cage-flux 0.15.1

Cage — a flux: deterministic attribution ledger for LLM token traffic and tool savings ($0, stdlib).

Cage

Cost dashboards tell you what your AI stack spent. Cage tells you what each tool actually saved you — and what a human would have cost instead.

You're paying for an agent, a graph tool, a rules engine, maybe Copilot. At the end of the month someone asks "is any of this worth it?" — and the honest answer is a shrug and a Slack thread. Cage meters every LLM call, collects a savings receipt from each tool in the stack, and turns the raw stream into an attribution ledger: what you spent, what each tool saved you, what every other combination of tools would have cost, and how much money and time the agent saved versus a person doing the same task. $0, deterministic, zero dependencies, no model in the maintenance path.

Named after John Cage. · Python ≥ 3.11 · stdlib only · MIT · sits beside fux, bach, wagner, orff.

▶ Demo GIF coming soon.

The story

Another README story. Yeah. Because nobody ever walked out of a meeting humming a feature table, and you will not remember mine. So forget the table. Here's ninety seconds about a conference room, a pile of money, and a bunch of people who have no idea what they're talking about. One of them is you. — Arpit

You ever notice how everybody's saving money now? Everybody. The agent's saving money. The graph tool's saving money. Copilot's saving money. Two tools you built over a weekend — saving money. Add it all up and you should be getting a check in the mail. Funny thing about that. The bill went up.

Here's the con. Nobody — and I mean nobody — can show you the number. They got slides. They got a roadmap. They got a guy named Kevin who "feels like it's a game-changer." What they don't got is one honest figure that says this tool saved this much on this task, and here's what it would've cost to do the boring old way, by hand. Ask for that number and watch the room go quiet and somebody suggest we "circle back."

And the kicker — you built half of it. So when finance points at you and says "is this worth it," you, the expert, the one who's supposed to know — you got a screenshot and a feeling. You're not in trouble for spending the money, folks. You're in trouble because you bought the same fog everybody else did.

Cage is the thing that ruins the fog. It's the itemized receipt nobody asks for and everybody needs: the graph tool saved 27,000 tokens here, fux saved 6,400, the agent did in four minutes what a person does in two hours — plus every other combo you could've run, priced out, each number stamped so you know which ones are real and which ones are some computer's best guess. It doesn't do synergy. It does arithmetic.

See it

$ cage matrix --task fix-handover-bug

Counterfactual matrix · task 'fix-handover-bug' · anthropic/claude-opus-4-8
  base 2,000 tok + output 1,500 tok held constant

graphify  fux  compressor   input tok    cost    source
   ✗       ✗       ✗           50,000   $0.1725   modeled
   ✓       ✗       ✗           23,000   $0.0915   modeled
   ✓       ✓       ✗           16,600   $0.0723   modeled
   ✓       ✓       ✓            8,600   $0.0483   measured   ← the run you actually made

  full stack vs all-off: 72% cheaper ($0.1725 → $0.0483)

Per-tool savings any meter can attempt. The part no cost dashboard does is the rest of that table — what each stack you didn't run would have cost — and the source column, so you always know which row is an invoice and which is a reconstruction. Only the configuration you actually ran is measured; no projection ever masquerades as an invoice. That discipline is the whole product.

Quickstart

pip install cage-flux           # the CLI, zero third-party deps
cd your-project
cage setup                      # guided wizard: defaults to all agents, wires skill + hooks + graphify
# non-interactively: cage setup --all   (or --claude / --codex / … for just one)
cage demo                       # seed the worked example
cage matrix                     # the counterfactual permutation table
cage human                      # agent-vs-human: $ and hours saved
cage query "how is human cost calculated"   # explain any number — live formula, $0

Adopting into a project — cage setup is the single front door: it offers Claude Code / Codex / Copilot / Kiro and defaults to wiring all of them (any agent's hook captures the whole stack, so there's no reason to pick just one). Drive it non-interactively with cage setup --all — or cage setup --claude for a single agent (--no-skill / --no-project / --no-graphify to skip parts). For finer control: cage setup --project-only scaffolds .cage/ + the bin/graphify interceptor without the global skill (agent wiring opt-in via --<agent>), cage setup --wire-only --claude wires just one agent's hooks + MCP, and cage setup --status reports what's already wired.

Metering from your own code is the library adapter — it targets the protocol, not any named client, and is fail-open (a metering error never breaks your call):

import cage

with cage.meter("code-edit", task="fix-bug") as m:
    resp = client.messages.create(...)            # any Anthropic/OpenAI client
    m.usage(provider="anthropic", model="claude-opus-4-8",
            tokens_in=8600, tokens_out=1500, cached_in=3200)

# A tool that shrank the context files a receipt for what it spared you:
cage.record_receipt(tool="fux", raw_alternative=8000, actual=1600,
                    call=m.call_id, task="fix-bug", method="modeled")

Explain it like I'm five

You and a robot helper did the chores. At the end of the day someone wants to know: did the robot actually help, or did it just look busy?

