v0.2.2 — auto-pair: kill the pairing step for new agents

One-tap onboarding: new telegram agents auto-pair, no "message the bot, paste a code" step

A customer's Telegram operator id is the same across their whole fleet, so 5dive now learns it once and auto-pairs every new agent.

• Shared operator allowlist (/var/lib/5dive/operator-allow.json): learned from any agent pair --user-id and from the CoS mint event (agent cos claim now returns the bot creator's id, per Bot API ManagedBotUpdated.user).
• agent create (telegram, no explicit --telegram-allowed-users) seeds the new agent's allowFrom from that store, so the bot accepts the operator's DMs immediately. Reports autoPaired in its result.
• Dashboard skips the pairing step when autoPaired is true (guarded — a bot that could not auto-pair still falls through to manual pairing, never left unreachable). CoS onboarding is now: create bot in one tap → done.
• Per-agent access is still fully customizable via the Telegram-access modal.

Builds on 0.2.1 (CoS managed-bot onboarding + cos set-avatar). Verified: test-vm smoke green; on-box end-to-end (create with no allowlist → autoPaired + access seeded from the store). Managed boxes update nightly; self-hosted: 5dive self-update.

v0.2.1 — fix CoS dashboard onboarding + per-agent bot avatars

Fixes the 0.2.0 CoS dashboard onboarding (it never rendered) + adds per-agent bot avatars

Bugfix (DIVE-320): In 0.2.0 the new-agent wizard's "Create a bot automatically" path never appeared — it silently fell back to pasting a token. agent cos verify/mint-link emitted a bare {ok:true,username,...} JSON, but the dashboard unwraps the standard {ok:true,data:{...}} envelope, so the result read as empty and the probe failed. Both subcommands now return the standard envelope. (claim/rotate stay bare — they're consumed by the create-path, not the dashboard.)

Feature (DIVE-325): 5dive agent cos set-avatar --agent=<name> --avatar=<png> sets an existing agent's Telegram bot profile photo (works for any telegram agent). Surfaced in the dashboard as a per-agent "Bot avatar" upload in the agent settings tab.

Managed boxes pick this up on the nightly refresh; self-hosted: 5dive self-update. The dashboard half ships via the web app. Note: the dashboard CoS onboarding + avatar upload require the box CLI to be >= 0.2.1.

v0.2.0 — Chief-of-Staff managed-bot onboarding (zero-token agents)

🚩 Flagship: zero-token agent onboarding via Chief-of-Staff managed bots (DIVE-320)

The headline of 0.2.0: spinning up a Telegram agent no longer requires pasting a BotFather token. A customer-owned Chief of Staff (CoS) bot (Bot Management Mode on) creates and manages a dedicated child bot per agent on the customer's behalf, via Telegram Bot API 9.6 "Managed Bots".

CLI

• 5dive agent cos set|verify|mint-link|claim|rotate — manage the CoS bot and mint/claim/rotate per-agent child bot tokens.
• 5dive agent create --telegram-cos=<child-username> — claim the CoS-minted child token at create time instead of --telegram-token (the paste path stays as a fallback). Auto-configures the child bot's name + description on claim.
• Per-customer CoS only (no central manager = no fleet single-point-of-failure); a leak is contained to one customer's fleet.

Dashboard

• New-agent wizard now offers a one-tap "Create a bot automatically" path: it detects an available CoS, mints the deep link, and provisions the agent once you create the bot in Telegram. Pasting a token still works.

This kills the top onboarding-friction step and the "my agent won't reply" tickets that came with mis-pasted tokens. Unblocks per-agent bot avatars (shipping next).

Also since v0.1.84

• Runtime GitHub-org resolution for the 5dive-com to 5dive-ai rename, plus migration of existing agents' marketplace refs (0.1.87, 0.1.88).
• agent auth status --auth-profile scopes the credential check (DIVE-296).
• openclaw auth tracks the 2026.6.6 flag rename (DIVE-297).
• Per-agent ~/.local/bin/claude shim at agent create (DIVE-283).
• README quickstart polish.

Managed boxes pick this up on the nightly refresh; self-hosted: `5dive self-update`.

v0.1.84 — first-run hardening + team-bot auto-attach (catch-up 0.1.78→0.1.84)

First-run hardening ahead of tonight's Show HN, plus the shared-team-bot work. Catch-up release covering 0.1.78 → 0.1.84.

