Providers cache the prompt prefix: cached tokens re-read at , rewritten
ones bill at . Most compressors optimise bytes removed and ignore that
multiplier. This one optimises the bill.
91.5% fewer tokens on tool output on Headroom's own benchmark, where Headroom removes 61.4%. Separately, the same SWE-bench tasks resolved at 0.71x the cost of plain Claude Code. Plus memory across sessions and subscription quota pacing. Claude Code, Codex, Copilot, Gemini and 12 more. Local, MIT.
TOKEN_OPTIMIZER_CACHE_DIR
Dernière vérification
2026-11-13
Étoiles
579
Outils
—
Disponibilité
—
0.1x
1.25x
Measured against their real implementation, on their own fixtures. Not a
reimplementation and not our fixtures: their harness compresses the payload,
dumps it, their resolver redeems their own markers, and ours is handed the
identical bytes. Their column is the best of the nine configurations that
capture holds, row by row.
Two of our arms are in the table, and they answer different questions. ours
is the block engine with nowhere to move a block, which is how the MCP tools and
the hooks run: everything it removes, it removes by compressing, and nothing
leaves the request. ours, proxy is the proxy a Claude Code session runs
through, which also moves a block it could not compress out to a file and
leaves a marker naming the path -- the same kind of trade their content-cache
reference makes. ours, preset is the block engine with that move switched on
at spillWholeBlockBelow: 0.9.
workload
payload
theirs
ours
ours, proxy
ours, preset
zero-turn ids
agentic-conversation
244,908
37.3% / 33.4%
42.9% / 30.8%
52.4% / 49.0%
100.0% / 100.0%
2099 / 2099 / 0
agent-loop
170,938
42.2% / 54.3%
69.3% / 59.1%
92.0% / 94.1%
93.6% / 94.3%
185 / 185 / 52
agent-loop-logs
327,500
51.5% / 53.0%
61.9% / 58.1%
88.7% / 91.5%
96.2% / 97.0%
2034 / 2034 / 81
api-responses
49,127
88.2% / 89.3%
66.6% / 59.1%
96.4% / 96.8%
96.6% / 96.8%
699 / 98 / 23
browser-session
781,774
21.7% / 13.1%
93.7% / 26.5%
55.5% / 90.2%
100.0% / 99.9%
334 / 334 / 0
code-search
131,264
47.4% / 59.7%
81.8% / 70.8%
97.7% / 97.9%
95.7% / 96.1%
201 / 201 / 43
database-rows
45,929
88.6% / 89.2%
65.6% / 56.5%
92.4% / 93.1%
92.6% / 93.4%
590 / 78 / 49
codebase-exploration
137,961
99.7% / 99.7%
16.0% / 12.7%
99.6% / 99.6%
100.0% / 99.9%
540 / 3 / 0
grep-output
75,219
99.6% / 99.6%
56.1% / 51.9%
99.5% / 99.6%
99.9% / 99.9%
1005 / 3 / 0
log-entries
91,385
80.8% / 83.2%
63.5% / 52.0%
95.4% / 95.7%
95.5% / 96.0%
1180 / 237 / 46
human-authored-json
37,425
30.6% / 30.0%
66.6% / 66.0%
81.6% / 83.7%
96.7% / 96.9%
154 / 154 / 11
issue-triage
109,354
53.1% / 55.2%
88.1% / 85.7%
97.2% / 97.6%
95.1% / 95.7%
59 / 59 / 21
rag-conversation
171,913
0.0% / 0.1%
84.6% / 83.2%
99.1% / 99.4%
100.0% / 100.0%
42 / 42 / 0
raw-build-log
158,057
99.8% / 99.8%
70.6% / 48.2%
99.7% / 99.8%
100.0% / 100.0%
430 / 3 / 0
relevance-probe
49,157
64.1% / 71.0%
82.4% / 77.2%
97.1% / 97.8%
95.8% / 96.6%
225 / 225 / 13
search-results
96,647
95.0% / 95.4%
74.3% / 59.4%
96.5% / 96.5%
96.6% / 96.6%
505 / 62 / 34
repeated-reads
158,285
37.4% / 30.4%
71.7% / 68.5%
99.5% / 99.6%
100.0% / 100.0%
394 / 388 / 0
sre-debugging
312,276
88.4% / 89.9%
67.4% / 62.4%
98.5% / 98.7%
97.1% / 97.5%
3116 / 506 / 82
Recorded 2026-10-17 at b1231a8c, by the command in the
record's regenerate field, over HeadRoom 0.37.0's own fixtures and harness. That arm runs
HeadRoom itself, so CI does not re-derive it -- it checks these figures and
this provenance against the record instead.
Over the corpus the proxy removes 82.5% of the characters and 91.5% of
the tokens; their best configuration removes 51.5% and 61.4%. It is cheaper than
their best arm on 18 of 18 rows when nothing is ever fetched back, which is
the end of the range their markers are built for, and on 17 of 18 when every
moved block is fetched back. The block engine alone removes
72.0% and 55.9%: more of the characters than theirs, fewer of the tokens.
Where they still win, and why. On codebase-exploration, grep-output,
raw-build-log, sre-debugging and the four raw-data rows their best arm takes
83.2-99.8% off the tokens and the block engine 12.7-62.4%. Their number there is
a content-cache reference: the block is not made smaller, it is taken out of
the request, put in a store, and replaced by a <<ccr:hash,kind,size>> marker.
An engine with nowhere to put a block cannot match that, so on those rows the
ours column loses on reduction, and on cost when nothing is fetched back. The
proxy makes the same move: it is ahead of them on five of those eight rows, level
on grep-output and raw-build-log, and a tenth of a point behind on
codebase-exploration. The one row where the proxy costs more than they do is
agentic-conversation, and only when every block it moved is fetched back:
they move none of that conversation out, so they cost the same at either end,
while the proxy moves part of it and pays a retrieval for each part it wants
back.
The other column is a turn. A marker is not the content: to read what it
stands for, the agent spends a request. Of the 13,792 retention units planted in
the corpus, the block engine keeps 5,223 in the text and 8,569 more rebuildable
from the text alone, so every one of them is there with no extra turn; theirs
keeps 6,711 in the text and 7,075 behind its store, one retrieval call away.
0 are unrecoverable on our side and 6 on theirs. Their best arms force 17
round trips across the corpus to get their content back; the block engine
forces 0. The proxy, like theirs, trades a turn for every block it moves, and
that turn is what the fetch-rate end of the cost model prices.
What their marker costs is priced, not just its size. Their marker is only
redeemable because their proxy also sends a headroom_retrieve tool with the
request -- on by default (ccr_inject_tool), and on every later request once a
session has used it -- so every row where their best arm leaves a marker is
charged for that tool's definition as well, recorded from their own package by
bench/compression/headroom/ccr-tool.py. The turn that first adds the tool also
rewrites the whole cached prefix behind it; that is not charged, which errs in
their favour. Ours needs no tool: our marker names a path the agent opens with
the Read it already has, and the path is priced at the length the proxy really
writes (<home>/.tok/<pid>/<15 digits>, src/compress/spill-path.ts) rather than
a shorter invented one.
browser-session is mostly the fixture. This payload holds four images, two
of them distinct, and the base64 in them is generated rather than photographic:
one distinct image alone folds from 160,032 characters to 5,572, which no real
PNG would do. What carries over to a real session is the duplication -- half the
image bytes are a second copy of an image already in the request, which is
44.3% of the whole payload. The run the marker names is still in the output
above it, so the reader rebuilds it without asking for anything, which is why
the characters fall much further than the tokens: the image tokens are counted
from pixels on both arms either way.
