Added

  • Delegation results arrive mid-turn, with live progress (ADR-88). Claude Code and Codex get finished delegations from a new PostToolUse hook (Codex asks you to trust it once in /hooks; managed-settings machines get it through one administrator prompt), omp from its extension's tool_result (omp 18.3.2 or later), pi as a steering message. Anything not taken that way still comes at the next prompt, and each result is still shown once. Running delegations report live progress (elapsed time, tool calls, last action, last line, tokens) in delegate / delegate_status and hnl run --list; delegate_status with since returns as soon as the run moves on, and delegate with notify: "progress" sends at most one progress update a minute. With nothing waiting, the tool-call hook exits after one directory listing (about 10 ms for the whole process).
  • Heavy accounts stay fast and bounded on the machine (ADR-88). Setup now restores only this machine's folders from the last 30 days, up to 500 MB, instead of every session in the account (hnl cloud pull-sessions --machine; --all shows the total and asks first, --yes to skip; settings cloud.restore.days, cloud.restore.maxBytes). The service keeps restored transcripts within that budget, least recently used out first, and logs what it frees; it never removes a file a harness wrote or a session you continued, and --all restores stay until you delete them. A restored copy is removed only after the cloud confirms it still holds that exact transcript; one the cloud has rolled out is kept and uploaded again. Copies the old setup pulled are recognised (by file details only, up to 2,000 per pass) and brought within the budget the same way. Gateway cache writes take about 0.1 ms on a large cache (were 100–300 ms), freed space returns to the disk without pausing requests, and gateway.compress.maxMB (default 128) caps the compressed originals. Every local database opens with WAL, a busy timeout and a bounded journal; a corrupt index is set aside and rebuilt, while stores holding unsent data are never rebuilt; a full disk pauses capture instead of failing requests or crash-looping, and a gateway that had to fall back to memory reopens its database every 5 minutes and keeps the usage it counted meanwhile. History substring search covers the newest 50,000 prompts and says when it stopped there. On a heavy account the service starts about 2.6× faster and peaks at about 314 MB instead of 720 MB.
  • The cloud stays fast as accounts and databases grow (ADR-88). Quota checks, /v1/quota, the dashboard's 10-second freshness check and the rolling storage caps read per-member counters instead of counting rows (the freshness check read up to 260,000 rows a call; it now reads 20). History, Memory and Runs browsing, filters, activity and totals read per-day counters; filtered and text views count the newest 1,000 matches and show "1,000+". Devices, Continue, session detail, org spend, team budgets, the token list (now paged) and the locked-sessions count read a fixed number of rows, and searches are scoped to your own data. The sync journal keeps only the newest change per row and idempotency keys expire after 30 days, which stops the tenant database growing with every gateway update. Hourly jobs (retention, cold storage, journal upkeep, a sweep for transcripts no row refers to, capacity checks) do bounded work per run; platform owners are emailed when a tenant database passes 70% or 85% of D1's 10 GB limit. Transient D1 errors are retried on sync writes without double-writing.
  • Download my data and backups work at any size (ADR-88). Large accounts get their archive built in the background and emailed as a signed link valid for 7 days (Settings › Account shows progress); small accounts still download directly. While an archive is being built, cold storage waits, so a month is never included twice or missed. Weekly backups run in bounded hourly steps with a manifest that proves they are complete, and owners are alerted if one is late or fails; before this, any table over 5 MiB could not be backed up. A backup in progress is resumed by the next weekly start (given up only after 28 days), tenant moves and cross-database merges wait for it, and a run whose source still holds unexported rows fails and alerts instead of claiming to be complete. Release deploys migrate every production database, refuse to ship if one is behind, and rebuild the memory search index after the new Worker is live so the brief overlap with the old one leaves nothing stale.

Update HarnessLink and every installed harness to its current release:

hnl update

New machine? Install with:

curl -fsSL https://harnesslink.sh/install.sh | bash

Newer: 17.14.0 · Older: 17.12.1 · All releases · Atom feed

Flag notifications