flowjet 2.1.9 — refreshed builtin skills bundle

Changed

• Refreshed builtin skills bundle. Shipped two new research skills and
  retired a stale one: added autoresearch (an autonomous research loop with an
  orchestrator, scoring, and a security checklist) and oh-my-research (a
  structured evidence → synthesis pipeline with scripts and pattern presets),
  and removed the superseded research-bootstrap skill. The platonic-coding
  skill was reworked — its spec/IMPL/WORKFLOW references were tightened, a new
  specs-partition.md reference added, and the old coding-plan-template.md /
  impl-guide-template.md drafts dropped in favor of the consolidated template
  namespace. The top-level builtin_skills/README.md (a stale pointer) was
  removed; skills are still auto-registered via register_builtin_skill_root.

flowjet 2.1.8 — PostgreSQL checkpointer on the server surface

Fixed

• PostgreSQL checkpointer is now wired for the server surface.
  Checkpointers.open() (the /v1/responses + ACP server path) used to warn
  “Checkpointer backend 'postgresql' is not wired for the flowjet server surface
  yet” and yield None, so thread state never survived a turn — a deployment
  whose persistence.default_backend is postgresql silently ran every turn as
  a fresh conversation. It now builds an AsyncPostgresSaver on nano's shared
  checkpoints pool (resolve_checkpointer() → (None, pool)), opens the pool,
  and runs setup() once per process under nano's advisory lock
  (SharedCheckpointerPool.setup_checkpointer), so concurrent workers cannot
  race LangGraph's DDL. The shared pool is deliberately not closed on exit
  (it is process-wide). Failure still degrades to “no persistence” with a
  warning instead of failing startup; unknown backends keep the old behavior.

flowjet 2.1.7 — configurable workspace watched subdirs + drop personal byline

Added

• Configurable workspace watched subdirectories. A new
  workspace_watched_subdirs setting (env FLOWJET_WORKSPACE_WATCHED_SUBDIRS,
  comma-separated) flows from Settings →
  SyncAwareWorkspaceResolver(watched_subdirs=...) → WorkspaceManager, which
  pre-creates the configured subtrees at workspace open and wires them into the
  DirtyTracker. None/empty keeps the engine defaults (input/working/output)
  — behavior unchanged — and unsafe entries fail loud at workspace open rather
  than silently dropping sync coverage.

Changed

• Default identity no longer carries a personal creator byline.
  FLOWJET_CREATOR is now empty (the soothe identity renderer omits
  "invented by {creator}" when the creator is blank) and
  FLOWJET_IDENTITY_FRAGMENT drops the byline lines, so identity answers name
  only ({assistant_name}). The vendor denylist (incl. Soothe/soothe-nano) is
  unchanged. Explicit agent.assistant_identity in deployment config still wins
  (e.g. MiraCloud branding).
• Soothe dependency pin raised to >=1.2.3,<2.0.0 to match the coordinated
  release.

flowjet 2.1.6 — ArcBox sandbox survives recycles, crashes and restarts

Fixed

• ArcBox tool sandboxes stay usable across recycles, crashes and restarts.
  The manager trusted its in-memory lease map, and the agent graph pinned its
  sandbox id at build time, so a guest that died behind a still-ready daemon
  record (host sleep, idle recycle, daemon restart) left the worker talking to a
  corpse: every file/shell tool failed and the agent reported an empty
  workspace even though the host tree was fully staged. open_workspace now
  probes readiness once per turn — a control-plane info() tier that catches a
  record outliving its VM plus terminal states, then a guest round-trip that
  catches a wedged guest — and kills, drops and replaces a lease that fails.
  Every new guest, whether spawned or adopted, is seeded from the host
  workspace, and in-flight run counts carry over so peer open/close pairs stay
  balanced. The agent cache key is now workspace + sandbox id, so a respawn
  rebuilds the graph against the live guest.
• Crash and restart recovery. VMs are created with a labelled orphan
  deadline (idle_timeout + on_idle=kill) that the reaper re-arms every 60 s,
  so a process that dies leaves nothing behind — the daemon reaps it, with no
  cross-process coordination and no startup kill-list. On first use the manager
  scans owner-scoped sandboxes and adopts a surviving VM, pulling guest state to
  the host before pushing (recovering a crashed turn's artifacts) instead of
  spawning a duplicate; a failed create reconnects the client and retries once.
• Every control-plane RPC is bounded (kill/set_lifecycle 15 s, connect
  by connect_timeout). The SDK's per-RPC deadline is opt-in and was
  unconfigured: one wedged call stalled the reaper loop, and a stalled heartbeat
  reads to the daemon exactly like a dead owner — which gets healthy, in-use VMs
  reaped at their deadlines.

