v0.1.0

First stable release. The whole ecosystem is aligned on this version: the core
CLI, the shared chatbot library, and every tagged agent / channel plugin now
carry v0.1.0, so costaff update --all --tag v0.1.0 pins a known-good set.
Functionally this is beta-3 plus the fixes below; plugins that were still on
v0.1.0-alpha-2 are re-tagged without code changes.

Fixed

• Fresh installs no longer break on mcp 2.0. requirements.txt asked for
  mcp>=1.9.4 with no upper bound, so any install done after mcp 2.0 shipped
  resolved to 2.0.x — which removed mcp.server.fastmcp, the import that
  mcp_servers/setup.py is built on. Every mcp_servers module failed to
  import, taking the mcp container image and the test suite down with it. The
  bound is now mcp>=1.9.4,<2, matching the release line
  agents/requirements.txt already pinned.
• WebChat async-push auto-wiring no longer breaks Enterprise stacks.
  costaff channel add/rebuild on a channel whose name contains "webchat"
  now reuses an existing WEBCHAT_ENT_INTERNAL_SECRET as
  WEBCHAT_INTERNAL_SECRET instead of minting a fresh one. The notifier
  prefers the WEBCHAT_* pair, so a fresh secret meant the sender and the
  Enterprise receiver (which validates against the _ENT_ value) disagreed
  after the next mcp recreate — async push silently died.
• costaff update now flags services/ changes for a core rebuild.
  services/ is baked into the mcp image and imported at container runtime
  (mcp_servers/tools/_shared.py → services.config); it was missing from
  _CORE_IMAGE_PATHS, so a services-only update ran stale code in the
  containers after a plain restart.
• Legacy coding_agent_enabled migration no longer fabricates a live
  agent. Any config read (get_config) used to turn the bare legacy flag
  into an ENABLED costaff-agent-coding entry pointing at a default
  container name — on hosts where that container never existed, the entry
  flowed into EXTERNAL_AGENTS_CONFIG and the Manager routed coding tasks
  into a dead endpoint. The migration now only keeps the agent enabled when
  an explicit CODING_A2A_URL/CODING_A2A_INTERNAL_URL proves real wiring;
  a bare flag migrates the entry as DISABLED (visible in costaff agent list, never routed) with a warning, and when a modern coding entry
  already exists the stale flag is dropped instead of registering the same
  agent twice.

v0.1.0-beta-3

Added

• Async task results reach WebChat OSS, not just Enterprise. Delivering a
  finished background task ("I'll notify you when it's done") is now a shared
  channel capability. The core prefers the generic WEBCHAT_PUSH_URL /
  WEBCHAT_INTERNAL_SECRET so each stack points at its own webchat container,
  falling back to WEBCHAT_ENT_PUSH_URL / WEBCHAT_ENT_INTERNAL_SECRET so
  existing Enterprise deployments are unaffected. Pairs with WebChat OSS
  beta-3, which receives these pushes over SSE.
• costaff channel add/rebuild webchat auto-wires async push — zero manual
  setup. Adding or rebuilding a WebChat channel now generates
  WEBCHAT_INTERNAL_SECRET (once, then preserved) and derives
  WEBCHAT_PUSH_URL from the core's container prefix, writing both to the core
  .env. The webchat container already mounts that .env, so the two sides
  line up automatically — "notify you later" works after a plain rebuild +
  costaff restart, without anyone hand-editing env files. Other channels are
  left untouched.

Fixed

• channel add/rebuild and agent add no longer break on a secondary core.
  The compose-fragment generator appended the core's docker network to
  whatever the source docker-compose.yaml hardcoded (typically
  costaff_default) instead of replacing it. On a non-default core (e.g. twk
  → costaff_twk) the service then referenced costaff_default, which the
  fragment doesn't declare, so the build failed with "refers to undefined
  network costaff_default". The service now references exactly the core's
  network. No change on the default core (the network already is
  costaff_default).
• costaff update now prompts a core rebuild when core/ changes. core/
  (notifiers, dispatcher, models, license) is baked into costaff-mcp-costaff
  via COPY . ., but it was missing from the update command's "needs a
  rebuild" path list — so an update whose only image-relevant change was under
  core/ (e.g. this release's async-push sender) would be silently left
  running the old code after a plain restart. core/ is now in that list, so
  the update offers costaff core-rebuild like it does for mcp_servers/ etc.

