AbstractRuntime 0.16.0__tar.gz → 0.17.0__tar.gz

{abstractruntime-0.16.0 → abstractruntime-0.17.0}/CHANGELOG.md

@@ -10,6 +10,55 @@ Unreleased changes live in [`changelog.d/`](changelog.d/README.md), one file per
 
 ## [Unreleased]
 
+## [0.17.0] - 2026-11-03
+
+### Added
+
+- **`EffectInterrupted`: a host stop interrupts an effect, it never fails it** (`core/effect_cancellation.py`,
+  exported as `abstractruntime.EffectInterrupted`; gateway backlog 1103). A host that stops on purpose (a gateway stop
+  or restart, a model-host restart for an update) raises it into the effect in flight. The runtime writes no terminal
+  record and re-raises: the ledger tail stays STARTED, the run stays RUNNING at its last durable step, and the next
+  tick (after the host is back) sends the effect again under the same idempotency key as attempt 1, so no retry is
+  consumed and nothing reaches an automation's outcome. A BaseException like `EffectKilled`, so no `except Exception`
+  on the way turns it into a failure. The interrupted call's live text ends `cancelled` with `detail: "interrupted"`.
+- **Public updates carry their type in replay** (`replay/view.py`). A `public_update` message has
+  `update_kind`: `"finding"` when the actor typed the event `kind: "finding"` (a fact or piece of
+  evidence it established), otherwise `"progress"` (what it is doing), including events written
+  before the field existed. Clients use it as given, for example to list the findings of a
+  conversation. See [docs/replay.md](docs/replay.md#public-conversation-messages).
+- **Hosts tell model errors apart through the discovery facade** (`integrations/abstractcore/discovery_facade.py`):
+  `ModelNotDownloadedError` and `ModelNotFoundError` are AbstractCore's own classes, so a host (the gateway's
+  "model not downloaded" sentence) catches exactly what Core raises and reads `model` / `provider` without importing
+  AbstractCore.
+- **Replayed tool calls keep the model's call id** (`replay/view.py`). An `activities[].tool_calls[]`
+  row keyed by a runtime call id (`id`) also carries `call_id`, the model's own id, which is what a
+  public update's `source_call_ids` cite; a client can now resolve a finding to the results it cites.
+
+### Changed
+
+- **Compare reads text next to a number or true/false exactly** (`visual/builtins.py` `_exact_from_text`, ADR 0049 as
+  amended 2026-11-03: "we must have automatic conversion"). When one side is a number or true/false and the other is
+  text, the text is read as that type by the Flow editor's literal grammar (a decimal number, ASCII, no `_`, finite,
+  a whole number kept exact beyond 2**53; or `true`/`false` in any case), so `5 == "5"` and `"9" < 10` are true. This
+  covers flows saved by AbstractFlow 0.13.0 (a typed value stored as text next to a wired number) and text arriving at
+  run time from a form, an API input or a model, with no re-save. `"1_000"`, `"0x10"`, `"inf"`, `"nan"`, full-width
+  digits and empty text are refused. Text against text is never converted (`"10" > "9"` compares text). The table is
+  `tests/fixtures/visualflow/compare_text_operands.json`.
+- **Compare refuses values it cannot compare instead of answering False** (`visual/builtins.py` `control_compare`,
+  ADR 0049). Text that is not a number next to a number, or not `true`/`false` next to true/false (`5 == "five"`,
+  every operator; text that is a number or true/false is read as one, see "Compare reads text next to a number or true/false exactly" above), and `<`/`<=`/`>`/`>=` between values Python cannot order
+  (`None > 3`, a list and a number), now fail the step with an error naming the node and pins: "Compare 'Is five'
+  (node same) ==: b is text ('five'), not a number; a is a number (5)." Before, these answered False silently.
+  Equality of same-type values and Python's equality across other types (`None == 5`) are unchanged; true/false
+  compares with numbers as Python's booleans do (`True == 1`).
+- **AbstractCore floor 2.33.0** in the base install and the `apple` and `gpu` extras (was 2.32.0).
+
+### Fixed
+
+- **An automation occurrence whose agent completed with `success: false` names the agent's reason**
+  (`automations/attention.py`). The attention item and the `automation_failed` email said "completed" (the run's status
+  word); the reason is now read from the Agent node's `meta.error`, and when there is none it says so plainly.
+
 ## [0.16.0] - 2026-11-02
 
 ### Added

{abstractruntime-0.16.0 → abstractruntime-0.17.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: AbstractRuntime
-Version: 0.16.0
+Version: 0.17.0
 Summary: AbstractRuntime: a durable graph runner designed to pair with AbstractCore.
 Project-URL: AbstractCore (website), https://www.abstractcore.ai/
 Project-URL: AbstractRuntime (GitHub), https://github.com/lpalbou/abstractruntime
@@ -20,7 +20,7 @@ Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Classifier: Programming Language :: Python :: 3.13
 Requires-Python: >=3.10
-Requires-Dist: abstractcore>=2.32.0
+Requires-Dist: abstractcore>=2.33.0
 Requires-Dist: abstractmemory>=0.4.0
 Requires-Dist: abstractsemantics>=0.0.5
 Requires-Dist: anyio<5.0.0,>=4.12.1
@@ -35,14 +35,14 @@ Requires-Dist: reportlab<5.0.0,>=4.0.0
 Requires-Dist: restrictedpython>=7.0
 Requires-Dist: unstructured[docx,odt,pptx,rtf,xlsx]<0.19.0,>=0.18.32
 Provides-Extra: apple
-Requires-Dist: abstractcore[apple]>=2.32.0; extra == 'apple'
+Requires-Dist: abstractcore[apple]>=2.33.0; extra == 'apple'
 Requires-Dist: setuptools<82.0.0,>=80.10.2; extra == 'apple'
 Provides-Extra: docs
 Requires-Dist: mkdocs-material>=9.0.0; extra == 'docs'
 Requires-Dist: mkdocs>=1.6.0; extra == 'docs'
 Requires-Dist: pymdown-extensions>=10.0; extra == 'docs'
 Provides-Extra: gpu
-Requires-Dist: abstractcore[gpu]>=2.32.0; extra == 'gpu'
+Requires-Dist: abstractcore[gpu]>=2.33.0; extra == 'gpu'
 Requires-Dist: setuptools<82.0.0,>=77.0.3; extra == 'gpu'
 Provides-Extra: test
 Requires-Dist: aiosmtpd<2.0.0,>=1.4.4; extra == 'test'
@@ -55,16 +55,32 @@ Description-Content-Type: text/markdown
 
 It is designed for long-running workflows that must survive restarts and explicitly model blocking (human input, timers, external events, subworkflows) without keeping Python stacks alive.
 
