Skip to content
Trazum

Trazum

Changelog

What changed, release by release, newest first. Each release has an address of its own.

Versioning policy: VERSIONING.md. Nothing breaks inside a major line, and 2.0.0 broke nothing either. A patch here can add a feature, and the policy says why. Below 1.0, minor versions could contain breaking changes, and said so in their first line.

What is merged into main but not yet tagged is in changes/, one file per change, and a release folds it in under its version. A change that alters nothing installable — a test, a document — still gets a file rather than nothing: the changelog is the record of what happened to this repository, and a merged commit with no entry is a change only git log remembers.

Every release149 releases

The 3 newest releases are drawn in full below. Every release has its own page.

  1. Unreleased
  2. 3.0.1
  3. 3.0.0
  4. 2.7.12
  5. 2.7.11
  6. 2.7.10
  7. 2.7.9
  8. 2.7.8
  9. 2.7.7
  10. 2.7.6
  11. 2.7.5
  12. 2.7.4
  13. 2.7.3
  14. 2.7.2
  15. 2.7.1
  16. 2.7.0
  17. 2.6.0
  18. 2.5.2
  19. 2.5.1A bill for a period
  20. 2.5.0Each request at its own size
  21. 2.4.3Each call at its own price
  22. 2.4.2The bill as an attachment
  23. 2.4.1The same door from the browser
  24. 2.4.0One door, and where the spend is
  25. 2.3.0The warning that cried wolf, and a skill written for one agent
  26. 2.2.1
  27. 2.2.0
  28. 2.1.0
  29. 2.0.0
  30. 1.86.0
  31. 1.85.0
  32. 1.83.0A receipt, and the surfaces that could not disagree
  33. 1.82.0The band was a measurement of its own training set
  34. 1.81.0The things nobody had checked
  35. 1.80.1The capital letter the grant keeps
  36. 1.80.0The doors somebody else walks through
  37. 1.79.0The dash the sweep left behind
  38. 1.78.0A sentence no locale could reach
  39. 1.77.1The folder name stands
  40. 1.77.0The agent's bill, told honestly
  41. 1.76.0The tour that does the work
  42. 1.75.0The readable terminal
  43. 1.74.0Any model's money
  44. 1.73.1The result follows the scenario
  45. 1.73.0The guided tour
  46. 1.72.0The playground
  47. 1.71.2The README the npm page never showed
  48. 1.71.1The help, in the language it was asked in
  49. 1.71.0The universal cost lens
  50. 1.70.0One drag
  51. 1.69.0The agent's own bill
  52. 1.68.0The browser catches up
  53. 1.67.1Ready to travel
  54. 1.67.0The month ends on a measured position
  55. 1.66.0One policy, three doors
  56. 1.65.0The format anyone can adopt
  57. 1.64.0The report somebody forwards
  58. 1.63.0Scale is measured, not assumed
  59. 1.62.1This machine, measured
  60. 1.62.0Held to its own standard
  61. 1.61.2An input nobody had tried
  62. 1.61.1Nothing was holding these
  63. 1.61.0A prompt this tool cannot improve
  64. 1.60.4You describe it, it asks
  65. 1.60.3A document nobody lists
  66. 1.60.2Checked by running
  67. 1.60.1What else does this fail on
  68. 1.60.0Our own medicine, measured
  69. 1.59.0A language needs a maintainer
  70. 1.56.2What this project was claiming about itself
  71. 1.56.1What the rules actually do
  72. 1.56.0Something that runs
  73. 1.55.0More than one machine
  74. 1.53.4What it says and what it does
  75. 1.53.3Two surfaces, two formats
  76. 1.53.2What the tool says about itself
  77. 1.53.1The band stays inside the family it was measured in
  78. 1.53.0Four of the seven, and why the other three are not here
  79. 1.52.1Two more providers, from facts already here
  80. 1.52.0The gateway in a real path
  81. 1.51.2The stream, and fourteen things nothing was checking
  82. 1.51.1A front door
  83. 1.51.0The record, and the minor
  84. 1.50.11The commitment
  85. 1.50.10Whose money
  86. 1.50.9The semantic pass
  87. 1.50.8The quality gate
  88. 1.50.7The experiment
  89. 1.50.6The ladder
  90. 1.50.5Cost per outcome
  91. 1.50.4The outcome
  92. 1.50.3The gateway
  93. 1.50.2The feedback loop
  94. 1.50.1The numbering
  95. 1.50.0The standard
  96. 1.49.0The live budget
  97. 1.48.0The cost review
  98. 1.47.0The browser sees the bill
  99. 1.46.0Five minutes
  100. 1.45.0The agent's budget
  101. 1.44.0The answer in milliseconds
  102. 1.43.0The watch
  103. 1.42.0The store
  104. 1.41.0The connector
  105. 1.40.0The long run
  106. 1.39.0Did it work?
  107. 1.38.0The plan
  108. 1.37.0The fleet
  109. 1.36.0The estimate stops guessing
  110. 1.35.0The reader who is not in the terminal
  111. 1.34.0Findings as policy
  112. 1.33.0The log it could not read yet
  113. 1.32.0The routing decision, priced whole
  114. 1.31.0The gate that explains itself
  115. 1.30.0The report as a diff
  116. 1.29.0The budget, the overlay and the small log
  117. 1.28.0The retry bill, the series and the standing word
  118. 1.27.0The ceiling, the drift and the tab in step
  119. 1.26.0The release that releases itself
  120. 1.25.0The retry, the archive and the shape in the tab
  121. 1.24.0How big, how uneven, and the day it spiked
  122. 1.23.0What if it were the other model?
  123. 1.22.0The gate, the window and the spreadsheet
  124. 1.21.0What the log does not say
  125. 1.20.0When, and what you can build on
  126. 1.19.0Which workload, and at what rate
  127. 1.18.0The bill, where the decisions are made
  128. 1.17.0What the report cannot see
  129. 1.16.0The worst case, on the record
  130. 1.15.0The same answer on every surface
  131. 1.14.0Drill-downs and drive-bys
  132. 1.13.0The bill learns to say no
  133. 1.12.0The log gets a clock
  134. 1.11.0What actually moves the bill
  135. 1.10.0Every hard edge, both sides
  136. 1.9.1The preflight
  137. 1.9.0The error band, measured
  138. 1.8.0No date in the file
  139. 1.0.0No date in the file
  140. 0.11.0No date in the file
  141. 0.10.0No date in the file
  142. 0.9.0No date in the file
  143. 0.8.0No date in the file
  144. 0.7.0No date in the file
  145. 0.6.0No date in the file
  146. 0.5.0No date in the file
  147. 0.4.0No date in the file
  148. 0.3.0No date in the file
  149. 0.2.0No date in the file
  150. 0.1.0No date in the file

Unreleased

Merged into main and not yet tagged. These are not in any release you can install.