Changed

• sandbox.orphan_timeout_seconds defaults to max(900, 3 × idle_timeout_seconds)
  instead of a fixed 900, and an explicit value is validated only when the
  sandbox is enabled — existing configs keep parsing unchanged. New keys:
  recover_orphans (adopt surviving VMs after a restart) and owner (label
  namespace, env FLOWJET_SANDBOX_OWNER) so planes sharing one ArcBox daemon
  never adopt or reap each other's sandboxes. See RFC-005 §5.1/§5.2.

v2.1.5

Fixed

• ask_user pauses render as Markdown on ACP/UI hosts. A clarification turn's reply reused the CLI terminal block (space-indented ⏸ Paused for input: followed by • option lines); Markdown-rendering hosts collapse the indentation, so the pause and its options displayed as one run-on paragraph. The ACP path now emits bold heading + bold question + - **label** — description bullets (Q1:/Q2: numbering for multiple questions, (Recommended) → ★ kept). The CLI keeps the plain-text block.

flowjet 2.1.4

Added

• Resumable ask_user pauses via shared thread-state checkpointer. The server now persists agent thread state (LangGraph checkpoint) to a shared SQLite database (WAL mode + busy_timeout) so that ask_user pauses survive pool-worker switches and server restarts. A second turn on the same session resumes the paused interrupt and delivers the user's input text as the answer. The server_checkpointer setting (auto by default, off for legacy stateless behavior) controls the policy end-to-end.

Changed

• CI now enforces mypy type compliance. A make typecheck (mypy) gate was added to the CI pipeline, and type annotations were fixed across the CLI, core, and server packages to achieve full compliance.

flowjet 2.1.3 — every failure path surfaces a message

Fixed

• CLI: every failure path now surfaces a message — no more silent exits.
  The CLI could exit with no output (SIGTERM/SIGHUP, plugin sys.exit, broken
  pipes, empty answers) and destroyed the original error stack even with -v.
  This closes every identified gap in the error-to-output flow:

  • RunFailed now carries exc_type + traceback; stream_query raises
    AgentRunError so the CLI reports the real exception type and, with -v,
    the original in-run stack instead of a re-raise site.
  • SIGTERM/SIGHUP cancel the run gracefully: "interrupted by SIGTERM" with
    conventional 128+signal exit codes (was: instant silent death).
  • Empty-answer runs warn on stderr ("agent produced no output · N tool calls
    failed · last tool error") and exit 3.
  • Partial narration is flushed to stderr before any failure/timeout/interrupt
    message — a provider drop no longer loses the whole answer.
  • --timeout SECONDS cancels a stuck run, reporting the last activity.
  • FJ_LOG_FILE=<path> keeps WARNING+ with full tracebacks in a rotating log
    for post-mortem diagnosis (incl. logger.exception from plugins).
  • Multi-line errors truncated to one line now hint "run with -v".
  • BrokenPipeError: exit 141 with a message; shutdown flush noise silenced
    via devnull dup2 (fj q | head).
  • Teardown suppresses CancelledError so it cannot mask the real result.
  • Server parity: pool worker_error, OpenAI-compatible and ACP failures
    populate/render exc_type + traceback (ACP shows "Error: : ").
  • Minor: thread-snapshot check failures are logged; config-load errors
    honor -v with a traceback; plugin sys.exit(0) still fails loudly.