-**Version:** 0.16.0 • **Python:** 3.10+
+**Version:** 0.17.0 • **Python:** 3.10+
 
 **Status:** pre-1.0 (API may evolve). For production use, pin versions and follow `CHANGELOG.md`.
 
+## What's new in 0.17.0
+
+- **A host stop suspends a run, it never fails it**: a host that stops on purpose raises `EffectInterrupted` into
+  the effect in flight; the runtime writes no terminal record, the run stays running at its last durable step, and
+  the next tick sends the call again as attempt 1 (no retry used, no automation outcome, nothing notified).
+- **Public updates are typed**: a `public_update` in the replay view carries `update_kind`, `"finding"` (a fact or
+  piece of evidence the actor established) or `"progress"` (what it is doing), so a client can list a conversation's
+  findings. See [docs/replay.md](docs/replay.md#public-conversation-messages).
+- **Replayed tool calls keep the model's call id**: an `activities[].tool_calls[]` row also carries `call_id`, the id
+  a public update's `source_call_ids` cite, so a finding resolves to the results it cites.
+- **Compare reads text next to a number or true/false exactly**: `5 == "5"` and `"9" < 10` are true, text it
+  cannot read as that type (`"five"`, `"0x10"`, empty) fails the step with an error naming the node and pins, and
+  text against text is never converted.
+- **An automation occurrence says why it failed**: when the agent finished without success, the attention item and
+  the notification carry the agent's own error, never "completed". Full list: [CHANGELOG.md](CHANGELOG.md).
+
 ## What's new in 0.16.0
 
 - **Calibrated decisions**: an `LLM_CALL` whose output says `modality: "decision"` returns typed answers
   (`answers`, `decision`) from a decision model instead of generated text; `run_facade.decide(...)` runs one as a
   durable child run and the `decide` VisualFlow node asks a choice, yes/no or score question. Requires AbstractCore
-  2.32.0. See [docs/integrations/abstractcore.md](docs/integrations/abstractcore.md).
+  2.32.0 or newer (AbstractRuntime 0.17.0 requires 2.33.0). See [docs/integrations/abstractcore.md](docs/integrations/abstractcore.md).
 - **Conversational questions and guidance**: `ASK_USER` can ask an essential question (waits indefinitely) or a timed
   clarification with a declared fallback; `Runtime.steer(..., display_text=)` records human guidance in its exact
   words; opt-in public updates and tool-approval `public_action` context appear in the replay view.
@@ -145,7 +161,7 @@ Remote-light runtime:
 pip install abstractruntime
 ```
 
-The base install includes AbstractCore 2.32.0 or newer (its light install:
+The base install includes AbstractCore 2.33.0 or newer (its light install:
 remote providers, tools, vision, voice, audio, music, and decision integration), plus the
 `abstractruntime-mcp-worker` entry point. It keeps inference remote/light by
 default: local engines such as MLX, vLLM, HuggingFace/Torch, Diffusers, and

{abstractruntime-0.16.0 → abstractruntime-0.17.0}/README.md

@@ -4,16 +4,32 @@
 
 It is designed for long-running workflows that must survive restarts and explicitly model blocking (human input, timers, external events, subworkflows) without keeping Python stacks alive.
 
-**Version:** 0.16.0 • **Python:** 3.10+
+**Version:** 0.17.0 • **Python:** 3.10+
 
 **Status:** pre-1.0 (API may evolve). For production use, pin versions and follow `CHANGELOG.md`.
 
+## What's new in 0.17.0
+
+- **A host stop suspends a run, it never fails it**: a host that stops on purpose raises `EffectInterrupted` into
+  the effect in flight; the runtime writes no terminal record, the run stays running at its last durable step, and
+  the next tick sends the call again as attempt 1 (no retry used, no automation outcome, nothing notified).
+- **Public updates are typed**: a `public_update` in the replay view carries `update_kind`, `"finding"` (a fact or
+  piece of evidence the actor established) or `"progress"` (what it is doing), so a client can list a conversation's
+  findings. See [docs/replay.md](docs/replay.md#public-conversation-messages).
+- **Replayed tool calls keep the model's call id**: an `activities[].tool_calls[]` row also carries `call_id`, the id
+  a public update's `source_call_ids` cite, so a finding resolves to the results it cites.
+- **Compare reads text next to a number or true/false exactly**: `5 == "5"` and `"9" < 10` are true, text it
+  cannot read as that type (`"five"`, `"0x10"`, empty) fails the step with an error naming the node and pins, and
+  text against text is never converted.
+- **An automation occurrence says why it failed**: when the agent finished without success, the attention item and
+  the notification carry the agent's own error, never "completed". Full list: [CHANGELOG.md](CHANGELOG.md).
+
 ## What's new in 0.16.0
 
 - **Calibrated decisions**: an `LLM_CALL` whose output says `modality: "decision"` returns typed answers
   (`answers`, `decision`) from a decision model instead of generated text; `run_facade.decide(...)` runs one as a
   durable child run and the `decide` VisualFlow node asks a choice, yes/no or score question. Requires AbstractCore
-  2.32.0. See [docs/integrations/abstractcore.md](docs/integrations/abstractcore.md).
+  2.32.0 or newer (AbstractRuntime 0.17.0 requires 2.33.0). See [docs/integrations/abstractcore.md](docs/integrations/abstractcore.md).
 - **Conversational questions and guidance**: `ASK_USER` can ask an essential question (waits indefinitely) or a timed
   clarification with a declared fallback; `Runtime.steer(..., display_text=)` records human guidance in its exact
   words; opt-in public updates and tool-approval `public_action` context appear in the replay view.
@@ -94,7 +110,7 @@ Remote-light runtime:
 pip install abstractruntime
 ```
 
-The base install includes AbstractCore 2.32.0 or newer (its light install:
+The base install includes AbstractCore 2.33.0 or newer (its light install:
 remote providers, tools, vision, voice, audio, music, and decision integration), plus the
 `abstractruntime-mcp-worker` entry point. It keeps inference remote/light by
 default: local engines such as MLX, vLLM, HuggingFace/Torch, Diffusers, and

{abstractruntime-0.16.0 → abstractruntime-0.17.0}/docs/adr/0002_execution_modes_local_remote_hybrid.md

@@ -30,6 +30,7 @@ effect payloads, client construction arguments, or Runtime-owned environment var
 - Thin-mode clients can run the workflow logic while delegating inference to a server.
 - Remote mode supports AbstractCore per-request `base_url` routing (dynamic endpoint selection).
 - Tool execution can be gated by trust/sandbox policy outside the router.
+- In every mode, a host stop interrupts an effect (`EffectInterrupted`: no terminal record, re-issued on the next tick); only a failure the effect itself met is recorded as failed.
 