Added#

  • bill says what a Bedrock or Vertex model id is when it leaves the id unpriced. us.anthropic.claude-sonnet-4-6, anthropic.claude-haiku-4-5-20251001-v1:0 and claude-sonnet-4-5@20250929 are named with their platform and, when the catalogue knows it, the Claude model inside. They stay unpriced, because each platform sets its own prices, and the hint says how to price them with an overlay. (#858)
  • trazum model-pricing reads the contracted rates your organisation already gave Claude Code in its modelPricing managed setting and writes them as a pricing overlay, so bill and every command that takes --pricing show the figure Claude Code shows. What the table cannot say (batch, fast mode) keeps the catalogue's figure and is named per model; a key it cannot place, a row Claude Code would drop and a rate of zero are named and priced nowhere. A receipt line priced by any overlay now carries the overlay's date rather than the catalogue's. (#856)
  • bill --prom-out writes the bill in the Prometheus text format, and a Grafana dashboard reads it. (#844) Teams that already run Grafana want the spend in the same place. The file is the --otlp-out payload written as text, sample for point, for the node exporter's textfile collector: Trazum pushes nothing. Each closed UTC day is its own series, labelled trazum_day, with no timestamp, because Prometheus refuses a scraped sample stamped a day back. Every cost sample keeps trazum_cost_basis="api-equivalent", what could not be priced is its own metric, and what a day still being measured held is a gauge of its own. docs/grafana/trazum-bill.json charts it, and says on every panel with money that it is API-equivalent. bill-prom.test.js holds that the samples add to the bill, that each matches its OTLP point, that a project name with a quote and a backslash survives, and that the dashboard queries no metric or label the file does not write.
  • Staging and production for the web app, a next channel on npm, and the three GitHub Environments stated in the repository. (#829)
    • The web app. Every merge to main is built by Vercel as a staged production deployment: "Auto-assign Custom Production Domains" is off on the Vercel project, so the build serves its own address until it is promoted, and promotion does not rebuild it. .github/workflows/promote-web.yml is the button: started by hand from main, under the production environment and its reviewer, it runs scripts/promote-web.mjs, which reads the deployment from Vercel's API and refuses anything but a ready, staged production build of the project (a preview would be rebuilt with production variables), then promotes it and waits until Vercel says the job succeeded. rollback puts production back on a build that served it before. After a rollback Vercel turns automatic assignment off, and the next promotion turns it back on; the job warns when that happens. The Vercel token, project id and team id are secrets of the production environment and nowhere in the repository; the token goes to api.vercel.com and to no other host. docs/setup.md, "Staging and production", has the flow, why a staged deployment and not a custom environment, how to roll back, and the Vercel pages it rests on, each with the date it was read.
    • npm. A version X.Y.Z-next.N publishes to the next dist-tag and never to latest. scripts/npm-channel.mjs decides the channel in release.yml's decide job and refuses a prerelease on main, a tag that does not name the manifests' version, and any other prerelease word. Every upload passes --tag from that answer; npm-serves.mjs waits on the same tag and fails at once if latest names a prerelease; the GitHub release, the attached files and the MCP registry listing run for latest only. prepare-release.mjs 3.1.0-next.1 moves the lockstep versions and nothing else, and the tests that hold a document to the current version read the release on latest on a next's tree (recordedVersion). docs/releasing.md, "A next prerelease", is the procedure: a branch, a tag, never a merge.
    • Environments. .github/environments.json states staging, production and release, each with a required reviewer, the branches that may use it and the names of its secrets. npm run environments compares GitHub with it and npm run environments -- apply --yes writes it, the way npm run ruleset does for the branch; the weekly Repository settings run compares and never writes. No job runs under staging yet: Vercel's Git integration builds it.
    • Tests. npm-channel.test.js, web-promotion.test.js and environments.test.js. Planted one at a time and put back, each caught by the test named: listed on the pull request.
  • The site publishes llms.txt and llms-full.txt, and a page for people building on the JSON. (#827) scripts/build-docs-site.mjs writes both through scripts/llms-txt.mjs into apps/web/public, from the documents /docs already publishes and in the manifest's order, and, like the rest of what that script writes, they are not committed. llms.txt follows llmstxt.org (read on 2026-11-19): an H1, a blockquote summary, H2 sections of markdown links with a one-line note each, and Optional last. llms-full.txt is the same documents, unchanged, in one file. trazum.config.json ignores the two files, which are copies of documents and not prompts. Nothing in either is typed: the summary, links and descriptions are read from the documents. docs/building-on-the-json.md, linked from docs/README.md and the README, links each command's contract in docs/json-output.md, says the contracts are versioned and held by tests, says what is stable and what is not, and says nothing uploads usage. llms-txt.test.js holds what the builder makes (twice, to the same bytes) to the documents and the page to json-output.md's sections. Planted and caught, results on the pull request: a builder that emits an extra line, a document deleted but still in the manifest and linked, and a link from the new page to a section of json-output.md that does not exist.
  • trazum bill --agent <name> bills one coding agent's records, and --timezone <zone> cuts days in a zone instead of UTC. (#826) --agent claude-code|codex|gemini-cli (or the name the bill prints, in any case) keeps that agent's records in the statement, the receipt, the side files and --otlp-out; a plain log or a provider export belongs to no agent and is left out, said on stderr with the count. An unknown agent is refused by name with the three known ones, and so is a known one with no record among what was read, which exits 1 like every refusal in bill. --timezone Europe/Madrid cuts --by day, --by week, --by month, the automatic month table, the span line and what a bare --since/--until day means in that zone, through Intl.DateTimeFormat with no new dependency; a day the clocks made 23 or 25 hours long is billed whole. The statement says which calendar it is on. An invalid zone is refused by name before any file is read. --timezone is refused, by flag, beside --otlp-out, --markdown-out, --html-out, --pdf-out and a --csv-shape other than slice, which cut their days in UTC, and both flags are refused beside --git-range; --receipt, -o and --send combine, because a receipt holds instants and no day. Without either flag the output is unchanged, held byte for byte by bill-agent-timezone.test.js for --by day, week and month, and the receipt gains no field, so docs/json-output.md has nothing to add. Documented in docs/commands.md, under bill; messages and help in English and Spanish. Planted one at a time and put back, each caught by a named test in that file: the agent filter ignored (in the statement, and in the OTLP file), the zone ignored, an unknown agent accepted, an invalid zone accepted, a zone line printed with no flag, a bare --until day ended 24 hours after it began instead of at the next local midnight, --otlp-out let through beside --timezone, Codex import evidence leaking onto another agent's receipt, a bare --since day read in UTC under --timezone, and --agent accepted beside --git-range. The results are recorded on the pull request.
  • trazum bill --git-range <range> --delivery sets the spend recorded in a repository beside what local git shows landed in the same range. (#825) It adds two lines at the end of the --git-range view and changes none above them: the commits of the range (merge commits counted and named), the pull requests git names by number, the UTC days on which a commit was committed, and the figure of the sessions inside the span divided by each. Pull requests come from local git alone, by the number in Merge pull request #N from ... on a merge commit or in ... (#N) closing the subject of a squash commit, counted once per number; a pull request merged by rebase names none, so the count is a floor and the line says so, and with none named there is no per-pull-request figure and never a $0. Only metadata is read: the log is asked for the parents, the commit time and the subject, the subject is read for a number and dropped, and no author name, email, message, branch or hash is read out or printed, and no pull request number is printed. The wording rule of #777 holds: the figures are what was recorded during a window divided by counts over the same window, and the line says they are not what any commit, pull request or day cost; nothing says a change caused the spend. One flag rather than a --by commit|pr|day, because --by already cuts a bill into days, weeks and months and a row per commit or pull request would be a figure for one change, which the transcripts cannot give. --delivery without --git-range is refused by name. Without the flag the output is unchanged, held byte for byte by bill-git-delivery.test.js against a statement written by hand, and checked once more by running the CLI built from origin/main and this one on one made-up repository: stdout, stderr and exit code identical in 10 runs (five invocations, both languages). Documented in docs/commands.md, under bill; help and messages in English and Spanish; the machine-readable output is not touched, so docs/json-output.md has nothing to add. Tests are in packages/cli/test/bill-git-delivery.test.js, on made-up repositories in temporary directories, with the money worked out by hand ($2 per million input tokens on Claude Sonnet 5.5: $4.00 over 8 commits, 2 pull requests and 2 days). Planted one at a time and put back, each caught by the test named: the author name asked of git (asks git for the parents, the commit time and the subject, and nothing else) or printed (names no author, message, branch, hash or pull request number), a line saying a commit cost the amount or caused it, in English and in Spanish (says what the figures are, and that none of them is what a commit cost), the delivery lines printed without the flag (adds lines and changes none..., prints, without the flag, exactly what it printed before the flag existed), a pull request counted once per commit (reads a pull request number from the two shapes GitHub writes, and only those), a number read from the middle of a subject (the same test), the merge shape read on a one-parent commit (the same test), a zero count divided to $0 (names no figure per pull request when none is named..., divides, and gives no figure where the count is zero), days cut by author time (counts a day by when the commit landed, not when it was written) and --delivery accepted without --git-range (is refused without --git-range, by name). The results are recorded on the pull request. Not measured: how long the second git log takes on a very large range.
  • trazum gateway --metrics-port <n> serves what the gateway counted, for Prometheus to scrape, and a Grafana dashboard reads it. (#824) GET /metrics on a second loopback port, off unless given: calls measured, the provider's tokens by kind, their cost at the catalogue's list rates with trazum_cost_basis="api-equivalent", calls with no list price, calls forwarded and not measured by cause, 402s by the reason the refusal body gave, substitutions, free paths and an unreachable provider, all since the process started. The cost is priced by the same function the live budget charges with, priceGatewayUsage, taken out of GatewaySpend.settle for it. A second port, not a path on the proxy's, because route is the one reader of a caller's URL there; the scrape port has no upstream, binds to 127.0.0.1 through listenGateway, and refuses a rebound page through refuseHost. Counts and names only: the labels are the provider, the model, metadata.trazum_label, a reason and a cause, and no session, prompt or response; nothing is pushed. A model and a label are the caller's text, so past 1,000 distinct pairs a call naming a new one is left out whole and counted in trazum_gateway_calls_dropped_total. docs/grafana/trazum-gateway.json charts it. Documented in docs/gateway.md and docs/commands.md; help and messages in English and Spanish. gateway-metrics.test.js drives calls through the proxy against a stub provider and reads the scrape; a guard in security.test.js holds the port to forwarding nothing, guarding the host first, recording no session, and listening on loopback.
  • The Claude Code status line says how much of your plan's window is used and when it resets. (#823) On a subscription the question is when you run out, not what a call cost. Claude Code hands the status line rate_limits on a Pro, Max or Ultra plan, or behind a gateway that sets a spend limit, and the line now shows it beside the session's figure: 5h 23%, resets 14:20 · 7d 41%, resets Sat 09:00, and spend 62% of $500/month, resets Mar 1 where a spend limit is set. Only what Claude Code sent is shown: a window without its percentage or its reset, or whose reset has passed, is left out, and a percentage is rounded down, so a window at 99.6% is not shown as used up. On an API key, and before a session's first response, Claude Code sends none and the line shows no plan figure: Trazum keeps no table of plan limits and works out no counter from tokens. Field names and units are from Claude Code's status line documentation, read on 2026-11-20. The status line suite runs the script on each case; planting a window whose reset has passed, a window without a percentage, a percentage rounded up, or the separator printed with nothing after it each fails it.
  • trazum waste points out the waste in a coding-agent session, priced, and changes nothing. (#822) trazum waste <session>.jsonl reads a Claude Code transcript and the transcripts of its subagents and names three costs: a file read again (the same path and range as an earlier read in the same conversation, no edit of it between, not an errored read), MCP servers loaded and never called, and a CLAUDE.md carried on every turn. It is a command of its own rather than a section of bill, because its answer is a short list of named costs each with the reason it is or is not priced, and bill answers whose spend a slice is. Prices are at the bundled, dated catalogue as API-equivalent cost, through profileUsage; it loads no config and refuses --pricing, --pricing-live and --config. A saving is computed from the session's own counts and never a percentage taken from elsewhere: a repeated read is priced by the prompt growth across the call that made it (P(next) - P(this) - output(this) tokens), carried on every later call of that conversation until the prompt shrinks (a compaction), as the difference between the conversation priced as it was and priced without those tokens. It is a lower bound, priced only where the call made no other tool call, and the rest are counted and said to be unpriced. MCP servers are named and not priced, because a server's tool-definition tokens are not recorded in a transcript; the loaded list is read from the deferred_tools_delta and mcp_instructions_delta attachments, whose field names were checked on 2026-11-19 against a real Claude Code transcript, and from an mcp_servers list on a system line, which that transcript did not have; --mcp-servers a,b names them instead and the output says so. CLAUDE.md is the one estimated figure: its size is a stat of the file in the session's directory (or --claude-md), its tokens are bytes over the estimator's divisor of 4, and the cost is shown with the estimator's published band, 33%, either side. It recommends and edits no file, and reads counts and paths only: no message text, command, tool result or file content reaches the report, held by a test that plants a secret in each. --min-reads <n> (default 3) chooses what is listed; --json prints a document with schemaVersion and kind: "trazum-waste", held in both directions to its table in docs/json-output.md and listed in docs/format.md, whose count of emitted documents moves from twenty-five to twenty-six. Documented in docs/commands.md; messages in English and Spanish; the command count moves from 58 to 59 wherever it is stated. Tests are in packages/core/test/session-waste.test.js and packages/cli/test/waste.test.js, on made-up transcripts in temporary directories with the money worked out by hand (two repeated reads of 3,000 tokens each on Claude Opus 5 are $0.042; a 4,000-byte CLAUDE.md over six calls is $0.0145). Planted one at a time and put back, each caught by the test named: a write between two reads ignored (does not count a read that follows a write to the same file), the read range ignored (does not count a different range of the same file as the same read), an errored read counted (ignores a read the tool reported as an error), every conversation merged (keeps every conversation apart...), no reset at a compaction and the output tokens not subtracted (prices the context a repeat added and carried...), a repeat priced although other tool calls shared its growth (counts a repeat made beside other tool calls and does not price it), a server given a price (does not price it: the definitions are not in the transcript), the band's sign reversed (prices the tokens a size implies...), --min-reads ignored (lists only files read at least --min-reads times...), a removed tool still loaded (forgets a tool the transcript later says was removed), an unreadable call counted as free (says it cannot tell when a call could not be read as usage...), the working directory copied into the report (carries nothing of what was said, run or returned into the report), the CLAUDE.md content opened or a file written (opens no file for its content except the transcripts), a pricing overlay accepted (refuses a pricing overlay, a missing path...) and a documented JSON field removed (documents every field in docs/json-output.md, and emits no other). Not planted: the check that the transcript pairs with its usage line for line, which has no input that can break it short of a defect in the converter. Not measured: how long it takes on a very large transcript, and what share of real sessions' repeats are priced rather than counted.
  • trazum session-guard is a Claude Code hook that warns once, then stops a session, at two dollar figures you set. (#821) trazum session-guard install --warn 5 --stop 15 --on-cannot-tell fail-open writes one hook to ~/.claude/settings.json, on the UserPromptSubmit and PreToolUse events. Each time it runs it reads the session's transcript and its subagents' transcripts, converts them as from-claude-code does, and prices them at the bundled, dated catalogue. Under --warn it is silent. At or over --warn it says so once per session, as {"systemMessage": ...} at exit 0. At or over --stop it blocks every time, with exit code 2 and the reason on standard error. Both are the forms Claude Code's hooks page (code.claude.com/docs/en/hooks, read 2026-11-19) documents, and the page and date are in a comment in session-guard.ts. Nothing has a default: a missing --warn, --stop or --on-cannot-tell is refused by name, and so is a --warn that is not below --stop. A transcript that cannot be read, a missing one, and a model the catalogue has no price for follow --on-cannot-tell, with the meaning it has on gateway, and never count as $0; calls that were priced count as a floor, so an unpriced model cannot hide a stop. Every figure is worded as API-equivalent cost, not what was paid, in English and Spanish. It reuses proactive's door: the same write function, backup, JSON-as-data reading and record in ~/.trazum/proactive.json, so uninstall removes exactly its own hook and the containers it made, and a hook of yours that mentions session-guard is never taken for its own. "The warning was given" is a file in the status line's cache directory and nowhere else. It loads no config and refuses --pricing and --pricing-live; nothing leaves the machine. Not available on Windows. The hook reads the whole transcript on every prompt and tool call, and how long that takes has not been measured. Documented in docs/commands.md; the command count in the README, the plugin page, docs/licensing.md and the web app's sample text moves from 57 to 58. Tests are in packages/cli/test/session-guard.test.js, on made-up transcripts with the money worked out by hand ($25 per million output tokens on Claude Opus 5, so 100,000 tokens is $2.50) and temporary homes only. Planted one at a time and put back, each caught by the test named: no warning given (warns between the figures..., warns at exactly the warning figure), warning repeated (gives the warning once per session...), no stop (blocks at exactly the stop figure...), stop below the threshold (does not block just under the stop figure), an unpriced model priced at $0 (never prices a model with no price at zero...), and the install touching another hook (adds the hook to both events and touches no other hook...).
  • A weekly cross-check of Google, DeepSeek, Moonshot, xAI and Mistral against LiteLLM's model list, as a signal to re-read a provider's page. (#820) scripts/price-crosscheck.mjs, run by the same scheduled CI job as price-drift.mjs, reads LiteLLM's public model_prices_and_context_window.json and reports, for the providers price-drift.mjs cannot read: a LiteLLM chat model the catalogue does not price (missing), an input, output or cache-read rate that differs from the catalogue's (differs, with the catalogue's source page and read date beside it), and a file with nothing for a provider (unread). LiteLLM is a signal, not a source: nothing is copied and the script never writes pricing.ts or any file. A LiteLLM rate that equals the row's base, its rate in force today or any of its tiers (DeepSeek's peak hours, a long prompt) is agreement, reported as a note naming the tier, and a row the catalogue has retired is a note and is not compared. Ids are matched exactly after taking off LiteLLM's provider prefix, never by resemblance; a catalogue model LiteLLM does not list is a note. On the weekly schedule a differing rate or an unread file fails the run and a missing model only reports; on other runs it never fails. CONFIRMED_DIFFERENCES records a row a person re-read and found right. raw.githubusercontent.com is now decided in trusted-hosts.test.js as public data, no credential. Documented in docs/releasing.md and docs/pricing-dataset.md. price-crosscheck.test.js runs a fixture LiteLLM file (a missing model, a differing rate, a matching rate, a provider out of scope) against a stand-in catalogue; the plants (a missing model not reported, a differing rate not reported, an equal rate reported, a write to pricing.ts, an out-of-scope provider reported, tiers ignored, retired ignored) are recorded on the pull request.
  • A plan can carry an edit to a tracked file, and the Action can open the pull request that makes it. (#796) plan wrote what to do and verify said whether it worked; for a change that is an edit to a prompt, the tool already produced the new text and stopped at printing it.
    • The action. trazum plan <log> --edits adds an edit-file action for each prompt file labels in trazum.config.json maps a workload to: the text the safe rules make of it (--reorder first moves stable blocks ahead of the first placeholder), recorded by the SHA-256 of the file before and of the text after, with the steps, the level and the rules left off. The text itself is not in the plan, so a hand-edited plan has nowhere to put other text; applyPlanEdit in the core is the only way to the new text, and it refuses a file that is not the one the plan saw and a result that is not the one it promised. Only a file git tracks gets an edit (isTracked in git.ts, the one module that runs git), never a source file, and a mapped label that gets none is named with the reason. An edit projects no saving: its figure is the workload's bill as the log measured it, and the file's tokens before and after are labelled as the estimator's. --min-usd keeps edits, which carry no money to be under. Nothing is written to the file.
    • Verify. verify hashes each edited file where it is now. Arrived needs both the promised text in the file and the workload's input tokens per call below the plan's baseline in the newer log. The file still as the plan found it did not arrive; a file holding anything else, or gone, is cannot-tell: file-changed and fails --gate; where nothing can read the file (the browser tab) it is cannot-tell: file-not-read and fails nothing.
    • The document. PlanDocument gains the edit-file kind, detail.edit, the assumption file-is-what-label-sends and the two cannot-tell reasons, and stays schemaVersion: 1: additions only, and a plan written before them parses and verifies as it did. parsePlanDocument refuses an edit-file action whose edit is missing, names a path that is absolute, climbs with .., has an empty or . segment, a backslash or a control character, or names .git, or carries a level other than safe. The HTTP savings API and the MCP savings tools keep refusing an edit-file action: they take text, not a repository, and could only answer that the file cannot be read.
    • The Action. A new input, plan-pull-request, off unless 'true', makes the plan from usage-log and runs one new step, Open the plan's pull request (action/open-plan-pr.mjs), the only step of the Action that writes to a repository. Through the Git Data API it reads each file from the default branch, makes the edit again there, writes one commit on a new branch trazum/plan-<hash of the edits> and opens one pull request whose body is the plan. It refuses the whole pull request when one file moved on; creates nothing when the plan has no edit or every edit is already on the default branch; reuses the open pull request for the same edits, does not reopen a closed one, and leaves alone a branch of that name with no pull request; creates no ref but its own, calls no merge endpoint, and sets no label, reviewer or assignee. It never fails the build for not opening. docs/ci.md shows it in a job of its own with contents: write and pull-requests: write; no other step gains a permission or the token. The mode refuses target, weekly-summary and every gate and window beside it.
    • Tests. packages/core/test/plan-edit.test.js, packages/cli/test/plan-edits.test.js (a real git repository in a temporary directory) and action/test/open-plan-pr.test.mjs, which drives the step against an in-memory stand-in for the GitHub API with no socket. Planted one at a time and put back, 20 caught and 0 escaped: a file the plan did not list added to the commit; text other than the promised text committed; plan --edits writing the file; a pull request opened with no change (both guards removed); a second pull request for the same plan, once by naming the branch after the whole document and once by not looking for the existing pull request; the default branch as the ref; a file's mode dropped; Markdown that is not the plan accepted as the body; a file that moved on accepted; an edit to a source file; a path climbing out of the repository; an edit projecting a saving; verify arriving on the file without the log; verify reading a file still as planned-from as arrived; an untracked file given an edit; verify not reading the files; --min-usd dropping an edit; the step run without plan-pull-request; an edit-file action with no edit parsed.
  • trazum bill --receipt --send <url> posts the receipt to an organisation's HTTPS endpoint, and nothing else leaves. (#789) The body is the bytes --receipt writes, one POST, one attempt, no retry and nothing in the background. The key is read from TRAZUM_RECEIPT_KEY as a bearer header, is not accepted as a flag and is removed from anything printed, a response body included. The URL goes through watch --webhook's own checkWebhook (https except on loopback, no credentials in the URL), and a redirect is refused. The receipt is written before the send, so a refusal or an error leaves it in place; the run then exits 3, and stderr carries the status and the response body, never the request headers. --send without --receipt is refused by name. receipt-send.ts is added to the outbound modules SUPPORT.md lists. Documented in docs/commands.md, under bill; messages in English and Spanish. bill-send.test.js receives the bytes on a local server, and the plants (a prompt in the request, the key printed, a redirect followed, plain http to a remote host accepted) are recorded on the pull request.
  • On a pull request, the Action's comment says what the change does to each prompt's cost against the base branch. (#785) The Action gated a prompt on its token budget and on drift past a recorded baseline; it did not say that this change makes these prompts cost so much more or less a month than they do on the base.
    • The command. trazum check <target> --base <ref> reads the prompts under the target at that commit and here, and gives every prompt that differs, was added or was removed a row: tokens on both sides, input cost on both sides and the difference. The base is read from the clone with git ls-tree and git show through git.ts (new: commitOf, filesAt), never fetched; a ref the clone does not have is reported as not compared, with the fix, and never replaced by the working tree. A ref starting with a dash is refused before git runs. Paths the directory walk would skip are skipped at the base too, through the new walkIncludes beside walkPrompts in the core. A prompt marked in a source file is compared prompt by prompt. --json carries it as againstBase; --markdown-out puts the table at the top of the report.
    • What a figure is. Input tokens at the model's input price, both sides at the same prices, so a repricing cannot read as a change the pull request made. Output is left out, because a prompt edit does not set it. The model is usage.model, or the default, and the report says which. Every row names its price per million input tokens and the day that price was read: the provider's review date in the catalogue (reviewedForModels), or a local overlay's own date.
    • The volume. A month is the calls somebody stated: a new config key, volumes, gives calls a month per label, and labels (the map plan --edits reads) says which prompt file the label sends. A prompt with no stated volume is priced per call and counted as such; several labels sending one file add up, and one of them without a volume leaves the file per call. A volume for a label labels does not map, or a negative one, is refused.
    • The gate. --max-monthly-increase-usd, and the Action input max-monthly-increase-usd, off unless set: fails when the prompts with a stated volume cost more than that much more a month on the head, names the changed prompts it could not judge, and fails when the base could not be read. The Action passes github.event.pull_request.base.sha to the check step through env:, so on a pull request the table is in the same report and the same comment, one per comment-key as before; it refuses the limit off a pull request, with usage-log, with weekly-summary and with plan-pull-request. actions/checkout needs fetch-depth: 0 for the base to be in the clone, which docs/ci.md shows. Without it and without the limit, the Action does not compare, so a default shallow checkout's report is unchanged.
    • Tests. packages/cli/test/check-base.test.js (a real git repository with a base and a head, and the expected table worked out by hand in its header), action/test/base-comparison.test.mjs (runs the check step's own script with bash against a stand-in CLI, and posts the rendered table twice through post-comment.mjs) and a walkIncludes agreement test in walk-ignore.test.js. Planted one at a time and put back, 15 caught and 0 escaped: the base read from the head; a missing volume treated as one call a month; a price without its date, once in the rendered row and once in the report; walkIncludes keeping a dot directory; a volume for an unmapped label accepted; the gate never breaching; the gate breaching at the limit exactly; a gate on an unreadable base passing; the Action not handing over the base; the Action running the limit off a pull request; a dashed ref reaching git; a marked prompt taking its file's volume; a label with no volume skipped in a sum; the base not held to the walk's rules.
  • The price catalogue is published as an open dataset, with every release. (#784) scripts/pricing-dataset.mjs writes packages/core/src/pricing.ts, read from the built @trazum/core, as one JSON document per release: one row per model and rate (input, output, each band, service tier, past period and promotion, the cache and batch ratios a model states, the geography multiplier and the web search fee), each with its provider, its unit, the pages that provider's prices were read from and the date. 3.0.1's document holds 734 rows for 91 models in 223,259 bytes, at https://trazum.vercel.app/datasets/pricing/3.0.1.json, beside schema-1.json (a JSON Schema) and changelog.json. MIT, the repository's licence. docs/pricing-dataset.md is the page for whoever consumes it.
    • Sources are per provider, because the catalogue keeps them so. The pages were prose above each date in PROVIDER_REVIEWED; PROVIDER_SOURCES in pricing.ts is that prose as data, the pages read on the date (Google's at aistudio.google.com, Mistral's at its inference page and three model cards), and provider-sources.test.js fails on a page the comment above the date does not name. No model is given a source of its own. No price or other catalogue value changed.
    • Patterns from literals. The validator compiles a schema pattern only from a table of literals, one for each pattern the schema names, and refuses any other, so no regular expression is ever built from the document it checks (Codacy, PR #809).
    • What it refuses. A price of zero, below zero or not a number stops the build, naming the model; so do an empty value anywhere but a batch ratio (the catalogue's null for "no batch API"), a provider with no date or no page, and a field of the catalogue the script does not know how to write. A ratio a model does not state gets no row: the catalogue's defaults are published once, in defaults. A model the catalogue does not price has no row at all. Nothing is fetched.
    • Releases. prepare-release.mjs writes the next version's document, the changelog and the schema from the catalogue as built, and refuses to rewrite a version already published. The assets job of release.yml copies the document as trazum-pricing-<version>.json only when it is byte for byte what the catalogue at the published commit builds, and SHA256SUMS covers it. 3.0.1's document was written after 3.0.1 was cut, from a pricing.ts whose only change since that tag is PROVIDER_SOURCES, so it is on the site and not on the release, and changelog.json names no release asset for it.
    • Held by test/pricing-dataset.test.mjs: the document read back, by code that shares nothing with the builder, is MODELS field for field; every row has its provider's pages and date; no value is zero; the published document is exactly what the script builds while pricing.ts is the file it records (by SHA-256); every published version satisfies the schema and lists what changed since the one before; the changelog and the schema are what the script writes. Planted and caught, each then restored: a row without a date (in the published file, and from the builder for xAI's rows), a value that differs from the catalogue (Fable 5.2's input 8 to 9 in the published file, and a band's output written as its input by the builder), a model priced at zero (Fable 5.2's input at 0 in pricing.ts, and a row at 0 in the published file), a page named nowhere in its review comment, a provider with no pages, a catalogue field left out, a ratio filled with the default, a schema keyword the validator does not read, a hand edit to changelog.json and to a document's changes, the upload removed from release.yml, prepare-release.mjs not writing the dataset, and write rewriting a published version. One plant was not caught, and is not a fault the guard claims: mistral.ai/pricing as a Mistral page, which the comment does name (the August review began there); the guard checks that a page is named, not that it was read on the date.
  • The two workflows that act outside the repository are held by a test. (#779) workflow-guards.test.js checks pr-guide.yml (the fetched script is pinned to a full commit and SHA-256, checked strictly before it runs, never run for a fork, and the job writes only to pull requests and issues) and mirror-to-n0.yml (it runs only on a push to main or by hand, never prints or stores the URL with its token, and pushes HEAD:main alone). The mirror now does nothing unless all three of its secrets are set; it checked the host only, so a missing user or token would have failed every push to main.
  • trazum bill --git-range <range> says what was recorded in this repository's sessions while a range of commits was written. (#777) trazum bill ~/.claude/projects --git-range main..HEAD, run from the repository, takes the times of the range from git log (the earliest author time to the latest author or commit time, so a rebased commit still starts the span at the day it was written) and sums the Claude Code records of this repository's project whose timestamps fall in it. The output says what the figure is: everything recorded in this repository during that span, not the cost of the change, in English and Spanish, and nothing in it says the change caused the spend.
    • Which project is this repository's, without reading a branch or a path out of a transcript. Claude Code names a project's folder after its absolute path, every character that is not a letter or a digit written as -. That encoding cannot be undone and is not: git gives the repository's root, the same encoding is applied to it, and a record belongs to the repository when the folder it was found in has that name (the label --label-from-project already puts on every record). gitBranch and cwd stay dropped by the reader. A session started in a subdirectory or a worktree sits in a folder of another name and is not matched, a session read from a single file has no folder to match, and when nothing matches the command says so instead of printing $0. Records from Codex, Gemini CLI and logs have no such folder and are counted as left out.
    • Sessions that cross an edge are named, not added. A session whose first and last call are both inside the span is in the figure. One that began before it or went on after it is listed with its first and last call and what each side cost, and is in neither the figure nor the left-out calls. A call with no timestamp is counted out; a call to a model the catalogue does not price is named, never $0.
    • Git. The log call is commitTimes in packages/cli/src/git.ts, the one module that runs a program. It reads times only (no author name, subject or hash), refuses a range that begins with - or holds a space or control character before it reaches the argv, and tells a range git refused from a git that could not run. The flag is refused beside --since, --until, --label, --no-label-from-project, --by, --receipt, --table, --stamp, -o and every --*-out, by name.
    • The Action does not take it. The sessions are on the developer's machine and a hosted runner usually has none; docs/commands.md says so and shows the step for a runner that does. No input was added to action.yml.
    • Tests. packages/cli/test/bill-git-range.test.js builds a real repository in a temporary directory with fixed author and commit dates (one commit rebased, so the two differ) and made-up sessions, and checks the figure worked out by hand: $3.00 + $1.00 = $4.00 over 4 calls in 2 sessions, two straddling sessions at $0.20 / $2.00 and $0.40 / $2.00, 2 calls of wholly outside sessions, 1 record of another project and 1 undated call left out. Planted one at a time and put back, 10 of 10 caught: a record of a wholly outside session counted; a straddling session counted as inside; a branch name from a transcript in the output (the reader passed gitBranch on and the straddle row printed it); the author time ignored; another project's records counted; a range reaching git as an option (--output= wrote a file); the edge instants excluded; the undated call counted; a worktree folder matched by prefix; the output reworded to say the change caused the spend.
  • trazum bill --otlp-out <file> writes the measured bill as OpenTelemetry metrics. (#776) Per UTC day, project, model and agent: trazum.bill.cost (USD, asDouble), trazum.bill.tokens (by trazum.token.type) and trazum.bill.calls, with what the catalogue cannot price as its own trazum.bill.unpriced.calls and trazum.bill.unpriced.tokens, which carry no dollar figure. Every cost point carries trazum.cost.basis=api-equivalent, because on a plan the figure is not what was paid. It is written through otlp.ts under doctor --otlp-out's encoding rules (64-bit integers as strings, money as a double) and it sends nothing: shipping the file is the collector's job. No OpenTelemetry dependency. A day still being measured is not a point: today, and any day the --since/--until window cuts, are left out of every metric and named in the resource (trazum.bill.open_days, open_calls, open_cost_usd, open_tokens, open_unpriced_calls), as are calls with no clock (trazum.bill.undated_*), all present at zero when empty, so the series plus those attributes add to the bill. bill-otlp.test.js round-trips a fixture to the totals the bill's own JSON prints, to within 1e-9 dollars, and by project and model. Planted and caught, with the results recorded on the pull request: a cost point with no trazum.cost.basis, a day written closed that is still open (at the end of the day, and at either end of a window), a cost 1% off the bill, cache reads dropped from the tokens, unpriced calls not carried, money written as an integer, a dropped agent attribute, and an open day not counted in the resource. Documented in docs/commands.md, under bill; help text in English and Spanish.
  • Every published price is in a committed snapshot, and a test fails when the catalogue moves without it. (#763) scripts/price-snapshot.mjs writes packages/core/test/fixtures/price-snapshot.json from the catalogue: the 734 rows the pricing dataset publishes, the default ratios, and each provider's date and pages, one row to a line. price-snapshot.test.js compares it with the catalogue. Measured on 2026-11-20 by moving each of the 681 published prices in turn: all 681 fail it, including the 19 a sample had found no test held (OpenAI's Flex and Fast tiers, four cache-write ratios of 1, three Mistral figures). It says a price did not move by accident, never that it is right. No price was changed.

