bundlebox

[bundlebox]

bundlebox

The zero-token software factory for AI coding agents.
Scan, pack, route and run work across Claude Code, Codex, Gemini CLI, Cursor, Copilot, OpenCode, Aider and any custom agent, and measure every token a session used and was spared.

[npm] [ci] [MIT]

An AI coding agent is billed for what it reads, and most of what it reads is orientation, not judgement: where a symbol lives, whether two tables agree, which command proves a change, what changed since last week. bundlebox answers every question a parse, a count, a path check or a set difference can answer before the agent opens, packs the answers into a brief sized to the agent's window, opens the agent under a measured-minimal flag stack, and measures afterwards what the session used and what it was spared. It never calls a model itself.

bb genesis   ->  a world     a doc becomes surfaces, rules and capabilities  0 tokensbb scan      ->  findings    what is wrong, with evidence                    0 tokensbb fix       ->  patches     local actuators close what they can             0 tokensbb cookbook  ->  a board     what the RUNNING system does, run by the kernel 0 tokensbb simulate  ->  limits      what it does at a hundred callers               0 tokensbb compile   ->  units       what to do about the rest, packed to one windowbb pinpoint  ->  a brief     where the work is, quoted and budgeted, before a model opensbb arc       ->  an index    every declaration in one file, read in microseconds  0 tokensbb route     ->  lanes       who does it, where, in what wavebb run       ->  sessions    the only verb that spendsbb finish    ->  a ledger    what proves this done, declared first and run after   0 tokensbb git       ->  PRs         commit, push, draft PR, review, gated merge     0 tokensbb runbook   ->  the system  is it up, is it ANSWERING, what broke since    0 tokensbb recom     ->  an answer   what was already driven, and whether it holds   0 tokensbb recom gate->  a decision  and therefore whether to drive at all           0 tokensbb dotty     ->  a screen    what it showed, as rows a session can diff      0 tokensbb sieve     ->  a window    a tool result shrunk before it is billed twice    0 tokensbb uptake    ->  a verdict   of everything wired in, what sessions reached for 0 tokensbb monitor   ->  the window  what the 5-hour block has left, and the guard   0 tokensbb session   ->  the bill    what a session used and saved, measured         0 tokensbb lathe     ->  automation  what this box did by hand more than twice, as scripts 0 tokensbb env       ->  a checklist what a complete .bundlebox holds, and whether this has 0 tokens

Measured on the reference workspace (regenerate with bb session, bb tokens profile --probe, bb buckmaster episodes):

Install

npm i -g bundlebox# orcurl -fsSL https://raw.githubusercontent.com/blackswanalpha/bundlebox/main/scripts/install.sh | sh

Needs Node ≥ 20 and git. Linux, macOS and Windows are supported: CI runs the full suite on all three against Node 20, 22 and 24. Two known gaps remain on Windows — a piped acceptance gate reports the last command's exit code rather than the first failure, because cmd.exe has no set -o pipefail (unpiped gates are exact), and bb session may not find transcripts, because the name Claude Code gives its projects directory there has not been verified. Three optional Rust and Python components make it faster and smarter, and everything degrades cleanly without them:

• Rust kernel (bbk): tree walks, fingerprints, token estimates, duplicate windows, symbol indexes, acceptance gates with real timeouts, worktrees. bb kernel install fetches a release binary; bb kernel build compiles it with cargo. Without it the JS implementations run, and a selftest pins both to identical answers.
• Rust index compiler (arc): compiles the derived symbol tables into one binary index the read and search guards answer from, 13x faster than scanning the tables in a cold hook process. bb arc build, recompiled by bb snapgen build when the tables move. Every reader falls back to scanning the tables when the index is absent, truncated or not an index, because "there is no index" and "nothing is declared" are different answers.
• Python expert system (bundlebox_expert, stdlib only): the rule engine with explainable derivations, confidence shrinkage, transcript signals, the process model and memory. bb buckmaster needs python3 ≥ 3.9; nothing in the zero-token path does.

There are no npm dependencies. A cron worker at 03:00 runs what is on disk or it does not run.

