v0.21.0

Added — Claude Agent SDK harness

Anthropic's Claude Agent SDK (claude-agent-sdk /
@anthropic-ai/claude-agent-sdk) — the programmable form of Claude Code —
wrapped as a first-class atomr-agents harness. It exposes Claude Code's full
feature set (slash commands, subagents, hooks, MCP, permission modes,
sessions, custom system prompts, built-in Read/Write/Edit/Bash/Grep/WebSearch
tools) and bills against your Anthropic API credits by driving the bundled
claude CLI. Distinct from the server-side "Managed Agents" API.

• atomr-agents-agent-sdk-core — the contract: AgentSdkConfig mirroring
  the SDK's ClaudeAgentOptions (round-trips to a JSON dict), a normalized
  message/result/event schema, and the pluggable AgentSdkBackend /
  AgentSdkSession trait seam with a deterministic in-memory MockBackend so
  the whole surface is testable without the SDK, the claude CLI, or
  network/credits.
• atomr-agents-agent-sdk-harness — orchestrator with a pluggable backend,
  an event broadcast, a session registry under an atomically-reserved
  concurrency quota, Anthropic-credit spend tracking via the shared
  SpendLedger, a .claude/ projection (skills / slash commands / MCP /
  settings), and a Callable impl. Behind the actor feature it exposes the
  bidirectional interactive session as an atomr_core::actor::Actor. The
  default permission_mode is bypassPermissions (max autonomy — overridable
  per request/spec).
• atomr-agents-agent-sdk-harness-web — axum REST + SSE companion (start
  runs/sessions, post messages, interrupt, change mode/model, stream events).
• PyO3 bridge — from atomr_agents.agent_sdk import harness. The Rust
  PythonAgentSdkBackend drives the real claude-agent-sdk over a reverse
  async-iterator bridge (Rust drives the Python __anext__); the interactive
  session is exposed as AgentSdkSession (the actor surface). atomr @tool
  guests and Rust tools are bridged into the agent as in-process SDK MCP tools
  via create_sdk_mcp_server.
• Per-session sandbox workspaces (Pattern C) — behind the harness sandbox
  feature, each interactive session (and each headless run) gets its own
  isolated microVM sandbox as a disposable
  workspace: the .claude/ projection is staged into it (write_file), the
  agent's exec/file is routed there via an in-process run_in_sandbox tool
  while the host Bash/Write/Edit are disabled (true containment), and the
  workspace is discarded — or snapshotted to the warm pool — on close
  (spec.workspace.{enabled,profile,backend,on_close,reuse_warm}). One
  SandboxClient is shared between the Rust harness and the Python tool so both
  resolve the workspace from a single registry; SandboxClient.attach(id)
  binds the tool. See Pattern C in docs/agent-sdk-harness.md.
• Host loader — <root>/agent-sdk/<id>/ (harness.yaml + commands/ +
  skills/ + mcp/) → AgentSdkHarnessSpec + .claude projection, reusing
  the host's on-disk conventions.
• Umbrella — new agent-sdk feature.
• Packaging — pip install atomr-agents[agent-sdk] pulls the SDK (which
  auto-bundles the claude CLI); the native extension imports it lazily.

Added — MicroVM sandbox capability (local + dev tiers)

Secure, instant-boot compute environments for executing untrusted agent
code (Python / Bash / JavaScript / Rust), per the Agentic MicroVM
Sandboxes PRD. This wave lands the backend-agnostic foundation plus the
locally-runnable tiers; the Firecracker/KVM backend and the Tier-3 gRPC
cluster are the documented next milestones.

• atomr-agents-sandbox-core — SandboxBackend / SandboxHandle
  traits, profile / budget / request types (with the security-relevant
  Rust 2 GB / 2 vCPU floor), SandboxEvent stream, and a deterministic
  in-memory MockBackend so the whole surface is testable without Docker
  or KVM.
• atomr-agents-sandbox-harness — orchestrator with a pluggable
  backend, a sandbox registry under an atomically-reserved concurrency
  quota, a best-fit bin-packing scheduler, a warm snapshot pool, lifecycle
  event broadcast, and a Callable impl. The Rust budget floor is enforced
  structurally at the harness boundary.
• atomr-agents-sandbox-tool — the execute_in_sandbox Tool.
• atomr-agents-sandbox-backend-docker — bollard-backed "insecure dev
  mode" backend (real exec, tar file I/O, commit-based snapshot/fork).
  Containers share the host kernel, so this is explicitly not an isolation
  boundary for untrusted code.
• atomr-agents-sandbox-harness-web — axum REST + SSE companion.
• atomr-agents-sandbox-proto — host↔guest wire protocol
  (length-prefixed postcard frames over AF_VSOCK).
• atomr-agents-sandbox-guest-agent — the in-VM PID-1 guest daemon.
• PyO3 bindings — from atomr_agents.sandbox import SandboxClient, SandboxConfig, SandboxProfile.
• Umbrella — new sandbox feature.

Changed — Coding-CLI vendor: Gemini CLI → Antigravity CLI (agy)

Google is transitioning the Gemini CLI into the Antigravity CLI; the
legacy gemini CLI stops serving requests on 2026-06-18. The
coding-cli-vendor-gemini crate is renamed to
atomr-agents-coding-cli-vendor-antigravity and now targets the
agy binary. Unlike the old Gemini CLI, Antigravity serves non-Gemini
models (Claude, etc.) selectable via the model flag.

• GeminiVendor → AntigravityVendor; GeminiParser →
  AntigravityParser; materialize_gemini_config →
  materialize_antigravity_config.
• New AntigravityConfig makes the binary name, headless flags, and the
  on-disk config directory (default .antigravity/) configurable — so
  operators can reconcile them with agy's real interface without a code
  change.
• Harness Cargo feature vendor-gemini → vendor-antigravity.
• Default Docker image gemini.Dockerfile → antigravity.Dockerfile
  (installs agy via the official install.sh).
• Breaking: CliVendorKind::Gemini → CliVendorKind::Antigravity;
  the serde wire value changes from "gemini" to "antigravity", so
  persisted CliRequest JSON with vendor: "gemini" no longer
  deserializes.

Added — Meetings harness: structured extraction from diarized transcripts

A downstream harness that turns a diarized SttConversation into a
structured MeetingAnalysis — attendee roster, linear append-only
notes/actions ledgers, and a tiered dynamic summary stack. See
docs/meetings-harness.md.

• crates/meetings-harness — the harness core: typed +
  Boxed* split mirroring stt-harness; RuleBasedExtractor
  default that is deterministic and LLM-free; pluggable
  MeetingExtractor trait for LLM-driven implementations;
  MeetingsToolSet of Tool impls (list_turns, upsert_attendee,
  append_note, append_action, update_action, revise_tail_segment,
  finalize_segment, regenerate_running, set_title, finalize).
  BatchExtractionLoop and StreamingExtractionLoop — the latter
  subscribes to an STT harness's broadcast and maintains an append-only
  ledger with a revisable in-flight tail segment summary. Persistence
  via MeetingsStore and (feature state) CheckpointerMeetingsStore,
  filed under the same conversation_id as the source transcript.
• crates/meetings-harness-web — an optional axum backend plus an
  embedded React + Vite SPA (port 7100) for reviewing and editing
  analyses: REST CRUD, attendee rename, action patch, run trigger / stop,
  WebSocket live feed of MeetingsHarnessEvents, SPA fallback.
• atomr-agents meetings — new CLI subcommand. meetings analyze --conversation-id <id> --model <model_id> runs the harness against
  an existing transcript; meetings serve boots the web UI. Gated by
  --features meetings / --features meetings-web.

Added — STT harness: agentic streaming speech-to-text pipeline

A new capability layer that diarizes, transcribes, and accumulates a
conversation record aligned to agentic structures. See
docs/stt-harness.md.

• crates/stt-harness — the harness core: audio sources (mic /
  file / bytes), a streaming session actor, diarization + voice-mode
  coalescing, the loop/termination strategies, and a ConversationStore
  with a Checkpointer-backed implementation (state feature).
  BoxedSttHarness mirrors the Agent/Harness boxed-handle pattern.
• crates/stt-harness-web — an optional axum backend plus an
  embedded React SPA (embed-ui feature, via rust-embed) for
  reviewing and editing diarized conversations: REST routes, a
  WebSocket live feed, and the SPA fallback handler.
• atomr-agents serve now runs the real review UI when the CLI is
  built with --features stt-web, persisting conversations through the
  configured crates/state checkpointer.
• xtask stt-web-build — builds the React SPA into ui/dist so
  stt-harness-web can be compiled with --features embed-ui.
• Python bindings — atomr_agents.stt_harness wraps the harness;
  the top-level facade now resolves native names defensively so a name
  not yet bound on the Rust side degrades to None instead of failing
  the whole package import.

