0.3.22

[0.3.22] - 2026-05-24

Fixed

• AttributeError: 'LocalBackend' object has no attribute '_read_bytes' at toolset get_instructions() time (#118, independently authored by @mcauthorn in #119). pydantic-ai-backend 0.2.8 promoted the bytes-read entry point on BackendProtocol from private _read_bytes to public read_bytes and (deliberately) kept no transitional alias. With pydantic-ai-backend>=0.2.7 unbounded, fresh resolutions pulled 0.2.8 transitively, so every toolset that reaches for bytes (context, memory, liteparse, skills/backend) blew up at instructions-load time. Word-boundary rename across all call sites and test mocks.

Changed

• Bumped pydantic-ai-backend>=0.2.7 → >=0.2.8 (#118, Renovate) — pulls in the exists() predicate, the read_bytes rename (see Fixed above), hashline_edit per-(backend, path) serialization, and the async_execute wire-up in the console toolset's execute tool.
• Bumped summarization-pydantic-ai>=0.1.4 → >=0.1.5 (#118, Renovate) — batched with the backend bump; pure CI-housekeeping release on the summarization side, no behaviour change.

0.3.21

[0.3.21] - 2026-05-24

Changed

• Bumped pydantic-ai-todo floor from >=0.2.1 to >=0.2.2 (#114, auto-opened by the new Renovate config). Brings in pydantic-ai-todo 0.2.2: the new AsyncRedisStorage backend (Redis Hash + companion List, session-scoped, multi-tenant, event-emitter integration) plus a follow-up batch of correctness fixes (remove_todo atomicity via single pipeline, Redis Cluster–safe hash-tagged keys, idempotent initialize()).

0.3.20

[0.3.20] - 2026-05-18

Fixed

• [WinError 2] crash on Windows when calling the execute tool (#108) — bumped the pydantic-ai-backend floor from >=0.2.4 to >=0.2.7. Releases before 0.2.7 hardcoded ["sh", "-c", command] in LocalBackend.execute(), so every shell invocation on Windows failed with FileNotFoundError: [WinError 2] The system cannot find the file specified, regardless of whether the target executable (e.g. powershell, pwsh.exe) was on PATH. pydantic-ai-backend 0.2.7 routes through _shell_cmd() (cmd /c on win32, sh -c elsewhere) and adds async_execute() with cancellation support.
• wait_tasks cancellation cascade in subagent orchestration — bumped subagents-pydantic-ai floor to >=0.2.4, which routes both mode="all" and mode="any" through asyncio.wait instead of asyncio.wait_for(asyncio.gather(...)). Previously, when pydantic-ai's _call_tools sibling-cancelled the wait_tasks tool call (or any outer cancel reached the orchestrator), the cascade silently killed every in-flight subagent — surfacing as TaskStatus.CANCELLED with an empty error even though the parent never requested it.

Changed

• Bumped minimum versions of all pydantic-ai-* sister packages so a fresh install pulls the latest releases by default: pydantic-ai-slim>=1.97.0 (was >=1.77.0), pydantic-ai-backend>=0.2.7 (was >=0.2.4, for both [console] and [docker] extras), summarization-pydantic-ai>=0.1.4 (was >=0.1.3), subagents-pydantic-ai>=0.2.4 (was >=0.2.1), pydantic-ai-shields>=0.3.2 (was >=0.3.1).

Internal

• # type: ignore[attr-defined] on five BinaryContent attribute accesses in pydantic_deep/processors/eviction.py. pydantic-ai-slim>=1.97 exposes BinaryContent as a PydanticDataclass whose fields and identifier property are invisible to pyright (the attributes exist at runtime — this is upstream type-info incompleteness). Keeps make typecheck green until upstream stubs catch up.

Infrastructure

• Renovate config (renovate.json) — opt-in dependency bot scoped to the five vstorm-co sibling packages (pydantic-ai-todo, pydantic-ai-backend, summarization-pydantic-ai, subagents-pydantic-ai, pydantic-ai-shields). All other dependencies are explicitly disabled. New releases of these packages will get grouped PRs that bump both pyproject.toml floors and uv.lock, with auto-merge on green CI. Activates once the Renovate GitHub App is installed on the vstorm-co organization.

0.3.19

[0.3.19] - 2026-05-14

Added

• PeriodicReminderCapability — periodic task reminders for long agent runs (#94) — injects a "what are you supposed to be doing" reminder into the message history every N model-request turns to prevent agent drift on long, tool-heavy runs. Uses before_model_request and per-run state isolation via for_run().