A second arm reimplements their published design from their own benchmark
generator's definitions — opaque hash markers, history compressed, a retrieval
tool and system message injected — so the four workloads with a published
comparator can be checked without their clone. node bench/compression/proof.mjs:
workload
payload
theirs
ours
verdict
code-search
17765
92.1%
97.8%
ours
sre-debugging
65694
92.2%
98.6%
ours
issue-triage
54174
72.8%
97.5%
ours
codebase-exploration
78502
47.4%
54.7%
ours
Four workloads, because those are the four this arm has a published comparator
for. The head-to-head harness scores eighteen; the other fourteen are not in
this arm, and are scored against their real implementation in the table above.
codebase-exploration was theirs and is now ours, by 7.3 points. An earlier
version of this table claimed 61.3% for us on that row, then parity, then a
published loss at 46.0% against their 47.4%. The first two were figures from a
branch rather than from here; the third was true on this tree when it was
written; four log engine commits then moved it to 54.2%, masking an interior
hashbang so a concatenated block still parses moved it to 64.9%, and
authenticating every marker with a six-character stamp took 0.2 of a point
back, and naming a spilled file by a digest of its contents, the way the proxy
sink really names it, instead of by its position in a counter, took another
half point, to reach 64.2%. Then the currency stopped being an estimate: every
figure in the table above is now the count Anthropic's own tokenizer returns
for the text, recorded into a committed fixture so this stays offline, and
under it the row reads 53.7%, and pricing the path the proxy's sink really
writes -- content-addressed and decimal, where this arm had invented a 32-digit
hex name -- moved it to 54.7%. That is the largest move in this history and
the only one that is not about the engine at all. Dividing characters by four
had been charging the control's hex markers too much and our own stamp far too
little -- the stamp was six characters of a vowelless base-31 alphabet then,
and a random string over a large alphabet is close to the worst case a BPE
tokenizer has, so it cost a measured 5.52 tokens where chars/4 billed 2. Those
three are the only figures here that moved the wrong way, and none of them is a
regression: each is the harness charging us for something it had been leaving
out. Being able to see that cost is what paid for the next move: the stamp is
now nine decimal digits, which carry slightly more entropy than those six
characters did and cost 4 tokens on every draw rather than a mean of 5.52 that
ranged from four to seven. Measured against the old stamp on this row -- the
same tree, the same fixture, only the stamp pinned back to its base-31 form --
that bought a tenth of a point, and a tenth is the honest size of it: a
marker got better than a quarter cheaper, but markers are a small part of what
the engine emits, so the headline moves far less than the per-marker saving
suggests. A
number in prose is a fact about the tree it was measured on, which is why
bench/compression/readme-table.check.mjs re-derives every figure in the block
above from the harness rather than trusting it.
Against the four comparators in this reimplemented arm: ours on all four.
Against their real implementation the answer is the table at the top of this
section, and it is a split rather than a sweep: the proxy is cheaper than their
best arm on every row when nothing is fetched back and takes more of the tokens
overall, while the block engine alone keeps every identifier without a turn and
loses on reduction wherever their best arm moves a block out of the request.
The two columns come from different arms of the same engine, and that is the
point. v3-history compresses history too and matches them byte for byte;
v1-frontier leaves the cached prefix alone and wins the invoice by 11–42%.
Maximising bytes removed is available and is not the default, because on a
cached prefix it costs money.
The competitor arm is this repository's reimplementation of their published
design — opaque hash markers, history compressed, a retrieval tool and system
message injected — not their binary, and it is held to the same signed-content
guard we hold ourselves to.
On whole sessions
node bench/compression/session-replay.mjs replays real recorded conversations
turn by turn and prices each the way a provider does. Four independent sessions,
whole request including system prompt and tool schema, against no proxy:
conversation length
ours
cheaper on
10 turns
0.953x
4 of 4
20 turns
0.915x
4 of 4
This is the region nothing else touches: conversation history is ~65% of a live
request and the only part that grows every turn. The saving grows with session
length, because model reasoning accumulates.
On end-to-end agent tasks
THOL, 16 real tasks against a no-proxy control:
cheaper on 11 of 16 tasks
median cost 0.926x, median turns 0.671x
zero quality cost: mean score 0.994 against control's 0.994,
lower on 0 of 16 tasks
One run per task, so the aggregate interval still spans 1.0; the per-task tally
and the score parity are the solid parts. Every figure here is regenerated from
the committed benchmarks, which ship in this repository.
The 30-second version
Your agent burns most of its context on work it already did: re-reading files
that have not changed, dumping a whole file to see three lines, running
unbounded searches, and re-deriving conclusions it reached last session and then
forgot.
Token Optimizer attacks that on four fronts.
1. It makes the expensive call impossible. Install the plugin and a built-in
Read of a 200 KB file is denied, with the refusal naming the cached,
diffed replacement. Same for Grep, Glob, Edit, Write, and cat /
head / grep -r through the shell. Re-reading a file you already read this
session returns only a diff — usually the single biggest win, and one that
size-based rules structurally cannot catch. There is no setting to turn on.
2. It remembers what your agent worked out. A per-project knowledge graph
accumulates findings, decisions and dead ends as a side effect of working, then
feeds them back the moment the agent touches the relevant file. A finding costs
~150 tokens to carry. Re-deriving it costs 5k–50k.
3. It measures itself, in public, and tells you when it is losing. A
materialized before/actual-return measurement for MCP progressive disclosure,
with later expansions debited from the same net. Modeled graph substitutions
and the randomized control arm for downstream graph effects remain separate.
Every number is measured, visibly collecting, excluded, or absent.
4. It attributes the traffic. Returned context, optional cost equivalents,
and net transport avoided are grouped by operation and MCP handshake identity. Codex, Claude Code,
Gemini, and any other connected client get separate rows. Old records without
identity remain explicitly unattributed instead of being assigned to whichever
agent happens to be open now.
No account and no hosted service. MIT, so it is usable at work.
Nothing is measured about you unless you ask for it. Anonymous usage data is
off by default and takes two separate opt-ins: TOKEN_OPTIMIZER_TELEMETRY=1
aggregates counts into a file on your own disk, and TOKEN_OPTIMIZER_BEACON=1
is what uploads any of it. Setting the first opens no socket. With both set,
the events already on disk are sent once per session at start-up -- in a batch,
never on the hot path -- and the file is cleared only after the receiver has
accepted them. install_doctor prints which of the two is on, how many events
are waiting and where they would go; it reports without sending. A package
built without the upload key cannot transmit at all, and the doctor says so
rather than letting an opt-in look like it is working.
DO_NOT_TRACK=1 overrides both, whatever else is set. The version check
npm run doctor and install_doctor make is a plain GET to the registry npm
installs from, carries nothing about you, and runs only when you ask for a
diagnosis -- never during a session. DO_NOT_TRACK=1 or
TOKEN_OPTIMIZER_UPDATE_CHECK=0 suppresses it before a socket is opened. An event may only
contain numbers and flags -- a string is dropped before it is written, so a
path, a prompt or an error message cannot travel even by mistake -- and the
machine identifier is a salted hash, never your hostname. What the counts are
about is the hook ledger your own project already keeps: how often the graph
answered instead of a tool, what each arm of the holdout was delivered and
withheld, and whether the saving exceeded the cost -- reduced to integers once
every six hours, never the ledger itself, which names your files. Usage is
counted as a rollup rather than a row per request: one event per 200 requests,
holding the request count, bytes in and out, how often compression paid, how
often knowledge was injected, and how many blocks were elided, plus one final
event when the proxy stops. That keeps the totals exact -- a row per request
would overflow the batch on a busy session and leave you reporting its first
five hundred -- at the cost of the per-request distribution, which is not
something anyone downstream needs. npm run doctor
prints exactly which of these is on and what has been recorded.