Added — Phase C: async scorers, guest adapters, boxed Agent/Harness, Python workflow runner

Closes the "Phase B" deferral comments scattered across crates/agent,
crates/harness, crates/eval, and crates/py-bindings/src/{agent, harness,workflow,eval}.rs. The five blocks:

• AsyncScorer trait (crates/eval/src/scorer.rs). Parallel to
  the existing sync Scorer; blanket impl<S: Scorer> AsyncScorer for S keeps existing sync impls compatible. LlmJudgeScorer,
  RubricScorer, and PairwiseScorer now implement AsyncScorer
  directly — dropping the tokio::task::block_in_place workaround
  the sync Scorer impl required. EvalSuite::scorer field type
  changed from `Arc<dyn Scor...

v0.20.0

Added — Claude Agent SDK harness

Anthropic's Claude Agent SDK (claude-agent-sdk /
@anthropic-ai/claude-agent-sdk) — the programmable form of Claude Code —
wrapped as a first-class atomr-agents harness. It exposes Claude Code's full
feature set (slash commands, subagents, hooks, MCP, permission modes,
sessions, custom system prompts, built-in Read/Write/Edit/Bash/Grep/WebSearch
tools) and bills against your Anthropic API credits by driving the bundled
claude CLI. Distinct from the server-side "Managed Agents" API.

• atomr-agents-agent-sdk-core — the contract: AgentSdkConfig mirroring
  the SDK's ClaudeAgentOptions (round-trips to a JSON dict), a normalized
  message/result/event schema, and the pluggable AgentSdkBackend /
  AgentSdkSession trait seam with a deterministic in-memory MockBackend so
  the whole surface is testable without the SDK, the claude CLI, or
  network/credits.
• atomr-agents-agent-sdk-harness — orchestrator with a pluggable backend,
  an event broadcast, a session registry under an atomically-reserved
  concurrency quota, Anthropic-credit spend tracking via the shared
  SpendLedger, a .claude/ projection (skills / slash commands / MCP /
  settings), and a Callable impl. Behind the actor feature it exposes the
  bidirectional interactive session as an atomr_core::actor::Actor. The
  default permission_mode is bypassPermissions (max autonomy — overridable
  per request/spec).
• atomr-agents-agent-sdk-harness-web — axum REST + SSE companion (start
  runs/sessions, post messages, interrupt, change mode/model, stream events).
• PyO3 bridge — from atomr_agents.agent_sdk import harness. The Rust
  PythonAgentSdkBackend drives the real claude-agent-sdk over a reverse
  async-iterator bridge (Rust drives the Python __anext__); the interactive
  session is exposed as AgentSdkSession (the actor surface). atomr @tool
  guests and Rust tools are bridged into the agent as in-process SDK MCP tools
  via create_sdk_mcp_server.
• Host loader — <root>/agent-sdk/<id>/ (harness.yaml + commands/ +
  skills/ + mcp/) → AgentSdkHarnessSpec + .claude projection, reusing
  the host's on-disk conventions.
• Umbrella — new agent-sdk feature.
• Packaging — pip install atomr-agents[agent-sdk] pulls the SDK (which
  auto-bundles the claude CLI); the native extension imports it lazily.

Added — MicroVM sandbox capability (local + dev tiers)

Secure, instant-boot compute environments for executing untrusted agent
code (Python / Bash / JavaScript / Rust), per the Agentic MicroVM
Sandboxes PRD. This wave lands the backend-agnostic foundation plus the
locally-runnable tiers; the Firecracker/KVM backend and the Tier-3 gRPC
cluster are the documented next milestones.

• atomr-agents-sandbox-core — SandboxBackend / SandboxHandle
  traits, profile / budget / request types (with the security-relevant
  Rust 2 GB / 2 vCPU floor), SandboxEvent stream, and a deterministic
  in-memory MockBackend so the whole surface is testable without Docker
  or KVM.
• atomr-agents-sandbox-harness — orchestrator with a pluggable
  backend, a sandbox registry under an atomically-reserved concurrency
  quota, a best-fit bin-packing scheduler, a warm snapshot pool, lifecycle
  event broadcast, and a Callable impl. The Rust budget floor is enforced
  structurally at the harness boundary.
• atomr-agents-sandbox-tool — the execute_in_sandbox Tool.
• atomr-agents-sandbox-backend-docker — bollard-backed "insecure dev
  mode" backend (real exec, tar file I/O, commit-based snapshot/fork).
  Containers share the host kernel, so this is explicitly not an isolation
  boundary for untrusted code.
• atomr-agents-sandbox-harness-web — axum REST + SSE companion.
• atomr-agents-sandbox-proto — host↔guest wire protocol
  (length-prefixed postcard frames over AF_VSOCK).
• atomr-agents-sandbox-guest-agent — the in-VM PID-1 guest daemon.
• PyO3 bindings — from atomr_agents.sandbox import SandboxClient, SandboxConfig, SandboxProfile.
• Umbrella — new sandbox feature.

Changed — Coding-CLI vendor: Gemini CLI → Antigravity CLI (agy)

Google is transitioning the Gemini CLI into the Antigravity CLI; the
legacy gemini CLI stops serving requests on 2026-06-18. The
coding-cli-vendor-gemini crate is renamed to
atomr-agents-coding-cli-vendor-antigravity and now targets the
agy binary. Unlike the old Gemini CLI, Antigravity serves non-Gemini
models (Claude, etc.) selectable via the model flag.

• GeminiVendor → AntigravityVendor; GeminiParser →
  AntigravityParser; materialize_gemini_config →
  materialize_antigravity_config.
• New AntigravityConfig makes the binary name, headless flags, and the
  on-disk config directory (default .antigravity/) configurable — so
  operators can reconcile them with agy's real interface without a code
  change.
• Harness Cargo feature vendor-gemini → vendor-antigravity.
• Default Docker image gemini.Dockerfile → antigravity.Dockerfile
  (installs agy via the official install.sh).
• Breaking: CliVendorKind::Gemini → CliVendorKind::Antigravity;
  the serde wire value changes from "gemini" to "antigravity", so
  persisted CliRequest JSON with vendor: "gemini" no longer
  deserializes.

Added — Meetings harness: structured extraction from diarized transcripts

A downstream harness that turns a diarized SttConversation into a
structured MeetingAnalysis — attendee roster, linear append-only
notes/actions ledgers, and a tiered dynamic summary stack. See
docs/meetings-harness.md.

• crates/meetings-harness — the harness core: typed +
  Boxed* split mirroring stt-harness; RuleBasedExtractor
  default that is deterministic and LLM-free; pluggable
  MeetingExtractor trait for LLM-driven implementations;
  MeetingsToolSet of Tool impls (list_turns, upsert_attendee,
  append_note, append_action, update_action, revise_tail_segment,
  finalize_segment, regenerate_running, set_title, finalize).
  BatchExtractionLoop and StreamingExtractionLoop — the latter
  subscribes to an STT harness's broadcast and maintains an append-only
  ledger with a revisable in-flight tail segment summary. Persistence
  via MeetingsStore and (feature state) CheckpointerMeetingsStore,
  filed under the same conversation_id as the source transcript.
• crates/meetings-harness-web — an optional axum backend plus an
  embedded React + Vite SPA (port 7100) for reviewing and editing
  analyses: REST CRUD, attendee rename, action patch, run trigger / stop,
  WebSocket live feed of MeetingsHarnessEvents, SPA fallback.
• atomr-agents meetings — new CLI subcommand. meetings analyze --conversation-id <id> --model <model_id> runs the harness against
  an existing transcript; meetings serve boots the web UI. Gated by
  --features meetings / --features meetings-web.

Added — STT harness: agentic streaming speech-to-text pipeline

A new capability layer that diarizes, transcribes, and accumulates a
conversation record aligned to agentic structures. See
docs/stt-harness.md.

• crates/stt-harness — the harness core: audio sources (mic /
  file / bytes), a streaming session actor, diarization + voice-mode
  coalescing, the loop/termination strategies, and a ConversationStore
  with a Checkpointer-backed implementation (state feature).
  BoxedSttHarness mirrors the Agent/Harness boxed-handle pattern.
• crates/stt-harness-web — an optional axum backend plus an
  embedded React SPA (embed-ui feature, via rust-embed) for
  reviewing and editing diarized conversations: REST routes, a
  WebSocket live feed, and the SPA fallback handler.
• atomr-agents serve now runs the real review UI when the CLI is
  built with --features stt-web, persisting conversations through the
  configured crates/state checkpointer.
• xtask stt-web-build — builds the React SPA into ui/dist so
  stt-harness-web can be compiled with --features embed-ui.
• Python bindings — atomr_agents.stt_harness wraps the harness;
  the top-level facade now resolves native names defensively so a name
  not yet bound on the Rust side degrades to None instead of failing
  the whole package import.