Start here

cd your-repobb init            # detect languages, agents and gates; write .bundlebox/config.jsonbb doctor          # what this box can run, which runtime serves each opbb wire --apply    # hooks, skills, instruction blocks and MCP entries for every agentbb env up --apply  # build everything a session reads, so nothing is derived by searchingbb scan            # the detectors. Seconds, 0 tokensbb findings        # what the store holds; bb explain <id> for onebb pinpoint gaps   # every open finding as a located, quoted, budgeted briefbb pinpoint next   # make one of them the ACTIVE brief the guards answer frombb compile         # findings -> units, each packed to one window (dry run)bb route           # units -> lanes (dry run)bb run             # writes the exact prompt and command per lane, spawns nothingbb run --apply     # spawns the sessionsbb finish check    # run the acceptance ledger; unproven never reads as greenbb session         # what the last session used and saved

From a document instead of from the code:

bb genesis docs/PRD.md --base http://127.0.0.1:4400   # surfaces, rules, capabilities, a seeded corpusbb genesis plan                                       # what nothing covers, ranked and tieredbb genesis pack                                       # one small brief per surfacebb genesis send calendar --run --spend                # the only step that costs anythingbb cookbook check && bb cookbook run                  # the kernel executes what came backbb mainboard gaps                                     # the first stage that does not hold, and its fixbb console                                            # one page: the pipeline, the window, every session

Every verb is a dry run until --apply. Only run and bridge send can spend.

Wiring into agents

bb wire --apply installs four things per agent detected on the box: an instruction block between markers, hooks where the agent supports them, skills where it loads them, and an MCP server entry. bb unwire removes only its own blocks, entries, hook rows and skills.

Two skills ship and cost nothing until they trigger, which is the property an instruction block does not have: it arrives in every system prompt and is billed whether the session was about it or not. bb-finish is completion discipline backed by runnable gates; antislop is the prose ruleset with what each rule costs and what to write instead.

bb mcp serves bb_pinpoint, bb_context, bb_snapgen, bb_findings, bb_scan, bb_oversight_brief, bb_explain, bb_tokens_estimate and bb_session over stdio as JSON-RPC 2.0, with no dependency.

bb wire --apply writes the entry for every agent in the table above. For an agent it does not manage:

{ "mcpServers": { "bundlebox": { "command": "npx", "args": ["-y", "bundlebox", "mcp"] } } }

Claude Code takes it as one line, or as a plugin, which carries the two skills and the hooks with it:

claude mcp add bundlebox -- npx -y bundlebox mcp# or/plugin marketplace add blackswanalpha/bundlebox/plugin install bundlebox@bundlebox

Full guide, including the agents not in this table, how to wire one bundlebox does not know about, and how to drive agents as lanes: docs/agents.md. bb uptake then reports which of the installed surfaces the sessions reached for.

Advisory lost; enforcement is the fix

bb uptake measured the gap. Over 15 sessions on this workspace the MCP tools fired in 0, pinpoint in 3 of the 13 sessions that opened five or more distinct files, and the reference tables in 5 of the 15 that ran a search. Those 13 sessions opened between 30 and 598 files each. Everything wired in front of the agent was a recommendation, and a recommendation loses to the model's own habit about three times in four.

So the wiring changed kind.

Run it, do not recommend it. UserPromptSubmit builds the brief itself on a task-shaped prompt — 0.47s against a 15s budget, because every input pinpoint reads is a stored artefact and it computes none of them. What enters the window is the map, about 300 tokens.

Serve what it found. PreToolUse denies a read whose region the brief already quotes and hands the quote back in the denial reason; denies an exact duplicate read; denies a declaration search the symbol tables already answer and hands back the rows; and asks before opening a file that was cut for budget. A file in scope is never fully blocked — Claude Code needs one successful read before it will edit, so the denial names the range and that read is allowed. Grep is the declared tool and Bash is where the measured transcripts search, so both are guarded.

Rank on specificity and path. Enforcing a brief raises the cost of a bad locate. Inverse document frequency over the candidate set stops an exact match on a name the tree uses everywhere from outscoring a loose match on a rare one, and a directory named wire now counts as evidence about what a file is for. An explicitly named file is pinned first rather than scored.

The verbs

The budget model

projected = overhead + brief + payload × churn + reserve

Overhead is probed, not assumed. Payload is the located region plus a widening allowance when a symbol is named, not the file. Churn is the term everyone forgets and it measures near 2.4. Reserve is per kind of unit. The verdict is FITS, TIGHT, SPLIT or HEAVY, and a split respects directory locality. Five levers act on those terms, in measured order: the lean flag stack, region not file, say it once, a cache-stable prefix, and optionally a local compression proxy on the wire reported on its own row. See EXPLANATION.md.

Measured vs estimated

Two kinds of number appear in every report and are never added: used and cache-saved are measured off the transcript; automation-saved is an estimate printed as a range. A model with no price is reported with tokens and no cost. A verb that could not look says unknown. A unit with no acceptance is unproven, not passed.