Cage is the chart on the fridge. It writes down how long each chore took with the robot, and how long it would have taken if you'd done it yourself — so you can see, in real minutes and real dollars, which helper earned its place and which one just made noise. And it's careful to mark which numbers it actually timed and which ones are its best guess, so nobody gets fooled by a confident-looking total. It does all of this for free, without ever phoning a friend for the answer.

Why it's different

It's not another cost dashboard. The difference is a set of properties, not features:

• Deterministic. Every derived view — report, attribution, the counterfactual matrix, ROI, the human axis — is pure parse/arithmetic over an append-only log. Same ledger + same policy ⇒ identical tables, every time. The numbers never drift because nothing guesses.
• Honest by construction. Every figure carries a method: measured (a real invoice), modeled (a reconstructed counterfactual), or estimated (a human/labor guess). A projection can never read as an invoice — the one property a "trust me, it paid off" slide can't offer.
• $0 and zero-dependency. Stdlib-only Python, dependencies = []. Heavy ML is an opt-in, off-by-default tier ([embeddings], [ml]), never on the default path. Portable as a tarball, auditable line by line.
• Agent-native. Every read command takes --json; the ledger is served over MCP. Built so an agent can pull its own cost numbers and verify them, not just read a chart.

The "so what" chain: deterministic → so the numbers never hallucinate → so each one carries a defensible method → so you can put the savings claim in front of finance, or an auditor. That last clause is the one a dashboard can't say.

Honest attribution — the part that survives the room

Anyone can sum a bill. Cage's job is to divide credit without lying about it, and it does that with three rules:

• Marginal-by-fixed-order. Each tool's receipt reports the saving it produced given the tools upstream of it in the canonical pipeline. The marginals sum exactly to the total — no overlap, no double-counting, $0 to compute, and defensible because the order is fixed and visible (not a black-box Shapley pass; that's a deferred opt-in audit mode).
• The counterfactual matrix. For a task whose tools each shrank a slice of context, Cage enumerates the 2ⁿ on/off permutations and prices each at the task's model — so "what would graphify-off + fux-on have cost?" is a row, not a hand-wave. Only the configuration actually run is measured; every reconstructed cell is modeled (or estimated if it leans on an estimate).
• Tier-1 — agent vs human. Beyond tool-vs-tool, Cage models the whole-task counterfactual: what a person would have cost in time and money. A human receipt is just a receipt whose tool is "human", priced in minutes → money at a configured rate ([human] in policy.toml, or CAGE_HUMAN_RATE). It is estimated unless you supply a real timesheet, and carries a confidence so round task-type guesses read as low-credibility instead of masquerading as precise. The time metric can go negative — if the agent thrashed longer than a human would have, the table says so.

Agent vs human · 14 tasks · rate source: policy ($80/hr)

agent     tasks   human $    agent $    saved $   saved hrs   conf   method
claude       9    $1,140.00    $4.12    $1,135.88     13.2     0.51   estimated
codex        3      $260.00    $1.55      $258.45      3.1     0.50   estimated
TOTAL       14    $1,530.00    $6.55    $1,523.45     17.9     0.51

The savings are anchored to the commit they produced — Cage snapshots a git-aware task record (SHA, branch, diff size, wall-clock) at task close, so a number can always be traced back to the change that earned it.

Authorship — who wrote which commit, and how sure are we

A different question than what did this cost: who is accountable for this diff. cage origin <sha> answers it from the same append-only substrate — a fourth record type that records which agent wrote which files in which commit, captured by a PostToolUse hook with a transcript fallback, never blocking an edit or a commit:

$ cage origin HEAD
sha 9f3c1a2 · origin: agent (claude-code) · confidence 0.83 · method hooked
  cage/origin.py        +118  -0
  cage/originrecord.py   +97  -0

The same honesty discipline as everywhere else, in a parallel namespace: a row carries a method (hooked > transcript > heuristic, never upgraded when fragments merge) and an origin (human / agent / agent-autonomous). unknown is never a stored row — a commit with no cage signal is unknown by absence, derived at read time, so the ledger stays sparse and pre-Cage history reads honestly without bloating it. origin=human is reachable only through an explicit human attestation (cage origin <sha> --attest human). Distribution is git-notes (refs/notes/cage-provenance, CI is the sole writer); cage verify is report-only and always exits 0 — visibility in CI, never a gate. Counts-never-content holds: file paths and line counts only, never a diff body or a commit message.

