Skip to content
Trazum

1.77.0

The agent's bill, told honestly

Added#

  • A dated id prices as the model it is. The 1.77 plan: ModelPricing gains aliases, and the catalogue indexes every alias beside its id, so claude-haiku-4-5-20251001 — the canonical API id, while the catalogue lists the short form a person recognises — stops landing in unpriced. Declared one line at a time, never derived: a rule that stripped anything resembling a date from an id would be a machine guessing that two ids bill alike, which is the one guess this product does not make. An id that merely resembles a known one is still unknown, and the guard says so in both directions.

  • stop_reason, carried at last. The transcript records it on the assistant message and the converter never read it, so truncation, the retry bill and the coverage gate all answered "cannot be measured" about a log whose own source knew. Read and emitted now, and nothing is inferred: a turn without the field still produces a record without it.

Changed#

  • A folder of projects is a folder of workloads. For a directory target, from-claude-code labels by project folder by default, with --no-label-from-project to decline and --label still winning. A single file is untouched: one file is one workload only if the caller says so. The flag has existed since the folder walk did and nobody found it, which is why a real forty-day run produced a label on 0 of 10,393 records and the report could then only describe a mixture. The web app's folder drop has labelled by project since 1.70; the two surfaces disagreed about the same gesture.

  • The label is the decoded project name. Claude Code names those folders by encoding the project's absolute path, / becoming -, so -Users-mac-Trazum is a path wearing a costume. The last segment of the encoding is the project's own directory name, which is the word a person would have chosen; the raw folder name stands whenever decoding would leave nothing.

Fixed#

  • <synthetic> is not a model. Claude Code writes it for turns it produced locally — interrupts, error notices — which carry a usage object of zeros that no provider ever billed. Priced, they were noise in the totals; they are excluded by name now, counted, and reported. Matching the exact string rather than anything in angle brackets is deliberate: a pattern that swallowed an oddly-named real model would delete spend from a bill without saying so.

Changed#

  • No em-dash reaches a visitor. Two hundred and seventy-one of them removed from copy somebody reads: 123 in the English dictionary, 94 in the Spanish, 45 across the landing page's five languages, nine in components and route bodies. Not a find-and-replace: an em-dash does four different jobs and each wanted a different repair, and about thirty strings had to be reshaped into two sentences because the alternative was a comma splice or a second colon in a line that already carried one. Spanish got Spanish punctuation rather than translated English punctuation. Scope is what a visitor can see; the em-dashes in code comments stay, because they are source rather than app. Verified from the rendered page rather than from a grep: the landing page and the Spanish app were loaded in a browser and their text searched. The whole suite stayed green without touching a single assertion, which is what "match on ids, not on prose" buys.

Fixed#

  • A sign-in that was built, working and invisible. With no GitHub app configured the header renders nothing, which is the honest rendering for a visitor but left the operator with no signal at all: no button, no error, no line to read, and therefore no way to tell a deployment deliberately running anonymous from one misconfigured. authConfig already computes a one-line reason for exactly this moment and /api/auth/session dropped it on the floor; it is returned now. It names environment variables, never their values, and those names are in the public documentation. The header still renders nothing: absence remains the right answer for a visitor, and only the endpoint an operator can curl gained one.

  • The CLA gate, corrected by the API rather than by reasoning. Its first live runs failed three pull requests, and each failure taught something the previous reading had got wrong. The action does not create the branch it writes signatures to, so cla-signatures had to exist first. Then: it resolves a committer as commit.author.user || commit.committer.user || commit.author || … and compares login || name against each allowlist pattern exactly, case included. The source suggested the capitalised commit name, because the agent's commit email looked like one no account would own — but GitHub resolves noreply@anthropic.com to the account claude, so the lowercase login is what arrives, and Claude !== claude. Asking the commits API what identity it actually reports is what settled it. Both spellings are listed now: the login for today, the commit name for the day that email is unlinked.

  • A default nobody chose. lock-pullrequest-aftermerge is true unless set, which locks the conversation on every merged pull request. It is explicitly false: a merged PR is still where somebody asks why a change was made.

Guarded#

  • The allowlist is pinned to the identities that actually commit here, and the locking value must be stated rather than inherited (security.test.js). A misconfigured gate does not degrade gracefully — it blocks every pull request and blames the contributor rather than the line that is wrong. Verified as a detector: removing an identity turns the test red.