Documentation#
- Align GitHub and npm around a no-global-install first report, keep advanced material expandable, and document the normalized-log plan/verify workflow.
- Distinguish API savings from subscription efficiency, with a context-growth, prompt-review and token-budget walkthrough that makes no quota promises, plus an optional compact agent-instruction template and evaluation method.
- Outline proposed task-level FinOps experiments that measure useful outcomes, quality and controller overhead across API and subscription workflows.
- Add a 30-day adoption playbook with interview prompts, explicit measurement denominators, a reproducible demo and an evidence-based case-study template.
Added#
Opt-in tool-log filter.
trazum filter compact --recovery-dir <directory>reduces consecutive identical display-log lines before agent context, preserving diagnostics and unique lines. The private local original is written before the recoverable view; failed storage and unsupported or oversized inputs pass unchanged.restorechecks the content hash;instructionsprints optional concise-output guidance. The Claude Code skill documents selective Bash integration with pipefail. No gateway rewrite, new process execution, network access, automatic interception or subscription-quota promise is added. Estimates include recovery overhead; the core transformation is browser-safe. The pipefail regression uses a literal shell command without environment-derived command arguments, and checks that stdout and stderr retain their channels.Local HTTP savings lifecycle.
POST /savings/planandPOST /savings/verifyaccept normalized usage text and portable plan JSON, price with the server catalogue, preserve incomplete evidence and expose OpenAPI 3.1 contracts. Strict input validation and the existing 1 MB body cap apply; no provider, filesystem or write capability is added.Native usage import over MCP.
import_usageaccepts bounded Claude Code or Codex transcript text and an optional explicit workload label. It returns normalized JSONL for profiling and savings plans, converter diagnostics and explicit evidence limits. Prompts, tool content and paths are excluded; session IDs remain. Empty output is not zero spend, imports cannot establish source completeness, and pricing remains API-equivalent rather than a subscription invoice. No filesystem, network or write capability is added.Local integration discovery.
GET /openapi.jsonpublishes the HTTP contract andGET /capabilitiesseparates native session import from generic CLI/MCP/HTTP access. An immutable browser-safe harness registry names supported formats without inferring billing, automatic interception or task quality.Agent workflow evaluation.
npm run test:agentsexchanges real CLI/MCP plans using synthetic usage, checks verification parity and preserves unknown results. The skill includes explicit paid-call boundaries and shared references.Acid neobrutalist interface. A typographic identity, local Space Grotesk, accessible color tokens and Light/Dark/System controls unify the dashboard, landing, playground and metadata. The bill adds summary metrics and section navigation while retaining its local analysis workflow.
Import explicit Codex rollout paths with
from-codex, CLI/web Bill detection, corroborated token deltas, separate conversion evidence and API-equivalent pricing caveats. Conversation content never enters normalized output.MCP savings planning and verification.
plan_savingsbuilds a CLI-compatible plan from normalized usage-log text;verify_savingschecks saved plan JSON against later usage. Both return JSON evidence, pricing gaps and assumptions, retain the three verification outcomes and accept optional pricing-overlay text. Inputs are bounded, unknown arguments rejected, and neither tool adds filesystem, network or model-call capabilities.Task-based documentation navigation. A quick start separates recorded usage from prompt projections; setup and integration recipes move out of the README while its existing section anchors remain available. The command reference gains a task map and chapter index, operational guides link their next steps, and historical plans are clearly separated from current usage. Account, licensing and scheduling descriptions now distinguish local CLI analysis from optional web storage; structured-output advice no longer promises that moving a schema to a parameter eliminates billed tokens.
The bill by month, in the documents and the browser. 2.5.1 gave the statement a table by month; the PDF, HTML and Markdown documents written by
billandprofilenow carry it too, every month rather than the newest twelve, with the part that has no timestamp said under the rows. The web app's Bill tab draws it above the daily chart: each month is a button that bills that month alone through the window, the newest twelve by default and every month on request.spendByMonthin@trazum/core. One function sums a report's days into UTC months and returns the undated remainder, so the months andundatedreconcile to the total. The statement now uses it too, and every surface that shows months goes through it, so no two can put a call in different months.
Fixed#
Harness identity is not a payment mode. Claude Code, Codex, Cursor and CI environment markers now report billing as unknown. Interactive harnesses keep a conservative token-only display default; explicit API cost estimates remain available with
--cost.Unknown model costs stay unknown in the dashboard. Unpriced model and workload rows no longer display zero dollars or zero percent as measured facts.
MCP reports its installed version during initialization. The handshake no longer announces 1.8.0 for a 2.5.1 package; a real stdio handshake test compares it with the package manifest, and the release checklist includes it.
Automatic Optimise demos cannot invoke paid passes. Automatic requests explicitly disable LLM rewriting and suggestions, even when those options were previously enabled. Manual requests keep the selected configuration.
Writer retains answers after failed requests. Answers commit only after a successful response, duplicate submissions are blocked while pending, and network failures are identified as connection errors rather than rate limits.
Bill clears an old workload filter when loading a new dataset. Uploads, pasted logs and tour samples no longer appear to cost zero because a project selected in the previous log is absent. Repricing and filtering the same dataset preserve its workload selection.
Gateway budgets now include usage from completed calls. The CLI adds process-local measured spend to its startup baseline and reserves concurrent estimates separately. Missing usage follows the selected failure policy; cache write TTLs are preserved for pricing. This is not a persistent or account-wide ledger. See gateway accounting.
A year of days drew the Bill tab's daily chart out of its card. Each bar kept a 2px floor and a 1px gap, so past about 160 days the bars ran across the column beside it. Long spans now draw without the floor and the gap, inside the card.
A window in the Bill tab had no visible way out once the log card folded. The report now names the window it shows and carries its own "Back to the whole log" button beside that line, named apart from the log card's Clear.
Changed#
- The Action pin moves to 2.5.1 in the README and
docs/running.md, the commit thev2.5.1tag points at.