Every number is reviewable — and you can ask it

Cage keeps its numbers in three layers, never mixed, so any figure is auditable in exactly one place:

Layer	Holds	Lives in
Contract	the closed enums (UNITS, METHODS) — the substrate's shape	schema.py
Policy	user-tunable economics: prices, the human rate, default minutes, budgets, pipeline order, confidence	policy.toml — the only place economic numbers live
Constants	code heuristics not meant as config but that must be reviewable: the token divisor, the matrix ceiling, the provenance ranks, the confidence fallback	constants.py

And because the math should explain itself, cage query prints the real formula for any value with its numbers read live from policy + constants — never a hard-coded literal, so an explanation can't drift from the code:

$ cage query "how is human cost calculated"
human-cost · how a human alternative is priced
  formula:  usd = minutes / 60 × rate     (rate = $80/hr, source: policy)
  chain: explicit usd > per-receipt minutes > task-type table > global default
  confidence: measured 0.9 · estimated 0.7 · type-table 0.5 · default 0.3
  method:   estimated — a labor guess; never 'measured' unless a real timesheet/quote.
  code:     cage/human.py · cage/convert.py · policy.toml [human]

Set CAGE_HUMAN_RATE=200 and that printed rate changes — proof it's the code's actual number, not a slide. It's deterministic and $0: a curated explainer registry, no LLM, no network. Try cage query --list for every topic, or --json for the agent-as-user.

cage query also explains how cage itself works, not just how a value is computed — cage query "how does cage work" walks the data flow, fail-open metering, attribution, method tags, receipts, and the rest, with the same live-fact guarantee (the printed ledger paths, pipeline order, and subcommand count are read from the running code, never typed in). cage query --list --kind concept lists just those topics.

How it works

One append-only log in, every view derived from it for $0:

record_call / record_receipt  →  .cage/ledger/{calls,receipts,tasks,provenance}.jsonl  (append-only)
        (meter, fail-open)                    │
                                              ▼  derive ($0, no model)
   policy.toml (prices/order/budgets/rates) → report · attrib · matrix · roi
                                             · human · trend · budget · why · origin

provenance.jsonl is a local buffer only — canonical authorship lives in refs/notes/cage-provenance, written by CI alone.

You meter at the provider boundary (library adapter, a reverse proxy for clients you can't edit, or by parsing a Claude Code / Codex transcript). Everything downstream is a deterministic projection. The ledger carries token counts, never prompt bodies — PII-safe by construction; point CAGE_LEDGER at a private store to keep even the counts off-disk.

A tool earns rows in attrib/matrix/roi by filing a savings receipt, and there are two ways in, by who owns the tool:

• In-tool (you own it) — e.g. fux carries a fail-open cage_receipt.py and emits its own tool="fux" receipt. Cage stays optional; fux runs unchanged with cage absent.
• External adapter (third-party) — e.g. graphify: cage graphify -- graphify query "…" runs graphify unmodified, passes its output through byte-for-byte, and files a tool="graphify" receipt by parsing the cited source_files. graphify is never edited; a metering error never alters its result.

The full command surface (ledger · attribution · human axis · ops · agents)

cage init                      # scaffold .cage/ (policy + gitignored ledger)
cage setup [--claude]          # guided onboarding: skill + init + wiring + graphify for one agent
cage setup --project-only --claude   # scaffold + graphify + PATH only (no global skill)
cage setup --wire-only --claude      # wire just one agent's metering hooks + MCP
cage setup --status            # report which agents are wired (changes nothing)
cage doctor --json             # verify this project's setup is correct (non-zero on failure)
cage report --by model         # ledger rollup: spend by route / model / day / agent
cage attrib --task ID          # per-tool marginal savings (sum of marginals = total)
cage matrix --task ID          # the counterfactual permutation table (2ⁿ on/off)
cage matrix --task ID --human  # …with a human anchor row + vs-human columns
cage roi --since 30d           # saved $ per tool vs its own cost + added latency
cage human [--agent claude]    # agent-vs-human: $ AND hours saved, per agent
cage human-record --task ID --type feature   # record a Tier-1 human alternative
cage trend --by week --metric both           # cost + time savings as a time-series
cage why <call-id>             # full provenance: a call + every receipt against it
cage origin <sha> [--attest human]   # who wrote which files in a commit (authorship)
cage notes-sync [--write]      # distribute authorship → refs/notes/cage-provenance (CI writes)
cage verify                    # report-only consistency pass over the ledger (always exits 0)
cage quality / cage outcome ID # cost per *successful* task (cost is honest with outcome)
cage regression                # alert when cost-per-call drifts up
cage recommend                 # cheapest-path: which tools to enable / skip
cage forecast                  # project monthly spend vs the budget
cage graphify -- graphify …    # meter a third-party graphify call (transparent passthrough)
cage setup                     # install /cage + /cage-doctor into every agent home
cage proxy --port 8788         # the universal meter for clients you can't edit
cage mcp                       # serve the ledger to agents over MCP (stdio)
cage serve                     # local dashboard over the ledger
cage query "how is X computed"  # explain any number deterministically, with live values
cage query "how does cage work" # …or the mechanism itself: data flow, attribution, method tags…
cage demo                      # seed the worked example that proves the thesis

Every read command takes --json for the agent-as-user (machine-readable, typed).