Added — Phase C: async scorers, guest adapters, boxed Agent/Harness, Python workflow runner

Closes the "Phase B" deferral comments scattered across crates/agent,
crates/harness, crates/eval, and crates/py-bindings/src/{agent, harness,workflow,eval}.rs. The five blocks:

• AsyncScorer trait (crates/eval/src/scorer.rs). Parallel to
  the existing sync Scorer; blanket impl<S: Scorer> AsyncScorer for S keeps existing sync impls compatible. LlmJudgeScorer,
  RubricScorer, and PairwiseScorer now implement AsyncScorer
  directly — dropping the tokio::task::block_in_place workaround
  the sync Scorer impl required. EvalSuite::scorer field type
  changed from Arc<dyn Scorer> to Arc<dyn AsyncScorer>.
• 6 Python→Rust guest adapters (crates/py-bindings/src/guest/).
  The single-file guest.rs (331 LOC) is split into a guest/
  directory (10 files, ~1400 LOC). New adapters: PyInstructionAdapter,
  PyMemoryStrategyAdapter, PySkillStrategyAdapter, PyPersonaAdapter,
  PyMemoryStoreAdapter, PyEmbedderAdapter. Each ships a
  build_guest_* factory mirroring the existing build_guest_toolset
  pattern.
• BoxedAgent + AgentSpec::into_agent (crates/agent/src/boxed.rs).
  Object-erased BoxedAgent for config-driven instantiation. The typed
  Agent<I,T,Ms,Sk> keeps hot-path monomorphization; both share a
  pub(crate) async fn run_turn_impl helper.
• HarnessDispatch + HarnessRef + BoxedHarness (crates/harness/).
  Mirror of the Agent boxed-han...

v0.19.0

Added — MicroVM sandbox capability (local + dev tiers)

Secure, instant-boot compute environments for executing untrusted agent
code (Python / Bash / JavaScript / Rust), per the Agentic MicroVM
Sandboxes PRD. This wave lands the backend-agnostic foundation plus the
locally-runnable tiers; the Firecracker/KVM backend and the Tier-3 gRPC
cluster are the documented next milestones.

• atomr-agents-sandbox-core — SandboxBackend / SandboxHandle
  traits, profile / budget / request types (with the security-relevant
  Rust 2 GB / 2 vCPU floor), SandboxEvent stream, and a deterministic
  in-memory MockBackend so the whole surface is testable without Docker
  or KVM.
• atomr-agents-sandbox-harness — orchestrator with a pluggable
  backend, a sandbox registry under an atomically-reserved concurrency
  quota, a best-fit bin-packing scheduler, a warm snapshot pool, lifecycle
  event broadcast, and a Callable impl. The Rust budget floor is enforced
  structurally at the harness boundary.
• atomr-agents-sandbox-tool — the execute_in_sandbox Tool.
• atomr-agents-sandbox-backend-docker — bollard-backed "insecure dev
  mode" backend (real exec, tar file I/O, commit-based snapshot/fork).
  Containers share the host kernel, so this is explicitly not an isolation
  boundary for untrusted code.
• atomr-agents-sandbox-harness-web — axum REST + SSE companion.
• atomr-agents-sandbox-proto — host↔guest wire protocol
  (length-prefixed postcard frames over AF_VSOCK).
• atomr-agents-sandbox-guest-agent — the in-VM PID-1 guest daemon.
• PyO3 bindings — from atomr_agents.sandbox import SandboxClient, SandboxConfig, SandboxProfile.
• Umbrella — new sandbox feature.

Changed — Coding-CLI vendor: Gemini CLI → Antigravity CLI (agy)

Google is transitioning the Gemini CLI into the Antigravity CLI; the
legacy gemini CLI stops serving requests on 2026-06-18. The
coding-cli-vendor-gemini crate is renamed to
atomr-agents-coding-cli-vendor-antigravity and now targets the
agy binary. Unlike the old Gemini CLI, Antigravity serves non-Gemini
models (Claude, etc.) selectable via the model flag.

• GeminiVendor → AntigravityVendor; GeminiParser →
  AntigravityParser; materialize_gemini_config →
  materialize_antigravity_config.
• New AntigravityConfig makes the binary name, headless flags, and the
  on-disk config directory (default .antigravity/) configurable — so
  operators can reconcile them with agy's real interface without a code
  change.
• Harness Cargo feature vendor-gemini → vendor-antigravity.
• Default Docker image gemini.Dockerfile → antigravity.Dockerfile
  (installs agy via the official install.sh).
• Breaking: CliVendorKind::Gemini → CliVendorKind::Antigravity;
  the serde wire value changes from "gemini" to "antigravity", so
  persisted CliRequest JSON with vendor: "gemini" no longer
  deserializes.

Added — Meetings harness: structured extraction from diarized transcripts

A downstream harness that turns a diarized SttConversation into a
structured MeetingAnalysis — attendee roster, linear append-only
notes/actions ledgers, and a tiered dynamic summary stack. See
docs/meetings-harness.md.

• crates/meetings-harness — the harness core: typed +
  Boxed* split mirroring stt-harness; RuleBasedExtractor
  default that is deterministic and LLM-free; pluggable
  MeetingExtractor trait for LLM-driven implementations;
  MeetingsToolSet of Tool impls (list_turns, upsert_attendee,
  append_note, append_action, update_action, revise_tail_segment,
  finalize_segment, regenerate_running, set_title, finalize).
  BatchExtractionLoop and StreamingExtractionLoop — the latter
  subscribes to an STT harness's broadcast and maintains an append-only
  ledger with a revisable in-flight tail segment summary. Persistence
  via MeetingsStore and (feature state) CheckpointerMeetingsStore,
  filed under the same conversation_id as the source transcript.
• crates/meetings-harness-web — an optional axum backend plus an
  embedded React + Vite SPA (port 7100) for reviewing and editing
  analyses: REST CRUD, attendee rename, action patch, run trigger / stop,
  WebSocket live feed of MeetingsHarnessEvents, SPA fallback.
• atomr-agents meetings — new CLI subcommand. meetings analyze --conversation-id <id> --model <model_id> runs the harness against
  an existing transcript; meetings serve boots the web UI. Gated by
  --features meetings / --features meetings-web.

Added — STT harness: agentic streaming speech-to-text pipeline

A new capability layer that diarizes, transcribes, and accumulates a
conversation record aligned to agentic structures. See
docs/stt-harness.md.

• crates/stt-harness — the harness core: audio sources (mic /
  file / bytes), a streaming session actor, diarization + voice-mode
  coalescing, the loop/termination strategies, and a ConversationStore
  with a Checkpointer-backed implementation (state feature).
  BoxedSttHarness mirrors the Agent/Harness boxed-handle pattern.
• crates/stt-harness-web — an optional axum backend plus an
  embedded React SPA (embed-ui feature, via rust-embed) for
  reviewing and editing diarized conversations: REST routes, a
  WebSocket live feed, and the SPA fallback handler.
• atomr-agents serve now runs the real review UI when the CLI is
  built with --features stt-web, persisting conversations through the
  configured crates/state checkpointer.
• xtask stt-web-build — builds the React SPA into ui/dist so
  stt-harness-web can be compiled with --features embed-ui.
• Python bindings — atomr_agents.stt_harness wraps the harness;
  the top-level facade now resolves native names defensively so a name
  not yet bound on the Rust side degrades to None instead of failing
  the whole package import.

Added — Phase C: async scorers, guest adapters, boxed Agent/Harness, Python workflow runner

Closes the "Phase B" deferral comments scattered across crates/agent,
crates/harness, crates/eval, and crates/py-bindings/src/{agent, harness,workflow,eval}.rs. The five blocks:

• AsyncScorer trait (crates/eval/src/scorer.rs). Parallel to
  the existing sync Scorer; blanket impl<S: Scorer> AsyncScorer for S keeps existing sync impls compatible. LlmJudgeScorer,
  RubricScorer, and PairwiseScorer now implement AsyncScorer
  directly — dropping the tokio::task::block_in_place workaround
  the sync Scorer impl required. EvalSuite::scorer field type
  changed from Arc<dyn Scorer> to Arc<dyn AsyncScorer>.
• 6 Python→Rust guest adapters (crates/py-bindings/src/guest/).
  The single-file guest.rs (331 LOC) is split into a guest/
  directory (10 files, ~1400 LOC). New adapters: PyInstructionAdapter,
  PyMemoryStrategyAdapter, PySkillStrategyAdapter, PyPersonaAdapter,
  PyMemoryStoreAdapter, PyEmbedderAdapter. Each ships a
  build_guest_* factory mirroring the existing build_guest_toolset
  pattern.
• BoxedAgent + AgentSpec::into_agent (crates/agent/src/boxed.rs).
  Object-erased BoxedAgent for config-driven instantiation. The typed
  Agent<I,T,Ms,Sk> keeps hot-path monomorphization; both share a
  pub(crate) async fn run_turn_impl helper.
• HarnessDispatch + HarnessRef + BoxedHarness (crates/harness/).
  Mirror of the Agent boxed-handle pattern. Harness::into_boxed,
  HarnessSpec::into_harness, and impl HarnessDispatch for Harness<L,T>
  added.
• PyAgent, PyHarness, PyWorkflowRunner Python wrappers with
  from_spec / from_dict constructors and async run_turn / run
  exposed as Python coroutines via
  pyo3_async_runtimes::tokio::future_into_py. WorkflowRunner::new
  was added as a public constructor.