v0.1.0-beta-2

Changed

• costaff channel is multi-core aware. add/list/remove/rebuild/tags
  now accept --core and resolve the target core's config, container
  prefix, workspace, .env, and compose project — matching costaff agent. Channel host ports are reserved across every registered core so
  allocation never collides. Single-install hosts are unaffected (the
  synthetic default core reproduces the historical layout exactly).
• costaff agent enable/disable/transfer recreate the Manager. They
  changed config and regenerated the env but only told the user to
  docker restart — which doesn't re-read env_file, so the change
  silently didn't apply. They now recreate the Manager like add/remove.
• Dashboard CORS is deny-by-default. ALLOWED_ORIGINS no longer
  defaults to * (which let any website script the dashboard). The
  bundled frontend is same-origin and needs no CORS; set ALLOWED_ORIGINS
  only when hosting the frontend on a different origin.
• costaff database backup/clean/restore fail friendly without a DB.
  They dumped a raw AttributeError traceback when no database was
  configured; now they print the same actionable message as db migrate
  and exit non-zero (restore also exits non-zero on a missing file).

Added

• Durable notification outbox. A channel push that fails (Telegram 5xx,
  WebChat secret unset, network) is now written to notification_outbox
  and retried by a background loop with exponential backoff, marked dead
  only after 8 attempts. A task result is no longer lost because a single
  push happened to fail. (migration 0003_notification_outbox)

Changed

• Dashboard passwords use PBKDF2-HMAC-SHA256 (600k iterations) instead
  of a single SHA-256 round. Existing auth.json records upgrade
  transparently on the next successful login. Token and password
  comparisons are now constant-time.
• costaff agent remove stops and removes the agent's containers (like
  channel remove / platform remove) and recreates the manager, instead
  of leaving an orphaned container holding its host port so the next
  agent add fails to bind.
• costaff update detects core-image changes (anything under
  mcp_servers/, migrations/, agents/, requirements.txt,
  Dockerfile) and guides — or, on a TTY, runs — costaff core-rebuild,
  because a plain restart recreates containers without rebuilding and
  would run the old code / skip a new migration.

Fixed

• Task lifecycle — no more stranded tasks. Startup orphan recovery now
  reaps EVERY doing row unconditionally (executors die with the process;
  the old 30-minute age gate stranded tasks when the container restarted
  within 30 minutes of a task starting), plus a periodic sweep catches
  workers lost mid-life. Reaped tasks now notify the user, finalize the
  live progress panel, and advance the agent's queue. A failed upstream
  task now cascades failed down its dependency chain instead of leaving
  downstream tasks spinning in queued forever.
• Per-agent serialization enforced in the executor. A second task for
  an agent that is already running one is deferred back to the queue —
  concurrent MCP sessions on one sub-agent trigger the anyio cancel-scope
  race, so this is now a structural guarantee instead of a convention.
• costaff agent model really persists for external agents. It now
  writes the agent's plugin .env (which wins in env_file order), backs
  the env-var names into new agent add entries, recovers them from the
  manifest for pre-v0.1.0 entries, and exits non-zero instead of printing
  a fake success when the agent has no model surface (url-type).
• send_message_now to WebChat actually sends. The webchat branch
  used to return "Sent." without delivering anything; it now pushes
  through the same channel endpoint the notification dispatcher uses.
• costaff update --all is no longer a silent no-op. Called directly
  (not via Typer), the rebuild helpers received the --core OptionInfo
  sentinel and every plugin failed in core resolution (Rebuilt 0/N).
• dispatch_plan links the chain correctly. It parsed the task id
  with a loose regex that a step title containing ID: could hijack,
  stranding the rest of the chain; it now anchors on the full UUID.
• Per-agent serialization survives name-spelling drift. coding vs
  coding_agent slipped past the busy-check and ran concurrently (the
  anyio race); every per-agent comparison now normalizes the name.
• Deliveries and failures no longer vanish. A LINE push with a missing
  token was recorded as success and skipped the outbox; the license gate
  failed a task without advancing the queue (stranding dependents); an
  empty ADK reply was stored as a done result. All three are fixed.
• The manager keeps its own MCP across add/remove. agent add no
  longer appends a sub-agent's MCP to the manager (which triggered the
  anyio cancel-scope race), and agent remove tears down the MCP wiring
  it created so a dead URL isn't fed back to the manager.
• Scheduler / delivery correctness. Poll no longer starves a ready
  task behind a dependency-blocked one; a finished immediate task isn't
  re-run; and the WebChat-Enterprise env guess can't override an
  explicitly-set channel.
• Plugin fragments join the core's real network (a non-default core
  runs on e.g. costaff_asst, not costaff_default), so the manager can
  reach agents/channels deployed with --core.
• CLI robustness. A corrupt config on one core no longer aborts port
  allocation on another; core use validates on single-install hosts;
  recreate_manager warns on failure instead of claiming success;
  docker up failures raise a catchable error (no traceback);
  costaff restart runs preflight before stopping; and agent add rolls
  back a source it cloned when the deploy fails.