Works with any agent — explicit capture over one global ledger

Cage meters whatever speaks the wire format, so all four agents share one ledger contract. Capture is pull-based and universal: cage import reads each agent's on-disk usage log into the ledger, and cage export refreshes then emits it — they need no hooks, no project, and work the same whether you run a CLI or a VS Code extension.

cage import                 # capture every agent's spend into the active ledger
cage export --format csv    # refresh, then emit (jsonl | csv | json)
cage report                 # where the spend went
cage watch                  # optional: a foreground loop you Ctrl-C (no daemon)

The ledger resolves --ledger/CAGE_BASE → project .cage/ → global ~/.cage — so a user with no project captures into the global ledger (cage setup --global to seed it). cage installs no background job (no launchd/systemd/cron); automate it, if you like, with your own cron line calling cage import.

Agent	Capture (universal)	Optional real-time	Read
Claude Code	cage import (transcript)	Stop hook (CLI only)	cage MCP
Codex	cage import (rollouts)	Stop hook (CLI only)	cage MCP
Copilot	cage import (session log)	agentStop hook (CLI only)	cage MCP
Kiro	cage import (token log)	agentStop hook (CLI only)	cage MCP
Your code / Orff	cage.meter() library	—	cage CLI / MCP

Hooks are an optional real-time add-on — they fire only under a CLI client, never under a VS Code extension — so cage import/cage export is the path that always works. cage report --project <name> slices the global ledger by working dir (exact for Claude; Copilot/Kiro/Codex logs carry no project, so they're excluded from that filter).

An agent's spend isn't showing up? cage doctor shows the active ledger, each agent's real capture state, and "last import: N ago"; the metadata-only debug log says per agent whether a hook fired or raised — see Debugging capture.

The $0 guarantee

Every derived view is parse / arithmetic over the log — no LLM call, ever, on the read or maintenance path. The only model spend is whatever your agent already does; Cage just meters it. The semantic cache and learned compressor ship behind opt-in [embeddings] / [ml] extras; the default install is model-free and dependency-free. 317 tests passing; cage demo reproduces the worked attribution example against a real ledger.

Honest limits. Cage doesn't decide your human rate — it prices minutes at a blended rate you set, and labels the result estimated so it never pretends to be a timesheet. Marginal-by-fixed-order is defensible and $0, but it is an ordering convention, not a Shapley value (that's a deferred audit mode). And a counterfactual cell is an honest reconstruction, never an invoice — the method column says so on every row, on purpose.

What's new

Latest release below — full history and detail in CHANGELOG.md.

• v0.15.1 — validation-pass fixes. Closes a concurrent-import double-count (two capture sweeps racing on the same ledger could land one turn twice — importcmd now holds an exclusive flock across the read-check-append section, fail-open when fcntl is absent), makes cage demo idempotent (re-running no longer doubles the §4.4 tables), lets cage setup --project-only scaffold with no agent flag, and adds cage export --json as an alias for --format json. Additive, fail-open preserved; +6 regression tests.
• v0.15.0 — meter dedup correctness + cage limits. A new cage limits view surfaces Codex rate-limit windows (latest local snapshot — remaining_pct + reset + age) and estimated AI-credit consumption (tokens × a per-model [credits] multiplier) for token-based providers only — every figure labelled estimated + reconcilable, Kiro never fabricated; cage limits --json debuts the cage.v1 envelope. Quota is a machine-local .cage/state/limits.json snapshot, not a ledger substrate. Plus a defensive meter dedup fix (deterministic id when a Claude turn lacks a uuid — empirically 0/29,714, uuid-present rows byte-identical). Additive: no CALL_FIELDS/make_call change, no ledger rewrite.

The name

Named after John Cage, whose 4′33″ framed four and a half minutes of "silence" so an audience would finally hear the ambient cost they'd been ignoring. Cage the tool does the same to your AI stack: it takes the spend and the savings everyone assumed were free or unknowable, and makes them something you can actually account for. It's part of a family of deterministic substrate → derived views tools — fux (decisions → rules) — and now Cage (LLM traffic + receipts → ledger). The names are deliberate, and they sit beside bach, wagner, and orff.

---

If you've ever been the one in the room with no numbers, run cage demo — pip install cage-flux.

License

MIT — see LICENSE.