What the dashboard proves on a real machine
The screenshots in this README come from the shipped server reading persisted
local data, not a design mockup. In the capture above it reports:
43,491 net verified MCP transport tokens avoided in the current live
proof: 54,037 gross reduction minus a deliberate 10,546-token expansion;
486,074,740 historical/tool-reported tokens quarantined, dominated by
repository scan volume that never entered model context;
a live Codex / Claude Code / Gemini stdio smoke against AiDotNet, with each
client attributed independently;
2,648 graph nodes, 6,527 edges, and 58 findings across 11 local projects;
more than 1,000 hook runs with zero failures and zero timeouts across six
active CLI clients in the selected rolling 24-hour window;
6,332 tokens of modeled graph-substitution potential, excluded from the
verified headline while the causal graph-reuse study remains Collecting.
The dashboard now reads native CLI usage receipts and prices uncached input,
cache reads, cache writes, and output with the exact captured provider, model,
route, request-time tier, and versioned official source. Ambiguous model ids stay
Not priced instead of receiving a blended guess. API/list-price equivalents
are kept separate from provider-reported charges and are never labeled as a
subscription invoice. See the token accounting contract.
Quick start
Claude Code — install the plugin, not the bare MCP server. The plugin is
what enforces; adding the server alone just gives the model tools it can ignore.
One ranked queue: what is costing the most per session, with an optional monthly
cost equivalent only after you configure your own effective rate. Each line
names how to fix it. Not a dashboard, not six reports — a queue.
Subagents and workflows
Agents are where token spend concentrates. A long-running subagent re-sends its
whole context on every turn, so every byte a tool adds is paid for again on each
later turn. In one measured session, 42 workflow subagents made 4,483 model
calls. The six longest made 218–348 calls each and peaked at 336K–609K tokens of
context. Together they spent 4.0M output tokens and 1.0B cache-read tokens. The bill is the re-reading, not the writing.
What works with no setup
The plugin's hooks run for subagents exactly as they do for the main session.
Nothing needs to be configured per agent.
Whole-file reads are outlined, through Read and through the shell.
Agents mostly read files with cat, type or Get-Content, rarely with
Read. In the measured session that was 2,071 shell reads against 9 Read
calls. On the first whole read of a large source file, either kind returns a
line-numbered outline of its symbols instead of the file. A second whole read
of the same file returns the file itself. A line range is never outlined:
sed -n 'A,Bp', head, Get-Content -TotalCount and a paged Read all pass
through untouched.
Each subagent is briefed once. A subagent never sees the SessionStart
guidance, so its first tool call carries a four-line briefing instead: load
the optimizer tools if they are listed, read ranges rather than whole files,
and keep the final report short. The tools are named only conditionally, so a
hooks-only install is never told to call something it does not have.
Advisories are said once. A routing hint such as "use smart_grep" appears
at most once per agent for each kind of call, not on every matching command.
Past findings arrive only when they name what you are running. A finding
with no explicit trigger is delivered with a command only when its claim names
that command's program and subcommand, such as npx jest or dotnet build,
not when the two merely share a word.
Repeated build and test output is bounded, as it is in the main session.
Through the proxy, a range you asked for comes back whole. A paged Read
or a shell slice is never folded, and a fold inside a whole-file read points at
the file's own line numbers.
What to add when you brief an agent
Claude Code gives subagents MCP tools as deferred tools: the schemas are not
loaded until the agent asks for them. An agent that never loads them falls back
to shell commands. Put this in the agent's prompt:
Before your first file read or search, load the token-optimizer tools:
ToolSearch("smart_read smart_grep smart_glob").
Use smart_read for files (it returns a diff on re-reads), smart_grep for
searching, and smart_glob for finding files. Do not print whole files or search
the tree through the shell. When you need part of a file, read a line range.
The server's own instructions now say the same thing to every client, deferred
or not.
Keep agents short
The plugin can only shrink what enters an agent's context. The number of turns
and the size of the context it carries are set by how the work is split up:
One bounded task per agent. Several short agents cost less than one agent
that runs for hundreds of turns, because each starts from a small context.
Give the agent the facts it needs (paths, line numbers, the exact
question) instead of asking it to rediscover them.
Ask for a structured result so the agent stops when it has the answer.
Measuring an agent
Each subagent writes its own transcript next to the session's, under
~/.claude/projects/<project>/<session>/subagents/. The usage block on every
assistant row records input_tokens, output_tokens,
cache_read_input_tokens and cache_creation_input_tokens.
Count each response once. One model response is written as several
assistant rows that share a message.id, and each row repeats the usage
recorded so far. Keep the LAST row per message.id, which holds the final
usage, and sum those. Summing every row overstates the cost: on the session
above it reported 1.8B cache-read tokens instead of 1.0B. Keeping the first row
instead understates output, at 0.5M instead of 4.0M.
The number of distinct message.ids is the number of model calls the agent
made. Read against its peak context, it shows whether the agent should have
been split.
The knowledge graph — the part nothing else has
Every agent session ends the same way: the reasoning evaporates. The next
session re-derives it, at full price, forever.
This builds a living per-project graph — nodes for files, symbols, tasks and
findings; edges for derived_from, contains, supersedes, contradicts,
related — and it fills itself in from real work. No ingestion job, no
embedding model, no rebuild step, no query to formulate.
you touch src/auth.ts
│
├─ verify() compares exp against the LOCAL clock (finding, 0.9)
├─ per-host retry budgets; global was rejected — deadlock (decision)
├─ ! the skew fix was reverted once already (dead end)
└─ [git] 47 changes in 90d, last three: "fix token expiry",
"revert skew fix", "fix token expiry again"
None of that is in your repository. It exists only because an agent once burned
tokens finding it out — and every other tool throws it away at the end of the
session.
What a default install actually produces. The structural graph — files,
symbols, tasks, and the edges between them — is captured from ordinary tool
traffic with no configuration at all. Findings are produced two ways. At session
end, derive reads evidence already on disk (command outcomes and exit codes,
red-to-green transitions, corrections, re-read churn) and writes findings from
it: no model call, no credential, nothing sent anywhere. And the active model
records durable conclusions itself through wiki_write.
The model-based semantic harvest is the third path, and the only one that
needs something you do not already have. It is not opt-in —
TOKEN_OPTIMIZER_HARVEST=0 turns it off — but its real gate is a credential:
with none it reports off:no-key, which is the state on CI, corporate machines,
and subscription-only logins. Point TOKEN_OPTIMIZER_HARVEST_ENDPOINT at a
local model and it runs free and private, with nothing leaving the machine.
npx token-optimizer-doctor states which of these is live.
Why this is not RAG
Classic RAG
This
Retrieves evidence; the model re-derives meaning each time
Retrieves verdicts — the reasoning already happened
Index built by a batch ingestion job
Accretes from real agent traffic — coverage follows attention
Staleness computed from content hashes, served with the invalidating diff
Returns only what is in the documents
Returns dead ends, which exist nowhere in your source tree
Traversal plus lexical search: deterministic, instant, explainable, and it works
offline.
The zero-turn refusal
A plain deny costs a full turn: the model calls Read, is refused, re-plans,
calls smart_read. But at refusal time we already hold the file and the
snapshot the graph stored — so the refusal carries the answer inside it.
Nothing to re-plan, no second call. Turn cost drops from one to zero.
And when the graph already holds the verdict a tool output would support, the
output never enters context at all. Not compressed. Absent.
The dashboard
npm install
npm run build
npm run dashboard # http://localhost:3100
The overview answers the questions a token optimizer should answer first:
How much MCP context did it avoid? The headline is gross materialized
payload reduction minus every later linked expansion. Graph estimates are
intentionally separate.