Changed

• EvalSuite::scorer field type: Arc<dyn Scorer> →
  Arc<dyn AsyncScorer>. The blanket impl<S: Scorer> AsyncScorer for S means callers holding a sync Arc::new(SomeSyncScorer)
  continue to compile via auto-unsizing into the new field type.
  Direct &dyn Scorer callers of LlmJudgeScorer / RubricScorer
  must migrate to &dyn AsyncScorer — Rust coherence forbids both
  the blanket AsyncScorer impl and a direct Scorer impl on the
  same judge (ContainsScorer is unaffected).
• Sibling deps (atomr, atomr-infer, atomr-accel) consume
  crates.io artifacts only — no path = "../sibling/..." links.
  Pinned to atomr 0.9.2 / atomr-infer 0.8.0 / atomr-accel 0.10.0.

v0.18.0

Changed — Coding-CLI vendor: Gemini CLI → Antigravity CLI (agy)

Google is transitioning the Gemini CLI into the Antigravity CLI; the
legacy gemini CLI stops serving requests on 2026-06-18. The
coding-cli-vendor-gemini crate is renamed to
atomr-agents-coding-cli-vendor-antigravity and now targets the
agy binary. Unlike the old Gemini CLI, Antigravity serves non-Gemini
models (Claude, etc.) selectable via the model flag.

• GeminiVendor → AntigravityVendor; GeminiParser →
  AntigravityParser; materialize_gemini_config →
  materialize_antigravity_config.
• New AntigravityConfig makes the binary name, headless flags, and the
  on-disk config directory (default .antigravity/) configurable — so
  operators can reconcile them with agy's real interface without a code
  change.
• Harness Cargo feature vendor-gemini → vendor-antigravity.
• Default Docker image gemini.Dockerfile → antigravity.Dockerfile
  (installs agy via the official install.sh).
• Breaking: CliVendorKind::Gemini → CliVendorKind::Antigravity;
  the serde wire value changes from "gemini" to "antigravity", so
  persisted CliRequest JSON with vendor: "gemini" no longer
  deserializes.

Added — Meetings harness: structured extraction from diarized transcripts

A downstream harness that turns a diarized SttConversation into a
structured MeetingAnalysis — attendee roster, linear append-only
notes/actions ledgers, and a tiered dynamic summary stack. See
docs/meetings-harness.md.

• crates/meetings-harness — the harness core: typed +
  Boxed* split mirroring stt-harness; RuleBasedExtractor
  default that is deterministic and LLM-free; pluggable
  MeetingExtractor trait for LLM-driven implementations;
  MeetingsToolSet of Tool impls (list_turns, upsert_attendee,
  append_note, append_action, update_action, revise_tail_segment,
  finalize_segment, regenerate_running, set_title, finalize).
  BatchExtractionLoop and StreamingExtractionLoop — the latter
  subscribes to an STT harness's broadcast and maintains an append-only
  ledger with a revisable in-flight tail segment summary. Persistence
  via MeetingsStore and (feature state) CheckpointerMeetingsStore,
  filed under the same conversation_id as the source transcript.
• crates/meetings-harness-web — an optional axum backend plus an
  embedded React + Vite SPA (port 7100) for reviewing and editing
  analyses: REST CRUD, attendee rename, action patch, run trigger / stop,
  WebSocket live feed of MeetingsHarnessEvents, SPA fallback.
• atomr-agents meetings — new CLI subcommand. meetings analyze --conversation-id <id> --model <model_id> runs the harness against
  an existing transcript; meetings serve boots the web UI. Gated by
  --features meetings / --features meetings-web.

Added — STT harness: agentic streaming speech-to-text pipeline

A new capability layer that diarizes, transcribes, and accumulates a
conversation record aligned to agentic structures. See
docs/stt-harness.md.

• crates/stt-harness — the harness core: audio sources (mic /
  file / bytes), a streaming session actor, diarization + voice-mode
  coalescing, the loop/termination strategies, and a ConversationStore
  with a Checkpointer-backed implementation (state feature).
  BoxedSttHarness mirrors the Agent/Harness boxed-handle pattern.
• crates/stt-harness-web — an optional axum backend plus an
  embedded React SPA (embed-ui feature, via rust-embed) for
  reviewing and editing diarized conversations: REST routes, a
  WebSocket live feed, and the SPA fallback handler.
• atomr-agents serve now runs the real review UI when the CLI is
  built with --features stt-web, persisting conversations through the
  configured crates/state checkpointer.
• xtask stt-web-build — builds the React SPA into ui/dist so
  stt-harness-web can be compiled with --features embed-ui.
• Python bindings — atomr_agents.stt_harness wraps the harness;
  the top-level facade now resolves native names defensively so a name
  not yet bound on the Rust side degrades to None instead of failing
  the whole package import.

Added — Phase C: async scorers, guest adapters, boxed Agent/Harness, Python workflow runner

Closes the "Phase B" deferral comments scattered across crates/agent,
crates/harness, crates/eval, and crates/py-bindings/src/{agent, harness,workflow,eval}.rs. The five blocks:

• AsyncScorer trait (crates/eval/src/scorer.rs). Parallel to
  the existing sync Scorer; blanket impl<S: Scorer> AsyncScorer for S keeps existing sync impls compatible. LlmJudgeScorer,
  RubricScorer, and PairwiseScorer now implement AsyncScorer
  directly — dropping the tokio::task::block_in_place workaround
  the sync Scorer impl required. EvalSuite::scorer field type
  changed from Arc<dyn Scorer> to Arc<dyn AsyncScorer>.
• 6 Python→Rust guest adapters (crates/py-bindings/src/guest/).
  The single-file guest.rs (331 LOC) is split into a guest/
  directory (10 files, ~1400 LOC). New adapters: PyInstructionAdapter,
  PyMemoryStrategyAdapter, PySkillStrategyAdapter, PyPersonaAdapter,
  PyMemoryStoreAdapter, PyEmbedderAdapter. Each ships a
  build_guest_* factory mirroring the existing build_guest_toolset
  pattern.
• BoxedAgent + AgentSpec::into_agent (crates/agent/src/boxed.rs).
  Object-erased BoxedAgent for config-driven instantiation. The typed
  Agent<I,T,Ms,Sk> keeps hot-path monomorphization; both share a
  pub(crate) async fn run_turn_impl helper.
• HarnessDispatch + HarnessRef + BoxedHarness (crates/harness/).
  Mirror of the Agent boxed-handle pattern. Harness::into_boxed,
  HarnessSpec::into_harness, and impl HarnessDispatch for Harness<L,T>
  added.
• PyAgent, PyHarness, PyWorkflowRunner Python wrappers with
  from_spec / from_dict constructors and async run_turn / run
  exposed as Python coroutines via
  pyo3_async_runtimes::tokio::future_into_py. WorkflowRunner::new
  was added as a public constructor.

Changed

• EvalSuite::scorer field type: Arc<dyn Scorer> →
  Arc<dyn AsyncScorer>. The blanket impl<S: Scorer> AsyncScorer for S means callers holding a sync Arc::new(SomeSyncScorer)
  continue to compile via auto-unsizing into the new field type.
  Direct &dyn Scorer callers of LlmJudgeScorer / RubricScorer
  must migrate to &dyn AsyncScorer — Rust coherence forbids both
  the blanket AsyncScorer impl and a direct Scorer impl on the
  same judge (ContainsScorer is unaffected).
• Sibling deps (atomr, atomr-infer, atomr-accel) consume
  crates.io artifacts only — no path = "../sibling/..." links.
  Pinned to atomr 0.9.2 / atomr-infer 0.8.0 / atomr-accel 0.10.0.

v0.17.0

Changed — Coding-CLI vendor: Gemini CLI → Antigravity CLI (agy)

Google is transitioning the Gemini CLI into the Antigravity CLI; the
legacy gemini CLI stops serving requests on 2026-06-18. The
coding-cli-vendor-gemini crate is renamed to
atomr-agents-coding-cli-vendor-antigravity and now targets the
agy binary. Unlike the old Gemini CLI, Antigravity serves non-Gemini
models (Claude, etc.) selectable via the model flag.