  • Four CLI modes via a new /remind command: off, first (zero-cost — re-states the first user message), context (zero-cost — compact transcript), llm (uses Claude Haiku / GPT mini / Gemini Flash to summarize progress).
  • LLMReminderGenerator with exception fallback to the zero-cost default.
  • Three render styles: system_reminder_tag (default), developer_note, user_prompt.
  • create_deep_agent() gains a periodic_reminder: bool | PeriodicReminderConfig | None = None parameter.
  • CLI: enabled by default in llm mode; configurable via periodic_reminder and reminder_mode keys in config.toml.
  • New top-level exports: PeriodicReminderCapability, PeriodicReminderConfig, ReminderGenerator, LLMReminderGenerator, make_config_for_mode.

• MessageQueue — mid-run message delivery (steering & follow-up) (#100) — lets external code (CLI keystrokes, webhooks, subagents) push messages into a running agent loop without cancelling and restarting it, preserving in-flight tool results and the prompt cache.

  • Steering messages are injected before the next LLM call via MessageQueueCapability.before_model_request. To avoid issues with downstream capabilities that strip lone trailing ModelRequest nodes, the steering UserPromptPart is merged into the last existing ModelRequest.
  • Follow-up messages are queued for delivery when the agent would otherwise stop, triggering a re-entry via the new run_with_queue() helper.
  • Two delivery modes: one_at_a_time (default) and all (drain entire queue based on the head message's mode).
  • DeepAgentDeps.message_queue field, propagated by reference through clone_for_subagent() so subagents can steer the parent.
  • create_deep_agent() gains a message_queue: MessageQueue | None = None parameter.
  • CLI: >>text mid-run = steering, plain text mid-run = follow-up, !cmd stays as shell command in all states. Side-panel QueuedWidget shows pending counts. Stale steering messages are surfaced as a warning when the run ends before reaching another LLM call, and follow-ups left over from a cancelled run are discarded with a count-only notification.
  • New top-level exports: MessageQueue, MessageQueueCapability, QueuedMessage, run_with_queue, format_steering, format_follow_up.

• Programmatic skills parameter on create_deep_agent (#97) — accepts list[Skill] instances directly, complementing the existing skill_directories= discovery path. Emits a UserWarning when skills= or skill_directories= are provided alongside include_skills=False.

• Docker sandbox environment variable support (#99 — fixes #98) — wires up the RuntimeConfig.env_vars plumbing that the programmatic DockerSandbox API already supported but the CLI and full_app example never exposed.

  • sandbox_env_vars: dict[str, str] and sandbox_env_file: str | None fields on CliConfig; matching parameters on create_cli_agent().
  • Three-level priority merge: config.sandbox_env_vars (lowest) → .env file → explicit sandbox_env_vars (highest).
  • examples/full_app/app.py auto-loads examples/full_app/.env into the SessionManager's default RuntimeConfig.
  • Uses RuntimeConfig(cache_image=False) to prevent secrets from being baked into cached Docker image layers.
  • _write_toml() extended to emit [table] sections for dict values, enabling round-trip persistence via set_config_value().

Fixed

• Esc to interrupt, tool spinner lifecycle, and empty message cleanup (#96 — closes #93):

  • Esc now interrupts a running agent (previously Ctrl+C only); focuses input when idle. Centralized on DeepApp.action_escape_key so any screen-level handler uses the same cancellation path.
  • Fixed a ToolCallWidget.on_mount race where complete() arriving before mount left the spinner timer uninitialized — the widget now renders its final state immediately when not in the pending state.
  • AssistantMessage.complete_tool_call is now idempotent (only acts on widgets still in the pending state), preventing the cancellation drain from overwriting correctly completed results.
  • /load session replay marks orphaned tool calls (from previously interrupted sessions) as "Interrupted" instead of leaving spinners hung; switched to isinstance checks against ToolCallPart / ToolReturnPart / UserPromptPart / TextPart and uses part.args_as_dict() for correct label display.
  • Empty assistant message bubbles are removed when a run is cancelled before producing any output, thinking, or tool calls.
  • The hints bar shows context-aware shortcuts (Esc interrupt) while the agent is running.

• create_deep_agent() rejected skills= kwarg (#97 — fixes #95) — callers using skills=[Skill(...)] previously got UserError: Unknown keyword arguments: 'skills' because the kwarg fell through to pydantic-ai's Agent() constructor.

0.3.18

[0.3.18] - 2026-05-05

Fixed

• EvictionCapability dropped BinaryContent (e.g. screenshots) from ToolReturn results — previously, any ToolReturn(return_value=..., content=[..., BinaryContent(...)]) was collapsed into a plain string before the size check, so the multimodal content (images, audio, PDFs) was silently discarded along with a text eviction message. The capability now only measures and evicts return_value; the content list and metadata are always preserved by re-wrapping the result. (#90)

Fixed

• DeepAgentDeps was missing the checkpoint_store field — the checkpointing middleware already resolved the store via getattr(deps, "checkpoint_store", None) and the docs instructed users to pass it at construction time, but the field was never declared on the dataclass so assignments were silently ignored and type checkers raised attr-defined errors. The field is now properly declared as checkpoint_store: Any = None and propagated as a shared reference through clone_for_subagent(). (#87)