How much context still reached the agents? Every successful current MCP
result records its actual returned text, even when no valid before-state
exists. That row is context-accounted but savings-unmeasured.
Which agent and action spent it? The client ledger and action table show
operations, returned context, optional cost equivalent, and net tokens
avoided. Lifecycle-only clients say Not measured; no zero is invented.
Did remembering cost more than it saved? Delivery and semantic-harvest
tokens are charged to the graph. A causal benefit is added only after the
treated/holdout evidence gate passes.
Walkthrough: get useful data, not an empty dashboard
Install the MCP server and the native adapter for your CLI. The MCP handshake
provides per-client accounting; native lifecycle hooks provide automatic
routing, capture, health, and delivery where the client protocol permits it.
Use smart_read, smart_grep, smart_glob, smart_edit, or any other MCP
operation normally. Every successful result records returned context; tools
with a comparable materialized before-state also record a gross reduction;
later expand calls debit that reduction.
Let the active model record durable conclusions with wiki_write. Before a
new agent re-derives work, call wiki_read for the project or the files it is
about to touch. Native clients can also deliver matching knowledge
automatically.
Use wiki_query to read the graph directly — one finding by key, a ranked
BM25 search over claims, a node with its neighbours, or the graph's own audit
— which is how a subagent that never sees the SessionStart briefing reaches
what previous sessions established.
Open http://localhost:3100. Use Overview for combined accounting and
What it knows for capture health, graph exploration, audits, and causal
evidence.
To register existing local repositories without reading their source, run
npm run projects:discover -- /absolute/path/to/repos. This makes coverage
gaps explicit; it does not fabricate findings.
For maintainers, this live smoke exercises the shipped stdio transport and
creates separately attributed rows without seeding the analytics database:
npm run dashboard:attribution-smoke -- /absolute/path/to/project /absolute/path/to/large-file
npm run dashboard:verify-live -- http://localhost:3100
The health panel is deliberately operational rather than a raw text dump. Each
client has activity, runtime failures/timeouts, policy outcomes, and observed
surface coverage. Diagnostics keep no prompts, commands, paths, or tool output.
Modeled substitution potential and causal graph effects are different claims.
The first is a full-file counterfactual and is never promoted to the verified
MCP headline. The second asks whether delivered knowledge prevented later reads;
it uses a control arm and remains Collecting until there are at least 20
treated file touches and 5 holdouts with valid downstream joins.
Drag to orbit, scroll to zoom, click a node for provenance, or switch to the
bounded one-hop focus view. The default All known projects scope pools local
graphs through opaque project IDs; filesystem paths never reach the browser.
Audit tab. Contradictions, stale findings, and low-confidence claims remain
reviewable instead of silently becoming model truth.
Evidence console. Client/model/task cohorts, matched effects with 95%
intervals, live outcome joins, harm feedback, and capability tiers for all 16
clients. Release and superiority claims fail closed while evidence is missing.
One-click Markdown export. The accumulated graph becomes documentation you
can inspect, edit, and commit.
Server-side by design: the browser asks for a neighbourhood, a search result or
a page. A mature graph holds thousands of nodes, and shipping it wholesale would
make every page load a multi-megabyte download for a view that shows twenty
things.
The default All known projects scope combines captured graphs through an
opaque machine-local project registry; filesystem paths never reach the
browser. Lifecycle hooks register repositories as they are used. To backfill
existing local checkouts without reading their source files, run the bounded
discovery command against one or more explicit roots:
npm run projects:discover -- /absolute/path/to/repos /absolute/path/to/worktrees
Coverage distinguishes repositories with graph data from known repositories
whose capture has not started. The balance cards are backed by persisted events:
Memory deliveries counts graph context actually supplied to an agent,
Kept back for comparison counts randomized control touches, and Cost of
remembering combines delivered-context tokens with measured semantic-write
payload cost. Reading avoided stays Collecting or Not measured until at
least 20 treated file touches and 5 holdouts exist with a downstream join; the
dashboard does not manufacture a savings estimate from missing data.
Every native hook writes the same bounded JSONL lifecycle record, including
Claude Code's custom router and compaction paths. Records carry the client and
plugin versions, event, hashed session/turn correlation, latency, outcome,
input/output byte counts, and response key shape. They deliberately retain no
prompt, command, tool output, file content, or raw working-directory path. The
fields include OpenTelemetry log severity and resource semantics so the local
files can be collected without inventing a second schema.
npm run diagnostics # last 24 hours, summary-first JSON
npm run diagnostics -- --hours 72 --output hook-summary.json
npm run diagnostics -- --include-events --limit 100 --output hook-report.json
Raw event rows are opt-in and capped at 1,000. The default report contains aggregate health and
at most twenty recent failures/timeouts, keeping routine troubleshooting output small enough for
CLI and model context windows.
The dashboard's Capture health panel shows runs, failures, timeouts and
p50/p95 latency by client. Logs rotate at 5 MiB, retain at most 40 files for 14
days, and live under .token-optimizer/logs when a state directory is set (or
~/.token-optimizer/logs otherwise). TOKEN_OPTIMIZER_LOG_DIR,
TOKEN_OPTIMIZER_LOG_MAX_BYTES, TOKEN_OPTIMIZER_LOG_MAX_FILES, and
TOKEN_OPTIMIZER_LOG_RETENTION_DAYS override those operational defaults.
What it does that other optimizers do not
Compression that does not break the cache, or lose the needle
An optional local proxy compresses tool results, search output, logs and
conversation history on the way to the model. A hook cannot do this: the
PostToolUse output schema is {hookEventName, additionalContext?, classifierContext?} and updatedOutput occurs nowhere in the client, so a hook
can add context but never replace a result. The proxy never tries to -- it
rewrites the outbound request, where those results already sit as history.
Two things make it different from simply compressing harder.
It never rewrites cached content. A cached prefix bills at 0.1x and a cache
write at 1.25x, so compressing history can cut tokens while multiplying the
bill. Measured on our own benchmark: compressing behind the cache breakpoint
removes more raw tokens on every workload and costs more money on every
workload. Compression stops at the frontier, and cache-weighted tokens are
reported beside raw ones so the trap is visible rather than inferred.
It keeps the rows that matter. Eliding a long array after the first few rows
scored 95.7% on a search payload here and destroyed both the UUID record and the
error record planted in it -- the only two rows anyone would have searched for.
The engine now keeps every row that departs from the shape, wherever it sits:
92.4% with the needles intact. A gate in the benchmark fails the build if a
planted needle disappears, because a size metric alone cannot tell compression
from truncation.
Reduction over the content each strategy is permitted to rewrite, on fixtures
matching the four workloads HeadRoom publishes, is the guarded table earlier in
this section -- 97.5%, 97.8%, 98.6% and 54.7% against their 72.8%, 92.1%, 92.2%
and 47.4%. It is stated once and checked there rather than restated here, which
is how this copy came to claim 98.9%, 98.2%, 92.8% and 61.3% long after the
tree had moved.
These are not their corpora, which are unpublished; the code workloads read real
files out of this repository and the rest are generated to the shape and scale
of their published ones, from their own benchmark generator's definition.
Reduction is not the bill, and the two disagree. Cache changes the
arithmetic: a cached prefix is re-read at 0.1x and a rewritten one is charged at
1.25x, so an arm can remove more bytes and still cost more. Run
node bench/compression/proof.mjs and that is exactly what happens -- the
HeadRoom-style arm removes a larger share of every workload and loses on
cache-weighted effective tokens on all of them:
workload
raw reduction, ours / theirs
effective tokens, ours / theirs
code search
84.6% / 95.8%
999 / 1,383
SRE debugging
86.2% / 97.1%
1,833 / 2,261
issue triage
89.2% / 95.3%
812 / 1,277
grep output
45.9% / 51.9%
7,146 / 9,514
raw build log
54.7% / 59.4%
14,448 / 16,769
browser session
45.7% / 49.4%
11,398 / 14,685
The competitor arm is this repository's own reimplementation of their published
design -- opaque hash markers, history compressed, a retrieval tool and system
message injected -- not their binary. It is implemented to win where it can: the
raw column is theirs on six of six.
