The first release since 2.0.0 that adds commands, and the plan that asked
for them starts from a number rather than a feature: about two hundred
downloads a month behind forty-nine commands. docs/plan-2.4.md names three
causes and a move for each, and every move is in this release or named below
as blocked on something this repository does not hold.
Changed#
changelog-coverage.test.jstells a release in preparation apart from an empty section. The guard was written after 2.3.0 and had never met a release: the pull request that cuts one foldsUnreleasedinto the new version's heading, so the section is empty while commits sit above the tag, which is exactly the state it fails on. It now skips when the manifests name a version the tags do not have and the changelog carries that version's heading, the two facts together, because either alone is the collision it exists for.The xAI and Moonshot price rows were reviewed and found orphaned.
docs.x.ai/docs/modelsandplatform.kimi.ai/docs/pricing/chatwere read on 2026-09-12 and neither lists the one id this table carries for it,grok-4andkimi-k2. Both rows keep their price, because calls in somebody's log really happened at it, and neither is marked retired, because that is recorded from the provider's own refusal to a request and needs a key this repository does not hold. Neither provider's newer models are added: xAI prices every model at two rates split at 200k prompt tokens, which this table has no way to express, and Kimi's page states no context window, which this table would have to invent. Each row says so in itsnotes, andtrazum modelsprints them.
Five merges landed on main with no entry here, and this section is where
they should have been. The rule three paragraphs above — a change that
alters nothing installable still lands there rather than nowhere — was stated
and unenforced, so four consecutive pull requests after 2.3.0 left Unreleased
empty and nothing failed. changelog-coverage.test.js now fails when there are
commits after the newest tag and nothing written under this heading. Backfilled
below rather than started clean, because a record that begins at the moment
somebody noticed is the kind of tidy history this file exists to refuse.
Added#
scripts/adoption.mjs: the adoption figures, read at release time. Plan 2.4's fifth move. Three public counters — npm downloads of@trazum/cliover the last thirty days, GitHub stars, the MCP registry's latest version — asked once and printed as the one linedocs/releasing.mdnow asks for under every new heading inRELEASES.md, so the next plan starts from a figure the way this one did. A counter that could not be read is printed as unavailable and never as zero: a zero says the product has no users, which is a claim, and an unavailable is a fact about the run.packages/core/test/adoption.test.jsholds that property without a network.The week's bill, every Monday: a recipe in
docs/running.md. Plan 2.4's fourth move, cut down to what was actually missing once the existing loop (connect,watch --once,pulse, and the four scheduler recipes) was read rather than assumed absent: the packaged spend gate on a cron, withsincecomputed by the job so a budget with no period assumed is a weekly one, and the report in the run summary because the Action's comment needs a pull request and a cron has none. The plan's section 4 is rewritten to say so.The README tells every MCP client how to install, not only Claude Code. It said two lines for Claude Code and "over stdio does the same" for everyone else. It now carries the
mcpServersJSON that Cursor, Windsurf and Claude Desktop all read, the registry listing it is published under, and thenpx @trazum/cli billline that works with nothing installed.docs/plan-2.4.mdsection 3 is rewritten from what was checked: the MCP registry already serves 2.3.0 as latest andrelease.ymlhas updated the listing on every release since 1.80.2, so that row needed nothing; the Marketplace listing and the Hugging Face Space are named as unchecked and blocked respectively rather than assumed.docs/plan-2.4.md: one door, and where the spend is. The first plan here that starts from a number rather than a feature: about two hundred downloads a month behind forty-nine commands. Three causes and a fix for each — one command that reads anything, converters for where AI money is actually spent, and presence on the surfaces that find tools — with the ones that could not be built from a schema read in this environment (Cursor, Hugging Face) named as blocked rather than written from memory.trazum bill, the 51st command: one door. The plan's first move. Reads a file or a directory, tells each file's shape from its own text with the sniffers the converters already shipped, converts it with the same converter the dedicated command uses, prices it, and ends on the receiptreceiptwrites. Every refusal a converter makes is made here; what differs is the telling — one line per file with its shape, its records and how many rows were left out, and the dedicated command named as the place that says why. Three refusals of its own: a file no shape claims is named and not guessed; a file two