• GeminiVendor → AntigravityVendor; GeminiParser →
  AntigravityParser; materialize_gemini_config →
  materialize_antigravity_config.
• New AntigravityConfig makes the binary name, headless flags, and the
  on-disk config directory (default .antigravity/) configurable — so
  operators can reconcile them with agy's real interface without a code
  change.
• Harness Cargo feature vendor-gemini → vendor-antigravity.
• Default Docker image gemini.Dockerfile → antigravity.Dockerfile
  (installs agy via the official install.sh).
• Breaking: CliVendorKind::Gemini → CliVendorKind::Antigravity;
  the serde wire value changes from "gemini" to "antigravity", so
  persisted CliRequest JSON with vendor: "gemini" no longer
  deserializes.

Added — Meetings harness: structured extraction from diarized transcripts

A downstream harness that turns a diarized SttConversation into a
structured MeetingAnalysis — attendee roster, linear append-only
notes/actions ledgers, and a tiered dynamic summary stack. See
docs/meetings-harness.md.

• crates/meetings-harness — the harness core: typed +
  Boxed* split mirroring stt-harness; RuleBasedExtractor
  default that is deterministic and LLM-free; pluggable
  MeetingExtractor trait for LLM-driven implementations;
  MeetingsToolSet of Tool impls (list_turns, upsert_attendee,
  append_note, append_action, update_action, revise_tail_segment,
  finalize_segment, regenerate_running, set_title, finalize).
  BatchExtractionLoop and StreamingExtractionLoop — the latter
  subscribes to an STT harness's broadcast and maintains an append-only
  ledger with a revisable in-flight tail segment summary. Persistence
  via MeetingsStore and (feature state) CheckpointerMeetingsStore,
  filed under the same conversation_id as the source transcript.
• crates/meetings-harness-web — an optional axum backend plus an
  embedded React + Vite SPA (port 7100) for reviewing and editing
  analyses: REST CRUD, attendee rename, action patch, run trigger / stop,
  WebSocket live feed of MeetingsHarnessEvents, SPA fallback.
• atomr-agents meetings — new CLI subcommand. meetings analyze --conversation-id <id> --model <model_id> runs the harness against
  an existing transcript; meetings serve boots the web UI. Gated by
  --features meetings / --features meetings-web.

Added — STT harness: agentic streaming speech-to-text pipeline

A new capability layer that diarizes, transcribes, and accumulates a
conversation record aligned to agentic structures. See
docs/stt-harness.md.

• crates/stt-harness — the harness core: audio sources (mic /
  file / bytes), a streaming session actor, diarization + voice-mode
  coalescing, the loop/termination strategies, and a ConversationStore
  with a Checkpointer-backed implementation (state feature).
  BoxedSttHarness mirrors the Agent/Harness boxed-handle pattern.
• crates/stt-harness-web — an optional axum backend plus an
  embedded React SPA (embed-ui feature, via rust-embed) for
  reviewing and editing diarized conversations: REST routes, a
  WebSocket live feed, and the SPA fallback handler.
• atomr-agents serve now runs the real review UI when the CLI is
  built with --features stt-web, persisting conversations through the
  configured crates/state checkpointer.
• xtask stt-web-build — builds the React SPA into ui/dist so
  stt-harness-web can be compiled with --features embed-ui.
• Python bindings — atomr_agents.stt_harness wraps the harness;
  the top-level facade now resolves native names defensively so a name
  not yet bound on the Rust side degrades to None instead of failing
  the whole package import.

Added — Phase C: async scorers, guest adapters, boxed Agent/Harness, Python workflow runner

Closes the "Phase B" deferral comments scattered across crates/agent,
crates/harness, crates/eval, and crates/py-bindings/src/{agent, harness,workflow,eval}.rs. The five blocks:

• AsyncScorer trait (crates/eval/src/scorer.rs). Parallel to
  the existing sync Scorer; blanket impl<S: Scorer> AsyncScorer for S keeps existing sync impls compatible. LlmJudgeScorer,
  RubricScorer, and PairwiseScorer now implement AsyncScorer
  directly — dropping the tokio::task::block_in_place workaround
  the sync Scorer impl required. EvalSuite::scorer field type
  changed from Arc<dyn Scorer> to Arc<dyn AsyncScorer>.
• 6 Python→Rust guest adapters (crates/py-bindings/src/guest/).
  The single-file guest.rs (331 LOC) is split into a guest/
  directory (10 files, ~1400 LOC). New adapters: PyInstructionAdapter,
  PyMemoryStrategyAdapter, PySkillStrategyAdapter, PyPersonaAdapter,
  PyMemoryStoreAdapter, PyEmbedderAdapter. Each ships a
  build_guest_* factory mirroring the existing build_guest_toolset
  pattern.
• BoxedAgent + AgentSpec::into_agent (crates/agent/src/boxed.rs).
  Object-erased BoxedAgent for config-driven instantiation. The typed
  Agent<I,T,Ms,Sk> keeps hot-path monomorphization; both share a
  pub(crate) async fn run_turn_impl helper.
• HarnessDispatch + HarnessRef + BoxedHarness (crates/harness/).
  Mirror of the Agent boxed-handle pattern. Harness::into_boxed,
  HarnessSpec::into_harness, and impl HarnessDispatch for Harness<L,T>
  added.
• PyAgent, PyHarness, PyWorkflowRunner Python wrappers with
  from_spec / from_dict constructors and async run_turn / run
  exposed as Python coroutines via
  pyo3_async_runtimes::tokio::future_into_py. WorkflowRunner::new
  was added as a public constructor.

Changed

• EvalSuite::scorer field type: Arc<dyn Scorer> →
  Arc<dyn AsyncScorer>. The blanket impl<S: Scorer> AsyncScorer for S means callers holding a sync Arc::new(SomeSyncScorer)
  continue to compile via auto-unsizing into the new field type.
  Direct &dyn Scorer callers of LlmJudgeScorer / RubricScorer
  must migrate to &dyn AsyncScorer — Rust coherence forbids both
  the blanket AsyncScorer impl and a direct Scorer impl on the
  same judge (ContainsScorer is unaffected).
• Sibling deps (atomr, atomr-infer, atomr-accel) consume
  crates.io artifacts only — no path = "../sibling/..." links.
  Pinned to atomr 0.9.2 / atomr-infer 0.8.0 / atomr-accel 0.10.0.

v0.16.3

Added — Meetings harness: structured extraction from diarized transcripts

A downstream harness that turns a diarized SttConversation into a
structured MeetingAnalysis — attendee roster, linear append-only
notes/actions ledgers, and a tiered dynamic summary stack. See
docs/meetings-harness.md.

• crates/meetings-harness — the harness core: typed +
  Boxed* split mirroring stt-harness; RuleBasedExtractor
  default that is deterministic and LLM-free; pluggable
  MeetingExtractor trait for LLM-driven implementations;
  MeetingsToolSet of Tool impls (list_turns, upsert_attendee,
  append_note, append_action, update_action, revise_tail_segment,
  finalize_segment, regenerate_running, set_title, finalize).
  BatchExtractionLoop and StreamingExtractionLoop — the latter
  subscribes to an STT harness's broadcast and maintains an append-only
  ledger with a revisable in-flight tail segment summary. Persistence
  via MeetingsStore and (feature state) CheckpointerMeetingsStore,
  filed under the same conversation_id as the source transcript.
• crates/meetings-harness-web — an optional axum backend plus an
  embedded React + Vite SPA (port 7100) for reviewing and editing
  analyses: REST CRUD, attendee rename, action patch, run trigger / stop,
  WebSocket live feed of MeetingsHarnessEvents, SPA fallback.
• atomr-agents meetings — new CLI subcommand. meetings analyze --conversation-id <id> --model <model_id> runs the harness against
  an existing transcript; meetings serve boots the web UI. Gated by
  --features meetings / --features meetings-web.

Added — STT harness: agentic streaming speech-to-text pipeline

A new capability layer that diarizes, transcribes, and accumulates a
conversation record aligned to agentic structures. See
docs/stt-harness.md.