On whole sessions.node bench/compression/session-replay.mjs replays real
recorded conversations turn by turn and prices each one the way a provider does,
charging the longest byte-identical leading run at 0.1x and everything after at
1.25x. Four independent sessions, whole request including system prompt and tool
schema, against no proxy at all:
turns
HeadRoom's design
ours (v4-substitute)
cheaper on
10
1.000x
0.953x
4 of 4
20
0.999x
0.915x
4 of 4
The replay uses stored transcripts, which discard reasoning text while retaining
signatures. Its savings therefore do not establish that useful reasoning survives
compression. It also uses our reimplementation of the competitor's design;
these are offline estimates, not results from their shipped proxy.
Live comparison against HeadRoom's shipped proxy. The balanced four-arm run
recorded in commit f712b4ec used bench/live/ab.sh, rotated each arm through
every position, and re-ran pytest independently after each task:
Weighted input is input + 1.25 * cache_creation + 0.1 * cache_read; it excludes
output cost. Both zero-floor Token Optimizer and HeadRoom reduced that measure
by about 64% versus control. Token Optimizer used about 0.9% more than HeadRoom.
This small arithmetic repair task does not establish superiority across tasks,
compression quality, or total cost. Rotation balances run position, but does
not guarantee that shared provider-cache effects disappear.
Codex three-task comparison against HeadRoom 0.37.0 (2026-09-18). Three
synthetic controlled-read tasks, three repetitions, three arms, every arm in
every position. All 27 runs returned the correct answer, so this compares cost
at equal correctness rather than accuracy. HeadRoom ran in --mode token, its
compression-first setting, not the cache default that exists to freeze the
cache hit rate.
arm
mean input tokens
of which cached
output
requests
agent seconds
uncompressed control
55,632
43,079
200
3.0
19.1
Token Optimizer
42,468
33,052
189
3.0
17.3
HeadRoom 0.37.0, --mode token
71,626
51,897
285
4.3
27.3
Input reduction against the control was 22.7% on the log task, 34.3% on the JSON
task and 11.5% on the code search; against HeadRoom, 54.9%, 10.0% and 42.2%.
Every task favoured this proxy on both comparisons, and the JSON column is where
HeadRoom is strongest and the margin thinnest.
The mechanism is visible in two numbers that move independently. Mean input per
REQUEST was 14,156 here against HeadRoom's 16,529 and the control's 18,544 --
both proxies compress, and this one compresses harder. Mean requests per RUN was
3.0, the same as the control, against HeadRoom's 4.3. A compressor that elides
content the agent then has to read back spends a turn recovering it, and on
these workloads a turn costs more than the elision saved; that is the failure
this project's own deferral floor exists to avoid, and it is what the extra
1.3 requests are.
WHAT THIS DOES NOT SHOW. These are token counts, not dollar costs, and
bench/live/report-codex.mjs deliberately declines to price them. Three short
synthetic workloads do not establish a universal win. Each task ran in a fresh
temporary workspace with no knowledge graph, so graph injection contributed
nothing and this measures compression alone. Cached tokens are a subset of the
input column, not an additional charge. HeadRoom's compression is scored from
provider usage rather than byte counts recorded at our own listener.
Reproduce with HEADROOM_MODE=token node bench/live/codex.mjs, then
node bench/live/report-codex.mjs <evidence-dir>; the figures above were read
only after that reporter returned valid: true with balanced: true and no
failures. Four earlier campaigns were discarded rather than published: one ran
without HeadRoom's [proxy] extra so every HeadRoom run errored before
readiness, one ran HeadRoom in its cache default, one was voided by the
reporter for an incomplete request capture, and one died on ENOSPC with the
host disk full.
The subsequent Claude Code default-versus-aggressive confirmation hit Claude's
weekly usage limit and is incomplete. Its missing usage cannot count as a win.
Codex validation runs independently through the Responses API.
Latest Codex joint confirmation (2026-09-16). The frozen twelve-family study
completed 120 pairs against installed HeadRoom 0.37.0. Our proxy passed 119/120
attempts; HeadRoom passed 120/120. One local allocation crash left unknown usage.
There were 77 strict joint wins (passing, cheaper, and faster), 20 cost-only
losses, 12 speed-only losses, nine losses on both measures, one equal-time pair
that cost more, and the failed attempt. Across the 119 fully measured pairs
only, estimated token cost was 26.9% lower and agent time 30.1% lower. JSON
agent time was 5.6% higher. Only logs and refactoring met the simultaneous
per-family criteria; the complete-study superiority gate failed. These results
do not establish that every task is cheaper or faster. The measured build
preserves explicit nulls, retains complete numeric tables, and avoids copying
unchanged payloads. See the full results, every loss, and retained crash.
Earlier expanded local Codex confirmation. A preregistered seven-family study ran
70 pairs / 140 attempts against installed HeadRoom 0.37.0. Our frozen proxy
passed 70/70; HeadRoom passed 69/70 with one upstream HTTP 503 and missing usage.
The complete-ledger gate therefore failed: superiority was not established.
Across the 69 fully measured pairs only, descriptive totals were 25.7% lower
estimated cost, 38.9% lower input, and 30.0% lower agent time for our proxy.
JSON remained a loss (35.1% higher estimated cost). See the
protocol, all attempts, and limitations.
A subsequent rare-string-group fix targets that JSON loss. In three fully measured
fresh pairs, input fell 17.7%, estimated cost 13.4%, and agent time 13.6% versus
HeadRoom, with three requests each. A fourth pair had a retained upstream 503 on
our arm; one measured pair still cost more. These are development results, not
confirmation of the new build or an all-workload win. See the
complete screen and exclusions.
Current amended Codex study. The next fixed 70-pair schedule completed:
our proxy passed 69/70 and HeadRoom 70/70. Our remaining attempt hit an upstream
503 with unknown usage, so strict superiority was not established. Across the
69 fully measured pairs only, estimated cost was 27.9% lower, input 44.2% lower,
and agent time 31.7% lower. Every family's mean cost favored our proxy, including
JSON at 27.5% lower, but 13 individual pairs still cost more. The run includes a
disclosed repair to a between-case file-hashing failure; the original product and
schedule stayed fixed. See the full results, amendment, and retained failure.
Exact search rows, subsequent development. Factoring shared declaration syntax
while preserving every identifier and value closed the code-size gap: 37.5% fewer
bytes than archived HeadRoom on the original 40 losing inputs. Eight fresh local
cases also sent fewer bytes. Four balanced live pairs all passed their audits;
our estimated token cost was 50.1% lower, input 44.8% lower, and agent time 51.1%
lower. This is a small development follow-up, separate from the broader study.
See the captured bodies, reconstruction checks, and live evidence.
Cost-loss follow-up (2026-09-16). We investigated all 13 losing pairs and
implemented exact ID-prefix factoring plus support for records inside truncated
shell envelopes. A fixed follow-up of those 13 cases passed all 26 attempts;
12 pairs favored our proxy, with 33.4% lower total estimated cost. The remaining
loss exposed the outer-envelope bug. After fixing it, two balanced controlled
pairs both favored our proxy: all four attempts passed, estimated cost was 34.3%
lower, and input was 15.9% lower. These development runs retain every original
loss; cache and model variation also affect results. See the case-by-case audit,
fixes, exact replay, and live evidence.