shapes claim is named as ambiguous and left alone; a provider's cost report is named as a bill rather than usage and pointed atreconcile.packages/cli/test/bill.test.jsruns it over a directory holding one of each.npx @trazum/cli bill ~/.claude/projectsis now the first thing that works with nothing configured.trazum from-openrouter, the 50th command, the plan's second move. OpenRouter's activity report (GET /api/v1/activity, management key, the last thirty days) read as a usage log, from the published schema with the endpoint's own example as the fixture. It meets the half Trazum already had:--pricing-liveprices hundreds of models from OpenRouter's catalogue, keyed by slug, and this report carries the same slugs. What OpenRouter charged (usage, andbyok_usage_inferencefor upstream charges on the operator's own keys) is summed over every row, refused or not, and printed beside Trazum's figure, never merged — the LiteLLMspendrule. Reasoning tokens are counted and deliberately not added, because the schema does not say whethercompletion_tokensalready holds them and adding them if it does would charge reasoning twice.--label-by-workspacemirrors the other mappings, exact match, nonullcase.packages/core/test/openrouter-activity.test.js(14 tests).trazum from-openai, the 49th command: the other provider's usage report, read as a log. The same door asfrom-anthropicand the same arrangement: the operator'scurl, the operator's admin key, and a command that reads only what came back, so Trazum still holds no provider credential. Every field is from the published OpenAPI schema ofGET /v1/organization/usage/completions, and the schema's own example is the fixture, because its numbers are an identity the converter leans on (input_tokens= uncached + cached + cache-write, each the sum of its text, audio and image parts). Three things it does that the Anthropic one does not need to. The record is written in the Chat Completions shape —prompt_tokenswith the cached half inside it — because that is how this report counts, and a record in the Anthropic shape would charge the cached half twice; a test parses a converted record back to prove the split. Audio and image tokens are never priced at a text rate: a row carrying any is reduced to its text part through the schema's own split and the rest is a named gap, with cache-write tokens on such a row left out too because the schema gives them no modality. And since the schema does not enumerate service tiers, any tier butdefaultis left out and named rather than counted.batch: truerows are refused like Anthropic's batch tier;--label-by-projectmirrors--label-by-workspace, exact match, with nonullcase because every OpenAI request belongs to a project with an id.reconcilenow reads OpenAI's cost report too, told apart by shape: itsamount.valueis dollars where Anthropic's is cents, so nothing is divided, and itsquantity_unitseparates what no token rate covers while the report's silence on batch is said out loud rather than papered over.packages/core/test/openai-usage.test.js(18 tests) andopenai-cost.test.js(14 tests); the core'sreconcilenow takes a provider-neutralBilledReading.trazum from-anthropic, the 47th command, and the surface was frozen at 46.plan-1.83-2.0.mdsays so and stays as written: the freeze was real and this is the decision to leave it, recorded rather than slipped in. What earned it is the one question no converter here could answer. Every other one reads a log or a proxy, and a log only knows the machines it was written on — the console, the other team, the laptop nobody instrumented are all spend that never reaches one.GET /v1/organizations/usage_report/messagesknows about all of it.It takes the answer, never the key. That endpoint needs an admin credential, which also manages the organisation's members and keys, and holding one is the thing this project's whole design exists not to do. The operator runs the
curlin their own shell and this reads what came back, so the converter stays a pure function of text and is tested without a network.It refuses three things rather than guessing them. A row with no model, because nothing on it says what answered. A row at any tier but standard, because batch is billed at a discount and a catalogue rate would overstate it by half while looking entirely right — this product's own definition of the unforgivable failure. And web search requests, billed per request, which no token rate reaches. Each is counted and named, and
has_moresays out loud that a bill built from one page is short.The identities on every row — the account, the service account, the API key, the workspace — are read by nothing: a fixture plants a marker in each and greps the output. A workspace id is not a project name, so
--labelis still where the project comes from.trazum reconcile, the 48th command: what you computed beside what you were billed. The payoff offrom-anthropic, and the one figure no other tool gives. Every door here answers what did this usage cost at these rates; the provider answers what did we charge you. The gap is the interesting part and there was nowhere to look at it.The two are never added.
