their full summary content — useful for verifying what the rolling summary captured and whether user goals/instructions survived compression.

Uninstall

Run the uninstall script — it handles both manual and marketplace installs, stops the proxy, cleans env vars, and removes all plugin registrations.

Linux / macOS:

cd ~/claude-rolling-context && bash uninstall.sh

Windows (PowerShell):

cd $HOME\claude-rolling-context; powershell -ExecutionPolicy Bypass -File uninstall.ps1

If you installed via marketplace and already deleted the repo, you can run it from the cache:

cd $HOME\.claude\plugins\cache\rolling-context-marketplace\rolling-context\<version>
powershell -ExecutionPolicy Bypass -File uninstall.ps1

License

MIT