Fixed#
Task exports remain accessible when downloads are blocked. Report JSON, comparison JSON and proposed instructions open a review panel with download, clipboard copy and selectable text. Download requests are not reported as saved files; denied clipboard access leaves the content selected for manual copying. English/Spanish feedback, keyboard dismissal and focus return keep the workflow usable without uploading report contents.
Recovery files cannot stall the tool filter. A FIFO in the recovery store is rejected without waiting for a writer.
restorefails without emitting content;compactpasses the original through and reports zero savings.
Added#
Task efficiency in the dashboard. The Tasks tab reads explicitly selected Claude Code/Codex sessions entirely in the browser, shows repetition evidence and coverage gaps, exports core-compatible JSON and optional instructions, and compares independent reports with declared acceptance. English/Spanish copy, bounded UTF-8 imports and a clear-workspace action keep the workflow local. No task execution, automatic policy application or causal savings claim.
Task efficiency across explicit agent sessions.
trazum taskand MCPanalyze_task/compare_tasksconnect Claude Code/Codex usage, repeated calls with matching recorded results, explicit acceptance and a fixed opt-in instruction policy. CLI apply/revert manages only its exact block in a selected existing file. Reports exclude raw tool content, paths and session IDs; source/line references and non-cryptographic reuse checks retain reviewability without authenticating or anonymizing evidence. Paired comparisons retain quality regressions and inconclusive results; repetition is not proof of waste and lower usage is not causal savings. No automatic hooks, task execution, model calls or quota forecast are added.Native usage import over local HTTP.
POST /usage/importaccepts explicit Claude Code or Codex transcript text and an optional label, returning normalized JSONL, conversion diagnostics and evidence limits for the savings workflow. Request and complete response each have a 1,000,000-byte UTF-8 cap; overflow returns400, never partial data. Session IDs remain; prompt, tool and path content is excluded. Empty output does not establish zero usage, and separate requests do not share deduplication state.Tool-filter FinOps measurements.
trazum filter compact --statsemits opt-in JSON on stderr with exact byte counts and net heuristic token savings for the view actually emitted, including recovery overhead. Failed recovery storage reports zero savings; unsupported token counts remain unknown. No log content, recovery path or hash enters the report. Stdout and producer exit status remain unchanged. These are per-invocation estimates, not billed savings, subscription capacity or task-quality measurements.
Verification#
- Synthetic stress probes now run in the agent workflow suite: the 16-source, 400,000-character task ceiling, bounded findings with inconclusive comparisons after truncation, CLI/core parity, 16 queued MCP calls with interleaved errors, and 24 concurrent local HTTP imports followed by oversize refusal and recovery. CI and the release workflow also require these probes against tarballs installed outside the workspace before publication. These tests establish bounded correctness, not a hosted-service throughput promise.