Changed#

  • Vercel no longer builds the agent's working branch, which only ever repeated main. (#849) After each merge that branch is reset to main and pushed, and Vercel built the same commit again as a preview: two of the nine deployments in the 24 hours to 2026-11-20 13:06 UTC. apps/web/vercel.json turns off deployments for claude/** and sets nothing else; pull request previews and production are unchanged.
  • A pull request adds its changelog entry as a file in changes/, and the release folds them into CHANGELOG.md. (#847) Every pull request used to edit the same ## Unreleased section of a 1.1 MB file, so two open pull requests always conflicted there, and a merge after a release once put an entry inside a published version (#666). Now each change is changes/<issue>.<kind>.md; prepare-release.mjs folds them under the new version, newest issue first, and deletes them; the site's changelog page shows them as pending. pr-readiness.mjs and npm run done require one such file and refuse a line added to CHANGELOG.md outside a release. The 27 entries that were under Unreleased moved to files verbatim: folded back, they are the same text.
  • The landing tells its whole story by scrolling, on one pinned desk with its one laptop (#837). The hero opens it, short enough that the laptop, open, already stands under it in the first screen (measured at scroll 0: 349px of its lid at 1440x900, 202px at 1280x720, all of it at 390x844 and its top 90px at 390x664); then one stage stays pinned to the end while the chapters' words scroll past beside it (under it, below 1100px wide), and each chapter changes the stage: the laptop lands and the objects arrive in front of it; bill is typed as soon as it has, and printed, and a paper receipt slides out from under the laptop's side, cut at that side while it comes out so that it never lies over the screen's text, and comes to rest lying flat beside it; each object takes its turn, once, in the order they lie (the CI gate with check and its exit status, the breaker with session-guard check warning then stopping, the clock with --agent and --timezone, the pencil with optimize, the calendar with models), lifting and turning pressed while the same screen types that command over the last output, which gives way to it (the screen is never blank between two), and prints its output, scrolling up as a terminal does; why the number holds up, on the stage: the receipt comes forward over the screen and beside each of its lines a note on paper of its own, with a leader rule from the line and shown whole, gives the page its rate was read from and the date, and the receipt is laid back down flat as the chapter ends; then bill examples/unpriced-model.jsonl, a new five-call example log, shows llama-3.3-70b-instruct named and kept out of the total, never counted as $0, the screen tagging the file as an example; what never leaves goes on the screen; and the closing, the lid's turn, closes the lid over it and ends on the command, the dashboard and GitHub. The budget check for agents is one compact band after the story, a single line on a wide window: the repository has no runnable example of spend_guard. Scrolling back walks it back. The seven objects are real buttons at every point and after the story: a press shows its output over whatever the scroll had (!important, which outranks the keyframes without stopping them), and scrolling into the next chapter lets it go. A press of an object whose chapter is not the current one also puts that object's card where the chapter's was, beside the desk or under it, a copy for the eye that assistive technology skips. The chapters are one table, apps/web/lib/story-timeline.ts, which sizes their slots and times the keyframes; which chapter is current is read by an IntersectionObserver, never a scroll listener. Every character on the screen and the receipt is the built CLI's: scripts/draw-landing-terminal.mjs runs bill, the gate, the hook (fed the input Claude Code sends), the clock, optimize, models (with the clock set to the day the newest rates were read) and the example log, then runs each again in a 60-column window (COLUMNS=60, the narrowest the CLI draws for), the transcript a screen too narrow for 80 columns shows instead, so a table is never wrapped mid-row (models, which the CLI does not narrow, is cut after its output column there, and a line says so); it writes the receipt from bill --receipt through the core's formatUsd, each line with the first of the core's PROVIDER_SOURCES for its provider beside the date. Depth within the identity: no drawn table, the floor shown only by hard shadows from one light at the top left and one shared perspective, the objects lying on it in front of the laptop and clear of its deck, the receipt lying flat beside it and unmoved when the lid closes, side faces on the objects, layers arriving at different rates, the pressed object lifting off its shadow. Without JavaScript the story is every output, the receipt and every chapter in reading order; reduced motion pins the desk and shows each chapter's last state at once, in order; a phone gives the laptop the gutters' width (its deck's front edge and its shut lid inside the window), its screen the 60-column transcript, the seven objects in one row, holds each chapter's card just under the desk with the next one right behind it, and brings a focused link below the desk. A chapter's card names its command at its prompt under the sentence, never as a small capital label over the heading; every word on the receipt is 11px or larger; its torn foot is one triangle repeated as a mask. Copy in the five marketing languages, French, German and Portuguese under the existing note. Measured at 4x CPU throttle over a wheel scroll of the whole page, three runs each on the scroll timeline and the script at 1440x900 and 390x844, after the fixes the review of recordings asked for: no long task, a layout shift of 0.0000 to four places in every run, the median frame 17 ms in every run, the 90th percentile 17 ms in eleven runs and 33 ms in one, the 99th 33 ms in eleven and 50 ms in one, the longest 67 ms in two runs. Each element's keyframes run only over its own stretch of the story, nothing tilted rises with an opacity, and the section's height and starting places are written down, because without them a fifth of the frames dropped at 1440x900 and the page showed a layout shift of 0.0203 at 390x844. The web suite's landing-story.test.mjs, the theme test and eight browser flows hold it, two of them at 1440x900 and 1280x720 for the desk in the window: the lid in the first screen, the objects and their names inside the window, the receipt never over the screen's text, each note whole and clear, the shut laptop inside the window, the objects clear of the laptop's deck, the receipt lying flat and the band after the story never under the stage. Planted one at a time and put back, each caught by the test named: a chapter whose object is not the pressed one, in the timeline (each object takes its turn once, in the order the objects lie), in the motion (while a chapter is current, its own object is lifted and pressed, and no other) and in the page (the browser flow landing: scrolling tells the whole story, "the check chapter: the object pressed"); a figure typed on the receipt (carries no figure of its own); a second laptop (the page draws one laptop and one screen); a chapter's output written in the page (the page shows every line of it, in order, once) or edited in the generated file (lib/landing-terminal.ts is what scripts/draw-landing-terminal.mjs writes and the model kept out of the total is the one the CLI named); a rate's date typed in the page (carries no figure of its own) or edited in the generated file, and a rate's page edited there (lib/landing-terminal.ts is what scripts/draw-landing-terminal.mjs writes and each line of the receipt carries the page its rate was read from and the date, both the core’s). And, for the review of recordings: the ghost total back behind the laptop (no ghost figure stands behind the stage); an output going off before the next comes on, the old blank screen (from the bill’s prompt to the end, some output is always whole on the screen); the laptop landing halfway through the arrival (the bill is typed as soon as the laptop has landed); the command as a label over a chapter's heading (no small label stands alone above a heading); the receipt at 9.5px on a phone and a 13-point torn edge (the receipt’s words are 11px or larger, and its torn edge is one tooth repeated); a calendar row as wide as 80 columns in the narrow transcript (the narrow screen’s transcript is the same commands in a 60-column window; the generator itself refuses to write one); the receipt not cut at the laptop's side while it prints (the browser flows at 1440x900 and 1280x720, "the receipt over the screen's text"); a press that leaves the chapter's card in place (the press flows on the desktop and the phone, "the card in view, the pressed object's"); and the phone's cards not held under the desk (the phone flow, "a card in view under the desk").
  • The roadmap carries eight issues from a review of what other AI cost tools do (#820 to #827). Each is our own take, held to the rules here: a LiteLLM cross-check that only tells a person to re-read a provider page, a per-session spend guard for Claude Code, priced waste findings, plan limits from the agent's own signals, Prometheus metrics, spend beside delivery from local git, bill --agent and --timezone, and llms.txt. .github/milestones.json places each in one milestone, and the open-work table and the ordered list in ROADMAP.md are redrawn by their scripts.
  • Four more counts the documents state are derived from what they count, and written in digits. (#778) The MCP server's tools, from TOOLS; the criteria npm run done checks, from CRITERIA; the browser flows, from FLOWS; and the languages of the token ground truth, from the names of its prose samples. derived-counts.test.js holds each against every page, as it already held commands, rules and samples, and refuses one spelled as a word. Writing them as digits found two that had drifted: the documents said 25 and 26 browser flows, and there are 28. The README's architecture diagram is held for its MCP line too.
  • The Action pin in docs/setup.md and docs/running.md moves to 3.0.1 (9b0b245). The 3.0.1 release PR was prepared without the 3.0.0 tag fetched, so prepare-release found no newer tag and left the pin at 2.7.12; the docs-pin guard in security.test.js caught the two releases of lag.
  • The Homebrew formula names @trazum/cli 3.0.1, with the SHA-256 59eb2a96…079a of the tarball npm serves, checked against sha256sum on that tarball.