Fixed

• 5dive init no longer dies on a fresh host with "bun not on PATH" (DIVE-265). All five channel-plugin prechecks now self-heal: if the agent user can't see bun, the CLI installs it system-wide (BUN_INSTALL=/usr/local) and only fails if that install fails. Found minutes after it bit a real first-run test.
• init fails fast without root — previously an unprivileged run let you answer the whole wizard before dying at agent create; now it exits up front with the sudo 5dive init hint.
• agent config set channels=telegram|discord stages the plugin synchronously (DIVE-250) — a bare channel attach with the token already on disk could restart into a session with no channel tool.
• agent list/agent info survive an absent per-type runtime config (DIVE-230) — a --defer-auth antigravity agent no longer kills the whole listing.
• Task-board jq: Argument list too long on big boards (DIVE-222).

Added

• agent list --json carries model + effort per agent (DIVE-211) — one call instead of an N×agent info fan-out.
• Shared team bot auto-attach (DIVE-248): new agents join the team group with their own forum topic automatically; --no-team-bot opts out. team-bot discover finds the group id itself (DIVE-247).
• Task gate alerts follow the conversation (DIVE-259) — pings land in the chat where the human actually is.

Full detail in CHANGELOG.md.

v0.1.77 — agent stats --all (batched) + DeepSeek v4-pro

• agent stats --all --json: every agent's stats in ONE invocation, so the dashboard/mobile collapse N per-agent box execs into a single call (the box shelld rate-limit is a shared 60/min per-box ceiling). Single stats <name> unchanged. (DIVE-206 / DIVE-202)
• DeepSeek provider maps: deepseek-chat → deepseek-v4-pro (deepseek-chat is deprecated 2026-07-24; v4-pro verified current).

v0.1.76 — team bot + heartbeat reliability + recurring tasks (catch-up 0.1.56→0.1.76)

Catch-up release: GitHub releases had stalled at v0.1.55 while main and the fleet advanced to 0.1.76. Boxes were never behind (they update via the nightly pull of main); this republishes the tag line. Spans 0.1.56 → 0.1.76.

Team bot & intercom

• One shared @FiveDiveBot for a whole fleet: forum-topic-per-agent in a team group, no-bot agents relayed through the shared bot (DIVE-159).
• Per-agent bot-to-bot write path + botToBotEnabled surfaced in agent list (DIVE-161).
• Intercom mirror: a dedicated topic mirroring inter-agent chatter, membership-based retarget, sender labels on shared-bot agents (DIVE-195).

Heartbeat reliability

• Orphan reclaim, no-clobber idle gate, wake-on-enqueue; same-account agents spread so they never start together (DIVE-131/132).
• Native claude agents --json for idle/blocked detection.
• Recurring-task materializer runs in the heartbeat tick (DIVE-138).

Tasks

• Recurring task kind + create path; terminal tasks dropped from the human inbox; in_progress clock resets on reassignment.
• Tap-to-answer buttons for codex/grok/antigravity gates (DIVE-118); task need --recommend; terse done/cancel pings (DIVE-148/150); unblock keeps a still-pending human gate blocked (DIVE-109).

Fleet / install / import

• team-import v1: 5dive.yaml v2 schema + role wiring + export/import (DIVE-98).
• CC downgrade floor via managed-settings requiredMinimumVersion (DIVE-133).
• Removed the orphaned gemini launcher type ahead of CLI sunset (DIVE-193).

v0.1.55 — tap-to-answer buttons on task need ping (DIVE-117)

Telegram inline buttons on the DIVE-105 gate ping: decision options + approval Approve/Deny, tap to answer. Gated to claude agents (plugin 0.4.59+ handles taps); others keep the text ping. Value re-resolved from DB on tap; text-only fallback if a keyboard is rejected.

v0.1.54 — instant Telegram ping on task need (DIVE-105)

Instant Telegram ping when an agent files a human gate via 5dive task need (DIVE-105). One DM to the paired human — 🙋 [DIVE-N] needs you: <ask> — dashboard-primary, with a task answer tail. allowFrom DMs first, forum-topic fallback. Self-gating: a failed notify never blocks the gate write.

v0.1.53 — Human Task Inbox (CLI)

DIVE-103: Human Task Inbox CLI data layer.

• 5dive task need <id> --type=decision|secret|approval|manual --ask="..." [--options=A|B] — park a task on a human gate.
• 5dive task inbox [--json] — list only tasks waiting on a human.
• 5dive task answer <id> --value="..." — record the answer, recompute status, ping the owning agent to resume.

5 additive columns (downgrade-safe migration). Secrets are never stored in the shared db and never embedded in the resume ping. Inbox decoupled from task-block status. Gated via independent ship-gate review.

v0.1.52

Added

• 5dive agent config <name> set effort=<low|medium|high|xhigh|max> — parity with set model=. Claude-only reasoning-effort setter (writes effortLevel to the agent's settings.json, same key the telegram plugin's /effort writes), validated against the five levels, errors clearly for non-claude types, applied via the existing deferred restart. xhigh/max are Opus-tier (Sonnet caps at high).
• 5dive agent info now surfaces effort (model · effort <level> in text, effort field in --json).