 ### See Also
 - Implementation: [`backlog/completed/005_abstractcore_integration.md`](../backlog/completed/005_abstractcore_integration.md)

{abstractruntime-0.16.0 → abstractruntime-0.17.0}/docs/api.md

@@ -139,8 +139,9 @@ tool-approval waits retain their own contracts.
 An opt-in actor may inherit `_runtime.conversation: {enabled: true}` through native child
 workflows and VisualFlow Agent nodes. This flag does not create an extra inference call.
 Public updates enter replay only through an explicit completed `EMIT_EVENT` named
-`public_update`, with payload `id`, `text`, `run_id`, optional `session_id`, `iteration` and
-`source_call_ids`. The event's owner must match `run_id`; private diary/thinking fences are
+`public_update`, with payload `id`, `text`, `run_id`, optional `session_id`, `iteration`,
+`source_call_ids` and optional `kind` (`"progress"` or `"finding"`, replayed as the message's
+`update_kind`). The event's owner must match `run_id`; private diary/thinking fences are
 excluded. See [replay.md](replay.md#public-conversation-messages).
 
 Actors may attach `public_action` to a proposed `TOOL_CALLS` payload with `text`,
@@ -458,13 +459,14 @@ Implementation: `src/abstractruntime/replay/` (ADR 0010). Guide: [replay.md](rep
 
 ### AbstractCore (LLM + tools)
 
-Requires: `pip install abstractruntime` (AbstractCore 2.32.0 or newer is part of the base install).
+Requires: `pip install abstractruntime` (AbstractCore 2.33.0 or newer is part of the base install).
 
 Implementation: `src/abstractruntime/integrations/abstractcore/*`.
 
 Entry points:
 - `create_local_runtime(...)`, `create_remote_runtime(...)`, `create_hybrid_runtime(...)` (`src/abstractruntime/integrations/abstractcore/factory.py`)
 - public discovery facade: `AbstractCoreDiscoveryFacade`, `get_abstractcore_discovery_facade(...)` (`src/abstractruntime/integrations/abstractcore/discovery_facade.py`)
+- model errors a host tells apart, AbstractCore's own classes: `from abstractruntime.integrations.abstractcore.discovery_facade import ModelNotDownloadedError, ModelNotFoundError` (`ModelNotDownloadedError` is a `ModelNotFoundError` whose `model` and `provider` name what to download); a host never imports `abstractcore` for them
 - public host facade: `AbstractCoreHostFacade`, `get_abstractcore_host_facade(...)` (`src/abstractruntime/integrations/abstractcore/host_facade.py`)
 - public voice library facade: `AbstractCoreVoiceFacade`, `get_abstractcore_voice_facade(...)`, `VoiceLibraryError` (`src/abstractruntime/integrations/abstractcore/voice_facade.py`): languages, cloned voices (list by language, read, rename or replace the transcript, delete), cloning from a recording, voice design, the capability catalog; local in process, remote over the AbstractCore server's `/v1` voice routes (see [Voice library](integrations/abstractcore.md#voice-library))
 - email facade: `email_facade` (AbstractCore's mail library for hosts; see [Email](#email))

{abstractruntime-0.16.0 → abstractruntime-0.17.0}/docs/architecture.md

@@ -259,6 +259,12 @@ ONE ordering law ([backlog 0045](backlog/planned/runtime_systemic_reliability/00
    completion/resume path increments `RuntimeHealth` (`record_error`) and
    logs a warning. Best-effort observability records (progress events,
    status events) may stay quiet; records that replay depends on may not.
+4. **A requested host stop is a crash the host announces: identical replay, no terminal record.**
+   A host that stops on purpose (a gateway stop or restart, a model-host restart for an update)
+   raises `EffectInterrupted` (`core/effect_cancellation.py`) into the effect in flight. The
+   runtime appends nothing terminal and re-raises; the run stays RUNNING, and the next tick sends
+   the effect again under the same idempotency key as attempt 1 (no retry consumed). Only a
+   failure the effect itself met is recorded as failed.
 
 Kill-and-replay coverage for this law lives in the crash harness
 (`tests/harness.py`, backlog 0045): for every persistence point, kill →

abstractruntime-0.17.0/docs/backlog/proposed/0865_model_loaded_ledger_event.md

@@ -0,0 +1,27 @@
+# 0865 — A `model_loaded` ledger event with the reserved context
+
+> Package: AbstractRuntime (with AbstractCore residency claims)
+> Type: task
+> Created: 2026-10-29
+> Priority: normal
+> Labels: residency, ledger, observability
+
+## Summary
+
+Operator ruling (2026-10-29): when a model loads it reserves the maximum context that fits (AbstractCore ADR 0010
+point 8), and "never silently: the choice is written to the ledger (a model_loaded event with model, advertised context,
+reserved context, free memory, buffer, reason) and to the log". AbstractCore already logs it and carries it on the
+provider's residency claim (`context_reservation`: advertised, reserved, free_bytes, buffer, reason). The ledger event
+is not written yet: a load happens inside a client build, outside any run's effect, so there is no step record to
+attach it to.
+
+## Scope
+
+- When a pooled client is built for a local model (`MultiLocalAbstractCoreLLMClient._get_client`), read the provider's
+  `context_reservation` and append one `model_loaded` record to the ledger of the run that triggered the build (or a
+  host-level ledger when no run did), with model, advertised context, reserved context, free memory, buffer, reason.
+- The replay view and Observer show it as one readable line.
+
+## Acceptance criteria
+
+- A chat that loads the default model leaves one `model_loaded` record with the numbers; a test is red without it.

