v0.10.2

Added

• Better subagent handling for Codex and Cursor (subagent tracking remains work in progress): Entire now consumes Codex's SubagentStart/SubagentStop hooks, and an agent can declare its own subagent transcript path instead of having the Claude Code layout guessed for it — which also fixes Cursor, which had parsed that path all along and dropped it. This is the in-session half only. Committed checkpoints still carry no per-subagent data for any agent, Codex does not aggregate subagent token usage, and only thread-spawned Codex subagents fire the hooks (#1958)
• entire repo create --object-format selects sha1 or sha256, validated client-side so a typo fails fast instead of coming back as an opaque 422 (#2043)

Changed

• The entire search TUI drops its Checkpoints tab and defaults to Sessions. Search now folds checkpoint hits into their owning sessions server-side, so the old default tab always showed zero results; a checkpoint stays reachable by ID via entire checkpoint explain (#2022)

Security

• redaction.openai_privacy_filter.command is honored only from a .entire/settings.local.json verified to be untracked. It becomes argv[0] of an exec on every push, and was read from the version-controlled .entire/settings.json with no validation — so a pull request could pair the setting with a payload committed alongside it and run code on every developer who pushes, including non-interactively in CI. A tracked settings.local.json is now ignored wholesale, since .gitignore does not apply to an already-tracked path. Rejection downgrades to resolving opf on $PATH rather than erroring, and the rewrite fails closed if that is missing (#2056)

Fixed

• Repo-scoped data could be read from the wrong region. Cell resolution scanned mirrors and silently fell back to the caller's home cell whenever a repo was mirrored in more than one region, so entire trail list --repo <multi-region repo> printed "No open trails found" instead of the repo's real trails. Resolution now goes through the repo's processing placement, and every failure is a returned error instead of a silent fallback. entire agent-help's trails probe and entire enable's post-success report had the same bug against the BFF, which does not proxy trails for CLI bearers (#2046, #2047)
• A repo that is not onboarded — or not visible to your login — is now named in one honest line across entire trail, entire experts, and entire checkpoint explain --repo, instead of an anonymous "This repository" on one surface and a raw internal resolution chain on the others. Single-repo cell resolution also tolerates the same catalog gaps the multi-repo fan-out already survived, and entire enable's trails probe gets its own deadline so a slow enable report can no longer starve it (#2054)
• entire trail update --body never worked: the description was sent as a field on the metadata PATCH, a route that does not serve body writes, and the rejection arrived as a bare Service Unavailable. Descriptions now go to the dedicated body route, and --body= clears one (#2055)
• Cloud-mode entire dispatch sends entire-cli/<version> as its User-Agent instead of Go's default, so its traffic is no longer invisible to server-side CLI version gating (#2049)

Housekeeping

• go-runewidth moves to 0.0.28 and the version pin from 0.10.1 is dropped — upstream builds its width table lazily now, so startup stays at ~18ms without pinning (#2040)
• Trail wire handling consumes entire-api's camelCase contract directly, dropping the remaining snake_case JSON and SSE normalization along with obsolete list-response compatibility fields (#2037)
• The Linux CI test jobs no longer apt-get install gnome-keyring: the token store has resolved to a temp file backend under test for some time, and the step's only remaining effect was to make three PR-blocking jobs depend on an Ubuntu mirror that intermittently hung them for the full 6-hour default timeout. Those jobs now also carry a 30-minute timeout (#2072)