Fixed#

  • The web app's guided tour and its phone menu are modal in fact, not only in name. (#860) The tour is a native dialog, so Tab no longer walks out onto the page under the dim and Escape gives focus back to the button that opened it; with the phone menu open, the page underneath is inert, so a screen reader no longer reads what the menu covers.
  • The triage workflow runs again: its sparse checkout carries scripts/changes.mjs. (#851) #848 made pr-readiness.mjs import that module, and the triage jobs, which check out only the scripts they run, stopped at ERR_MODULE_NOT_FOUND before labelling anything. sparse-checkout.test.js now follows the imports of every script a sparse job runs and fails on a file its checkout leaves out.
  • The landing's story clears its stage as it ends, and its laptop reads as a thin laptop. (#841) At the end of the story the stage let go with the shut laptop, the receipt and the seven objects still on it, and the closing card stopped with its foot on the rule of the band after the story (1440x900, and in a screenshot of production). Now each thing leaves once the story is done with it: the receipt after its chapter, and in a new last chapter with no words the closing card, the objects and the shut laptop, while the band after the story comes up over the emptying stage, so the stage is never a window of nothing. What has left is hidden, so it is not reached or read out; a press or focus inside the story brings it back, and with reduced motion it is gone at once. The laptop is drawn as a thin aluminium laptop, with no vendor's logo: thin even bezels, a camera, rounded corners, dark keys with a half-height function row, speaker grilles and a wide trackpad. On a wide window each chapter's words had also drifted 16px further from their moment on the stage per chapter, 176px by the end, from a gap the column kept from its still layout; it has none now. The web suite holds the clearing, and the browser flows check that each thing is gone at the end.
  • The landing's laptop is on from the first screen, and a tall window no longer leaves the stage half empty. (#839) A recording of the deployed landing at 1788x1176 showed three faults. The laptop's screen stayed blank black until the reader scrolled; the bill's prompt and its cursor are now drawn from the story's first frame, and the command stays untyped until its chapter types it. The pinned stage kept the laptop and the objects at the top of the window and left the floor under them empty; it now shares what the window leaves over above the laptop and under the objects. The seven objects came up strung down a diagonal, each taking 30% of their beat; each now takes 90%, so the row arrives nearly as one. The web suite holds the first frame: planting the blank screen back fails it.
  • Three low-severity findings from a security audit are closed. (#818) None is a demonstrated loss of a secret or of code execution.
    • trazum serve and trazum gateway no longer answer a web page that rebinds its own hostname to 127.0.0.1. Both bound to loopback and read neither Host nor Content-Type, so a page could send text/plain (no preflight) to POST /cost and read the measured spend, the budget and the policy verdicts, and the gateway's 402 body carried the standing. A request whose Host is not localhost, 127.0.0.1 or [::1] on the bound port is now a 403, and a POST that is not application/json a 415 (after routing, so an unknown path is still a 404). A client that posted without content-type: application/json now gets 415: curl -d needs -H 'content-type: application/json', which docs/commands.md now shows. Provider SDKs and the documented clients already sent it. The checks are in packages/cli/src/loopback-guard.ts; a Unix socket skips the Host check.
    • The web app's operator endpoint is no longer a free choice of model. With TRAZUM_LLM_BASE_URL set to a server that checks no key (Ollama), llm: { apiKey: "x", model: "any-model" } ran any model the operator had pulled. On that endpoint a caller's llm.model is now used only if the operator listed it in the new TRAZUM_ALLOWED_LLM_MODELS; otherwise TRAZUM_LLM_MODEL runs. A key still checks nothing on such a server, which SECURITY.md now says. The route's comment, which said the endpoint costs the operator nothing, is corrected.
    • mirror-to-n0.yml declares permissions: contents: read and checks out with persist-credentials: false. It pushes to N0 with its own secret in the URL, so the GitHub token was unused and sat in .git/config. security.test.js now fails on any workflow without a top-level permissions:.
    • Tests. loopback-guard.test.js sends the rebound request over node:http (attacker Host, Origin, text/plain) to both servers; api.test.mjs names a model on a keyless operator endpoint; security.test.js reads the workflows. Planted and put back: the Host check removed from each server, the content-type check removed from each server, the caller's model honoured again, the permissions: block removed, persist-credentials set to true; each failed the test named for it.
  • The Homebrew formula waits until npm's packages are a day old. (#816) Homebrew installs with npm --min-release-age=1, and on 2026-11-19 brew install davmunrey/trazum/trazum failed on a Mac with ETARGET: No matching version found for @trazum/core@3.0.1 with a date before ..., because the tap had moved to a version npm had published seven hours earlier. scripts/homebrew-formula.mjs now reads the publish time of the CLI and of every @trazum/ dependency it pins, and refuses to write the formula before the youngest is a day old, naming the moment it may. RELEASES.md says so in the Homebrew step.
  • A line of backticks no longer takes quadratic time to optimise. (#814) The three code-fence patterns in structure.ts opened with `{3,}, and on a line with no newline after it the engine gave the run back one backtick at a time: 11.6 s at 100,000 backticks, and one anonymous POST /api/optimize at the 400,000-character cap held the web app for 187 s. The opening run is now atomic (a lookahead and a backreference), so it is never given back, and the same 200,000-backtick input takes well under a second. security.test.js gains a backtick run and a tilde run with no newline; with the old pattern planted back, the backtick fixture took 48 s and failed.
  • A release's MCP registry check reads every page. (#806) The registry's search answers 30 entries a page and names the next one in metadata.nextCursor. The check read only the first page, so 3.0.0, the thirtieth version, passed, and 3.0.1, the thirty-first, failed a minute after it had published correctly and was listed with its npm package and its .mcpb. The check now follows the cursor until there is none. security.test.js fails when the cursor is removed from the step.
  • A web session no longer starts with package-lock.json changed. The session-start hook installed with npm install, which rewrote the lockfile in the layout of the container's npm; npm 10.9 dropped 48 libc lines a newer npm had written. It now runs npm ci, which installs the lockfile as it is. (#788)

Added#

  • Trazum can be downloaded from the project's own places, with no marketplace account. (#774, #780, #781; advances #640, which stays open because no listing exists) The owner asked on 2026-11-19 for everything needed so it can be downloaded. Six parts:
    • Release files. A new assets job of release.yml, which needs the job that publishes (so after npm actually serves every package and after the release page exists; the publish order is untouched), builds trazum-vscode-<version>.vsix with @vscode/vsce@4.1.1 and trazum-mcp-<version>.mcpb with @anthropic-ai/mcpb@2.1.2 (mcpb validate first), makes SHA256SUMS over them and the bill of materials, and attaches the three with the workflow's own token. The job holds contents: write and no npm credential, id-token or environment, because it runs two packagers fetched by npx; the step that holds the token runs none of them. No marketplace, no new secret. The workflow makes no build-provenance attestation, so none is claimed. A manual run builds the files and attaches nothing. Built once here on 2026-11-19: the .vsix is 456,522 bytes in 132 files, the .mcpb 492,983 bytes in 132 files.
    • A fault the first .vsix had. npm run package:vscode produced 8 files and 8.39 kB, and none was @trazum/core, which dist/extension.js imports: the ignore file excluded node_modules. The extension would have failed to load for want of it. scripts/release-assets.mjs now lays the extension out with the closure the lockfile names and .vscodeignore lets exactly that package through. Not run inside VS Code or Cursor: only that the built dist/reading.js resolves its import from the unpacked file.
    • The bundle. manifest.json uses only fields of the format's MANIFEST.md (spec 0.3) and claims no platform: its compatibility is the Node version the package already declares. It holds the server's built files, @trazum/core's built files, the licence and the icon, and nothing the lockfile does not name. Its server was started from the unpacked bundle under plain Node and answered initialize and tools/list. Not opened in Claude Desktop. The bundle is unsigned.
    • The registry. The committed server.json stays npm only: the digest of a file built after the release PR merged cannot be in it. The mcp-registry job now waits for assets, downloads the bundle and SHA256SUMS from the release, and scripts/mcpb-registry-entry.mjs adds the mcpb entry (address and fileSha256) to the copy it publishes, refusing when the file and the sums disagree. A failed attachment therefore holds the listing back until the job is re-run. Not run: it runs for the first time at the next release.
    • Homebrew. packaging/homebrew/trazum.rb (@trazum/cli 3.0.0, its npm tarball's SHA-256, depends_on "node" and std_npm_args) is written by scripts/homebrew-formula.mjs, which compares the tarball's sha512 with the registry's before hashing it; a test holds the file to a fixture tarball with no network. Pushing it to the tap is a manual step, named in RELEASES.md: docs/compliance/secret-rotation.md covers no token for it and #781 forbids asking for one. Not tried with brew install.
    • Gemini CLI, Codex and the other clients. gemini-extension.json starts @trazum/mcp at the exact version with the skill as its context file (fields from Gemini CLI's own extension reference); docs/agents.md gains a table and snippets for Claude Code, Claude Desktop (bundle and JSON), Cursor, VS Code, Codex (command and TOML) and Gemini CLI, written by scripts/install-snippets.mjs from packages/mcp/package.json and held to it by a test. prepare-release.mjs moves the manifest and the table with the version. Not tried with gemini extensions install.
    • The page. /download, in the header, footer and sitemap and in the site's five languages: the CLI (npx, npm install -g, the tap) at the exact version, the MCP snippets drawn from the same source as docs/agents.md, the .vsix with how to install it, the Action's pinned line read from docs/setup.md, and the checksums. Links go to releases/download/v<version>/<name> for the version the site is built at, never releases/latest; so for a short while after a release merges, and for 2.7.12 itself, whose release carries none of these files, those addresses do not answer, and the page says so. It names no store, loads nothing and sends nothing: a twenty-fifth browser flow watches the network, measures contrast in both themes and checks a phone, and took 2,145 ms on 2026-11-19 against a local production build.

Added#

  • trazum weekly writes the week down, and the Action posts it where the owner names. (#757) bill and plan answered when somebody asked, and the README's answer to "tell me weekly" was a cron line and a mailbox nobody read. trazum weekly [<file|dir>...] [--until <when>] [-o <file>] [--notify] reads what bill reads (the agents' own session folders with no path) and writes Markdown: the bill for the seven days before --until, what moved against the seven before (both measured, same catalogue, half-open windows so no call is in both), and the three changes plan ranks first. A change that has not happened is projected, not measured and is printed beside what that workload cost this week, which is measured; a loss already paid is already spent; verify is named as what measures it afterwards, and nothing is projected beyond the week. Every dollar amount and percentage in the text is read back against the two weeks' reports, worked out from the reports and not from the summary, and a summary holding one that is in neither is refused and not written. A week with no call is said to have none, a week before with no call has no comparison and no percentage, and unpriced and undated calls are counted out loud. --notify shows one line of figures (no name, no path) through the notifier watch --notify uses. The CLI holds no token and makes no request. For a repository the Action gains weekly-summary (issue or comment) and weekly-summary-issue: with usage-log it writes the summary and action/post-weekly.mjs posts it to the one place named, with the workflow's own token and issues: write. A week is posted once: a re-run replaces its own post (found by a marker naming the week, authored by a bot) and never an issue or comment a person wrote. It sets no label or assignee, refuses target and every gate in the same run, and never fails the build for not posting. Planted and caught (16): a total in the summary scaled, a movement taken the wrong way round, weeks that overlap, the figure check switched off in the core and in the command, a figure printed that is not in the bill, a projection worded as done, a notification carrying names, the CLI reaching for a token, a second issue for the same week, a comment on an issue the owner did not name, a contributor's issue overwritten, a place the owner did not name accepted, a file that is not a summary posted, a third place for the step, and a gate allowed beside the summary. Not in this change, and split into #796: the mode that makes the deterministic edits on a branch and opens one pull request whose body is the plan.

  • trazum watch --sessions judges the budgets against the sessions on this machine, with no store, and trazum watch --notify tells the person at the screen. (#756) watch judged the store, which is empty until trazum connect <provider> --store fills it, while trazum bill reads the sessions Claude Code, Codex and Gemini CLI already keep on disk. So the command that needs no setup could not alert, and the one that alerts needed the most. --sessions reads the folders bill reads with no path, through the same readers and the same catalogue, and judges spend.maxUsd, spend.maxDayUsd and spend.maxCacheLossUsd as it always has: a measured crossing and never a projection, a day still being measured not yet judgeable rather than passed. A crossing is reported once and then prints as STILL OVER, and a restart names the stretch nobody was watching, the same as for the store. The two sources keep their record apart in .trazum/watch.json (a new optional source field, written only by --sessions), so what one reported never quiets a crossing only the other shows. --notify is a desktop notification through what the platform already has (osascript, notify-send, a PowerShell toast), nothing installed and no dependency added, off unless asked for. It fires for what is new this cycle, and carries the sentence the terminal prints and nothing else: no prompt text, no file path, no project name. The program is started without a shell and the text is handed over as data, never spliced into a script. Where the machine has no notifier the run says so and the exit code, the output and --webhook carry on. What watch does with the store, a payload and --webhook is unchanged. The notifier was tested here with a stand-in on Linux; the macOS and Windows calls are tested as the arguments they build and were not run on those systems.

Removed#

  • Breaking, for 3.0.0: tier is gone from a model in the pricing catalogue. Use capability. It was deprecated in 1.8.0, the entry that deprecated it said it would go in 2.0, and 2.0.0 removed nothing. The owner decided on 2026-11-18 that it goes in 3.0.0, and nothing else goes with it. The migration, repeated here as VERSIONING.md asks:

    • A pricing overlay that writes "tier" on a model is now refused, with an error that names the key ("models.<id>.tier") and says to write "capability". It is not ignored, because a key somebody wrote and nothing reads is a setting that looks set. Delete "tier" and write "capability" in its place, mapping "haiku" to "small", "sonnet" to "mid", "opus" to "large", and "frontier" and "unknown" to themselves. A model the bundled catalogue does not have still has to be complete, and capability is the field it needs; tier is no longer asked for. The same refusal applies to an overlay handed to applyPricingOverlay as an object.
    • A consumer of the JSON reads capability where it read tier, with the same mapping. Every document that carries a model whole stops carrying tier: the models GET /api/optimize serves, the overlay --pricing-live writes for a model the catalogue does not have, and the snippet trazum ownrate prints. No other document carried it.
    • In code, ModelPricing has no tier, so model.tier and the type ModelPricing['tier'] do not compile, and PRICING_MODEL_KEYS no longer lists it. recommendTier, recommendTierDetailed, cheapestOfTier and cheapestOfTierIn keep the names they always spoke (haiku, sonnet, opus, frontier, unknown) and choose by the model's capability through the same mapping; where an annotation used ModelPricing['tier'] for those names, Parameters<typeof cheapestOfTier>[0] is the same five. The model-downgrade advisory chooses by capability too.