{abstractruntime-0.16.0 → abstractruntime-0.17.0}/docs/getting-started.md

@@ -12,7 +12,7 @@ Remote-light runtime:
 pip install abstractruntime
 ```
 
-This installs AbstractCore 2.32.0 or newer with remote provider, vision,
+This installs AbstractCore 2.33.0 or newer with remote provider, vision,
 voice, audio, music, decision, tool, and MCP-worker support. The base install is
 remote-light: it can route multimodal workflows to hosted or OpenAI-compatible
 endpoints, but it does not select local inferencer stacks such as MLX, vLLM,

{abstractruntime-0.16.0 → abstractruntime-0.17.0}/docs/integrations/abstractcore.md

@@ -16,7 +16,7 @@ Implementation pointers (this repo):
 pip install abstractruntime
 ```
 
-The base install includes AbstractCore 2.32.0 or newer. That is the supported baseline for the `decision` output modality (the `output.decision` route and `/v1/decisions`), the current server auth split (`Authorization` for server auth, `X-AbstractCore-Provider-API-Key` for provider overrides), generated-media contracts, image upscaling, capability catalog, prompt-cache control-plane endpoints (including session attribution via `/acore/prompt_cache/key_meta`), host memory snapshots, the host-wide loaded-model sweep used by residency listings, the process-wide MLX residency report and eject used by model unload, durable bloc prompt-cache helpers, bindings and lifecycle operations, task-aware model residency for text/image/video/TTS/STT, current tool catalog, AbstractCore's public output-selector contract, async/sync text-generation output-selector parity, video generation endpoints, the public local vision-cache catalog helper used by Runtime discovery, vision adapter discovery plus batch/LoRA media controls, and the released shared workspace/file-filter utility surface used by Runtime packaging and integration checks.
+The base install includes AbstractCore 2.33.0 or newer. That is the supported baseline for the `decision` output modality (the `output.decision` route and `/v1/decisions`), the current server auth split (`Authorization` for server auth, `X-AbstractCore-Provider-API-Key` for provider overrides), generated-media contracts, image upscaling, capability catalog, prompt-cache control-plane endpoints (including session attribution via `/acore/prompt_cache/key_meta`), host memory snapshots, the host-wide loaded-model sweep used by residency listings, the process-wide MLX residency report and eject used by model unload, durable bloc prompt-cache helpers, bindings and lifecycle operations, task-aware model residency for text/image/video/TTS/STT, current tool catalog, AbstractCore's public output-selector contract, async/sync text-generation output-selector parity, video generation endpoints, the public local vision-cache catalog helper used by Runtime discovery, vision adapter discovery plus batch/LoRA media controls, and the released shared workspace/file-filter utility surface used by Runtime packaging and integration checks.
 
 The base install also includes the remote-light media/capability plugins needed
 for AbstractCore's multimodal `generate(..., output=...)` path. Local
@@ -486,7 +486,7 @@ When both hold, every `LLM_CALL` of the run streams from the provider and the ru
 - `call_id` is the `step_id` of the `LLM_CALL` ledger record that later holds the final answer, so a client can replace its live text with the durable result.
 - `seq` counts the events of one call from 0, `llm.delta_end` included.
 - `channel` is `content` (the answer) or `reasoning` (the model's thinking). Reasoning arrives from the provider's reasoning stream, and inline `<think>…</think>` markup is split out of the content live, so clients can hide or fold it. The markup itself is never sent.
-- `reason` is `completed`, `failed`, `cancelled` or `unavailable`. Every live fragment of a call is sent before the call's durable record is written (the last batch is flushed first, and later fragments are dropped); every call then ends with exactly one `llm.delta_end`, sent after that record is in the ledger. Each retry attempt is its own call with its own `call_id`; so is the rare re-run of an attempt interrupted by a stray kill (`<step_id>:reinvoke`), whose interrupted first run ends `cancelled` with `detail: "reinvoked"` (the run goes on; a client can say "reply restarted" rather than "stopped").
+- `reason` is `completed`, `failed`, `cancelled` or `unavailable`. Every live fragment of a call is sent before the call's durable record is written (the last batch is flushed first, and later fragments are dropped); every call then ends with exactly one `llm.delta_end`, sent after that record is in the ledger. Each retry attempt is its own call with its own `call_id`; so is the rare re-run of an attempt interrupted by a stray kill (`<step_id>:reinvoke`), whose interrupted first run ends `cancelled` with `detail: "reinvoked"` (the run goes on; a client can say "reply restarted" rather than "stopped"). A call interrupted because its host stopped on purpose (`EffectInterrupted`) ends `cancelled` with `detail: "interrupted"`: the run is suspended, not stopped, and the call is sent again when the host is back.
 - Tool-call markup a local model writes as text (`<tool_call>`, `<|tool_call|>`, `<function_call>`, …) never reaches the live text. AbstractCore already removes it from streamed content; the runtime also stops the live content of a call at the first such marker. The final record carries the tool calls.
 - Harmony transcripts (gpt-oss) are split live: the `final` channel streams as content, `analysis` as reasoning, and a tool call (`commentary to=functions.…`) is held back. The framing tokens are never sent.
 - Text is coalesced: the first fragment is sent immediately, later fragments are grouped over about 40 ms.

{abstractruntime-0.16.0 → abstractruntime-0.17.0}/docs/replay.md

@@ -25,7 +25,10 @@ client_copy = apply_delta(doc, delta)      # == view.to_json()
 