• crates/stt-harness — the harness core: audio sources (mic /
  file / bytes), a streaming session actor, diarization + voice-mode
  coalescing, the loop/termination strategies, and a ConversationStore
  with a Checkpointer-backed implementation (state feature).
  BoxedSttHarness mirrors the Agent/Harness boxed-handle pattern.
• crates/stt-harness-web — an optional axum backend plus an
  embedded React SPA (embed-ui feature, via rust-embed) for
  reviewing and editing diarized conversations: REST routes, a
  WebSocket live feed, and the SPA fallback handler.
• atomr-agents serve now runs the real review UI when the CLI is
  built with --features stt-web, persisting conversations through the
  configured crates/state checkpointer.
• xtask stt-web-build — builds the React SPA into ui/dist so
  stt-harness-web can be compiled with --features embed-ui.
• Python bindings — atomr_agents.stt_harness wraps the harness;
  the top-level facade now resolves native names defensively so a name
  not yet bound on the Rust side degrades to None instead of failing
  the whole package import.

Added — Phase C: async scorers, guest adapters, boxed Agent/Harness, Python workflow runner

Closes the "Phase B" deferral comments scattered across crates/agent,
crates/harness, crates/eval, and crates/py-bindings/src/{agent, harness,workflow,eval}.rs. The five blocks:

• AsyncScorer trait (crates/eval/src/scorer.rs). Parallel to
  the existing sync Scorer; blanket impl<S: Scorer> AsyncScorer for S keeps existing sync impls compatible. LlmJudgeScorer,
  RubricScorer, and PairwiseScorer now implement AsyncScorer
  directly — dropping the tokio::task::block_in_place workaround
  the sync Scorer impl required. EvalSuite::scorer field type
  changed from Arc<dyn Scorer> to Arc<dyn AsyncScorer>.
• 6 Python→Rust guest adapters (crates/py-bindings/src/guest/).
  The single-file guest.rs (331 LOC) is split into a guest/
  directory (10 files, ~1400 LOC). New adapters: PyInstructionAdapter,
  PyMemoryStrategyAdapter, PySkillStrategyAdapter, PyPersonaAdapter,
  PyMemoryStoreAdapter, PyEmbedderAdapter. Each ships a
  build_guest_* factory mirroring the existing build_guest_toolset
  pattern.
• BoxedAgent + AgentSpec::into_agent (crates/agent/src/boxed.rs).
  Object-erased BoxedAgent for config-driven instantiation. The typed
  Agent<I,T,Ms,Sk> keeps hot-path monomorphization; both share a
  pub(crate) async fn run_turn_impl helper.
• HarnessDispatch + HarnessRef + BoxedHarness (crates/harness/).
  Mirror of the Agent boxed-handle pattern. Harness::into_boxed,
  HarnessSpec::into_harness, and impl HarnessDispatch for Harness<L,T>
  added.
• PyAgent, PyHarness, PyWorkflowRunner Python wrappers with
  from_spec / from_dict constructors and async run_turn / run
  exposed as Python coroutines via
  pyo3_async_runtimes::tokio::future_into_py. WorkflowRunner::new
  was added as a public constructor.

Changed

• EvalSuite::scorer field type: Arc<dyn Scorer> →
  Arc<dyn AsyncScorer>. The blanket impl<S: Scorer> AsyncScorer for S means callers holding a sync Arc::new(SomeSyncScorer)
  continue to compile via auto-unsizing into the new field type.
  Direct &dyn Scorer callers of LlmJudgeScorer / RubricScorer
  must migrate to &dyn AsyncScorer — Rust coherence forbids both
  the blanket AsyncScorer impl and a direct Scorer impl on the
  same judge (ContainsScorer is unaffected).
• Sibling deps (atomr, atomr-infer, atomr-accel) consume
  crates.io artifacts only — no path = "../sibling/..." links.
  Pinned to atomr 0.9.2 / atomr-infer 0.8.0 / atomr-accel 0.10.0.

v0.16.2

Added — Meetings harness: structured extraction from diarized transcripts

A downstream harness that turns a diarized SttConversation into a
structured MeetingAnalysis — attendee roster, linear append-only
notes/actions ledgers, and a tiered dynamic summary stack. See
docs/meetings-harness.md.

• crates/meetings-harness — the harness core: typed +
  Boxed* split mirroring stt-harness; RuleBasedExtractor
  default that is deterministic and LLM-free; pluggable
  MeetingExtractor trait for LLM-driven implementations;
  MeetingsToolSet of Tool impls (list_turns, upsert_attendee,
  append_note, append_action, update_action, revise_tail_segment,
  finalize_segment, regenerate_running, set_title, finalize).
  BatchExtractionLoop and StreamingExtractionLoop — the latter
  subscribes to an STT harness's broadcast and maintains an append-only
  ledger with a revisable in-flight tail segment summary. Persistence
  via MeetingsStore and (feature state) CheckpointerMeetingsStore,
  filed under the same conversation_id as the source transcript.
• crates/meetings-harness-web — an optional axum backend plus an
  embedded React + Vite SPA (port 7100) for reviewing and editing
  analyses: REST CRUD, attendee rename, action patch, run trigger / stop,
  WebSocket live feed of MeetingsHarnessEvents, SPA fallback.
• atomr-agents meetings — new CLI subcommand. meetings analyze --conversation-id <id> --model <model_id> runs the harness against
  an existing transcript; meetings serve boots the web UI. Gated by
  --features meetings / --features meetings-web.

Added — STT harness: agentic streaming speech-to-text pipeline

A new capability layer that diarizes, transcribes, and accumulates a
conversation record aligned to agentic structures. See
docs/stt-harness.md.

• crates/stt-harness — the harness core: audio sources (mic /
  file / bytes), a streaming session actor, diarization + voice-mode
  coalescing, the loop/termination strategies, and a ConversationStore
  with a Checkpointer-backed implementation (state feature).
  BoxedSttHarness mirrors the Agent/Harness boxed-handle pattern.
• crates/stt-harness-web — an optional axum backend plus an
  embedded React SPA (embed-ui feature, via rust-embed) for
  reviewing and editing diarized conversations: REST routes, a
  WebSocket live feed, and the SPA fallback handler.
• atomr-agents serve now runs the real review UI when the CLI is
  built with --features stt-web, persisting conversations through the
  configured crates/state checkpointer.
• xtask stt-web-build — builds the React SPA into ui/dist so
  stt-harness-web can be compiled with --features embed-ui.
• Python bindings — atomr_agents.stt_harness wraps the harness;
  the top-level facade now resolves native names defensively so a name
  not yet bound on the Rust side degrades to None instead of failing
  the whole package import.

Added — Phase C: async scorers, guest adapters, boxed Agent/Harness, Python workflow runner

Closes the "Phase B" deferral comments scattered across crates/agent,
crates/harness, crates/eval, and crates/py-bindings/src/{agent, harness,workflow,eval}.rs. The five blocks:

• AsyncScorer trait (crates/eval/src/scorer.rs). Parallel to
  the existing sync Scorer; blanket impl<S: Scorer> AsyncScorer for S keeps existing sync impls compatible. LlmJudgeScorer,
  RubricScorer, and PairwiseScorer now implement AsyncScorer
  directly — dropping the tokio::task::block_in_place workaround
  the sync Scorer impl required. EvalSuite::scorer field type
  changed from Arc<dyn Scorer> to Arc<dyn AsyncScorer>.
• 6 Python→Rust guest adapters (crates/py-bindings/src/guest/).
  The single-file guest.rs (331 LOC) is split into a guest/
  directory (10 files, ~1400 LOC). New adapters: PyInstructionAdapter,
  PyMemoryStrategyAdapter, PySkillStrategyAdapter, PyPersonaAdapter,
  PyMemoryStoreAdapter, PyEmbedderAdapter. Each ships a
  build_guest_* factory mirroring the existing build_guest_toolset
  pattern.
• BoxedAgent + AgentSpec::into_agent (crates/agent/src/boxed.rs).
  Object-erased BoxedAgent for config-driven instantiation. The typed
  Agent<I,T,Ms,Sk> keeps hot-path monomorphization; both share a
  pub(crate) async fn run_turn_impl helper.
• HarnessDispatch + HarnessRef + BoxedHarness (crates/harness/).
  Mirror of the Agent boxed-handle pattern. Harness::into_boxed,
  HarnessSpec::into_harness, and impl HarnessDispatch for Harness<L,T>
  added.
• PyAgent, PyHarness, PyWorkflowRunner Python wrappers with
  from_spec / from_dict constructors and async run_turn / run
  exposed as Python coroutines via
  pyo3_async_runtimes::tokio::future_into_py. WorkflowRunner::new
  was added as a public constructor.