    No price changed and no recommendation changed. Every field of all 91 models, multipliers included, was dumped from the build before the change and after it, and the two are identical once tier is taken out of the first. Every advisory for every model, at eight prompts, three dates, with and without the Batch API and in both languages (8,736 calls), and every cheapestOfTier and cheapestOfTierIn answer by provider and date, are identical before and after. No model's tier and capability disagreed. tier-removed.test.js holds the recorded answers (the cheaper model each of the 91 is offered at three prompts, and the cheapest of each tier by provider), fails if tier comes back in the type, the catalogue or a document that carries a model, and holds the overlay's refusal; api.test.mjs in the web app calls GET /api/optimize and holds its models to the same. VERSIONING.md lists nothing as deprecated and still working, and versioning.test.js still sees its search find a @deprecated member, on the text types.ts had. The recorded answers are a file a price change can move, so scripts/record-recommendations.mjs writes it again from the build, the test holds the file to what that script writes, and CONTRIBUTING.md, "Changing prices", says to read the diff and say which answers moved. Twenty-one faults were planted and each failed a test: tier back in the type, plain and marked @deprecated as it was, on one catalogue model, and in each of the three documents; the refusal removed with the key kept, with it silently dropped, falling back to the unknown-key message, unchecked on an object, no longer naming the key, and reading a patch that is not an object; six ways of choosing other than by capability; the VERSIONING.md entry put back; and the recording script skipping a prompt, and skipping a provider. The refusal is in English only, as every error of the overlay is. (#764)

  • Breaking, for 3.0.0: recommendTier answers in small, mid and large, and the lookups that take its answer are renamed to match. (#798) #764 took tier out of a model and left the five old names (haiku, sonnet, opus, frontier, unknown) as the inputs and outputs of four functions and of the model-downgrade advisory, with a table, CAPABILITY_OF_TIER, translating them to the capability a model carries. The owner decided on 2026-11-19 that one vocabulary is spoken in 3.0.0, the model's, and that no release keeps both. A name of the old ladder was a vendor's, said about a task that may be sent to any vendor's model; recommendTier reads a prompt and has three answers, so frontier and unknown were never among its outputs. The migration, repeated here as VERSIONING.md asks:

    WasIs
    recommendTier(prompt, tokens) returns'haiku', 'sonnet', 'opus''small', 'mid', 'large'
    recommendTierDetailed(...) returns{ tier, complexSignals, simpleSignals, conflicted }{ capability, complexSignals, simpleSignals, conflicted }
    cheapestOfTier(tier)takes haiku ... unknowncheapestOfCapability(capability), takes small, mid, large, frontier, unknown
    cheapestOfTierIn(catalogue, tier, on, provider?)takes haiku ... unknowncheapestOfCapabilityIn(catalogue, capability, on, provider?), same arguments after the second
    CAPABILITY_OF_TIER, TierNameexportedremoved: nothing translates, and a capability is a Capability
    ModelDowngradeParams.tier (a locale catalogue's modelDowngrade)tier: stringcapability: string
    haiku / sonnet / opus as the name of an answerhaiku, sonnet, opussmall, mid, large; frontier and unknown are unchanged
    • A consumer of the JSON reads small, mid or large where the detail of a model-downgrade advisory quoted "haiku", "sonnet" or "opus": in trazum optimize --json, in the response of POST /api/optimize and in the text the report, the web page and the playground print, in English and Spanish. No field carried the name and none is added or removed; the detail is a sentence for display, and docs/json-output.md now says so under "The capability an advisory names". The advisory's id, severity and estimatedMonthlyUsd are unchanged.
    • The MCP server changes nothing, and neither do the GitHub Action and the VS Code extension: optimize_prompt prints an advisory's id and title and not its detail, no tool schema named a tier, and the other two never read the advisory. A client written against them needs no migration.
    • In code, a call to cheapestOfTier('sonnet') is cheapestOfCapability('mid'); an if (tier === 'opus') on what recommendTier returned is === 'large'; a custom locale catalogue reads capability in modelDowngrade. A name of the old ladder passed to cheapestOfCapability or cheapestOfCapabilityIn throws, as an unknown capability always did, and is not translated. recommendTier, recommendTierDetailed and the tier-signals-conflict advisory keep their names, which the owner decided.

    No price changed and no model chosen changed. recommendations.json was written again by scripts/record-recommendations.mjs from the build, and a script that maps the names of the file as it stood (haiku to small, sonnet to mid, opus to large, and the two keys cheapestOfTier and cheapestOfTierIn to their new names) produced the new file byte for byte: 11,695 bytes, the cheapest model of each capability by provider, and the 273 answers (91 models at three prompts) that name a model or none, all the same. docs/assets/demo.svg and the landing's terminal were drawn again and differ in the one word of one sentence. Twenty faults were planted and each failed a test: an old name emitted by optimize --json alone, by the web route alone, by the English advisory and by the Spanish one, returned beside capability, exported as an alias and accepted by the lookups; mid read as small, large read as mid, small read as large, frontier ranked level with mid, the advisory quoting another capability than it judged, offering the model of another capability, and the lookup ignoring the capability; the cheapest small model repriced, Haiku 5.5 restated as mid, one recorded model changed and one recorded name changed; and the recording script leaving large out and writing the old key. Three surfaces could not be planted in because nothing on them emits the name: the MCP server, the Action and the extension.

Added#

  • /pro: a page for Trazum Pro, which says it is in development and not open yet. Linked from the header and the footer, in the five languages the site speaks. In order: what Pro answers for an organisation (where the model spend went, whether it is inside its budget, what changed when it moved, a figure an auditor can trace); the one property it is built around, that no prompt text reaches its server and what travels is a receipt, made by trazum receipt in this repository and held by receipt-redaction.test.js, with the label named as the one field that travels as it was written; what stays free in the open tool and what Pro adds, in the words of docs/licensing.md; the tiers, by managed spend a month and not by seat (Team €149 up to €5,000, Growth €499 up to €25,000, Scale €1,499 up to €100,000, Enterprise by agreement) with the sentence that makes the share the break-even; and the status, with one call to action. That is a mailto: link to the address CODE_OF_CONDUCT.md publishes, the subject filled in, and a link to the repository's releases: no form, no field, no cookie, no request and no dependency. The tiers and the status are one module, apps/web/lib/pro.ts, whose comment names Trazum Pro's own docs/pricing.md as where the owner decided them; the page formats them and types no figure. The site gains a sitemap.xml of its pages, which it did not have, named by robots.txt, and the layout's site address moves to lib/site-origin.ts so the two read one. pro-page.test.mjs holds the page: the tiers parsed back out of the rendered table in every language and compared with the module, and the module with the owner's figures written again in the test; each share equal to its price over its ceiling; the not-open sentence and the status while PRO_OPEN is false; no claim an open service would make; the one mailto equal to the one address the repository publishes; the page in the header, the footer and the sitemap; every pair of text and ground it uses at 4.5:1 or better in both themes, worked out from globals.css; and nothing sent. A twenty-third browser flow follows the header's Pro entry from the landing, reads the four tiers written by hand, finds the one mailto, measures every text's contrast on paper and then with Dark chosen, and fails on any request that leaves the origin; the documents that count the flows say twenty-three. Six faults were planted and each failed a test: Growth printed at a price the module does not hold, the not-open sentence removed, the status above the title removed, Pro taken out of the header, a fetch to another host on load, and an image requested from another host with its address built from parts, which the browser flow caught by the request it saw. (#793)

  • Codacy and DeepSource read the repository from a configuration that is in it, and Codacy is sent the coverage. The owner installed both on 2026-11-18. .codacy.yaml and .deepsource.toml leave out the two directories a script writes (wiki/, plugin/skills/), so a finding is reported once and against the file a person edits, and .deepsource.toml says what the code is: TypeScript, ES modules, for Node, with React. Without it DeepSource's first review of a pull request reported the CLI's own printing as "console in code that runs on the browser" and every function at the top of a module as a declaration in global scope. The Coverage workflow's last step sends Codacy the reports it already writes: Codacy's reporter at a pinned version, 14.1.3, refused unless its SHA-512 is the one its publisher lists, never its install script and never an action; the token from the repository secret CODACY_PROJECT_TOKEN, read from the environment and never on a command line; nothing fetched where there is no token, as on a fork's pull request. Codecov stays where the record is kept, and the step cannot fail the job because of Codacy: no token, a failed download and a refusal each leave a line in the log. A reporter whose checksum is not the pinned one does fail it, and is not run. Neither service is a required check, and a test holds the ruleset to that. CONTRIBUTING.md, "The machines", names both and holds them to the rule already there for a review bot: a finding is traced in the code first, and a wrong one is not quieted by changing code that was right. The first reading of Codacy's 794 findings, group by group, is in the issue; none read so far is a defect in shipped code. (#767)

  • One command turns on everything that speaks without being asked, and one turns it off: trazum proactive. Trazum can speak first in three places, each off until somebody wired it by hand: what the session has cost, in Claude Code's status line, with a Stop hook that refreshes it; trazum watch, which says when measured spend crossed a budget, from a cron entry; and spend_guard over MCP, from the server entry in each client. trazum proactive reports which are on (one line each: on, off, not ready with the reason, not available, could not tell), trazum proactive on offers each one that is off and says every file and every line first, and trazum proactive off removes exactly the lines it added. Nothing is written without a yes for that item: --dry-run writes nothing, and input that is not a terminal without --yes is a dry run and says so. It edits ~/.claude/settings.json, the user's crontab, and .mcp.json or .cursor/mcp.json in the directory it runs in (the two clients whose files the documents state), and its own ~/.trazum/. Every write goes through one function that refuses a path off that list, a symbolic link and a file that changed since it was read, writes a temporary file and renames it, keeps the mode, and keeps one backup beside the file the first time it changes. A file is opened once, without following a link, and its kind and its text both come from that open. An item is written whole or not at all: if one of its writes fails, what it had already written is put back and nothing is recorded. A crontab that cannot be read again just before the write is not taken for an empty one. The status line script is this command's only while it is the copy it wrote (the record keeps its sha256): one you changed, or a file already there, is not written over and not removed. JSON is edited as data, keeping key order and indentation; a file that does not parse, or whose hooks is not an object, is left alone and named; a status line somebody else configured is never replaced; the hook is appended to Stop and never replaces what is there, and is never wired to SessionStart, whose output is handed to the model. Its own lines are recognised by the exact command written (the hook and status line), a trailing # trazum watch (managed) (the cron line) and a manifest of what it added (the MCP entry, so one pasted from the README is not removed). The status line script is copied from the CLI package into ~/.trazum/statusline/, because plugin/ is not in an npx install, and a CLI run from the npx cache is refused for the status line and the schedule: it would not be there tomorrow. A CLI file that is not executable is refused for the status line as well, because the hook starts it by its path and fails without a word, and so is a home directory with a quote in its name, which would end the quoting of the script's path. watch is scheduled only when the store has something in it and trazum.config.json has a spend limit; otherwise it is not ready, naming the command or the key, and the store is never filled nor a budget written. Sends nothing anywhere. A command of its own rather than a mode of init, because init proposes a config for one repository and its --dry-run and --yes mean that; this edits a home directory. The CLI now has 56 commands, and --json is documented in docs/json-output.md. (#755)

  • The documentation is on the site, at /docs. The header's Docs entry, the footer's and the application rail's lead there instead of to docs/README.md on GitHub, and the landing's links to the CI and agent documents lead to their pages. /docs is the index, grouped as docs/README.md groups it and opening with that file's own table of tasks; /docs/<slug> is one page per document, with a contents list built from its headings (sticky on wide screens, a disclosure on narrow ones), the document's own anchors as GitHub gives them so old links keep landing, View on GitHub for the same file, and previous and next within its group. Thirty-six documents are published: the README as an overview, the guides, references and trust documents under docs/, CONTRIBUTING.md and VERSIONING.md. Plans, the audit, the compliance records, the CLA and the policy and history files are not, and the index says they exist and links each on GitHub. The list is one file, apps/web/lib/docs/manifest.ts, and docs.test.mjs fails on a document docs/README.md links that has no page and is not left out, on a file under docs/ in neither list, and on a group docs/README.md does not agree with. Search over titles and headings runs in the page from a small index the build writes: no server, nothing sent, nothing typed put in the address. A relative link between published documents becomes a link between their pages, any other relative link an absolute GitHub link, and an anchor that does not exist stops the build; images are copied and served from the site's own origin, and an image on another host (the README's badges) is never loaded: the page shows its description as a link. The Markdown is read by a reader written for the subset the documents use, about twelve hundred lines with no dependency, which builds a tree React draws, so nothing in a document can become markup; it stops the build, naming the line, at a construct it does not know, raw HTML included, and was checked against markdown-it on all 37 files (same headings, lists, tables, code blocks, links and text). The documents stay English in every language and each page says so. Two documents had a paragraph run straight into ---, which GitHub and this reader both take as a heading of the whole paragraph; each gained the blank line it was missing, in README.md and docs/commands.md. The Dockerfile now carries the documents into its build stage and the app's public/ folder into the runtime image, which it had not. A fifteenth end-to-end flow goes from / through Docs, a search, a result, and the heading. (#735)

  • Mutation testing reaches the code that chooses a rate and multiplies by it. A target in scripts/mutate.mjs can now be part of a file, { file, functions }, found by name among the built file's top-level declarations; a name the file does not declare, or a listed function with nothing to change, stops the run. pricing.ts is mutated in the ten functions that choose a rate and give the cache and batch ratios, 212 mutants and none of the catalogue's numbers, and usage.ts in add, 79. Their first runs caught 178 of 212 and 62 of 79. Two new test files, rate-choice.test.js and bill-lines.test.js, each from a rule the code's own comments state, killed 12 and 16 of the survivors: a Sunday billed at DeepSeek's weekday peak, a promotion ended a millisecond early, a period of the price history without ratios of its own priced at the defaults, the undecided band billed at a cheaper band than the dearest, and the batch, band, cut-off and unpriced-search lines of a bill, none of which a test read. The other 23 are recorded with their reasons, so the record holds 62 survivors and 739 of 801 mutants caught, 92.3%, 5.1% of the core's 15,800. docs/mutation-testing.md also counts the catalogue's own question: 681 published prices, of which a sample of 69 was measured and 50 are held by a test. --list counts each target's mutants and runs nothing. Eight faults planted against the selection, eight caught. No price, rate or date changed (#745).

  • A brand page: the wordmark, the colours, the type and how to use them. /brand, linked from the footer, offers the wordmark and the plain T, each for a light ground and a dark one, as SVG and as PNG: eight files in apps/web/public/brand, written by scripts/draw-icon.mjs and held byte for byte by brand.test.js like the icons. The wordmark is outlines of Space Grotesk at weight 700, placed with the font's own advances, pair kerning and the header's tracking, read from the font the app ships by scripts/wordmark.mjs (the WOFF2 container, gvar, HVAR and the kern feature, with no dependency); against a browser's rendering of the same text it has the same bounding box, and 314 of 878,005 inked pixels differ, the soft edges. The PNGs are filled from the same path by a scanline rasteriser in scripts/brand-files.mjs. Every value on the page is read when the site is built: the colours from globals.css by the role each row of DESIGN.md's token table names, the typeface from where the stylesheet declares it, and the voice, in three sentences, and the list of what not to do from DESIGN.md, which gained a ## Using the brand section to hold them. Each token has a copy button that writes the clipboard and sends nothing. The page states no trademark policy: it points at the licensing page, which says there is none. The social card each page draws now sets the name as that wordmark, docs/assets/social-preview.html carries the same outlines between two markers the icon script rewrites, and brand.test.js holds both to the palette. A seventeenth end-to-end flow follows Brand from the footer, fetches every file it offers and copies a colour. (#737)

  • Models and prices, the roadmap and the changelog are pages of the site. /models is the catalogue the product bills with: every row of @trazum/core's MODELS (91 at this commit) by provider, its rates in formatUsd, cache and batch as the multiple the catalogue states, and in the row's details its bands, service tiers, earlier periods, promotion and note. Nothing is computed. A retired model, a promotion and a model with no price are each drawn in form as well as words, the seven ids providers list and the catalogue leaves unpriced are listed with their reasons and no figure, and each provider shows the date it was last read and the page the catalogue's source names. A provider and a text filter run in the page and send nothing. /roadmap is .github/milestones.json in order, with issues, the live milestone's address and no date, and ROADMAP.md's ## Under consideration. /changelog lists all 147 releases newest first, one anchor each, draws the unreleased changes and the newest three, and gives every release /changelog/<version>; a title in place of a date is shown as a title. Of the file's 1,047,825 bytes the index module is 109,256, and the page sent 326,008 bytes, 74,382 compressed. They are built at build time by scripts/build-site-pages.mjs, with no request to GitHub, and site-pages.test.mjs holds each to its source. The Markdown reader gained ~~strikethrough~~, and a sixteenth end-to-end flow follows Models. (#736)

  • Mutation testing reaches the code that prices a call. billing.ts is the sixth target: it says, for every call the bill and its findings price, the moment it was billed at, the size of prompt that picks its band, how it ran, and what its web searches cost. Its first run made 35 mutants and caught 28, 80.0%. Two new tests in billed-at.test.js and one in web-search.test.js killed 3 of the 7 survivors: a call with no clock could be billed at 1970-01-01 instead of the report's date, a call's one-hour cache writes could be left out of the prompt length that picks its band, and a single web search could be billed at nothing. The other 4 change nothing a caller can see and are recorded, so the file stands at 31 of 35, 88.6%, and the whole record at 471 of 510, 92.4%, with 39 survivors on record (#731).