 Explicit completed `public_update` events become assistant messages with
 `kind: "public_update"`, stable run-scoped `id`, `text`, `run_id`, optional `session_id`,
-`iteration`, `source_call_ids`, and normal ledger provenance. Replay does not turn reasoning,
+`iteration`, `source_call_ids`, `update_kind`, and normal ledger provenance. `update_kind` is
+the producer's type for the update: `"finding"` (a fact or piece of evidence the actor
+established) or `"progress"` (what it is doing); an event without a `kind` of `"finding"`
+replays as `"progress"`. Replay copies the type and never infers it from the text. Replay does not turn reasoning,
 raw tool output or a marker inside model text into public speech. Diary/thinking fences
 are excluded. These are actor-authored updates; their wording alone is not proof of a fact
 or of completed work.
@@ -102,7 +105,7 @@ Each rule below was once a divergence between clients:
 | `interaction` | the person-facing wait, or `null`: `kind` (`tool_approval`, `ask_user`, `event`), `run_id`, `wait_key`, `step_id`, `seq`, `prompt`, `choices`, `allow_free_text`, `details`, `tool_calls` (resolved), `activity_id`, `answer_payload`, and `command` (the resume a client sends to answer it) |
 | `open_waits[]` | every open wait of the tree, with `person_facing` |
 | `messages[]` | verbatim turns: `role`, `kind` (`prompt`, `answer`, `question`, `final`, `event`), `text`, provenance (`run_id`, `step_ids`, `seqs`, `artifact_ids`, `source`) |
-| `activities[]` | one per execution: `kind`, `name`, `status`, `node_id`, `execution_id`, `seqs`, `step_ids`, `tool_calls[]` (each call's status, output, error), `wait` (`kind`, `wait_key`, `ended_by`: `resume`, `later_step`, `terminal`), `usage`, `child_run_id`, `progress` |
+| `activities[]` | one per execution: `kind`, `name`, `status`, `node_id`, `execution_id`, `seqs`, `step_ids`, `tool_calls[]` (each call's `id`, status, output, error, and `call_id`, the model's own call id, when `id` is a runtime call id; public updates cite `call_id`), `wait` (`kind`, `wait_key`, `ended_by`: `resume`, `later_step`, `terminal`), `usage`, `child_run_id`, `progress` |
 | `artifacts[]` | every `{"$artifact": id}` reference, with the run, step, seq and JSON path that produced it |
 | `usage` | tokens and LLM calls summed over the tree |
 | `warnings[]` | what the view could not show (records before a window, an unresolved `$slim`) |

{abstractruntime-0.16.0 → abstractruntime-0.17.0}/docs/troubleshooting.md

@@ -20,7 +20,7 @@ Fix:
 - Install or upgrade the base Runtime package. LLM/tools integration, common
   remote-light multimodal dependencies, and the MCP worker entry point are part
   of the base install.
-- The current AbstractCore integration expects `abstractcore>=2.32.0`.
+- The current AbstractCore integration expects `abstractcore>=2.33.0`.
 
 Verify:
 
@@ -383,6 +383,21 @@ Cause and fix:
 
 See [automations.md, `feed.item@1`](automations.md#feeditem1).
 
+## A Compare step fails with "is text (...), not a number"
+
+Symptom:
+- A run stops at a Compare node: "Compare 'Is five' (node same) ==: b is text ('five'), not a number; a is a number
+  (5)."
+
+Cause and fix:
+- One side is a number (or true/false) and the other is text that is not one. Compare reads text next to a number as
+  that number (`"5"`, `" 5 "`, `"-2.5"`, `"1e3"`) and text next to true/false as `true` or `false` (any case), so
+  `5 == "5"` is true. Anything else is refused: words (`"five"`), `"1_000"`, `"0x10"`, `"inf"`, `"nan"`, full-width
+  digits, and empty text. Give that input a number (or true/false), or, if both sides are meant to be text, wire them
+  so both carry text. See [Pin defaults are typed values](workflow-bundles.md#pin-defaults-are-typed-values).
+- An error ending in "these cannot be ordered" means `<`, `<=`, `>` or `>=` met values Python cannot order, often an
+  unset value (`None`): give the input a value, or compare with `==`.
+
 ## MCP worker command is not found
 
 Symptom:

{abstractruntime-0.16.0 → abstractruntime-0.17.0}/docs/workflow-bundles.md

@@ -100,6 +100,18 @@ Authoring guidance:
 - keep route keys unique per target node
 - use one normal data edge or `pinDefaults` for the fallback value, then add `inputRouteOverrides` only for routes that need a different value
 
+## Pin defaults are typed values
+
+A `pinDefaults` value is a JSON value with its type (`5`, `"5"`, `true`, `[]`), and the runtime uses it as stored.
+AbstractFlow stores a value typed on a node card with the type of its input (ADR 0049 in the AbstractFramework
+repository). Compare (`compare`) makes one conversion: when one side is a number or true/false and the other is text,
+the text is read exactly as that type, by the editor's grammar (a decimal number: `"5"`, `"-2.5"`, `".5"`, `"1e3"`; or
+`true`/`false` in any case), so `5 == "5"` and `"9" < 10` are true, whether the text was saved by AbstractFlow 0.13.0
+or arrives at run time from a form, an API input or a model. Text that is not such a value (`"five"`, `"1_000"`,
+`"0x10"`, `"inf"`, empty), or an order between values Python cannot order (`None > 3`), fails the step with an error
+that names the node, instead of answering `False`. Text against text is never converted (`"10" > "9"` compares text).
+true/false compares with numbers as Python's booleans do (`True == 1`).
+
 ## See also
 
 - `architecture.md` — VisualFlow → WorkflowSpec compilation path

{abstractruntime-0.16.0 → abstractruntime-0.17.0}/llms-full.txt

@@ -55,16 +55,32 @@ See llms.txt for the linked index.
 
 It is designed for long-running workflows that must survive restarts and explicitly model blocking (human input, timers, external events, subworkflows) without keeping Python stacks alive.
 
-**Version:** 0.16.0 • **Python:** 3.10+
+**Version:** 0.17.0 • **Python:** 3.10+
 
 **Status:** pre-1.0 (API may evolve). For production use, pin versions and follow `CHANGELOG.md`.
 
+## What's new in 0.17.0
+
+- **A host stop suspends a run, it never fails it**: a host that stops on purpose raises `EffectInterrupted` into
+  the effect in flight; the runtime writes no terminal record, the run stays running at its last durable step, and
+  the next tick sends the call again as attempt 1 (no retry used, no automation outcome, nothing notified).
+- **Public updates are typed**: a `public_update` in the replay view carries `update_kind`, `"finding"` (a fact or
+  piece of evidence the actor established) or `"progress"` (what it is doing), so a client can list a conversation's
+  findings. See [docs/replay.md](docs/replay.md#public-conversation-messages).
+- **Replayed tool calls keep the model's call id**: an `activities[].tool_calls[]` row also carries `call_id`, the id
+  a public update's `source_call_ids` cite, so a finding resolves to the results it cites.
+- **Compare reads text next to a number or true/false exactly**: `5 == "5"` and `"9" < 10` are true, text it
+  cannot read as that type (`"five"`, `"0x10"`, empty) fails the step with an error naming the node and pins, and
+  text against text is never converted.
+- **An automation occurrence says why it failed**: when the agent finished without success, the attention item and
+  the notification carry the agent's own error, never "completed". Full list: [CHANGELOG.md](CHANGELOG.md).
+
 ## What's new in 0.16.0
 