Local proxy performance (2026-09-16). Against installed HeadRoom 0.37.0,
2,160 completed local-upstream requests covered repeated and unique logs, JSON,
and code-search output in three rotated arm orders. Our mean latency and process
CPU were lower in all six groups. Sampled peak process-tree private memory was
80.0 MiB versus 1,909.9 MiB; this is footprint, not allocation volume. Unique
code still forwarded 5.7% more bytes. HeadRoom's rate limit was disabled for this
throughput measurement; compression settings stayed default. These measurements
do not measure model quality or billed cost. See the
raw samples, reproduction commands, and limitations.
Live Codex comparison (2026-09-15). Local Codex CLI 0.154.0, gpt-6-astra,
and installed HeadRoom 0.37.0 completed a balanced 27-run campaign. All three
arms passed all nine tasks; every initial read was complete, and all provider
ledger totals matched Codex's reported usage.
workload
mean input, ours
mean input, HeadRoom
reduction vs HeadRoom
JSON outlier lookup
39,318
44,043
10.7%
code search
44,427
82,697
46.3%
log diagnosis
39,203
84,308
53.5%
These counts include cached input and extra retrieval turns; they are not raw
compression percentages or dollar costs. Output tokens did not improve on every
workload. See the Codex evidence and limitations
and reproduction instructions. These three synthetic tasks
do not establish superiority across all workloads.
A separate 27-run natural-workflow development screen passed all runs and
measured 35.5%, 43.5%, and 16.6% less total input than installed HeadRoom for retry
bug fixes, multi-file refactors, and configuration refreshes. These are small
synthetic repositories with natural tool selection. Uncached input and latency
have separate results; see the workflow evidence and limitations.
The proxy is enabled by default and binds loopback only.
It is now on without launching anything through us. Until 7.1.0 the only way
to be routed was to start the client through our wrapper, so anyone who installed
with /plugin and then opened Claude Code from a shortcut, an IDE or the desktop
app saved nothing -- and the doctor reported that as a broken installation rather
than as a feature that never applied. Installation now writes the endpoint into
Claude Code's own settings.json, which is the one thing a session we did not
start will read, and a small background proxy serves it. Four rules make that
safe to switch on:
Recorded. Every value written is stored with whatever was there before.
token-optimizer-uninstall --apply restores it exactly, and refuses when the
value is no longer the one we wrote -- if you change it, it is yours again.
Never written on hope. The entry appears only after the proxy has actually
served the route. A settings file naming a dead port does not degrade politely:
the client cannot reach its provider at all.
Self-healing. The check runs again at every session start, before the first
model request, and removes the entry the moment the route cannot be served.
Hands off what is not ours. A client already pointed at another local proxy
is left alone, and a settings file we cannot parse is never rewritten.
Set TOKEN_OPTIMIZER_DEFAULT_ROUTING=0 to keep the wrapper-only behaviour, or
TOKEN_OPTIMIZER_PROXY_AUTOSTART=0 if you do not want a local background service
at all. Either one also undoes an entry already written.
Managed commands. Global npm installs activate Claude hooks and wrap the
client commands you actually have -- claude, codex, gemini, opencode,
qwen, crush, droid, cn (Continue), copilot and amp -- in Command
Prompt, PowerShell, Bash, and Zsh when lifecycle scripts are enabled. Only
commands already on your PATH are wrapped, so installing this does not claim a
name for a client you have not installed; TOKEN_OPTIMIZER_MANAGED_CLIENTS takes
a comma-separated list (or all) when a client lives somewhere PATH cannot see,
and 0 skips shell activation entirely. Windows installs add owned .cmd
launchers to the User PATH, so profiles and PowerShell execution-policy changes
are not required. For local installs or disabled lifecycle scripts, run
token-optimizer-install. Reopen the terminal to load the updated PATH and shell
activation. Each managed session starts its own proxy, registers the packaged
core MCP tools (including wiki), and shuts the proxy down on exit.
token-optimizer-run <client> ... works without shell activation.
Which clients can be routed, and which cannot. Ten have a supported way to
redirect model traffic and are listed above. Claude Code and Gemini have exactly
one provider each, so they are routed on installation. The rest choose a provider
in their own configuration, and are routed once that configuration names an
endpoint -- because forwarding on a guess would deliver one provider's
credentials to another company. Six clients cannot be routed at all: Cursor,
Cline, Windsurf, Kilo, Roo and Zed run the assistant inside the editor process
and expose no documented redirect. Everything else -- the optimizer tools, the
graph, the hooks -- still applies to them, and the doctor states the limitation
instead of counting it as a failure.
token-optimizer-route shows what is running and what configuration has been
written on your behalf. Zed is the one client that can be routed and cannot be
routed for you -- its provider is a named entry you pick inside the editor, and
only you know which endpoint and model it should use:
token-optimizer-route # what is routed right now
token-optimizer-route zed --upstream https://api.openai.com/v1 --model gpt-4o
token-optimizer-route zed --remove
Zed's provider schema here comes from its published settings documentation and
has not been exercised against an installed Zed in this repository. It is
written into a file the command can take back out, and nothing is written unless
the route is already being served.
Existing custom provider authentication stays in the client, and your own
endpoint stays yours: the proxy is inserted in front of whatever you configured,
never in place of it. TOKEN_OPTIMIZER_PROXY=0 disables routing and
TOKEN_OPTIMIZER_MODE=off disables optimization. npm lifecycle scripts may be
blocked, so package installation alone is not proof that activation ran.
OpenCode's session plugin uses its resolved configuration and project directory;
account credentials and provider files stay in place. Unsupported endpoints and
provider modes retain native routing. Claude routing controlled by local
managed-policy files or Windows registry policy also stays native. Remote/MDM
policy can arrive after launch; the launcher reports observed model traffic, not
just listener startup.
doctor checks routing configuration.
Request-body capture is opt-in via TOKEN_OPTIMIZER_PROXY_CAPTURE; when enabled,
it writes plaintext request content to the named directory.
Capture queues are bounded to 128 requests and 16 MiB of queued snapshots and
metadata. Excess captures are rejected with a warning while requests continue;
live audits reject incomplete capture evidence. See the
allocation hardening and follow-up results
for the reproduced OOM fix, constrained-memory replay, and remaining cost losses.
Your provider key is forwarded in the request headers and is never read, stored
or written by the proxy. That is a narrower claim than "nothing sensitive is
written", and the difference matters: the proxy does write message content to
disk, described next, and it does not inspect that content for secrets. If a
secret is in your conversation, it can reach a spill file the same way any other
text does.
It will not reach an unencrypted network, though: an upstream that is neither
https nor loopback is refused rather than forwarded to, and with no upstream
configured the proxy serves only Anthropic's own routes -- so a client for a
different provider is told to name its provider rather than having its key sent
to the wrong company.
It does write some payload to disk, and you should know exactly when. An
elision has to name a way back to what it removed, and content that arrived in a
tool result has no file of its own -- so that content is written to a spill
file and the marker names its path. This happens only when an engine actually
elides something recoverable-by-path: a JSON array tail, a set of function
bodies, a passage of prose. Requests below the size floor, requests nothing
claims, and every elision that is lossless are all written nowhere.
Spills go under your OS temp directory in token-optimizer-spill/, one file per
distinct content, created 0600 (owner read/write only) and named by an HMAC of
the content under a salt generated fresh in each proxy process -- so the name
discloses nothing and two runs do not collide. They are never read back by us;
the agent reads them with the Read tool it already has, which is the whole
point of a path instead of a hash.
They are not deleted while the proxy is running, and that is deliberate: a
marker whose spill has been swept is exactly the dangling reference this design
exists to avoid, and the agent may follow a path many turns after it was
written. So retention within one run is bounded only by the distinct content you
elide -- there is no quota, and a very long session that elides constantly can
accumulate.
