Skip to content
Trazum

2.3.0

The warning that cried wolf, and a skill written for one agent

Added#

  • reviewedForModels(models, catalogue) in @trazum/core. The review date behind one report rather than behind the whole table: the oldest provider among the models actually priced, falling back to the catalogue's own date in the three cases where that cannot be established — an overlay in effect, a model the catalogue does not carry or carrying no provider, and a provider with no recorded date. The fallback is inside the function rather than at each call site, for the reason isOffered records: the fifth call site is always the one written with only the first half of a two-part rule.
  • pricing.reportReviewed and pricing.reportAgeDays in profile --json. The provenance of these figures, beside lastReviewed and ageDays, which go on meaning the table's own oldest provider. New keys are additions the contract allows; a key cannot change meaning under a minor, so neither did.
  • ## Through MCP and ## Before you spend, whichever door you have in the agent skill. The seven tools, the client-agnostic stdio wiring, and the one rule that decides whether any of it works — the server never opens a file, every tool takes the text.
  • The availability check can ask xAI and Moonshot. PROBES in scripts/check-model-availability.mjs covered five of the catalogue's seven providers; grok-4 and kimi-k2 were the two ids nothing in this repository could ask about, and the record said so in every run rather than leaving them out. Both probes are written now, so the moment a key exists the question is one command.

Fixed#

  • The stale-price warning named a date belonging to models the report never used. PRICING_LAST_REVIEWED is the oldest provider's, which answers how old is this table and not how old are the prices in front of me — and the CLI profile, the MCP report and the browser bill all decided staleness from it. The sentence they print says, in these words, that the table behind every dollar here was reviewed on that date.

    On 2026-08-31 that was false on a report of Claude and OpenAI calls: it named 2026-06-24, the date belonging to grok-4 and kimi-k2, whose providers no longer list them so neither price can be re-read and neither date can honestly move. The prices actually used had been read four days and zero days earlier.

    Two costs, and the second is worse. The figure was wrong, and a warning that fires on every run is one people stop reading — so the day the table really is stale, nothing has been said that was not said yesterday. trazum models had already worked this out and prints its dates per provider; the fix reached one surface and not the three that qualify a figure.

  • Three converters were documented as unbuilt for a whole arc after they shipped. The agent skill described from-litellm, from-helicone and from-langsmith as "named as next but not built", so an agent asked about a LiteLLM export offered to write a converter that had been there for releases. All five now sit in one table, with from-otel named as the one to offer when you do not know what somebody runs.

  • A refusal about the credential was being recorded as a refusal of the model, and it nearly retired a live one.

    xAI answers 400 "Incorrect API key provided" to a bad key — not 401 — and the first version of the probe above duly classified grok-4 as gone, with the provider "quoting" a sentence about the key. It was caught by running it with a credential that turned out to be a key id rather than a key: the script printed GONE grok-4 400, which is a confident wrong answer of exactly the shape this file exists to prevent.

    What it would have cost is worth stating, because "a script mislabelled a row" understates it: pricing.ts copies that sentence verbatim into retired.because, retired is a stronger statement than recommendable: false and is honoured everywhere that one is, and the reports quote it. A typo in an environment variable would have retired a working model, in the provider's own words, on the record, and every surface downstream would have repeated it.

    The status alone cannot decide it, so what decides it is what the provider said: an answer naming the key is about the key, whatever number it arrives with. Those are now recorded as unasked, which is where an unasked question belongs — the same rule the rate-limit branch already followed. The match is deliberately broad: a false "could not ask" costs a re-run once somebody fixes the key, and a false "gone" costs a retired model that still works.

    Held at the record rather than at the classifier, by model-availability.test.js: no refused entry may quote a credential error, and no retired.because may either. A probe for a provider nobody has thought of yet inherits the guard; a unit test on one function would not have given that. Proved by planting the exact 400 this was found from and watching three assertions fail by name.

  • The README's Action pin was two releases stale, and CI said so before a reader could. It sat at 2.1.0 while 2.2.0 and 2.2.1 shipped; docs/releasing.md allows exactly one release of lag, because the pin can only advance to a release commit once that commit exists. It now points at 2.2.0.

    Worth recording how it was caught, because the guard is younger than the habit it protects: this failed only in CI, and passed locally, because a fresh checkout has every tag and a working copy has whatever it last fetched. The check derives the allowed lag from git tag rather than from anybody remembering, so a stale tag list is a quieter guard rather than a louder one — and reproducing it locally meant fetching tags first, which is now the first thing to try when a security.test.js failure will not reproduce.

  • api.x.ai and api.moonshot.ai are declared in trusted-hosts.test.js under a kind of their own — availability check, not a customer path — rather than as gateway upstreams. The gateway fronts calls a customer makes, and nothing routes customer traffic to either; naming the distinction is what keeps it true by inspection. That guard is what caught them, which is the second time this release it has been the thing that noticed.

Changed#

  • The agent skill is written for any agent, not for one. Every command is spelled trazum <command>, and one table says how to spell that for whichever door the reader has: a shell in this checkout, a shell anywhere else, MCP tools and no shell, a library import, or a browser. It used to open on npm install and repeat node packages/cli/dist/index.js fifteen times, which is a document about a shell handed to readers who do not have one.

    The derivation to the plugin copy is one table row instead of fifteen paths as a result, and scripts/build-plugin-skill.mjs says so.

Guards#

  • Three over the skill, all derived, because its tables are hand-written and that is the shape of every defect this repository has found in itself. Every command the skill tells an agent to run is one COMMAND_FLAGS dispatches; the tool table under ## Through MCP equals the set the MCP server registers, in both directions, because a tool the skill omits is one an agent with no shell will never call; and every from- converter that exists is mentioned somewhere, which is the false claim above pinned from the other side. All three plants fire.

    The tool check began by scanning the file for snake_case and needed an exclusion list to stop flagging ANTHROPIC_API_KEY and pull_request_target — two lines after a comment about how this repository keeps paying for exclusion lists. It reads the table now, bounded with sectionOf() rather than by naming the heading that follows, which publish.test.js fails a test for doing and which this test did until it ran.

  • reviewedForModels is held from four directions in pricing.test.js, including the property that makes its fallback safe to rely on: whatever it returns is never older than the catalogue-wide date, for every model in the catalogue.

  • stale-pricing.test.js was asserting the defect. It derived staleness from PRICING_LAST_REVIEWED while its fixture priced one Anthropic model. It derives it from the fixture's own model now, and gains the case that pins the distinction — a clean log and one carrying grok-4 — skipped rather than faked on a day the two dates agree.

Prices#

  • Nothing moved, and two dates could not. OpenAI was read on 2026-08-31 and Anthropic, Google, DeepSeek and Mistral within four days of it, all inside STALE_PRICING_DAYS. xai and moonshot stay at 2026-06-24: grok-4 is absent from the model list xAI's own docs serve, and kimi-k2 from the llms.txt Moonshot publishes, so there is no price to re-read and moving the date would say a check happened that did not. Neither is marked retired — that needs the provider refusing a real request in its own words, and an absence from a page has refused nothing.