 - **Calibrated decisions**: an `LLM_CALL` whose output says `modality: "decision"` returns typed answers
   (`answers`, `decision`) from a decision model instead of generated text; `run_facade.decide(...)` runs one as a
   durable child run and the `decide` VisualFlow node asks a choice, yes/no or score question. Requires AbstractCore
-  2.32.0. See [docs/integrations/abstractcore.md](docs/integrations/abstractcore.md).
+  2.32.0 or newer (AbstractRuntime 0.17.0 requires 2.33.0). See [docs/integrations/abstractcore.md](docs/integrations/abstractcore.md).
 - **Conversational questions and guidance**: `ASK_USER` can ask an essential question (waits indefinitely) or a timed
   clarification with a declared fallback; `Runtime.steer(..., display_text=)` records human guidance in its exact
   words; opt-in public updates and tool-approval `public_action` context appear in the replay view.
@@ -145,7 +161,7 @@ Remote-light runtime:
 pip install abstractruntime
 ```
 
-The base install includes AbstractCore 2.32.0 or newer (its light install:
+The base install includes AbstractCore 2.33.0 or newer (its light install:
 remote providers, tools, vision, voice, audio, music, and decision integration), plus the
 `abstractruntime-mcp-worker` entry point. It keeps inference remote/light by
 default: local engines such as MLX, vLLM, HuggingFace/Torch, Diffusers, and
@@ -361,7 +377,7 @@ Remote-light runtime:
 pip install abstractruntime
 ```
 
-This installs AbstractCore 2.32.0 or newer with remote provider, vision,
+This installs AbstractCore 2.33.0 or newer with remote provider, vision,
 voice, audio, music, decision, tool, and MCP-worker support. The base install is
 remote-light: it can route multimodal workflows to hosted or OpenAI-compatible
 endpoints, but it does not select local inferencer stacks such as MLX, vLLM,
@@ -764,6 +780,12 @@ ONE ordering law ([backlog 0045](backlog/planned/runtime_systemic_reliability/00
    completion/resume path increments `RuntimeHealth` (`record_error`) and
    logs a warning. Best-effort observability records (progress events,
    status events) may stay quiet; records that replay depends on may not.
+4. **A requested host stop is a crash the host announces: identical replay, no terminal record.**
+   A host that stops on purpose (a gateway stop or restart, a model-host restart for an update)
+   raises `EffectInterrupted` (`core/effect_cancellation.py`) into the effect in flight. The
+   runtime appends nothing terminal and re-raises; the run stays RUNNING, and the next tick sends
+   the effect again under the same idempotency key as attempt 1 (no retry consumed). Only a
+   failure the effect itself met is recorded as failed.
 
 Kill-and-replay coverage for this law lives in the crash harness
 (`tests/harness.py`, backlog 0045): for every persistence point, kill →
@@ -1215,8 +1237,9 @@ tool-approval waits retain their own contracts.
 An opt-in actor may inherit `_runtime.conversation: {enabled: true}` through native child
 workflows and VisualFlow Agent nodes. This flag does not create an extra inference call.
 Public updates enter replay only through an explicit completed `EMIT_EVENT` named
-`public_update`, with payload `id`, `text`, `run_id`, optional `session_id`, `iteration` and
-`source_call_ids`. The event's owner must match `run_id`; private diary/thinking fences are
+`public_update`, with payload `id`, `text`, `run_id`, optional `session_id`, `iteration`,
+`source_call_ids` and optional `kind` (`"progress"` or `"finding"`, replayed as the message's
+`update_kind`). The event's owner must match `run_id`; private diary/thinking fences are
 excluded. See [replay.md](replay.md#public-conversation-messages).
 
 Actors may attach `public_action` to a proposed `TOOL_CALLS` payload with `text`,
@@ -1534,13 +1557,14 @@ Implementation: `src/abstractruntime/replay/` (ADR 0010). Guide: [replay.md](rep
 
 ### AbstractCore (LLM + tools)
 
-Requires: `pip install abstractruntime` (AbstractCore 2.32.0 or newer is part of the base install).
+Requires: `pip install abstractruntime` (AbstractCore 2.33.0 or newer is part of the base install).
 
 Implementation: `src/abstractruntime/integrations/abstractcore/*`.
 
 Entry points:
 - `create_local_runtime(...)`, `create_remote_runtime(...)`, `create_hybrid_runtime(...)` (`src/abstractruntime/integrations/abstractcore/factory.py`)
 - public discovery facade: `AbstractCoreDiscoveryFacade`, `get_abstractcore_discovery_facade(...)` (`src/abstractruntime/integrations/abstractcore/discovery_facade.py`)
+- model errors a host tells apart, AbstractCore's own classes: `from abstractruntime.integrations.abstractcore.discovery_facade import ModelNotDownloadedError, ModelNotFoundError` (`ModelNotDownloadedError` is a `ModelNotFoundError` whose `model` and `provider` name what to download); a host never imports `abstractcore` for them
 - public host facade: `AbstractCoreHostFacade`, `get_abstractcore_host_facade(...)` (`src/abstractruntime/integrations/abstractcore/host_facade.py`)
 - public voice library facade: `AbstractCoreVoiceFacade`, `get_abstractcore_voice_facade(...)`, `VoiceLibraryError` (`src/abstractruntime/integrations/abstractcore/voice_facade.py`): languages, cloned voices (list by language, read, rename or replace the transcript, delete), cloning from a recording, voice design, the capability catalog; local in process, remote over the AbstractCore server's `/v1` voice routes (see [Voice library](integrations/abstractcore.md#voice-library))
 - email facade: `email_facade` (AbstractCore's mail library for hosts; see [Email](#email))
@@ -2114,7 +2138,7 @@ Fix:
 - Install or upgrade the base Runtime package. LLM/tools integration, common
   remote-light multimodal dependencies, and the MCP worker entry point are part
   of the base install.
-- The current AbstractCore integration expects `abstractcore>=2.32.0`.
+- The current AbstractCore integration expects `abstractcore>=2.33.0`.
 
 Verify:
 
@@ -2477,6 +2501,21 @@ Cause and fix:
 
 See [automations.md, `feed.item@1`](automations.md#feeditem1).
 