Changed

• ArcBox tool sandbox: guest network is now off by default
  (sandbox.network defaults to false, was true). With true the guest is
  NAT'ed through the ArcBox host and the daemon only knows ENABLED|NONE, so
  everything the host can route to — internal network and corporate VPN
  included — was reachable from agent shell commands. With false the guest
  boots without a network device (no route, no DNS): shell/code execution can
  reach neither the public internet nor internal services. Web search/crawl is
  unaffected — web_search/fetch_url are host-placement tools and never run
  in the guest. Opt back in per deployment (sandbox.network: true in the
  nano.yml) when the guest really needs egress.

---

Full changelog: https://github.com/mirasoth/flowjet/blob/main/CHANGELOG.md

flowjet 2.1.2 — upgrade soothe to 1.2.1

Changed

• Bump soothe to 1.2.1. The runtime dependency floor moves from >=1.2.0 to >=1.2.1,<2.0.0 to track the upstream fix release. No flowjet API or behavior changes; the locked dependency set is otherwise unchanged.

---

Verification: All 493 unit + 230 server tests pass; ruff lint/format clean. make build produces flowjet-2.1.2 wheel + sdist; twine check passes. Wheel metadata confirms Requires-Dist: soothe<2.0.0,>=1.2.1.

flowjet 2.1.1 — ACP ask_user surfacing + transitive arcbox

Highlights

• ACP clarification turns no longer end empty. When the agent paused with ask_user questions over ACP, the questions were only surfaced as an intermediate agent thought — but hosts accumulate message chunks as the turn's reply, so a clarification turn ended with no content at all. The ACP agent now also emits the questions block as a message chunk (rendered via format_questions_block), so the user sees the questions in the host's reply surface and answers there.
• format_questions_block supports non-CLI hosts. The resume footer ("Resume with: fjf -t ") was hardcoded for the CLI. ACP and other non-CLI hosts now pass resume_hint="" to omit the CLI-specific footer; the user answers in the host's own input surface. The CLI path is unchanged.
• Drop the direct arcbox dependency from the server extra. The arcbox SDK ships transitively via soothe-nano (a dependency of soothe), so the explicit arcbox>=0.1.0 pin was redundant and could drift out of sync. The sandbox import error message is updated to point at the transitive path.

Full changelog: https://github.com/mirasoth/flowjet/blob/v2.1.1/CHANGELOG.md

flowjet 2.1.0 — core restructure + ArcBox tool sandbox (RFC-005)

Highlights

• flowjet.core layout restructure. Runtime-protocol package regrouped by responsibility (runtime/, sandbox/, agent/, streaming/, bridges/); duplicated nano/soothe adapter lifecycle collapsed into AgentAdapterBase + IsolatingBackendBuilder; InteractionMode.from_metadata() is the single request-mode parser. Behavior unchanged.
• ArcBox tool sandbox (RFC-005). Tool calls (shell + workspace file ops) run inside a per-workspace ArcBox Linux microVM via soothe-nano 2.0's ArcBoxBackend. Host workspace stays the integration point via SandboxAwareWorkspaceResolver (host→guest at open, guest→host at close, skipping unchanged files).
• Dynamic max_sandboxes compute-sizing. None = 70% of CPU / vcpus (min 8); 0 = unlimited (no LRU eviction); env overrides FLOWJET_SANDBOX_MAX and FLOWJET_SANDBOX_MAX_PERCENT.
• Removed whole-worker arcbox isolation (FLOWJET_ISOLATION_MODE=arcbox and all FLOWJET_ARCBOX_* settings) — superseded by the tool sandbox; thread pool is the only isolation pool.

Full changelog: https://github.com/mirasoth/flowjet/blob/v2.1.0/CHANGELOG.md