When the proxy stops, its spill directory is removed. Each proxy process gets
its own directory under token-optimizer-spill/, so stopping one never sweeps
another's live paths, and the moment it stops is also the moment no agent can
still be following one of them. If a run ends without that cleanup -- a kill
signal, a power cut -- the directory is left behind in your OS temp directory
and is safe to delete by hand. Running with the proxy off writes no spills at
all.
Four switches, and what each one trades
variable
default
what it does
TOKEN_OPTIMIZER_PROXY
on
the compression proxy itself; set 0 to opt out
TOKEN_OPTIMIZER_COMPRESSION
balanced
balanced, aggressive, conservative, lossless
TOKEN_OPTIMIZER_PROXY_KNOWLEDGE
on
put what this project already learned in the cached prefix
TOKEN_OPTIMIZER_PROXY_NET_SAVING
off
never send upstream more bytes than the client gave
Both of the on rows said off here until 7.1.0, which was wrong about the
shipped code rather than a change of default: an unset value has always meant
enabled, and 0, false, no and off are what turn either one off.
TOKEN_OPTIMIZER_PROXY_NET_SAVING is off because the knowledge block is supposed
to grow the request: it is charged once as a cache write and then read at 0.1x for
the rest of the session, which pays on a long conversation and does not on a short
one. Turn it on and the block is dropped on any turn where sending it would hand
upstream more bytes than the client handed us -- the wire never inflates, and the
prefix knowledge is what you give up for that.
Rollout channels
Each switch above is its own decision, which is fine until there are ten of them.
A channel is one decision instead: stable is what is on by default today, and
each wider channel carries everything the narrower ones do plus the features still
being measured.
variable
default
what it does
TOKEN_OPTIMIZER_ROLLOUT_CHANNEL
stable
stable, beta, canary, dev
TOKEN_OPTIMIZER_FEATURES
unset
a comma list of feature names to turn on within the channel
TOKEN_OPTIMIZER_DISABLE_FEATURES
unset
a comma list to turn off, whatever else is set
beta turns on install self-repair; canary adds compact tool definitions. Those
two are staged: the channel is what enables them.
Everything else a wider channel reaches is request-only, and a channel never turns
one on for you -- it only lets you ask. beta makes the output shaper and the
shared knowledge graph askable, canary adds identifier substitution and
thinking-block dropping, and dev holds nothing back. Each of those either changes
what the model writes or forces an answer the code otherwise works out from the
directory it is running in, which is a decision worth naming rather than inheriting:
A variable you already set still wins over the channel. If
TOKEN_OPTIMIZER_PROXY_SUBSTITUTE=1 is in your settings, substitution is on even
on stable, because that switch shipped before channels existed and moving you to
a channel is not supposed to quietly turn your own configuration off.
TOKEN_OPTIMIZER_DISABLE_FEATURES is the one thing nothing overrides.
A name this does not recognise turns nothing on, and a value it cannot read is not
treated as yes. Both are reported rather than guessed at:
npm run doctor # the channel and what it turned on
npm run doctor -- --features # every feature and the reason for each
The telemetry switches
variable
default
what it does
TOKEN_OPTIMIZER_TELEMETRY
off
aggregate counts into a file in your home directory
TOKEN_OPTIMIZER_BEACON
off
upload that file; requires the one above as well
DO_NOT_TRACK
unset
overrides both, whatever else is set
TOKEN_OPTIMIZER_BEACON_URL
project
send somewhere else, for a fork or a self-hosted table
TOKEN_OPTIMIZER_BEACON_KEY
packed
the key to send with; empty in a build without one
TOKEN_OPTIMIZER_BEACON_TABLE
telemetry_events
the table to insert into
TOKEN_OPTIMIZER_UPDATE_CHECK
on
ask npm for the latest version, in doctor only
The key is not in this repository. It is public and the package is published, so
a committed key would be handed to everyone who runs npm view and could not be
rotated out of the copies already installed; the release workflow stamps it into
the tarball instead. A build without it still works and still aggregates locally
-- it just cannot upload, and install_doctor prints packed without a key
rather than letting an opt-in look like it is working.
lossless is worth knowing about: it forbids every transform that removes
something the output cannot reconstruct -- function bodies, array tails,
lower-signal prose -- and keeps the ones that can, which is whitespace, null
keys, folded duplicate lines with their timestamps listed, repeated path
prefixes, and back-references to content still in the request. Measured on the
codebase-exploration workload it still removes 12.2%, with zero lossy
elisions. For review or audit work where "the model can read the path" is not
an acceptable answer, that is the setting.
Presets are a starting point, not a ceiling: the library takes a
CompressionOptions object layered over the preset, so "aggressive, but keep
six head rows" is expressible.
The graph, in the cached prefix
The expensive failure here is turns, not tokens. This project measured a
posture that cut nothing and cost 1.471x through extra turns alone. So a
finding that prevents one wasted turn pays for a great deal of context.
The knowledge graph knows things that would prevent them, but until now it
reached the model two ways and both arrive too late or too dear: a SessionStart
index chosen once from the opening task text, and a PreToolUse advisory that
fires after the model already decided to make the call it is advising about --
so acting on it costs the very turn it was meant to save.
Verified project findings enter the cached prefix by default, with a
2,000-character budget. The proxy freezes the selected block for a conversation
to preserve the cache; new sessions can select newly recorded findings. This
includes the Codex Responses API path. Shared graphs exclude project-specific
claims. Set TOKEN_OPTIMIZER_PROXY_KNOWLEDGE=0 to disable injection.
Added knowledge is reported separately as injectedChars. It adds input tokens;
net cost improvement requires measuring whether it prevents enough work.
Earlier head-to-head results do not establish that the newly enabled combination
wins every workload.
Images
An image block was invisible to all of this until recently -- every walker keyed
on block.text. That matters because an image costs about width * height / 750 tokens, so one 1456x816 screenshot is roughly 1,585, re-sent as history on
every later turn. A browser-driving agent sends the same frame repeatedly, and
those repeats now collapse to a reference to the copy already in the request.
Dimensions come from the file header, so there is no codec dependency; resizing
is deliberately not done for the same reason.
Relevance: BM25 by default, a model if you want one
Retention is ranked by BM25, which is why this needs no Python, no model
weights and no RAM floor, and why it is deterministic enough to keep a cached
prefix byte-stable. Lexical has a ceiling, though: "the database ran out of
handles" shares no token with "connection pool exhausted", and worse, BM25
does not go quiet on that question -- it confidently picks the wrong line.
So there are two ways to do better. registerRanker installs any synchronous
ranker. Or supply a SemanticEncoder and let the proxy warm an embedding
cache: a request-level async pre-pass embeds the candidate units in one batch,
and the engines then read vectors synchronously, exactly as before. An
onnxruntime-node adapter is included, loaded dynamically so the runtime is
never a dependency of this package.
Anything the pre-pass did not see falls back to BM25 rather than scoring zero
-- a missing vector is our omission, not evidence about the line.
The adapter is verified against real inference, not mocked:
npm run verify:onnx runs nine checks against a 2.3 KB ONNX graph with real
weights, committed alongside the generator that builds it. It is a script
rather than a jest test because onnxruntime checks instanceof Float32Array
in native code and jest's per-suite VM realm has its own typed arrays; CI
installs the optional runtime and runs it.
Not claimed: that an embedding model actually beats BM25 on these
workloads. The mechanism is proved and the measurement is not done.
Compaction is consolidation, not loss
Everyone else checkpoints and restores what you had — which spends the
scarcest budget in the session replaying context you already paid for.