Changed

• EvalSuite::scorer field type: Arc<dyn Scorer> →
  Arc<dyn AsyncScorer>. The blanket impl<S: Scorer> AsyncScorer for S means callers holding a sync Arc::new(SomeSyncScorer)
  continue to compile via auto-unsizing into the new field type.
  Direct &dyn Scorer callers of LlmJudgeScorer / RubricScorer
  must migrate to &dyn AsyncScorer — Rust coherence forbids both
  the blanket AsyncScorer impl and a direct Scorer impl on the
  same judge (ContainsScorer is unaffected).
• Sibling deps (atomr, atomr-infer, atomr-accel) consume
  crates.io artifacts only — no path = "../sibling/..." links.
  Pinned to atomr 0.9.2 / atomr-infer 0.8.0 / atomr-accel 0.10.0.

v0.13.0

Added — Meetings harness: structured extraction from diarized transcripts

A downstream harness that turns a diarized SttConversation into a
structured MeetingAnalysis — attendee roster, linear append-only
notes/actions ledgers, and a tiered dynamic summary stack. See
docs/meetings-harness.md.

• crates/meetings-harness — the harness core: typed +
  Boxed* split mirroring stt-harness; RuleBasedExtractor
  default that is deterministic and LLM-free; pluggable
  MeetingExtractor trait for LLM-driven implementations;
  MeetingsToolSet of Tool impls (list_turns, upsert_attendee,
  append_note, append_action, update_action, revise_tail_segment,
  finalize_segment, regenerate_running, set_title, finalize).
  BatchExtractionLoop and StreamingExtractionLoop — the latter
  subscribes to an STT harness's broadcast and maintains an append-only
  ledger with a revisable in-flight tail segment summary. Persistence
  via MeetingsStore and (feature state) CheckpointerMeetingsStore,
  filed under the same conversation_id as the source transcript.
• crates/meetings-harness-web — an optional axum backend plus an
  embedded React + Vite SPA (port 7100) for reviewing and editing
  analyses: REST CRUD, attendee rename, action patch, run trigger / stop,
  WebSocket live feed of MeetingsHarnessEvents, SPA fallback.
• atomr-agents meetings — new CLI subcommand. meetings analyze --conversation-id <id> --model <model_id> runs the harness against
  an existing transcript; meetings serve boots the web UI. Gated by
  --features meetings / --features meetings-web.

Added — STT harness: agentic streaming speech-to-text pipeline

A new capability layer that diarizes, transcribes, and accumulates a
conversation record aligned to agentic structures. See
docs/stt-harness.md.

• crates/stt-harness — the harness core: audio sources (mic /
  file / bytes), a streaming session actor, diarization + voice-mode
  coalescing, the loop/termination strategies, and a ConversationStore
  with a Checkpointer-backed implementation (state feature).
  BoxedSttHarness mirrors the Agent/Harness boxed-handle pattern.
• crates/stt-harness-web — an optional axum backend plus an
  embedded React SPA (embed-ui feature, via rust-embed) for
  reviewing and editing diarized conversations: REST routes, a
  WebSocket live feed, and the SPA fallback handler.
• atomr-agents serve now runs the real review UI when the CLI is
  built with --features stt-web, persisting conversations through the
  configured crates/state checkpointer.
• xtask stt-web-build — builds the React SPA into ui/dist so
  stt-harness-web can be compiled with --features embed-ui.
• Python bindings — atomr_agents.stt_harness wraps the harness;
  the top-level facade now resolves native names defensively so a name
  not yet bound on the Rust side degrades to None instead of failing
  the whole package import.

Added — Phase C: async scorers, guest adapters, boxed Agent/Harness, Python workflow runner

Closes the "Phase B" deferral comments scattered across crates/agent,
crates/harness, crates/eval, and crates/py-bindings/src/{agent, harness,workflow,eval}.rs. The five blocks:

• AsyncScorer trait (crates/eval/src/scorer.rs). Parallel to
  the existing sync Scorer; blanket impl<S: Scorer> AsyncScorer for S keeps existing sync impls compatible. LlmJudgeScorer,
  RubricScorer, and PairwiseScorer now implement AsyncScorer
  directly — dropping the tokio::task::block_in_place workaround
  the sync Scorer impl required. EvalSuite::scorer field type
  changed from Arc<dyn Scorer> to Arc<dyn AsyncScorer>.
• 6 Python→Rust guest adapters (crates/py-bindings/src/guest/).
  The single-file guest.rs (331 LOC) is split into a guest/
  directory (10 files, ~1400 LOC). New adapters: PyInstructionAdapter,
  PyMemoryStrategyAdapter, PySkillStrategyAdapter, PyPersonaAdapter,
  PyMemoryStoreAdapter, PyEmbedderAdapter. Each ships a
  build_guest_* factory mirroring the existing build_guest_toolset
  pattern.
• BoxedAgent + AgentSpec::into_agent (crates/agent/src/boxed.rs).
  Object-erased BoxedAgent for config-driven instantiation. The typed
  Agent<I,T,Ms,Sk> keeps hot-path monomorphization; both share a
  pub(crate) async fn run_turn_impl helper.
• HarnessDispatch + HarnessRef + BoxedHarness (crates/harness/).
  Mirror of the Agent boxed-handle pattern. Harness::into_boxed,
  HarnessSpec::into_harness, and impl HarnessDispatch for Harness<L,T>
  added.
• PyAgent, PyHarness, PyWorkflowRunner Python wrappers with
  from_spec / from_dict constructors and async run_turn / run
  exposed as Python coroutines via
  pyo3_async_runtimes::tokio::future_into_py. WorkflowRunner::new
  was added as a public constructor.

Changed

• EvalSuite::scorer field type: Arc<dyn Scorer> →
  Arc<dyn AsyncScorer>. The blanket impl<S: Scorer> AsyncScorer for S means callers holding a sync Arc::new(SomeSyncScorer)
  continue to compile via auto-unsizing into the new field type.
  Direct &dyn Scorer callers of LlmJudgeScorer / RubricScorer
  must migrate to &dyn AsyncScorer — Rust coherence forbids both
  the blanket AsyncScorer impl and a direct Scorer impl on the
  same judge (ContainsScorer is unaffected).
• Sibling deps (atomr, atomr-infer, atomr-accel) consume
  crates.io artifacts only — no path = "../sibling/..." links.
  Pinned to atomr 0.9.2 / atomr-infer 0.8.0 / atomr-accel 0.10.0.

v0.12.0

Added — Meetings harness: structured extraction from diarized transcripts

A downstream harness that turns a diarized SttConversation into a
structured MeetingAnalysis — attendee roster, linear append-only
notes/actions ledgers, and a tiered dynamic summary stack. See
docs/meetings-harness.md.

• crates/meetings-harness — the harness core: typed +
  Boxed* split mirroring stt-harness; RuleBasedExtractor
  default that is deterministic and LLM-free; pluggable
  MeetingExtractor trait for LLM-driven implementations;
  MeetingsToolSet of Tool impls (list_turns, upsert_attendee,
  append_note, append_action, update_action, revise_tail_segment,
  finalize_segment, regenerate_running, set_title, finalize).
  BatchExtractionLoop and StreamingExtractionLoop — the latter
  subscribes to an STT harness's broadcast and maintains an append-only
  ledger with a revisable in-flight tail segment summary. Persistence
  via MeetingsStore and (feature state) CheckpointerMeetingsStore,
  filed under the same conversation_id as the source transcript.
• crates/meetings-harness-web — an optional axum backend plus an
  embedded React + Vite SPA (port 7100) for reviewing and editing
  analyses: REST CRUD, attendee rename, action patch, run trigger / stop,
  WebSocket live feed of MeetingsHarnessEvents, SPA fallback.
• atomr-agents meetings — new CLI subcommand. meetings analyze --conversation-id <id> --model <model_id> runs the harness against
  an existing transcript; meetings serve boots the web UI. Gated by
  --features meetings / --features meetings-web.

Added — STT harness: agentic streaming speech-to-text pipeline

A new capability layer that diarizes, transcribes, and accumulates a
conversation record aligned to agentic structures. See
docs/stt-harness.md.

• crates/stt-harness — the harness core: audio sources (mic /
  file / bytes), a streaming session actor, diarization + voice-mode
  coalescing, the loop/termination strategies, and a ConversationStore
  with a Checkpointer-backed implementation (state feature).
  BoxedSttHarness mirrors the Agent/Harness boxed-handle pattern.
• crates/stt-harness-web — an optional axum backend plus an
  embedded React SPA (embed-ui feature, via rust-embed) for
  reviewing and editing diarized conversations: REST routes, a
  WebSocket live feed, and the SPA fallback handler.
• atomr-agents serve now runs the real review UI when the CLI is
  built with --features stt-web, persisting conversations through the
  configured crates/state checkpointer.
• xtask stt-web-build — builds the React SPA into ui/dist so
  stt-harness-web can be compiled with --features embed-ui.
• Python bindings — atomr_agents.stt_harness wraps the harness;
  the top-level facade now resolves native names defensively so a name
  not yet bound on the Rust side degrades to None instead of failing
  the whole package import.