  • Merging to main publishes the wiki. wiki/ was generated and checked, and the last step was two commands the owner had to remember, so the published wiki was only as current as the last time somebody did. scripts/publish-wiki.mjs takes a checkout of the wiki's repository (--into <dir>), copies the pages that differ and says which it wrote, which were unchanged and which exist only on the wiki. A page only on the wiki is reported and never deleted, the way a label only on GitHub is left alone. --check prints the differences, exits 1 when a page of wiki/ differs or is missing there, and writes nothing. It refuses, with exit 2, a wiki/ that scripts/build-wiki.mjs --check does not accept, a missing --into, a directory that is not a git checkout and any other argument. It commits nothing, pushes nothing, reads no environment variable and makes no network call. The new Wiki workflow does the git: on a push to main that changes wiki/, the script or the workflow, the publish job clones the wiki with the job's own token, runs the script, commits as github-actions[bot] with a message naming the commit of main it was built from, and pushes, or says nothing changed and commits nothing. It is the only job with contents: write, which GitHub's wiki needs because it has no narrower permission; the token reaches git as an Authorization header in the environment of a step, never in a URL or an argument, and its encoded form is masked. Every Monday at 10:13 UTC, and on a manual run from any branch, a read-only compare job runs --check, so a page edited in GitHub's editor, or a merge that was never published, is noticed. There is no pull request trigger. The wiki needs one page saved in the web UI before the first publish, because a wiki with no page cannot be cloned; the two commands build-wiki.mjs prints stay as the fallback. wiki/ has 9 pages, as build-wiki.mjs --check counted. The push itself was not run from the session that wrote it, which could not reach the wiki's repository, so the workflow's first run on main is the first real one. What was measured here: git sends a header given through GIT_CONFIG_* for one URL to a local server. wiki-publish.test.js holds the script, the workflow's permissions and its token handling, and the documents; SECURITY.md says what the workflow may write and what limits it (#726)

  • The site opens on the landing, and the application moved to /app. / is the landing; the application is unchanged at /app and still opens on the bill. /?tab=<name> redirects permanently (308) to /app with every parameter it carried, /landing to /, and a bare / is never redirected: redirectFor in apps/web/lib/site.ts decides, the middleware answers before anything else, and the target is always this origin. Share links, the badge, /admin and the API keep their addresses. Every public page has one header and one footer, built from one list of entries: the full wordmark home, Playground and Docs (the docs index on GitHub until the site has its own), the language and appearance controls, a GitHub button drawn inline with an accessible name, and the primary action into the dashboard; below 1280px the entries move into a menu that is a disclosure and holds no focus. The footer adds the licence, the changelog, the security policy and the npm package. The landing is rebuilt from the README's own sentences: its first line as the headline, npx @trazum/cli bill with a copy control, the README's illustrative bill from examples/claude-projects (labelled as the README labels it, and held to what the CLI prints), why the number holds up, what never leaves the machine, the CI gate, the agents' budget check and prompt trimming, and the command again. The application's rail and phone bar gain the way back to the site and the same GitHub button, and the bill panel opens on one control that loads the sample month. Each public page has its own title, description, canonical address and social card, drawn in the colours scripts/brand.mjs names (which gains the light tokens, held to globals.css by brand.test.js), and an unknown address gets a branded page with the way home. Every link to the web app in the README, the docs, the skill and the wiki uses the new addresses, and site-links.test.js fails on one that would only redirect. A fourteenth end-to-end flow follows a visitor from the landing to a priced sample month (#734).

  • Discussions has a form for each category a visitor writes in, and one address. GitHub Discussions was switched on on 2026-11-18, and the only thing that sent anybody there was a line in config.yml. .github/DISCUSSION_TEMPLATE/ now holds q-a.yml, ideas.yml and show-and-tell.yml. Each opens by asking for no usage log, no prompt that is not public and no key, then asks what an answer needs: for a question, the goal, the command, what happened and what was expected, the version and how Trazum was installed; for an idea, the problem first, what is done today, and a pointer to the feature request form for an idea that is clearly a thing to do; for show and tell, what was built or measured and how to repeat it. There is no form for announcements, which only the maintainer posts, nor for polls. The file names are GitHub's default category slugs, which could not be read from the repository, and each form says so at its top; a category that was renamed loses its form until the file is. Whether GitHub accepts the forms was not tested. README.md now says a question goes to Discussions and a bug to an issue, through its form. issue-intake.test.js holds the forms (a first block that says what not to paste, unique field ids, the version where a run is described, the slugs) and requires every page that sends somebody to Discussions to use https://github.com/Davmunrey/Trazum/discussions and nothing else. The Repository settings check now also compares has_discussions with true, reports it as not read when GitHub does not show it, and writes it with apply --yes; switched off, three pages would point nowhere. Whether the workflow's token is shown that field was not measured. (#732)

  • Vercel Speed Insights, off unless the operator switches it on. Vercel's bot opened #722 with @vercel/speed-insights rendered in the layout after <Analytics />, always on. That contradicted what the README, SECURITY.md and the wiki promise: analytics is off by default, nothing records page views on its own, and every URL that leaves has its share token replaced and its query dropped. #722 was the bot's version, and this one differs in four ways. The switch is the build-time variable NEXT_PUBLIC_VERCEL_SPEED_INSIGHTS, on only for the exact value 1, read in lib/analytics.ts beside the PostHog key. The package is imported dynamically inside the on branch of SpeedInsights.tsx, the way posthog-js is, and a test fails if any file under apps/web imports it statically. Every measurement passes beforeSend through redactUrl, the rule PostHog events use, and one whose URL does not parse is dropped. And the lockfile gains the package and nothing else: npm 10.9.9 on this machine also removed sixteen libc fields, as the bot's lockfile did, so the entry was restored by hand and npm ci accepts it. Read from the package, not from Vercel: the script that sends the measurements is served at /_vercel/speed-insights/script.js and is not in the package, so what it sends was not read, and it only works on a Vercel deployment. The Content-Security-Policy does not change and does not read the variable: the script is on the page's own origin and is added by a script the nonce vouches for. In a production build served with the variable set to 1, Chromium requested the script, got the server's 404 because this machine is not Vercel, and reported no policy violation. A test builds the policy in two processes, one with the variable and one without, and requires the same bytes. With the variable unset the package is in one 3,471 byte chunk that the page never requests; the same was true of posthog-js before this, because Next here does not fold the environment variable away at build time. The README, SECURITY.md, .env.example, docs/setup.md, the rotation page and the wiki say all of this (#723).

  • The roadmap's order is held in a file, and GitHub's milestones are made to match it. .github/milestones.json is an ordered array, first is next: each entry has a title, a description and the issues that belong to it, and none has a date, because the order is the commitment and the calendar is not. npm run milestones reads GitHub's milestones and open issues through gh and prints each difference as a sentence, exiting 1 on one: a milestone the file has and GitHub lacks, a different description, an issue in another milestone or in none, an accepted or blocked issue the file places nowhere, an issue in two milestones of the file, and a due date. A milestone only on GitHub is printed and left alone. apply says what it would change; apply --yes creates, corrects and places, never deletes, then reads GitHub again and judges that reading. The Milestones workflow applies the file when a change to it merges, from the one job that holds issues: write, reports on a pull request, and fails every Monday when GitHub has drifted. ROADMAP.md carries the ordered list, each title linking its live milestone, drawn from the file by node scripts/milestones.mjs roadmap, and the wiki gains a Roadmap page generated from its Next section (#720).

  • Codecov groups the coverage by package. codecov.yml now defines one component for each package the coverage run measures, core, cli, mcp and tokenizer-openai, each by the path of its sources: packages/<name>/src/**, which is the shape every path in an uploaded report has. A component is defined in the file alone, so nothing changes at upload time. There are no component statuses and none in default_rules: the plan does not include them and a status nobody can read is noise. Codecov's validator accepted the file. Codecov's settings page lists Components as available with its Pro plan, as read on 2026-11-18, so the configuration is in place and the plan decides whether it is shown. coverage.test.js holds the components to the packages the script measures, and docs/coverage.md says what they are and where they show (#717).

  • Test results and the web app's bundle size go to Codecov. Two things Codecov offers were left unconfigured on purpose when coverage was set up, and the owner asked for both. Test analytics: the coverage run now also writes each package's results as coverage/<package>.junit.xml, which Node's runner produces itself, and the Coverage workflow uploads them with the uploader it already pins and checks. It answers what the required job cannot: which test fails some of the time, as one in the core did on two of five runs on 2026-11-18. Bundle analysis: a new Bundle workflow builds the web app and scripts/bundle-report.mjs sends the size of everything under apps/web/.next/static. Measured that day: 22 files, 1,851,354 bytes, 552,234 compressed. Codecov's Next.js plugin is for webpack and the web app builds with Turbopack, so the standalone analyzer is used and nothing is added to the web app's build; it is installed in CI outside the workspace at one exact version with install scripts off, the script refuses any other version, and its telemetry is switched off. Neither is a required check, both statuses are informational, a report Codecov does not take is a warning on a pull request and a failure only on main, and the token is read from the environment by name. The build names chunks by a hash of their contents, so the total is the figure to read and the per-file difference is not. docs/coverage.md says what is sent in each case: test names, durations and failure messages; file names and sizes. No source (#714).

  • The ruleset GitHub enforces is compared with the one in the repository. .github/rulesets/main-branch.json is the branch policy, and GitHub enforces a copy an admin puts there. Nothing read the copy, and on 2026-11-18 it had drifted: it lacked PR readiness, which the file had required for weeks, and allowed merge commits, which the file did not. Every pull request that had changed the file was green. npm run ruleset reads the live ruleset through gh and prints each difference, rule by rule, naming the entry only one side has; it exits 1 when there is one. It also reads the five merge settings the ruleset leans on, each with why, and says "not read" for one the token is not shown, which is neither fine nor a failure. npm run ruleset -- apply says what would change, and with --yes, from an admin's own gh, writes the file to GitHub, sets the settings and reads both back. The file is held exactly for what it states: a parameter GitHub reports and the file is silent about is printed and left alone, on reading and on writing. The Repository settings workflow runs the comparison every Monday and fails on a difference; on a pull request that touches the file it reports and does not fail, since such a pull request differs from GitHub by design until an admin applies it. The workflow reads and never writes, and no admin token is stored for it. Its first run on a runner showed what a local run with an admin's token could not: GitHub leaves bypass_actors and the four allow_* settings out of its answer to the workflow's token. Comparing that absence with the file's empty list would have failed every Monday for something nobody changed, so those five are printed as "not read", and any other field that goes missing is still a difference. SECURITY.md said "nothing syncs it"; it now says what does, and what the weekly run cannot see (#708).

  • Line coverage, measured and sent to Codecov. The repository said how many tests it had and, on five files, whether they would notice a change; nothing said which lines the tests run at all. npm run coverage runs the four published packages' tests with Node's own coverage, mapped back through the build's source maps so every path is a TypeScript file, and writes one report for each package. Measured on 2026-11-18 with every suite together: 94.5% of 70,633 lines in 152 source files (core 95.9%, cli 94.0%, mcp 77.1%, tokenizer-openai 98.0%), figures rounded down. Two source files are in no report: one holds only types, and the MCP server's entry point is started and killed by its tests, and a killed process writes no count. The Coverage workflow uploads the reports on every pull request and on main. It is not a required check and is built so it cannot block a merge: a test that fails under coverage does not fail it, an upload that fails does only on main, and both of Codecov's statuses are informational, with no target. The uploader is Codecov's command-line tool at a pinned version, refused unless its checksum matches; the token is a repository secret that reaches it through the environment. The workflow's first run was refused by Codecov and went red: the secret had not reached the job, Codecov takes an upload without a token only from a fork, and the uploader exits 1 on a refusal whether or not it was asked to. A refusal on a pull request is a warning now, as it was meant to be, and the log says whether a token was set. The web app and the editor extension are not measured yet, each for a stated reason. npm run done stopped the change once before the push: a guard requires a rotation procedure for every secret a workflow reads, and CODECOV_TOKEN had none. It has one now, in docs/compliance/secret-rotation.md. docs/coverage.md (#706).

  • One command that says whether a change is done, and why not. The pieces of the loop were all here: npm run verify, the required checks, PR readiness. What joined them was one session's habit, and on 2026-11-17 it missed twice: a new test file invisible to a test that reads the list of tracked files until it was staged (#685), and a release reported finished before its packages could be downloaded (#688). npm run done runs sixteen criteria in order and answers each with passed, FAILED and the reason, or not run and what is missing: every new file tracked, every commit signed off, a changelog line under Unreleased, an issue named, then each step of CI's required job as CI writes it (the command of a one-line step is read out of ci.yml when it runs), then mutation testing when a changed file brings it into play. It adds no criterion of its own, fixes nothing, and exits 0 only when everything that applies passed; not run is not passed, and part of the list is never a verdict. done.test.js fails when a step is added to the required job that is neither a criterion nor left out with its reason, when a criterion names a step, a heading or a test that is not there, and when the script gains a command that writes. docs/done.md states the loop, and a new AGENTS.md points a coding agent at the command, the document and the review skill. Its first whole run, on this change, answered "not done": three lint findings in its own test. Thirty-eight faults were planted against done.test.js; all were caught, one only after it was planted again, because it had removed one of the two places AGENTS.md names the command and the file still named it (#694).

  • A review skill that carries this repository's own rules. What a review here looks for was written for a person in CONTRIBUTING.md and for a bot in .coderabbit.yaml, and a coding agent in a checkout had neither as something it could run. .claude/skills/trazum-review/SKILL.md is that: sixteen rules, each with the paths that bring it into play, where it is stated, what holds it and how to check it, and the form a finding is reported in. It restates no rule. review-skill.test.js holds every pointer: a path that exists, a heading that is in CONTRIBUTING.md, a test file that is in the tree, an npm script that is a script. It fails when the Review section gains a question the skill does not point at, when the bot is given a path no rule applies to, and when the skill names a command that would change the repository: it reviews and does nothing else. One rule the skill could only point at a comment for now has a test: no test file bounds a section of a document by looking for the heading after it (#693).

  • End-to-end tests of the web app. The web app had 51 test files and none of them opened the page. The one thing that drove a browser, the privacy probe, checks a single promise, and a page can keep it and print the wrong total. scripts/e2e.mjs runs twelve flows in a real browser against the production build, each from an empty browser: a pasted log priced to the cent by part, label and model; a paste that is not a log refused with the lines it could not read; the same bill at phone width with no sideways scroll; a prompt optimised and two compared; the prompt writer's next question; a playground command; a recorded run read into a report on the Tasks tab; every tab by click and by address; dark mode and Spanish surviving a reload; the phone's menu. A flow also fails on any error the page throws. Every expected figure is worked out by hand in a comment ($1.43 for the bill, $1.575 for the run), and e2e.test.mjs holds the three prices that arithmetic assumes, so a price change says to re-price the log before a browser says it waited for a total. The same test fails when a tab a visitor can open has no flow, when a flow claims a tab it never names, when the step leaves CI or is made unable to fail, and when the runner is asked for a flow that does not exist. The library tab is left out by name: it needs an account. Twelve faults were planted in the app in two builds and every one was caught, eleven by the flow written for it and the last by every flow that had nothing else to fail on: a cache read at a fifth of input ($1.47), a theme not remembered, a lang attribute not following the switcher, a tab that no longer rewrites the address, a paste box wider than a phone, a comparison without its percentage, a count the wrong way round, one answer too many asked for, unreadable lines counted and not named, a chosen file dropped, a menu button that opens nothing, and an error thrown after mount. Eighteen more were planted against e2e.test.mjs. Two escaped and each has a test now: a tab opened only by the flow that walks the rail counted as covered, and a build step removed outright passed a check that it came first. The flows are a step of the required Build and tests check, with the browser driver the probe already installs; they took 17 seconds locally. npm run e2e, docs/end-to-end-tests.md (#692).