+## A Compare step fails with "is text (...), not a number"
+
+Symptom:
+- A run stops at a Compare node: "Compare 'Is five' (node same) ==: b is text ('five'), not a number; a is a number
+  (5)."
+
+Cause and fix:
+- One side is a number (or true/false) and the other is text that is not one. Compare reads text next to a number as
+  that number (`"5"`, `" 5 "`, `"-2.5"`, `"1e3"`) and text next to true/false as `true` or `false` (any case), so
+  `5 == "5"` is true. Anything else is refused: words (`"five"`), `"1_000"`, `"0x10"`, `"inf"`, `"nan"`, full-width
+  digits, and empty text. Give that input a number (or true/false), or, if both sides are meant to be text, wire them
+  so both carry text. See [Pin defaults are typed values](workflow-bundles.md#pin-defaults-are-typed-values).
+- An error ending in "these cannot be ordered" means `<`, `<=`, `>` or `>=` met values Python cannot order, often an
+  unset value (`None`): give the input a value, or compare with `==`.
+
 ## MCP worker command is not found
 
 Symptom:
@@ -2515,7 +2554,7 @@ Implementation pointers (this repo):
 pip install abstractruntime
 ```
 
-The base install includes AbstractCore 2.32.0 or newer. That is the supported baseline for the `decision` output modality (the `output.decision` route and `/v1/decisions`), the current server auth split (`Authorization` for server auth, `X-AbstractCore-Provider-API-Key` for provider overrides), generated-media contracts, image upscaling, capability catalog, prompt-cache control-plane endpoints (including session attribution via `/acore/prompt_cache/key_meta`), host memory snapshots, the host-wide loaded-model sweep used by residency listings, the process-wide MLX residency report and eject used by model unload, durable bloc prompt-cache helpers, bindings and lifecycle operations, task-aware model residency for text/image/video/TTS/STT, current tool catalog, AbstractCore's public output-selector contract, async/sync text-generation output-selector parity, video generation endpoints, the public local vision-cache catalog helper used by Runtime discovery, vision adapter discovery plus batch/LoRA media controls, and the released shared workspace/file-filter utility surface used by Runtime packaging and integration checks.
+The base install includes AbstractCore 2.33.0 or newer. That is the supported baseline for the `decision` output modality (the `output.decision` route and `/v1/decisions`), the current server auth split (`Authorization` for server auth, `X-AbstractCore-Provider-API-Key` for provider overrides), generated-media contracts, image upscaling, capability catalog, prompt-cache control-plane endpoints (including session attribution via `/acore/prompt_cache/key_meta`), host memory snapshots, the host-wide loaded-model sweep used by residency listings, the process-wide MLX residency report and eject used by model unload, durable bloc prompt-cache helpers, bindings and lifecycle operations, task-aware model residency for text/image/video/TTS/STT, current tool catalog, AbstractCore's public output-selector contract, async/sync text-generation output-selector parity, video generation endpoints, the public local vision-cache catalog helper used by Runtime discovery, vision adapter discovery plus batch/LoRA media controls, and the released shared workspace/file-filter utility surface used by Runtime packaging and integration checks.
 
 The base install also includes the remote-light media/capability plugins needed
 for AbstractCore's multimodal `generate(..., output=...)` path. Local
@@ -2985,7 +3024,7 @@ When both hold, every `LLM_CALL` of the run streams from the provider and the ru
 - `call_id` is the `step_id` of the `LLM_CALL` ledger record that later holds the final answer, so a client can replace its live text with the durable result.
 - `seq` counts the events of one call from 0, `llm.delta_end` included.
 - `channel` is `content` (the answer) or `reasoning` (the model's thinking). Reasoning arrives from the provider's reasoning stream, and inline `<think>…</think>` markup is split out of the content live, so clients can hide or fold it. The markup itself is never sent.
-- `reason` is `completed`, `failed`, `cancelled` or `unavailable`. Every live fragment of a call is sent before the call's durable record is written (the last batch is flushed first, and later fragments are dropped); every call then ends with exactly one `llm.delta_end`, sent after that record is in the ledger. Each retry attempt is its own call with its own `call_id`; so is the rare re-run of an attempt interrupted by a stray kill (`<step_id>:reinvoke`), whose interrupted first run ends `cancelled` with `detail: "reinvoked"` (the run goes on; a client can say "reply restarted" rather than "stopped").
+- `reason` is `completed`, `failed`, `cancelled` or `unavailable`. Every live fragment of a call is sent before the call's durable record is written (the last batch is flushed first, and later fragments are dropped); every call then ends with exactly one `llm.delta_end`, sent after that record is in the ledger. Each retry attempt is its own call with its own `call_id`; so is the rare re-run of an attempt interrupted by a stray kill (`<step_id>:reinvoke`), whose interrupted first run ends `cancelled` with `detail: "reinvoked"` (the run goes on; a client can say "reply restarted" rather than "stopped"). A call interrupted because its host stopped on purpose (`EffectInterrupted`) ends `cancelled` with `detail: "interrupted"`: the run is suspended, not stopped, and the call is sent again when the host is back.
 - Tool-call markup a local model writes as text (`<tool_call>`, `<|tool_call|>`, `<function_call>`, …) never reaches the live text. AbstractCore already removes it from streamed content; the runtime also stops the live content of a call at the first such marker. The final record carries the tool calls.
 - Harmony transcripts (gpt-oss) are split live: the `final` channel streams as content, `analysis` as reasoning, and a tool call (`commentary to=functions.…`) is held back. The framing tokens are never sent.
 - Text is coalesced: the first fragment is sent immediately, later fragments are grouped over about 40 ms.
@@ -6543,7 +6582,10 @@ client_copy = apply_delta(doc, delta)      # == view.to_json()
 
 Explicit completed `public_update` events become assistant messages with
 `kind: "public_update"`, stable run-scoped `id`, `text`, `run_id`, optional `session_id`,
-`iteration`, `source_call_ids`, and normal ledger provenance. Replay does not turn reasoning,