docs/commands.mdalready stated the rule for LiteLLM'stotal_cost— a provider-billed figure sits beside Trazum's and is never merged into it, because two price tables summed into one number is how a report becomes quietly wrong. This is that rule with arithmetic attached: nothing corrects one figure with the other, and nothing decides which is right.The difference comes apart, when the report can take it apart. With
group_by[]=descriptionthe cost report says whether a row was tokens, a web search, a code execution or a session, and whether it was batch-tier. So the difference splits into what no token rate covers, what was billed at a discountfrom-anthropicrefuses to misprice, and a remainder — the same standard-tier tokens priced two ways, or usage the log never saw. The remainder is never folded into the other two, and a negative one is reported as it is: Trazum priced more than the provider charged, which is a stale rate in the direction that costs somebody money. Without that grouping the run says the difference cannot be attributed rather than naming a remainder that is really the whole difference.The unit is the trap.
amountis a decimal string in the currency's lowest unit:"123.45"in USD is $1.2345. Read as dollars it overstates a bill a hundredfold and looks plausible doing it. Divided once, at the end, with a test asserting the factor against the schema's own example.Three refusals rather than three plausible numbers. Windows that do not line up (a receipt for one day against a bill for others is a wrong number under a right title), a currency with no rate to convert it by, and an
amountthat is not a number — counted, never read as zero, because a zero quietly shrinks a bill.--label-by-workspace <rules.json>onfrom-anthropic. The gap the command shipped with, named in its own doc comment and closed here: the provider knows workspaces and does not know what you call your projects, so the mapping is one the operator writes. The same shape--label-by-cwdestablished, matched exactly rather than by prefix — paths nest and opaque identifiers do not.The
nullthat means two things. The report's schema usesnullforworkspace_idboth when the caller did not group by workspace and for the default workspace. Nothing on the row separates them, so it is derived from the report: if any other row carries an id, grouping was on and anullis the default workspace; if none does,nullrules do not apply and the run says the split asked for was not made. Guessing either way would put the default workspace's money on a label nobody chose.scripts/build-wiki.mjsmatches the commands heading by pattern. It held## The 46 commandsas a literal, which is a second copy of a count the product decides, and it went stale the first time a command was added.--label-by-cwd <rules.json>onfrom-claude-code, andlabelForCwdin@trazum/core.--label-from-projectlabels by the transcript's own folder, which says nothing when one folder held two projects — somebody who moved between repositories without starting a new session has one transcript and no field in it saying which work was which. This repository's own bill was the example. The transcript carries acwdper line, so the rules file maps a path prefix to a label the operator wrote, longest prefix wins, and a directory no rule covers is left unattributed rather than guessed. A malformed rule is refused by number rather than skipped, since a rule that quietly labelled nothing is the silent version of the guessed label.scripts/build-wiki.mjsand thewiki/pages it generates. GitHub indexes a wiki separately from code, so somebody searching for "cache TTL" reaches a page rather than line 812 of a 1,200-line README. No page is written by hand: each one is a section ofREADME.mdcopied verbatim, with relative links rewritten to absolute blob URLs and images to raw ones, because a wiki is served from/wiki/and a repository-relative path is broken there.wiki.test.jsruns the script with--check. The script never pushes — a wiki is a second git repository with its own credentials.scripts/prepare-commit-msg. Adds theSigned-off-bytrailer from the identity git is already recording as the author, after a missing one stranded a branch mid-review.signoff-hook.test.jsreads the required regex out of the workflow rather than restating it, and runs the hook against a throwaway repository so the test reads the hook and not the machine it runs on.