  • A linter, and the dead code it found. Nothing linted this repository: strict TypeScript and CodeQL catch type errors and known security patterns, and neither says a word about a name nobody reads. npm run lint runs Biome 2.6.4 over every tracked file, 886 of them in under a second, and fails on a warning as it does on an error. It is the first thing npm run verify does and a step of CI before the build. With its recommended rules and the style group off it reported 90 errors and 103 warnings on 2026-11-18. What was a mistake is fixed: 14 names nobody read (a PUBLISHED_BAND constant the token-band test defined and never compared anything with, among them), 7 imports of nothing used, 16 parameters nothing read (six functions lose one, with every call site; the rest are dropped, or marked unused where a caller still passes them), three Object.prototype.hasOwnProperty.call that are Object.hasOwn, and seven test patterns written [^] that say [\s\S] now. What the code does on purpose everywhere is a rule switched off in biome.jsonc, 16 of them, each with its reason on the lines above. What it does on purpose in one place carries a biome-ignore and its reason next to the line: seven, in the web app and the editor extension's declarations. Three React rules are off without an answer: 28 findings about effect dependencies, index keys and components defined inside components, which wait for the end-to-end tests of #692 and are tracked in #696. It lints and does not format. lint.test.js holds the configuration: the pinned version, a warning failing, the count of rules that are off, a reason above each, and a reason of some length on every excused line. A development dependency of the repository only; no published package gains one. (#691)

  • Mutation testing for the core, and the 79 tests it showed to be missing. scripts/mutate.mjs changes the built code of five files one small thing at a time (a < into a <=, a * into a /, a call taken out) and reports every change no test noticed, at the line of TypeScript it came from. The files are the ones where a wrong line costs most: optimize.ts, segment.ts and rules.ts, which rewrite a prompt, and savings.ts and session-cost.ts, which turn token counts into money. The first whole run, on 2026-11-17, caught 277 of 446 mutants: 62.1%. A cache read could be divided by its multiplier instead of multiplied in the cost of a conversation; the median could take the wrong element of any list whose middle values differ; the percentage a prompt was reduced by could be a hundredth of itself; two protected spans that touch could lose one. All of it passed 2,994 tests. Five new test files, each written from the survivors of one target, bring the recorded figure to 408 of 446: 91.5%. The 38 left are in mutation.baseline.json and read one by one in docs/mutation-testing.md: 24 change nothing a caller can observe, and 14 are not settled. dist is never edited: a loader serves each mutant to one process. Each mutant is given only the test files whose coverage reaches it. Before any result is trusted the loader is checked both ways, every mutant is parsed so a syntax error is never counted as a caught bug, and --audit runs recorded survivors against the tests they were not given: the first day's run chose tests wrongly for every a ?? b whose right side never ran, and that is how it was found. The parser it needs is installed outside the workspace, pinned, as the privacy probe's browser driver is: nothing published and nothing in the lockfile depends on it. The Mutation testing workflow asks a pull request about the lines it changes and runs every target on Mondays. What is not mutated is said on the same page as the score: regular expressions, the contents of strings, a for loop's counter, and every file but those five, which are 446 of the core's roughly 15,700 possible mutants. (#690)

Changed#

  • The terminal carries the brand: one module draws the reports, the answer comes first, and it fits the window. packages/cli/src/style.ts is now the one module that draws a terminal report: the commands hand it sentences and rows, and it decides the colour, the width and the order. bill, profile, plan, optimize, check and proactive draw through it. Each opens with its question as a ruled heading and the figure that answers it on the line under it, then the breakdown, then the notes and what to do next: bill moves the period and the subscription caveat below its tables, and on a terminal writes the files it read to stderr after the statement rather than five lines above it, dimmed on a terminal; profile moves the period below the four parts of the bill; check says All 3 within budget. under its heading instead of after the files; optimize draws Start here as that heading. Colour comes from the terminal: true colour where COLORTERM says so, the nearest of the 256 where TERM does, the 16 otherwise, and none for NO_COLOR, a pipe or a file; FORCE_COLOR names a depth (1, 2, 3, or 0 for none). The answering figure is marked in ink on the acid ground, once per report; the bars and the lever arrows take the accent, #70831b, the brand's olive at a lightness that keeps 3.5:1 on light and dark terminals alike; success, caution and error keep the terminal's own green, yellow and red, and every one of them is also a word or a mark (!, ?, →, OK, OVER), so the caveats that were whole yellow paragraphs now carry a !. With a window (or COLUMNS), reports are drawn between 60 and 76 columns: paragraphs reflow, a table narrows its bars and then cuts its name, and a figure never leaves its row. A pipe keeps the layout it had. No figure, exit code or --json document changed: terminal-report.test.js compares every number the five reporting commands print on the examples with what they printed before, at no width and at 60, 80 and 120 columns. The README's two pictures and the landing's screen are redrawn from the new output; the picture draws the mark in the acid. (#758)

  • The landing tells the product as you scroll: a laptop, the CLI at work, and a lid that closes. Below the headline a laptop drawn in the brand arrives; scrolling pushes the camera into its screen until the terminal fills the window, and inside it npx @trazum/cli bill examples/claude-projects types itself and prints the bill line by line, its share bars filling, then check examples/ gates the repository's own example prompts and echo $? prints the status the process returned, then optimize examples/support-prompt.txt prints its whole report. The camera pulls back out, what never leaves the machine settles onto the screen, and the lid closes over it in 3D, ending on the closed machine with the wordmark and what does leave beside it. Every character on the screen is the built CLI's output, written by scripts/draw-landing-terminal.mjs into apps/web/lib/landing-terminal.ts and compared by the web suite; the MCP spend_guard answer has no scene because the repository has no example call of it to run. Scroll is the only clock: each moving element runs keyframes of transform, opacity and clip-path written from the timeline, on the section's CSS scroll timeline where the browser has one and held paused at the scroll position by a passive, frame-throttled script where it has not, so scrolling back plays it back. On a full scroll at 4x CPU throttle no task over 50 ms ran while scrolling, and the page's layout shift was 0.0086 at 1440x900 and 0.013 at 390x844, once the headline was given a fallback face sized to Space Grotesk. With reduced motion or without JavaScript every scene is ordinary content in reading order and the laptop is still. The public site is now light first: with no choice stored it is paper even on a dark system, until Dark is picked in the header; the application keeps System as its default. The boot mark that stood in for the application while its scripts loaded moved under /app, because at the root it hid the whole landing from a reader without JavaScript. DESIGN.md records the rule, which the owner asked for on 2026-11-18. Five browser flows hold it, in docs/end-to-end-tests.md. (#744)

  • ROADMAP.md opens with what comes next, and the release history is its own file. The roadmap was 3,394 lines and ## Next began on line 2,927, after the story of every release from 0.1.0. That story is in HISTORY.md now, the 2,903 lines moved without a word changed, and ROADMAP.md is 493 lines with ## Next on line 26 and one line linking the history. The tests that read the history read it there: every version it calls released is in CHANGELOG.md, every version CHANGELOG.md released is in it or inside a range (1.9.1, which it never carried, is named as the one exception), its first run still goes oldest first, and the pages that record what was true then (derived counts, the band, the protected kinds, the forecast) treat it as a record. A new test holds that ## Next is within the first 40 lines, that no release entry is back in the roadmap, and that a link into a heading of either file arrives at one. The open-work table is redrawn from the open issues, and #744, #745 and #747 are in the milestones. The release steps in VERSIONING.md and docs/releasing.md say HISTORY.md where they said the roadmap's Released (#725).

  • A pull request has more to pass before it can merge. On 2026-11-18 somebody outside the project asked to take an issue for the first time, and the rules had been written for pull requests that all came from the owner. Three things changed. Signed-off-by on every commit and cla ran on every pull request and neither was required, so an unsigned commit or an unsigned contributor could be merged by a click made in a hurry; both are in the ruleset's required checks now. PR readiness asked for an issue number and never for what stood behind it; it now reads the issue and refuses the pull request unless a maintainer labelled it accepted, says which label it carries instead, refuses a closed issue whatever it carries, and refuses to judge when it cannot read the issue. Dependabot and a pull request labelled no-issue are not asked. And AGENTS.md says what an agent may do with somebody else's work, which is nothing the owner has not decided: no assigning, accepting, approving, merging, closing or answering. It is written down because it happened that day. CONTRIBUTING.md has the same rule from the contributor's side, with what a pull request from outside is read for. The ruleset still asks for zero approving reviews: with one maintainer, merging is already that maintainer's act (#718).

  • The three React rules are on, and each of their 27 findings is answered where it is. biome.jsonc switched useExhaustiveDependencies, noArrayIndexKey and noComponentHookFactories off until something drove the pages. The end-to-end tests do now, so each finding was read. Of the seven effects behind the fifteen dependency findings, four get the dependency: the playground's cancelTypist and the share control's refresh become one function for the life of the component, and the optimiser's demo names its setter. Three carry a reason on their own line, because the dependency the rule asks for is the bug: the bill re-prices only when the price card changes, the optimiser re-runs only when a pricing input moves, and the prompt writer asks its first question once. Of the nine index keys, the hours of the hourly chart are data and are keyed as data; eight stay, each with why its list has nothing else to key by. The three Eyebrow components defined inside Comparer, Optimizer and Plan are one, in components/ui/eyebrow.tsx: defined inside, each heading was unmounted and remounted on every render of its panel. The issue counted 28; #697 had already removed a fourth Eyebrow nobody used. Two of the suggested fixes were planted as the rule words them, and both break the page: naming send in the prompt writer's effect puts a reader back on the first question when the language changes, and naming an unstable cancelTypist stops the tour's typing hand before it runs a command. The end-to-end tests gain what catches each: a language change in the middle of the prompt writer, and a thirteenth flow that walks the tour's ten steps and requires each of the terminal's three commands to be typed and run once. One existing test pinned the share control's dependency list as text and failed on the change; it now checks what it meant, that the effect depends on the sign-in state and that refresh is one function, so naming it cannot ask the server again on every render. The count of rules switched off goes from 16 to 13 (#696).

Fixed#

  • A schema with thousands of keys no longer holds optimize for half a minute. The advisory that notices prose restating an output schema asked one pattern of every sentence for every key, and built the pattern again each time: the square of the text. A prompt of 40,000 characters, half a schema of 1,778 keys and half prose, made optimize take 6.8 seconds, and one of 80,000 characters took 37.8, measured on 2026-11-18; the HTTP API accepts 400,000. The prose is now read once, every key looked for in one pass, and the key's own pattern only confirms the place: 80,000 characters take 0.37 seconds. The answer is the same, in the same order: restated-format-scaling.test.js compares the two on 3,000 generated cases, holds the case folding to the engine's own over every cased character, and fails when four times the text takes ten times as long. Found by reading a Codacy finding that called the pattern a regular-expression denial of service; the pattern was linear, the loop around it was not. (#772)

  • Every finding Codacy lists is answered, and a script says when one is not. #772 read the 204 findings outside the tests one by one: 6 fixed, 92 from two tools and two patterns switched off, 106 dismissed with a reason a reader can check, in .github/codacy.json. The 710 in the tests are left out by .codacy.yaml, and a test holds that no shipped source is. npm run codacy reads Codacy's public list and prints what the file does not account for, including a new finding of a kind already dismissed; apply --yes, with an account token, switches off what the file says and dismisses the rest, a hundred at a time, without printing the token. Each regular expression a finding named was timed at a text and one four times as long; two in the docs build's markdown reader grow faster on hostile input and only ever read the repository's own documents. Also fixed while reading: the terminal pictures' accessible name now escapes a quotation mark, the image runs as user 1000 by number, and the two plans' repeated headings name their arc. docs/coverage.md, "Codacy", says what is off and why. The nine findings Codacy raised on this change's own code, in the new script and the new detector, were read and recorded the same way, and a tenth, the record quoting the CI database's throwaway login, was reworded. (#772)

  • The landing's terminal stood in the top of its window; it is centred now. Measured in a browser at five sizes, light and dark, through the whole story: at 1440x900 the bill's output had 72 px above it and 525 below halfway through its scene, its middle 226 px above the window's, and once the output filled the column the lines scrolled on up to the window's edge. Two things put it there. The column printed from its top and moved only to scroll, and it scrolled and was sized by a line height worked out for 19 px type while the lines were set in the still page's 15 px. Now the line height is read from the column's own type, the column is centred across the free part of the window (beside the scene's plate on a wide screen, under it and above the marker on a tall one) and down it, and lines scrolled past its top are clipped, so a long output has the same room above as below: 72 and 72 at 1440x900, 16 and 16 at 390x844. On a tall screen the free part is under the showing scene's own plate, so Story.tsx measures each plate's foot when the stage or a plate changes size, never while scrolling. As one plate gives way to the next the output moves to the next scene's middle while only the plate it moves away from is up, so it is never under a plate; on a tall screen the plates fade where they stand instead of sliding 24 px. Across the swap from the gate to the trimming, whose plate is 97 px taller at 390x844, the output's top moves at most 16.5 px between two frames 0.5% of the story apart (24.5 px of scrolling); there and at 768x1024 it comes no closer than 15.8 px to a plate. The plate on a phone keeps the page's 16 px gutter (it was 15.6), and the closing band's padding is the other bands' 80 px at 1024 px and up (it was 96). The type, the copy, the colours, the scenes and their timing are as they were. A browser flow measures where the output stands after each run of output at 1440x900 and 390x844 and fails beyond 3 px or on a plate taller than its words, and holds the output clear of the plates through each swap at 390x844 and 768x1024; ten faults were planted in the page and three in the flow's own test, the alignment as deployed and plates stretched to the tallest one's height among them, and each was caught. (#770)

  • The pull-request guide no longer asks Dependabot to link an issue. A bot cannot, so on a green dependency update (#655) the guide named @dependabot[bot] and never reached the step that moves it: the owner's approval, which GitHub accepts on a bot's pull request and records as a review. pr-guide.yml now pins the version from Davmunrey/Zolts#116: a bot's pull request is not asked for an issue, a green one tells the owner to approve and merge, and the owner's review is requested once, which notifies them. The pinned file was fetched from the new commit and matches the new SHA-256. (#773)

  • The wiki's Home page closed a <div> it never opened. The page is a section of the README, and the cut fell after the closing tag of a block whose opening tag is above it. scripts/build-wiki.mjs now pairs the tags of every page it writes: a closing </div> with nothing open is dropped, one left open is closed, and a code fence is left alone. wiki.test.js holds every page to closing what it opens, and checks that the README section still hands Home the orphan, so it cannot pass because the input stopped exercising it. (#765)

  • PR readiness told a pull request with twenty-two lines under Unreleased to add one. scripts/pr-readiness.mjs read this file with git show through execFileSync and Node's default limit on a child's output, which is 1 MiB. This file passed it, the read failed, and the script's fileAt turned every failure into an empty file, so it compared nothing with nothing and blamed the author. It is a required check, so every pull request after that one would have failed it. The limit is set now, and fileAt has two answers, the file or "this commit does not have it"; anything else ends the run with the code for "could not judge" and says what could not be read. The tests run the script on a real repository whose changelog is over 1 MiB. Four faults were planted and it failed on each, among them the script exactly as it was (#754).

  • The linter was 1,071 bytes from failing every pull request. Biome warns about any file over 1 MiB that it is given, a warning is an error here, and this file was 1,047,505 bytes on main. The next entry longer than what was left would have failed the required check for a pull request that had done nothing wrong, and so would every one after it. It was found when a change with an ordinary entry did exactly that. Biome read 102 Markdown files and has no rule that reports on any of them: without them the same run checks 857 files in place of 959 and finds the same things. So the linter does not read Markdown any more, and biome.jsonc says why. The limit itself is not raised, because a source file of a mebibyte is worth the warning. lint.test.js now fails by name on the next file the linter reads that reaches the limit, and on a changed limit. Three faults were planted and it failed on each: Markdown read again, the limit raised, and a source file of 1.1 MiB added (#751).

  • The versioning policy described a line that ended in August. VERSIONING.md still had "The 1.x freeze", "a breaking change waits for 2.0" and ^1.50.0 as the pin to copy, with the manifests at 2.7.12, and this file opened with the rule for versions below 1.0 and no other. It now says what holds inside 2.x, that a breaking change waits for the next major, and what 2.0.0 was: a major that broke nothing, as ROADMAP.md records. One promise had come due unnoticed. The 1.8.0 entry that deprecated tier on a model said it would be removed in 2.0, and 2.0.0 removed nothing. The policy now lists it under what is deprecated and still working, says the removal did not happen, and leaves whether it goes in 3.0 undecided, because nobody has decided it. The field's own comment said "removed in 2.0" and says the same now. Nothing is removed and no behaviour changes. The release steps said the type-check covers four workspaces and it covers six. versioning.test.js holds the policy to the major in the manifest, the pins to that major, the list of deprecations to every @deprecated in the packages' sources, and the two workspace counts to the workspaces. It found a second ~1.50.0 in the first run. Sixteen faults were planted, among them the manifest moved to 3.0.0 with the policy untouched, and it failed on each (#747).