There is a third number and it is measured on both sides. bb bench runs the same task twice — once with the factory in front of it and once without — and reports the difference:

bb bench init                 # a suite from the open findings that name a filebb bench run                  # both arms, per task, with the losses printedbb bench show --json

BARE greps the tree for the task's terms, reads the search output and opens a range around the first hit in each of the top files, which is what a session does on a box where nothing is installed. PACKED is the one bb pinpoint prompt for the same task. Both arms are counted by the same estimator over text on disk, so neither calls a model and a second run on the same tree returns the same number. A task that costs more packed than bare is printed in the table, not dropped.

Every bb pinpoint brief carries one more section, and it is the one worth reading first: what this brief does not settle. No gate detected, nothing in the symbol tables matched, no region located, no evidence on file, a scope that had to be cut to fit — each is stated with its reason rather than filled in by guessing, and scored so two briefs can be compared. A question costs one turn; a wrong assumption costs the review that catches it.

The console

bb console serves one read-only page for the workspace on 127.0.0.1:7788, and bb console build writes the same page as a single file with the state embedded. It binds to loopback, has no write route, and nothing on it calls a model.

It pushes rather than polls. The store is files, so the filesystem already knows when something changed: /api/stream watches it, recomputes the state once per change however many tabs are open, and sends it. A 10s timer was wrong in both directions — a run that finished was invisible for nine seconds, and an idle workspace recomputed everything six times a minute per open tab. The header always says which state the connection is in (live, polling, reconnecting, static), because a dashboard that cannot tell you it has lost the server is worse than one that is plainly offline: the stale numbers still look like numbers.

Architecture

bin/bb.js            the entrypointsrc/                 Node, ESM, zero dependencies — every verb, the adapters, the store, the consoleskills/              the skills bb wire installs: bb-finish, antislopkernel/              Rust — bbk: walk, fingerprint, estimate, dupes, symbols, anchor, gate, worktree,                     and the scenario runner, the load simulator and the health probearc/                 Rust — the declaration index compiler. src/arc/read.js is the in-process reader,                     which is where the speed-up lands; the binary is the reference implementationexpert/              Python (stdlib) — rule engine, triage, confidence, signals, rules, graph, model, memory,                     the world derivation, coverage planning, scenario selection and board verdicts.bundlebox/          per-repo: config.json, cookbook/ (corpora), genesis/ (world models),                     var/ (store, boards, simulations, calibration), out/ (tables, briefs, packs, the page)

The kernel runs a corpus on threads with one connection per worker and a shared pacer; a 130-step corpus that takes two minutes paced against a real limiter takes under a second against a local mirror. It speaks http only, because TLS would be a dependency it must build without — the JavaScript engine runs when the base is https, when the kernel is absent, or when a corpus uses a pattern outside the kernel's documented subset, and the board names which engine ran and why. test/scenario.test.js pins the two to identical answers.

The three runtimes agree by test: test/kernel.test.js and test/expert.test.js pin the kernel and the expert system to the JS answers. One implementation per fact where only one runtime has it; parity where two do.

Configuration

One file, .bundlebox/config.json, holding only the keys you changed. Every default lives in src/core/config.js with its reason. Measured values live in .bundlebox/var/calibration.json and are written by read-merge.

{  "lanes": { "agent": "claude", "max_parallel": 4, "model": "sonnet", "daily_budget_usd": 20 },  "budget": { "max_tokens": 160000, "churn_factor": 2.4, "anchor_widen": 0.45 },  "kernel": { "gates": { ".": { "quick": "npm test", "full": "npm run lint && npm test" } } },  "detectors": { "promote_at": "medium" }}

Documentation

• docs/agents.md: integrating bundlebox into every agent, and running work through them
• docs/index.html: the reference site
• EXPLANATION.md: the thesis, the flow, what changed from the original
• whitepaper/bundlebox-whitepaper.md
• docs/review-of-the-original.md: the defects found in the reference implementation
• docs/prices.md: price table provenance
• docs/prior-art.md: seven projects read, what came back and what did not
• CHANGELOG.md and changelogs/
• CONVENTIONS.md: how the code is organised, for contributors

Releases

Tags vX.Y.Z on main build the kernel for five targets, publish to npm with provenance, and create a GitHub release with the binaries and the changelog. bb update checks the registry at most once a day; bb update --apply installs.

License

MIT. src/finish/vendor/ is vendored from Leonxlnx/unlazy 2.1.0 (MIT, commit 1667149), renamed and otherwise untouched; src/finish/vendor/PROVENANCE.md records every rename and why nothing else moved. bundlebox began as a stdlib-Python factory wired into one workspace; this package is its portable core, rebuilt and corrected. The optional wire integration talks to headroom over the process boundary and never imports it.