Added — Phase C: async scorers, guest adapters, boxed Agent/Harness, Python workflow runner

Closes the "Phase B" deferral comments scattered across crates/agent,
crates/harness, crates/eval, and crates/py-bindings/src/{agent, harness,workflow,eval}.rs. The five blocks:

• AsyncScorer trait (crates/eval/src/scorer.rs). Parallel to
  the existing sync Scorer; blanket impl<S: Scorer> AsyncScorer for S keeps existing sync impls compatible. LlmJudgeScorer,
  RubricScorer, and PairwiseScorer now implement AsyncScorer
  directly — dropping the tokio::task::block_in_place workaround
  the sync Scorer impl required. EvalSuite::scorer field type
  changed from Arc<dyn Scorer> to Arc<dyn AsyncScorer>.
• 6 Python→Rust guest adapters (crates/py-bindings/src/guest/).
  The single-file guest.rs (331 LOC) is split into a guest/
  directory (10 files, ~1400 LOC). New adapters: PyInstructionAdapter,
  PyMemoryStrategyAdapter, PySkillStrategyAdapter, PyPersonaAdapter,
  PyMemoryStoreAdapter, PyEmbedderAdapter. Each ships a
  build_guest_* factory mirroring the existing build_guest_toolset
  pattern.
• BoxedAgent + AgentSpec::into_agent (crates/agent/src/boxed.rs).
  Object-erased BoxedAgent for config-driven instantiation. The typed
  Agent<I,T,Ms,Sk> keeps hot-path monomorphization; both share a
  pub(crate) async fn run_turn_impl helper.
• HarnessDispatch + HarnessRef + BoxedHarness (crates/harness/).
  Mirror of the Agent boxed-handle pattern. Harness::into_boxed,
  HarnessSpec::into_harness, and impl HarnessDispatch for Harness<L,T>
  added.
• PyAgent, PyHarness, PyWorkflowRunner Python wrappers with
  from_spec / from_dict constructors and async run_turn / run
  exposed as Python coroutines via
  pyo3_async_runtimes::tokio::future_into_py. WorkflowRunner::new
  was added as a public constructor.

Changed

• EvalSuite::scorer field type: Arc<dyn Scorer> →
  Arc<dyn AsyncScorer>. The blanket impl<S: Scorer> AsyncScorer for S means callers holding a sync Arc::new(SomeSyncScorer)
  continue to compile via auto-unsizing into the new field type.
  Direct &dyn Scorer callers of LlmJudgeScorer / RubricScorer
  must migrate to &dyn AsyncScorer — Rust coherence forbids both
  the blanket AsyncScorer impl and a direct Scorer impl on the
  same judge (ContainsScorer is unaffected).
• Sibling deps (atomr, atomr-infer, atomr-accel) consume
  crates.io artifacts only — no path = "../sibling/..." links.
  Pinned to atomr 0.9.2 / atomr-infer 0.8.0 / atomr-accel 0.10.0.

v0.11.0

Added — Meetings harness: structured extraction from diarized transcripts

A downstream harness that turns a diarized SttConversation into a
structured MeetingAnalysis — attendee roster, linear append-only
notes/actions ledgers, and a tiered dynamic summary stack. See
docs/meetings-harness.md.

• crates/meetings-harness — the harness core: typed +
  Boxed* split mirroring stt-harness; RuleBasedExtractor
  default that is deterministic and LLM-free; pluggable
  MeetingExtractor trait for LLM-driven implementations;
  MeetingsToolSet of Tool impls (list_turns, upsert_attendee,
  append_note, append_action, update_action, revise_tail_segment,
  finalize_segment, regenerate_running, set_title, finalize).
  BatchExtractionLoop and StreamingExtractionLoop — the latter
  subscribes to an STT harness's broadcast and maintains an append-only
  ledger with a revisable in-flight tail segment summary. Persistence
  via MeetingsStore and (feature state) CheckpointerMeetingsStore,
  filed under the same conversation_id as the source transcript.
• crates/meetings-harness-web — an optional axum backend plus an
  embedded React + Vite SPA (port 7100) for reviewing and editing
  analyses: REST CRUD, attendee rename, action patch, run trigger / stop,
  WebSocket live feed of MeetingsHarnessEvents, SPA fallback.
• atomr-agents meetings — new CLI subcommand. meetings analyze --conversation-id <id> --model <model_id> runs the harness against
  an existing transcript; meetings serve boots the web UI. Gated by
  --features meetings / --features meetings-web.

Added — STT harness: agentic streaming speech-to-text pipeline

A new capability layer that diarizes, transcribes, and accumulates a
conversation record aligned to agentic structures. See
docs/stt-harness.md.

• crates/stt-harness — the harness core: audio sources (mic /
  file / bytes), a streaming session actor, diarization + voice-mode
  coalescing, the loop/termination strategies, and a ConversationStore
  with a Checkpointer-backed implementation (state feature).
  BoxedSttHarness mirrors the Agent/Harness boxed-handle pattern.
• crates/stt-harness-web — an optional axum backend plus an
  embedded React SPA (embed-ui feature, via rust-embed) for
  reviewing and editing diarized conversations: REST routes, a
  WebSocket live feed, and the SPA fallback handler.
• atomr-agents serve now runs the real review UI when the CLI is
  built with --features stt-web, persisting conversations through the
  configured crates/state checkpointer.
• xtask stt-web-build — builds the React SPA into ui/dist so
  stt-harness-web can be compiled with --features embed-ui.
• Python bindings — atomr_agents.stt_harness wraps the harness;
  the top-level facade now resolves native names defensively so a name
  not yet bound on the Rust side degrades to None instead of failing
  the whole package import.

Added — Phase C: async scorers, guest adapters, boxed Agent/Harness, Python workflow runner

Closes the "Phase B" deferral comments scattered across crates/agent,
crates/harness, crates/eval, and crates/py-bindings/src/{agent, harness,workflow,eval}.rs. The five blocks:

• AsyncScorer trait (crates/eval/src/scorer.rs). Parallel to
  the existing sync Scorer; blanket impl<S: Scorer> AsyncScorer for S keeps existing sync impls compatible. LlmJudgeScorer,
  RubricScorer, and PairwiseScorer now implement AsyncScorer
  directly — dropping the tokio::task::block_in_place workaround
  the sync Scorer impl required. EvalSuite::scorer field type
  changed from Arc<dyn Scorer> to Arc<dyn AsyncScorer>.
• 6 Python→Rust guest adapters (crates/py-bindings/src/guest/).
  The single-file guest.rs (331 LOC) is split into a guest/
  directory (10 files, ~1400 LOC). New adapters: PyInstructionAdapter,
  PyMemoryStrategyAdapter, PySkillStrategyAdapter, PyPersonaAdapter,
  PyMemoryStoreAdapter, PyEmbedderAdapter. Each ships a
  build_guest_* factory mirroring the existing build_guest_toolset
  pattern.
• BoxedAgent + AgentSpec::into_agent (crates/agent/src/boxed.rs).
  Object-erased BoxedAgent for config-driven instantiation. The typed
  Agent<I,T,Ms,Sk> keeps hot-path monomorphization; both share a
  pub(crate) async fn run_turn_impl helper.
• HarnessDispatch + HarnessRef + BoxedHarness (crates/harness/).
  Mirror of the Agent boxed-handle pattern. Harness::into_boxed,
  HarnessSpec::into_harness, and impl HarnessDispatch for Harness<L,T>
  added.
• PyAgent, PyHarness, PyWorkflowRunner Python wrappers with
  from_spec / from_dict constructors and async run_turn / run
  exposed as Python coroutines via
  pyo3_async_runtimes::tokio::future_into_py. WorkflowRunner::new
  was added as a public constructor.

Changed

• EvalSuite::scorer field type: Arc<dyn Scorer> →
  Arc<dyn AsyncScorer>. The blanket impl<S: Scorer> AsyncScorer for S means callers holding a sync Arc::new(SomeSyncScorer)
  continue to compile via auto-unsizing into the new field type.
  Direct &dyn Scorer callers of LlmJudgeScorer / RubricScorer
  must migrate to &dyn AsyncScorer — Rust coherence forbids both
  the blanket AsyncScorer impl and a direct Scorer impl on the
  same judge (ContainsScorer is unaffected).
• Sibling deps (atomr, atomr-infer, atomr-accel) consume
  crates.io artifacts only — no path = "../sibling/..." links.
  Pinned to atomr 0.9.2 / atomr-infer 0.8.0 / atomr-accel 0.10.0.