v2.6.0

2.6.0 — Compare accepted tasks, keep the evidence

Adoption, read 2026-10-13: 1815 downloads of @trazum/cli in the last 30 days, 1 GitHub stars, MCP registry latest unavailable (The operation was aborted due to timeout).

Trazum can now group explicit Claude Code and Codex sessions into a task report,
show repeated tool results, and compare a later run against the same acceptance
criterion. The CLI supports review, reversible instruction application and
comparison; MCP adds analyze_task and compare_tasks. Neither interface runs
the task or decides that it succeeded.

npx -y @trazum/cli@2.6.0 task --help
npx -y @trazum/mcp@2.6.0

The task walkthrough covers the complete workflow.
The Tasks dashboard analyzes selected files locally and exports portable JSON
and instructions through preview, download, clipboard copy or text selection.
Reports retain coverage gaps and caller-declared outcomes. Repeated calls can
be deliberate verification; fewer tokens do not establish causal savings or
better quality. Truncated or incomplete evidence yields an inconclusive comparison.

The local HTTP API gains /usage/import for Claude Code/Codex transcripts.
filter compact --stats adds opt-in byte/token estimates without changing
stdout or the producer's exit status. Recovery rejects FIFOs instead of waiting
indefinitely for a writer.

Release validation now includes dense inputs at the 16-source/400,000-character
task ceiling, bounded reports, queued MCP calls with errors, and 24 concurrent
local HTTP requests with subsequent oversize refusal and recovery. The probes
also run against a packed installation. These are synthetic correctness checks,
not a hosted-service capacity promise. Evaluation at scale must include accepted
outcomes, retries, failures and controller overhead, not just token reductions.

No breaking change, new cloud service, automatic hook or default policy is added.
All packages, the plugin and MCP discovery advance together to 2.6.0.