  • Nine places where two documents disagreed now agree, and two were not contradictions. An issue listed eleven candidates; each was read on both sides against the code and this file. Real: (1) ROADMAP.md offered the editor extension as open and scheduled while saying it shipped in 1.86.0. Both are true of different facts, and each place now says which: it is built and packages to a .vsix, and it is not on the Marketplace, which is #640. (2) The Under consideration opening said nothing there was built, above an entry struck through as shipped; it now says the struck entries are the record of what the file went back on. (3) docs/mutation-testing.md said 38 survivors for the audit and 35 everywhere else: the audit ran over the 38 that were on record that day, and the page now says so and gives the 35 the record holds now. (4) and (5) The plan index called the 1.83 to 2.0 plan written and nothing delivered, and the 2.4 plan in progress; 2.0.0 and 2.4.0 are in this file, and the rows say what each delivered and, for 2.4, what waits on something outside. (6) docs/commands.md called LangSmith, Helicone and LiteLLM unbuilt above the three sections that document them, and ROADMAP.md called from-langsmith unreleased; it is in 1.83.0. The plan itself stays as written and the index row says what became of it. (8) The 1.52 to 1.60 plan says Mistral is unfronted; the 1.82.0 entry in this file has the gateway fronting it, and the index row says so, leaving xAI and Moonshot. (10) The roadmap's forecast that the tokenizer question lands in 1.54.0 now carries what happened: 1.85.0 answered it with a separate package, #642 asks about DeepSeek and Mistral ones, and 1.54.0 stays open for Google, xAI and Moonshot. (11) The roadmap's last section stated the rule for versions below 1.0; it now gives the three meanings VERSIONING.md states and calls the below-1.0 rule history. Not contradictions: (7) the README says xAI retired grok-4 on 2026-05-15, which is xAI's notice, and the catalogue has no retired mark on it, which records a refusal nobody here has seen; and (9) docs/our-own-medicine.md says Moonshot and xAI were not probed for availability while the roadmap says Google's estimator error is unmeasured, which are two questions. Both pairs are left as written.

    Three guards keep them from coming back. docs.test.js reads the plan index and fails a row that says a plan is unfinished when the release the plan ends at has a heading here. mutation.test.js holds every count of survivors on record in docs/mutation-testing.md to mutation.baseline.json. roadmap-forecast.test.js read ROADMAP.md only, and now reads every page that says what exists, plans and records excluded, for a shipped converter called unbuilt. Planted, fourteen faults, all caught. (#727)

  • Segmenting a text of many protected spans no longer takes the square of them. For every match, the segmenter asked whether it overlapped a range already reserved by looking at every reserved range, and asked whether a quotation held a reservation by looking at every quotation. A text of little but placeholders, tags and code spans has as many reservations as matches. Measured on 2026-11-18: optimize took 0.5 s on 70,000 characters, 3.3 s on 140,000, 16 s on 280,000 and 26 to 30 s on 399,000, which is under the 400,000 the HTTP API accepts. Both questions are now answered by halving over ranges that are in order and do not overlap, and the same 399,000 characters take 1.1 to 1.5 s; the segmenter's share went from 10 s to under 50 ms. What found it was a test with a five-second budget failing one coverage run in three on a runner, which was read as a flaky test until the run could name it (#712). The suite's own comment had excused this very fixture as "doing exactly the work it is supposed to do". The new tests hold it without a budget in seconds: the two searches are compared answer for answer with the loops they replace, the number of ranges a search reads is counted, and the segmenter is timed against itself at a quarter of the cap and at the cap. The file is under mutation testing, and the record was made again: 92.6% of 475 mutants caught, where it was 91.5% of 446, with none of the 29 the searches added left alive (#716).

  • A test that fails under coverage is named. npm run coverage said "a test failed in this run" and nothing else: the report was the run's only output, so which test, where and why were in no log and no file. On 2026-11-18 two of the five Coverage runs that completed on a runner said it of the core, on commits the required job had passed, and there was no way to learn which test it was. The runner now writes its own account of each run beside the report, as coverage/<package>.tap, and the output lists every failing test with its file and line and the first line of its error. A group that failed only because a test inside it did is not listed twice, and a test marked todo or skipped is not listed at all. The job still does not fail on it: the verdict stays the required job's (#712).

  • The coverage document no longer says its figure is Codecov's. The script's comment and docs/coverage.md said the combined figure "is the figure Codecov shows". That was written before Codecov had taken an upload. The first, for main on 2026-11-18, came out at 93.3% (65,907 lines run, 111 partly, 4,615 not) where the same run's npm run coverage printed 94.5% (66,809 run), over the same 70,633 lines in the same 152 files. The document now gives both and what is known: the script counts a line by its line record; Codecov's ratio is hits over hits, partials and misses, and it judges a line that carries a branch record by the branch, so a line that ran 108 times with an if nobody takes is, to Codecov, not run. Of 1,044 lines the two disagree on, 1,023 carry a branch record. It also says what is not known: Codecov called 353 lines run although every report says a branch on the line was never taken, and why was not tested. The README has a coverage badge, showing Codecov's figure (#709).

  • A quick edit to a description no longer leaves a cancelled required check. PR readiness runs when a pull request is opened and again when its description is edited, and the second run cancelled the first. Edited seconds after opening, #699 was left with a cancelled PR readiness check and a guide comment that said to fix a failing check while nothing was failing. It was intermittent: the same edit on #700, #701 and #702 did not show it. A run in progress now finishes, as the triage workflow's has since #689; the runs for one pull request stay in order, so the newest description still has the last word (#703).

  • The whole mutation run can start on a runner, and says why when it cannot. Dispatched on main after #695, the Mutation testing workflow stopped in 46 seconds: four of the core's test files failed before a single mutant. The workflow built only the core and checked out one commit; own-gate.test.js and security.test.js need the CLI built, and changelog-coverage.test.js needs the tags. It builds everything and fetches the history now, as CI's test job does. The fourth, courtesy.test.js, needs neither: it asserted that hostile input is handled within 1,500 ms, and its slowest case takes 795 ms on an idle machine. That budget was a benchmark where the same tests elsewhere use a cliff detector, and it failed in CI's ordinary test job too, on this very change: guillemet: 1697ms on 180000 characters. It is 5,000 ms now, as in security.test.js; the fault it guards against took 47 seconds. And since any budget can be missed under enough load, a failure from a test file that reads a clock is believed by the mutation run only when it fails again with nothing else running, before any mutant and against one, and those files are tried last. The tool printed which files failed and not why; it prints each failing test and its error. And #695 merged green with a workflow that could not run because a pull request that changes no target returns before running a test: one that changes the tool, its record or the workflow now runs one small target whole. Dispatched on this change, the whole run passed on a runner in 1,259 seconds and caught the same 408 of 446 mutants the record holds, file by file. (#698)

  • A release waits until its packages can be downloaded, not until npm knows their names. scripts/npm-serves.mjs waited for npm's latest dist-tag to name the new version and stopped there. On 2.7.12, 2026-11-17, that check passed at 22:42:49 UTC and @trazum/core's tarball answered 404 at 22:43:05 and 22:45:23, then 200 at 22:45:55: for about three minutes after the release reported success, npx @trazum/cli@2.7.12 failed with E404, because the CLI pins that exact version of the core. The check now asks for each tarball as well, with a HEAD request to the address an install fetches, and when it gives up it says which of the two it was still waiting for. The Triage workflow no longer cancels a run in progress: on #687 a description edited seconds after opening left a cancelled check that the guide comment reported as a failure, and both runs can simply finish. And the README's caption under the optimize picture says where its dollars come from: the token counts and the $31.00 saved are the same anywhere, while the first lever's $416.99 is what the command prints in this repository, whose configuration assumes 300 output tokens a call. Outside it the published CLI assumed 500 and printed $661.99 for the same prompt. (#688)

Added#

  • Triage that happens whatever opened the issue, and a roadmap made of the issues. CONTRIBUTING.md said every new issue starts as needs-triage. On 2026-11-17 eleven issues were open and none carried a label: the forms apply theirs, and none of the eleven had come through a form. The Triage workflow now gives every new issue a state, and gives a pull request from this repository an area for each package its paths touch and the kind of the issue it closes; it only adds, so a label a person changed stays changed. New labels: blocked, for work that is accepted and waits on something held outside the repository, and area:core, area:cli, area:mcp, area:web, area:action and area:vscode. The open issues are triaged: seven of the ten backlog items turned out to be waiting on a key, an account or an export only the owner holds, and say so now. ROADMAP.md gains "Open work, as issues", a table of the accepted and blocked issues drawn by node scripts/triage.mjs roadmap; the weekly run fails when the page and the issue list disagree. CONTRIBUTING.md gains a "Review" section, what a review here looks for and how a machine's finding is handled, and .coderabbit.yaml configures a review bot that is free on public repositories: it is told those same questions, reviews drafts because that is the state a pull request is reviewed in here, and has every feature that writes to the repository switched off. It does nothing until the owner installs the app. Five issues are open for a first contribution: the French, German, Portuguese, Italian and Dutch dictionaries each need one afternoon from somebody who reads the language (#680 to #684). (#679)

  • The command reference documents every command. docs/commands.md had no chapter for six of the 55 commands trazum --help lists: init, where, schema, rules, filter and models. Each has one now, with output transcribed from a run, and command-reference.test.js holds the page to the CLI's own list: a command with no chapter fails, as does a link on the page to a chapter that is not there. Four commands (optimize, check, baseline, eval) stay under chapters named for the job, each held to the chapter that runs it. Found by checking the living docs against the CLI: every command and flag they mention exists, every model id is in the catalogue, and the package READMEs' examples import exported names. docs/setup.md and docs/running.md pin the Action at 2.7.11 where they said 2.7.9, and scripts/freshness.mjs no longer says a stale price warns "every user": it warns on reports that priced that provider. (#673)

  • The web app answers for /favicon.ico, an Apple touch icon and a web manifest. Measured on the deployed app on 2026-11-17, each of the three answered 404 and the page carried one icon tag, to icon.svg: a client that asks for /favicon.ico without reading the page got nothing, and a phone's home screen got a screenshot. The app now ships favicon.ico (16, 32 and 48 pixels), a 180-pixel touch icon and a manifest with 192 and 512-pixel icons and a maskable one, all of them the T of icon.svg, written by npm run draw:icon and held byte for byte by brand.test.js. The manifest declares display: browser: it names and draws a shortcut and does not claim an installed app. (#675)

Fixed#

  • The safe level no longer rewrites a thanks, a quotation or three ordinary words. Found by drawing the README's picture from a real run (#675): the shortened support prompt contained "So much for your help with this." and "In advance." The politeness rule deleted each courtesy phrase wherever it stood, and measured on 2026-11-17 that did three kinds of damage at the level that promises not to change what a prompt asks for. A fragment where a sentence was: "Thanks a lot for doing this." became "A lot for doing this.", and the same in Spanish, French and German. A broken instruction: "Tell the customer: thank you for your order." became "Tell the customer: for your order.", and "Thanks to the new policy refunds take five days." lost its first word. A changed instruction: "Speak kindly to the customer." became "Speak to the customer.", "Aim to please the customer." became "Aim to the customer.", and "If the user says thanks, reply politely." became "If the user says reply politely." A thanks is now removed only as a whole sentence that does nothing but thank, with what it is for, and is left exactly as written when it follows a colon, sits in a list, opens a clause or is what the sentence is about. One courtesy is deliberately no longer removed: a thanks hung on the end of a clause by a comma ("summarise it for me, thank you very much") cannot be told from the last item of a list of things to say, so it stays. "please" after "to", "kindly" anywhere but in front of what is asked, and German "Bitte" after a determiner are left alone. Text inside quotation marks is protected, as code, URLs and placeholders already were: Reply "Please hold on" used to become Reply " hold on", so the model was told to output a different string and the report counted the difference as a saving. ProtectionKind gains quoted. This also corrects a figure this project published about itself: docs/our-own-medicine.md said the rules recover 12 tokens from the four system prompts the package ships, credited to emphasis. All 12 were the rules rewriting "..." to "." inside the JSON example in two of those prompts. The figure is 0. On examples/support-prompt.txt the shortened prompt is now clean and the README's picture is redrawn: 326 tokens down to 202, $31.00 a month, and a first lever worth $416.99, 13 times more. (#677)

  • The README's pictures show what the CLI prints, in the product's own identity. docs/assets/demo.svg was typed by hand under a comment calling it a transcript, for a prompt that was in no directory of the repository. Its arithmetic was held to the price table and stayed right; the report around it moved on. On 2026-11-17 the CLI opened with its "Start here" line where the picture closed with it, suggested Claude Haiku 5.5 where the picture named Haiku 4.5, and assumed 300 output tokens where the picture said 500. It is now drawn by npm run draw:demo from a run on examples/support-prompt.txt, shows the first 26 of the report's 51 lines and says so in the picture, and demo-image.test.js fails when the file is not that output. The README's figures follow: 326 tokens down to 214, $28.00 a month saved by the rules, and a first lever worth $419.93, 15 times more. The same picture separated its columns with plain spaces, which an SVG shown through <img> collapses, so GitHub drew the transcript with its indentation and alignment gone; both transcripts now write no-break spaces and are as wide as their longest line. Both are coloured by the escape codes the CLI emits rather than by guessing from the shape of a line: the bill's share bars and its subscription caveat are drawn the way a terminal shows them. And the three pictures, with the editor extension's icon, were still in the palette and the mark the product had before DESIGN.md settled on acid yellow and a plain T, so the first screen of the repository, the editor's listing and the browser tab showed different products. All of them now take their colours and the T from scripts/brand.mjs, which brand.test.js holds to the web app's globals.css and icon.svg. docs/setup.md shows the web app from a production build, in the light or the dark theme to match the reader's; the previous screenshot was taken from a development server and had the framework's badge in its corner. The example prompt is one this repository's own gate reads, so trazum.baseline.json is re-recorded with it: three prompts, 877 tokens. (#675)

  • A batch call's cache reads on three Google models were billed at half of what Google charges. A batch call is priced at the row's batch ratio, and its cache reads at the row's cache ratio of that halved input, which is what most Batch tables print. Google's Batch tab for Gemini 3.1 Pro Preview reads "Context caching price: Same as Standard", $0.20 and $0.40 against batch inputs of $1.00 and $2.00, and the tabs of Gemini 2.5 Pro ($0.125) and 2.5 Flash ($0.03) print the standard figures. 2.7.11's notes named this as not changed. The three rows now carry multipliers.batchCacheRead: 0.2, the cache price over the batch input, and the bill, a session's cost, a truncated retry, a repeat, a conversation's growth and the usage-report connector all ask for the ratios of the call they priced. One batch call on 3.1 Pro Preview with 50,000 fresh input tokens, 100,000 read from the cache and 2,000 out was billed $0.072 and is $0.082. For the two 2.5 models the figure ends on 2026-12-01, when Google bills them at the rates of Gemini 3.8 Flash and 3.5 Flash-Lite, whose Batch tabs halve the cache price; 2.5 Flash gains a dated period for that, at the same $0.30/$2.50. The other six Google rows print a halved cache price and are unchanged, as is every other provider. A pricing overlay can set batchCacheRead. Read from Google's pricing page on 2026-11-17, for the Batch tabs' cache rows alone; Google's review date does not move. Seventeen planted faults were each caught. (#671)

Changed#

  • The npm token's expiry date is recorded: 2027-02-15. The token that published 2.7.11 was created on 2026-11-17 with ninety days to run, by its owner's account of it; npm shows the date only when a token is made, and this repository cannot ask for it. .github/release-credentials.json held null, which scripts/freshness.mjs could only print as a note. With the date, the weekly scheduled run goes red on main from 2027-02-01, fourteen days ahead, and scripts/prepare-release.mjs refuses to prepare a release from 2027-02-15. The token is counted as gone from the first minute of that day. Nothing installable changes. (#669)