Security#
next16.3.2 to 16.3.3, which closes two critical unauthenticated remote code execution advisories: one on Windows-hosted servers (GHSA-p293-qw3h-jr36) and one in the Image Optimization API when AVIF files are used (GHSA-2xp9-vwfh-vxw4). Only the second reachesapps/web, which is deployed on Linux and serves/opengraph-imageand/icon.svgthrough Next. Dependabot grouped it underroutineand its title says nothing about any of this, which is the reason this entry exists: a security fix that arrives looking like a version bump is one that waits behind the other version bumps. Carried in with the rest of that group —@types/node26.4.0,lucide-react1.35.0,posthog-js1.422.5,@types/react-dom19.2.5 — because splitting a resolved lockfile to isolate one entry is a worse risk than the four routine bumps beside it.
Fixed#
- Two test suites were measuring the calendar. Both went red at midnight on
the 1st of a month, on a dependency bump that touched neither, after being
green all month; nothing was wrong with the product.
serve.test.jscompared a window against a literal August instant while its fixture moved with the calendar — in a file already carrying a helper written to prevent exactly that, which is worse than carrying no helper, because the file then reads as a problem already dealt with.position-tool.test.jspinned eight days in August and asserted forty dollars on the month scope. Both fixtures are now relative to the month the suite runs in. - The agent skill described the
labelsconfig block as mapping a raw label to a workload name. It does not: it maps a usage-log label to the prompt file that label sends, and the schema validates every value as a file path. An agent following the old sentence sent somebody to write a mapping that fails to load. The row now says what the block is for, and the prose says what to do instead — choose the label at the source with--label, one transcript at a time. - The packaged Action pin in
README.mdadvanced to 2.3.0's own release commit, from a 2.2.0 commit, in all three places it appears. - Both CodeQL entry points advanced to v4.37.9 in one commit. Dependabot
raised
initandanalyzeas two pull requests, and the workflow's own comment says neither may be merged alone:analyzerefuses a configuration fileinitdid not write, so a mismatched pair turns the security job red with an error naming a version nobody chose.security.test.jsalready held that rule and either pull request would have failed it, which is the guard working rather than an obstacle to route around.
Added#
scripts/assert-commit-identity.sh, run by the web session hook. The platform re-asserts its own bot identity in global git config on every session start, so the first commit after any re-clone was authored "Claude" with noSigned-off-by— refused by the owner's own rule and then by the required DCO check. Twice in one day, fixed by hand both times. The script derives the identity instead of hardcoding one: the email is the account that opened the session (so a contributor's fork session asserts the contributor, never this repository's owner), and the name is whatever that email already calls itself here — as an author, or in aSigned-off-bytrailer, which is often all a squash-merge main preserves. Written repo-locally, the one scope the platform's re-assertion does not touch, and it installs the sign-off hook while it is there. A session with no human account changes nothing: the bot identity is then the true one.commit-identity.test.jsholds all of it — and its own first guard was satisfied by the hook's comment mentioning the script after the call was deleted, found by planting exactly that deletion; it matches the invocation now.
Changed#
changelog-coverage.test.js. Fails when commits exist after the newest tag and this section is empty. Derived fromgit describeand this file rather than from a list of pull requests somebody keeps in step, for the reason every other guard here is: a list nothing binds to what it describes is the defect this repository keeps finding in itself.calendar-fixtures.test.js. Fails any test file that both builds an instant from the clock and asserts an absolute one, since those are two claims about the same clock that cannot both stay true. It strips comments first, or the bug report reads as the bug, and it exempts one file — its own, by resolving its own path rather than by an entry in a list — because it has to hold the planted violation that proves it fires.