hivemind v1.1.0

Features

• hivemind usage — new command for server-backed token and cost reporting. Supports daily/monthly/session granularity, date ranges, per-agent and per-model breakdown, and --json output. Covers all machines you sync from, not just the local one (#959).
• Session labels — tag sessions with arbitrary KEY=VALUE pairs, links, and freeform tags via hivemind tag. Labels in the user.* namespace are yours and survive re-processing; hivemind.* labels are derived automatically from session data (#962, #972).
• Auto-tagging skill — hivemind install-agent now also installs a hivemind-tag skill into Claude Code, Codex, and Cursor so agents can label their own sessions. Run hivemind config set tag.nudge true to have a SessionStart hook point every new session at the skill automatically (#990, #995).
• daemon.default_repo — set a fallback repository slug (in config or via HIVEMIND_DEFAULT_REPO) for sessions that run outside a git repo, e.g. sandboxes or /tmp checkouts. Never overrides a real repo detected by git (#792).
• hivemind serve upgrade — upgrades a self-hosted stack to the latest (or a pinned) release, waits for the app to return healthy, and warns if the CLI is older than the target server (#810).
• hivemind serve setup --config — non-interactive setup from a JSON config file, enabling infrastructure-as-code deployments. Re-running setup now preserves JWT_SECRET and HIVEMIND_SECRET_KEY instead of rotating them and invalidating every issued token. Terraform modules for GCS are included (#927).
• Disable the daemon — hivemind config set daemon.disabled true pauses all syncing without uninstalling the service. The service stays loaded (so launchd/MDM presence checks stay green), hivemind status shows Disabled, and upgrade scripts don't fail (#1125).
• Multi-agent workflow tracking — sessions spawned as Claude Workflows, Codex collaboration runs, or cross-agent MCP hand-offs are linked to their parent session and visible in the timeline (#876, #899, #964, #996).
• Claude thinking tokens — extended thinking tokens (including those from sub-agents) are now tracked and attributed correctly (#1042).
• Skills leaderboard — skills are a first-class concept with per-session tracking and a leaderboard view on the dashboard (#743).
• GitHub Enterprise — the daemon authenticates against GHE instances for PR attribution and org membership (#926).
• Google Vertex AI — available as an LLM and embeddings provider for self-hosted setups (#925).
• Tailscale integration — self-hosted installs can expose hivemind over Tailscale with automatic HTTPS, no public domain required (#754).
• Forager improvements — Forager's remote MCP server enables agent-to-agent session delegation; --org now accepts readable slugs and is optional when only one workspace exists (#1098, #1131).
• New model pricing — Opus 5.5, GPT-6 Sol, and GPT-6 Luna added at launch rates; offered in Forager model selection (#1120, #1142).

Agent Adapter Fixes

• Cursor — recovers dropped reasoning steps and the final assistant response after a tool call (#768); discovers sessions from the composerHeaders table so new-format databases aren't missed (#887); restores cost enrichment with TTL jitter and SCIM email-alias lookups (#895); fixes double-counted usage snapshots from evolving in-progress records (#1016); fixes subagent cost attribution (#955); detects session changes via header fingerprints (#956).
• Codex — recognizes the exec tool so skills, commits, and PRs resolve correctly (#1000); surfaces tool-result images and stops mangling tool output (#999); stops charging reasoning tokens twice (#1033); detects compaction events in the current harness (#1052); resumes syncing correctly after Codex rewrites rollout files in place during background_paginated_rollout_migration (#1143).
• Claude — stops counting CLI-rendered artifacts as conversational turns (#880); workflow-spawned sub-agents are now captured and displayed (#808).
• OpenCode — reasoning tokens are counted as disjoint from output tokens, not additive (#1045).
• General — stops pricing sessions against model names no rate table recognizes, falling back to a safe default instead of a silent mis-price (#1044); fixes diff extraction that treated a partial Read as the file's full content, producing spurious diffs (#979).

Cost & Pricing

• Cost is now attributed to the time it was incurred rather than session-end, so daily spend totals line up with activity (#844).
• Session cost breakdown now reconciles with the session total for agent-billed sessions (#903).
• Cursor sessions use Cursor's billed cost from the Admin API instead of a local estimate (#764).
• Session tier (free/pro/etc.) is derived from the full sub-agent spend tree, not just direct cost (#986).
• Fixed model-name normalization so Cursor sessions are no longer priced at the _default fallback rate (#939).
• Lazy-load pricing tables so the daemon CLI starts without loading the full rate table into memory (#987).

Reliability & Performance

• Daemon CPU at idle cut dramatically: session log reads now use a sparse line-offset index (28× faster for large logs); resolve_git_repo is memoized with a 60 s TTL (593× faster); state is only written every five minutes or on shutdown instead of every sync cycle (#1093).
• OpenCode no longer pins a CPU core — the watcher no longer recurses the full database directory on every write; only sessions active in the last 15 minutes are re-synced per cycle (#1123).
• Cursor sync no longer spikes CPU on large databases — legacy bubbles without timestamps are now fingerprinted so unchanged history isn't re-queued; Cursor's SQLite key index is used instead of a full table scan (#950).
• Codex polling on macOS uses FSEvents for fast detection of new sessions, falling back to polling only live files; idle databases go cold after 6 h (#1123).
• Daemon upgrade prompt targets the latest stable release rather than just the minimum acceptable floor (#729).
• Self-hosted systemd service uses Restart=always on the stable-launcher ExecStart so auto-upgrades are self-healing (#837).
• MCP dependency bumped to 1.28.1 to address upstream security advisories (#873).

Other Fixes

• hivemind api no longer crashes with UnicodeDecodeError on binary responses (images, downloads); --raw round-trips bytes correctly (#1001).
• hivemind doctor no longer incorrectly flags package/cask daemon installs against brew services (#821).
• OIDC-federated daemons no longer lose credentials on disk between restarts (#984).
• PRs created through the Codex GitHub connector are now detected for PR attribution (#951).
• Remote-controlled Claude sessions (web API) are no longer synced, as they don't represent local agent work (#983).
• Dev-mode re-exec message is sent to stderr so it doesn't pollute piped output (#812).