Selection here is derived, not a category list: cost-to-rederive × irrecoverability × reuse-probability, with dead ends and decisions on a floor,
because cheap-to-find is not the same as cheap-to-find-again. Restoration then
adapts to the situation — mid-problem, cold resume, or in-flow — within a
measured budget:
Where you were: does clock skew explain the 401s?
ruled out: token signing, clock drift on the client
untested: NTP skew on the server
That is resuming a thought. A summary describes one.
Progressive disclosure that knows what you asked
A large tool result becomes a preview chosen by the session's actual
question, after parsing the output's shape (test report, diff, stack trace,
log, JSON) — not the first 40 lines because they are first.
[selected against: "which shard fails?"]
--- failures ---
FAILED DBNetTests.BceOnRelu -- expected 0.0 got NaN
FAILED TftGradientFlow -- gradient did not reach the encoder
---- omitted: 1,760 lines of passing tests (expand 8bb6bd66) ----
Every cut is named, because a model reasoning over a silent truncation
cannot know it is missing anything. expand serves from a content-addressed
store — it never re-runs your test suite — and expanding both teaches the next
preview and promotes what you needed into the graph, so the second expansion
never happens.
Prompt-cache economics, measured from your own transcript
Provider caches are billable and provider-specific; a cache hit is not a free
input token. For example, Anthropic publishes separate cache-read and
cache-write multipliers, while OpenAI and Gemini expose their own cached-input
usage and pricing rules. Token Optimizer reads native cache fields when the
client supplies them and keeps reads, writes, uncached input, and output
separate. It never applies one provider's cache multiplier to another client.
The attribution view then does the part a hit rate cannot:
! CLAUDE.md:2 has an embedded timestamp, invalidating everything after it
about 329,421 tokens re-written per session
Attributed to a line, priced by what sits behind it. Keep-warm is decided by
expected value from your observed gaps, per TTL tier — and when neither tier
pays, it says so.
Model routing decided by outcomes, not by task size
Everyone guesses from task shape and never checks. This reads which model ran
each episode and what happened — retries, errors, turns — and prices both
mistakes: what an overpowered model wastes, and what an underpowered one costs
in retries. A tier that needs a retry in more than half its episodes is excluded
at any price, because four cheap turns that fail are not cheap.
Waste detection that becomes a ratchet
A report is read once and forgotten. Here a detection produces a durable,
measured, reversible fix — a skip rule, a composite touch — plus a ~50-token
session-start briefing so the waste never starts. Detectors are a shipped floor
plus patterns derived from your project's own history, each carrying what it
has actually saved:
generated/schema.d.ts: read in 9/9 sessions, never the source of a finding
3,400 tokens/session; cost equivalent not priced; apply: waste_audit action="apply"
Anything that touches your files is proposed as a diff and never applied.
One audit across every project
fleet_audit ranks your whole machine by measured cost, and does something a
per-project scanner cannot: a fix proven in one project is offered to the others
containing the same file contents, carrying the evidence from where it was
measured. Matching is by content hash, never by filename.
It also runs the natural experiment nobody else can — enforcing clients versus
directive ones — and reports it whichever way it falls, with the confound
stated.
It learns from what already went wrong, offline
Every failed tool call you have already paid for is sitting in your agent's own
transcripts. npx token-optimizer-learn reads them locally, groups the failures
that keep repeating, and writes the ones worth remembering into the instructions
file that agent already reads -- CLAUDE.md for Claude Code, AGENTS.md for
Codex.
npx token-optimizer-learn # what it found, printed, nothing written
npx token-optimizer-learn --projects # which projects have transcripts
npx token-optimizer-learn --since 14 --write
--agent <claude|codex|auto>, --project <path>, --since <days>,
--max-sessions <n>, --json. Nothing is written without --write, and a
block it writes is delimited, so running it again replaces its own advice instead
of appending another copy -- nothing outside those markers is ever touched.
Three things it refuses to do, because each one produced a confidently wrong rule
on a real corpus:
It does not blame a chained command line. A transcript records the line and
the exit code, not which of cd X && git grep a && ls b failed, so those are
counted as unattributable and reported as a number rather than pinned on the
first command.
It does not turn your failing build into advice. A failing test or compile
is the work going wrong, not a habit to avoid.
It says what it could not read. Transcripts it could not parse and failures
it could not categorise are printed as counts, because they bound how much of
the result to believe.
It reads local files only -- no network, no upload -- and reduces your home
directory to ~ before any recommendation text is written. TOKEN_OPTIMIZER_LEARN_PLUGINS
takes a comma-separated list of module specifiers to add another agent's format;
nothing is discovered or imported unless you name it.
Trust: we ship hooks that refuse your tool calls
That is a bigger ask than a normal dependency makes, so:
Verify the release. Published from CI with npm provenance — npm audit signatures ties the artifact to the workflow run and the commit, without
trusting us. CHECKSUMS.sha256 ships alongside for offline checking.
If nothing seems to be happening, the lifecycle bundle is probably not
installed. An MCP server process cannot modify the host that launched it, and
npm 11 gates lifecycle scripts behind allow-scripts. Install the native
plugin/hook bundle listed for your client below; adding only the MCP server gives
the model tools but no pre-execution veto. For a legacy global Claude Code
installation, recovery is one line:
npx token-optimizer-install # wire Claude Code hooks
npx token-optimizer-doctor # prove the Claude hooks work
Check that it works — not that files exist.
npx token-optimizer-doctor # or npm run doctor, from a clone
It feeds a synthetic payload to the real hook binary and asserts a large read
is refused and a small one is not. A checklist would have passed on the exact
bug this project once shipped, where the plugin was connected, visible in
/mcp, and saving nothing. Every failure names its own fix.
Every refusal carries its own off switch. Enforcement that hides its disable
is coercive, and the person who needs it is mid-refusal, not reading a README:
auth.ts is 91 KB. Call smart_read instead.
(Not what you wanted? TOKEN_OPTIMIZER_MODE=assist stops refusals and keeps everything else on.)
Removal is exact. The installer records every file it wrote, with hashes.
Uninstall removes only what still matches; anything you edited since is left in
place and named. Your own hooks are never touched, and we never rewrite your
settings.json — we m
T
Token Optimizer
ooples·579 étoiles·
91.5% fewer tokens on tool output on Headroom's own benchmark, where Headroom removes 61.4%. Separately, the same SWE-bench tasks resolved at 0.71x the cost of plain Claude Code. Plus memory across se
91.5% fewer tokens on tool output on Headroom's own benchmark, where Headroom removes 61.4%. Separately, the same SWE-bench tasks resolved at 0.71x the cost of plain Claude Code. Plus memory across sessions and subscription quota pacing. Claude Code, Codex, Copilot, Gemini and 12 more. Local, MIT.
Outils
Outils
Pas encore de liste d'outils. Les serveurs distants en obtiennent une après le premier contrôle d'état.
FAQ
Questions
Comment installer Token Optimizer ?
Copiez sur cette page la configuration stdio pour Cursor, Claude Desktop, Claude Code ou VS Code. Elle utilise l'installation de l'éditeur, pas un proxy PlugMCP.
L'éditeur indique ces variables d'environnement : TOKEN_OPTIMIZER_CACHE_DIR.
Quel transport Token Optimizer utilise-t-il ?
Token Optimizer s'exécute en local via stdio.
Quand Token Optimizer a-t-il été vérifié pour la dernière fois ?
Dernière vérification : 2026-11-13. 91.5% fewer tokens on tool output on Headroom's own benchmark, where Headroom removes 61.4%. Separately, the same SWE-bench tasks resolved at 0.71x the cost of plain Claude Code. Plus memory across sessions and subscription quota pacing. Claude Code, Codex, Copilot, Gemini and 12 more. Local, MIT.