Security

• Every published port now binds 127.0.0.1 by default (manager ADK web,
  bundled Postgres, dashboard, channel fragments) with per-host opt-out
  env vars; fresh installs generate a random Postgres password; corrupt
  config.json aborts loudly instead of being overwritten with defaults;
  retry-exhausted agent runs are recorded as failed, not done.
• External-agent a2a_url registration now DNS-resolves and rejects
  hostnames that map to loopback or the cloud-metadata / link-local range
  (SSRF), while still allowing private-LAN federation nodes.
• GET /api/config no longer returns bot tokens in plaintext — it reports
  only whether each is configured.

CI / tooling

• Tests run across the Python 3.10 / 3.11 / 3.12 matrix (the documented
  floor); wheels package the CLI subpackages; Dependabot keeps pip and
  Actions patched; a pushed v* tag cuts a GitHub Release from the
  matching CHANGELOG section.

v0.1.0-alpha-2

The second tagged pre-release of the CoStaff platform core. This release focuses on a smoother first run — preflight validation, a richer onboard wizard, install fixes — and a new tag-pinning workflow for reproducible agent / channel deployments.

The sister agent and channel repos are tagged v0.1.0-alpha-2 in lockstep — pin them together with --tag v0.1.0-alpha-2.

Highlights

🚀 Tag-aware CLI — pin plugins to a release

• costaff agent add / channel add accept --tag (alias --ref) to pin a clone to a release tag, branch, or commit.
• costaff agent rebuild / channel rebuild honour the persisted pin from config.json, switching the working tree via git fetch --tags && git checkout <ref> instead of pull --ff-only; --tag <new> overwrites the pin.
• costaff update --tag pins the core repo itself.
• agent list / channel list gained a Ref column showing the current pin.
• New costaff agent tags <name> / costaff channel tags <name> list available release tags on a plugin's origin (newest first, current pin marked ✓ pinned).
• Git wrapper gained clone(..., ref=...), fetch_tags(), checkout(), and current_ref() — covered by 10 + 8 new tests.

🩺 Safer startup & diagnostics

• costaff start preflight check validates .env (model API key, DB URI, security secrets, workspace dir) before touching Docker; fatal issues abort with the exact fix instead of a crash-loop. Skippable via --no-preflight.
• costaff doctor suggested fixes — detected problems (Docker unreachable, network missing, dead agent port, env issues, missing channel sources, DB unreachable) are replayed at the end as a deduplicated problem → fix list.

🧭 Onboarding upgrades

• Re-running costaff onboard now defaults every prompt to the existing .env value (safe re-entry).
• The Gemini API key is live-verified against the Gemini API, with an immediate warning on rejection.
• WebChat is pre-selected; already-deployed channels are kept instead of re-cloned; the wizard can create the dashboard admin account; a "next steps" panel closes the flow.
• costaff agent add now seeds agent_mcp_filters with the 4-core-tool whitelist (send_message_now / add_task_comment / move_to_shared / list_data_files), so fresh sub-agents no longer inherit the manager's full ~40-tool MCP spec.

Fixed

• install.sh no longer aborts on Ubuntu 24.04 (installs python3.12-distutils only where it still exists); on macOS it launches Docker Desktop and waits up to 90s; on Ubuntu it starts the Docker daemon via systemd when stopped.
• costaff bootstrap now generates MCP_SECRET_KEY / API_HEADERS_KEY / ID_SALT like the interactive wizard — CI deploys no longer run with the template salt and unauthenticated internal APIs. Default Gemini model bumped gemini-2.5-flash → gemini-3-flash-preview.
• .env.template documents COSTAFF_WORKSPACE_DIR and adds worked LiteLLM examples for Ollama / OpenAI / Anthropic.
• costaff agent rebuild / channel rebuild now force_remove each declared container before compose up --force-recreate, fixing Conflict. The container name "/X" is already in use across mixed deploy histories.
• rebuild --tag <ref> no longer persists the new ref to config.json when the underlying git checkout fails.

Changed

• external_agents[name] / dynamic_channels[name] entries in config.json may now carry an optional ref field; absence preserves the legacy "track default branch" behaviour.

---

Full changelog: CHANGELOG.md · Compare: v0.1.0-alpha-1...v0.1.0-alpha-2