+`iteration`, `source_call_ids`, `update_kind`, and normal ledger provenance. `update_kind` is
+the producer's type for the update: `"finding"` (a fact or piece of evidence the actor
+established) or `"progress"` (what it is doing); an event without a `kind` of `"finding"`
+replays as `"progress"`. Replay copies the type and never infers it from the text. Replay does not turn reasoning,
 raw tool output or a marker inside model text into public speech. Diary/thinking fences
 are excluded. These are actor-authored updates; their wording alone is not proof of a fact
 or of completed work.
@@ -6620,7 +6662,7 @@ Each rule below was once a divergence between clients:
 | `interaction` | the person-facing wait, or `null`: `kind` (`tool_approval`, `ask_user`, `event`), `run_id`, `wait_key`, `step_id`, `seq`, `prompt`, `choices`, `allow_free_text`, `details`, `tool_calls` (resolved), `activity_id`, `answer_payload`, and `command` (the resume a client sends to answer it) |
 | `open_waits[]` | every open wait of the tree, with `person_facing` |
 | `messages[]` | verbatim turns: `role`, `kind` (`prompt`, `answer`, `question`, `final`, `event`), `text`, provenance (`run_id`, `step_ids`, `seqs`, `artifact_ids`, `source`) |
-| `activities[]` | one per execution: `kind`, `name`, `status`, `node_id`, `execution_id`, `seqs`, `step_ids`, `tool_calls[]` (each call's status, output, error), `wait` (`kind`, `wait_key`, `ended_by`: `resume`, `later_step`, `terminal`), `usage`, `child_run_id`, `progress` |
+| `activities[]` | one per execution: `kind`, `name`, `status`, `node_id`, `execution_id`, `seqs`, `step_ids`, `tool_calls[]` (each call's `id`, status, output, error, and `call_id`, the model's own call id, when `id` is a runtime call id; public updates cite `call_id`), `wait` (`kind`, `wait_key`, `ended_by`: `resume`, `later_step`, `terminal`), `usage`, `child_run_id`, `progress` |
 | `artifacts[]` | every `{"$artifact": id}` reference, with the run, step, seq and JSON path that produced it |
 | `usage` | tokens and LLM calls summed over the tree |
 | `warnings[]` | what the view could not show (records before a window, an unresolved `$slim`) |
@@ -6865,6 +6907,18 @@ Authoring guidance:
 - keep route keys unique per target node
 - use one normal data edge or `pinDefaults` for the fallback value, then add `inputRouteOverrides` only for routes that need a different value
 
+## Pin defaults are typed values
+
+A `pinDefaults` value is a JSON value with its type (`5`, `"5"`, `true`, `[]`), and the runtime uses it as stored.
+AbstractFlow stores a value typed on a node card with the type of its input (ADR 0049 in the AbstractFramework
+repository). Compare (`compare`) makes one conversion: when one side is a number or true/false and the other is text,
+the text is read exactly as that type, by the editor's grammar (a decimal number: `"5"`, `"-2.5"`, `".5"`, `"1e3"`; or
+`true`/`false` in any case), so `5 == "5"` and `"9" < 10` are true, whether the text was saved by AbstractFlow 0.13.0
+or arrives at run time from a form, an API input or a model. Text that is not such a value (`"five"`, `"1_000"`,
+`"0x10"`, `"inf"`, empty), or an order between values Python cannot order (`None > 3`), fails the step with an error
+that names the node, instead of answering `False`. Text against text is never converted (`"10" > "9"` compares text).
+true/false compares with numbers as Python's booleans do (`True == 1`).
+
 ## See also
 
 - `architecture.md` — VisualFlow → WorkflowSpec compilation path

{abstractruntime-0.16.0 → abstractruntime-0.17.0}/llms.txt

@@ -3,10 +3,10 @@
 > Durable workflow runtime (interrupt → checkpoint → resume) with an append-only execution ledger.
 
 - Python: 3.10+ (see `pyproject.toml`)
-- Version: 0.16.0
+- Version: 0.17.0
 - Ecosystem: [AbstractFramework](https://github.com/lpalbou/AbstractFramework) umbrella, integrates with [AbstractCore](https://github.com/lpalbou/abstractcore)
 - Hard invariant: `RunState.vars` must be JSON-serializable (`src/abstractruntime/core/models.py`)
-- Current AbstractCore baseline: base `abstractruntime` requires AbstractCore 2.32.0 or newer with its light install (remote providers, tools and the MCP worker, media routing, the `decision` output modality); `abstractruntime[apple]` and `abstractruntime[gpu]` require AbstractCore's `apple` and `gpu` settings (`pyproject.toml`)
+- Current AbstractCore baseline: base `abstractruntime` requires AbstractCore 2.33.0 or newer with its light install (remote providers, tools and the MCP worker, media routing, the `decision` output modality); `abstractruntime[apple]` and `abstractruntime[gpu]` require AbstractCore's `apple` and `gpu` settings (`pyproject.toml`)
 - Install profiles: base `abstractruntime` is remote-light with MCP, generated media routing, and Runtime-owned permissive PDF read/write nodes; `abstractruntime[apple]` and `abstractruntime[gpu]` add local inferencer stacks.
 - Config boundary: Runtime does not read `ABSTRACTGATEWAY_*`; Gateway-owned config should be translated into explicit Runtime inputs or Runtime-owned env names. Hosts can attach `resolve_provider_endpoint_profile(provider_id)` to the LLM client; Runtime treats it as an optional resolver hook, injects transient provider/base URL/key params for the call, and redacts secret-like params from persisted observability.
 

{abstractruntime-0.16.0 → abstractruntime-0.17.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "AbstractRuntime"
-version = "0.16.0"
+version = "0.17.0"
 description = "AbstractRuntime: a durable graph runner designed to pair with AbstractCore."
 readme = "README.md"
 requires-python = ">=3.10"
@@ -69,7 +69,8 @@ dependencies = [
   # AbstractVoice/Vision/Music/3D plugin floors.
   # 2.32.0: the `decision` output modality (`output.decision` route, `/v1/decisions`,
   # `MultimodalGenerateResponse.decision`).
-  "abstractcore>=2.32.0",
+  # 2.33.0: the release wave's Core (typed `ModelNotDownloadedError`, AbstractVoice 0.20.0 floor).
+  "abstractcore>=2.33.0",
   "openai<3.0.0,>=1.109.1",
   "httpx<1.0.0,>=0.28.1",
   "anyio<5.0.0,>=4.12.1",
@@ -86,12 +87,12 @@ dependencies = [
 
 [project.optional-dependencies]
 apple = [
-  "abstractcore[apple]>=2.32.0",
+  "abstractcore[apple]>=2.33.0",
   "setuptools<82.0.0,>=80.10.2",
 ]
 
 gpu = [
-  "abstractcore[gpu]>=2.32.0",
+  "abstractcore[gpu]>=2.33.0",
   # vLLM 0.9-0.19 caps setuptools below 80; 77.0.3 is still a modern floor.
   "setuptools<82.0.0,>=77.0.3",
 ]