Added

• Binary content pruning via max_binary_content — a new before_model_request hook in EvictionCapability bounds the number of multimodal binary parts kept in model-visible history. Older BinaryContent values (from both UserPromptPart.content and ToolReturnPart.content) are written to the backend at deterministic paths (/large_tool_results/binary_{id}.{ext}) and replaced with a compact read_file-able text reference so the agent can still retrieve them on demand. If a backend write fails, the binary is left in place to avoid data loss.
  • EvictionCapability gains a max_binary_content: int | None field (default 3).
  • create_deep_agent() gains a matching max_binary_content: int | None = 3 parameter across all overloads; pass None to keep every binary in history.
  • New constants exported from the top-level package: DEFAULT_MAX_BINARY_CONTENT, BINARY_PRUNED_TEMPLATE.
  • Storage paths use BinaryContent.identifier (a stable SHA1 of the bytes), making re-pruning the same binary idempotent.

0.3.17

[0.3.17] - 2026-04-22

Added

• LiteparseToolset — document parsing via LiteParse

  • New toolset at pydantic_deep.toolsets.liteparse
  • Tools: parse_document (text extraction) and screenshot_document (per-page images)
  • Reads files from any backend as bytes — works with StateBackend, LocalBackend, DockerSandbox
  • Optional OCR via built-in Tesseract or pluggable HTTP server (PaddleOCR, EasyOCR)
  • Lazy parser initialization — the Node.js CLI is found/installed on first tool call
  • Configurable: ocr_enabled, ocr_language, ocr_server_url, dpi, max_pages
  • Graceful error messages when the liteparse package or Node.js CLI is not installed
  • Enabled via include_liteparse=True in create_deep_agent()

• liteparse optional extra in pyproject.toml

  • pip install pydantic-deep[liteparse] installs the Python wrapper
  • Node.js >= 18 and npm install -g @llamaindex/liteparse are required separately

0.3.16

[0.3.16] - 2026-04-22

Changed

• instructions now replaces BASE_PROMPT instead of appending to it — previously, passing instructions="..." to create_deep_agent() produced a system prompt of BASE_PROMPT + "\n\n" + instructions. Now instructions is used verbatim as the full system prompt. instructions=None (the default) keeps the existing behaviour — BASE_PROMPT is used automatically. Users who want to extend the default rather than replace it can do so with an f-string: instructions=f"{BASE_PROMPT}\n\nYour extra text". BASE_PROMPT is exported from the top-level package. (#84, reported by @rremilian)
• Subagent and team-member factories always prepend BASE_PROMPT — agents spawned automatically by the task() tool or spawn_team() continue to receive BASE_PROMPT followed by their task-specific instructions, so subagent behaviour is unchanged.
• apps/deepresearch double-prompt bug fixed — MAIN_INSTRUCTIONS already contained BASE_PROMPT; it was previously duplicated in the final system prompt because the old append logic prepended it again. The new semantics resolve this without any change to the deepresearch app itself.

0.3.15

[0.3.15] - 2026-04-17

Fixed

• PatchToolCallsCapability caused ValidationException: duplicate Ids on Bedrock when tools raised ModelRetry — when a tool raised ModelRetry, pydantic-ai records the retry as a RetryPromptPart (carrying the original tool_call_id) on the following ModelRequest, not as a ToolReturnPart. The patch processor only scanned for ToolReturnPart when deciding whether a ToolCallPart was orphaned, so it injected a synthetic ToolReturnPart with the same id — leaving the request with two parts sharing one tool_call_id. Strict providers (Bedrock minimax.minimax-m2.5 and others) rejected the request with The toolResult blocks at messages.N.content contain duplicate Ids. The processor now treats RetryPromptPart as a valid answer to a ToolCallPart, so no synthetic return is injected and the history remains valid. (#79, reported by @thatGreekGuy96)

0.3.14

[0.3.14] - 2026-04-16

Fixed

• Subagents ignored parent web_search/web_fetch settings — the default subagent factory in create_deep_agent hardcoded web_search=True and web_fetch=True, overriding the parent agent's configuration. On Bedrock and Vertex Anthropic models this produced a 400 error (web_fetch_20250910 not accepted), because the beta web tools are not supported there. The factory now propagates the parent agent's web_search and web_fetch flags to spawned subagents. (#77, reported by @SvdR82)

0.3.13

[0.3.13] - 2026-04-13

Fixed

• User-provided tools lost metadata when passed via tools= parameter — tools registered through create_deep_agent(tools=[...]) were previously added via agent.tool(tool.function) after construction, which hardcoded takes_ctx=True and discarded all Tool-level metadata (name, description, prepare, max_retries, requires_approval, timeout). Tools are now passed directly to the Agent constructor, preserving all metadata and correctly honouring the original takes_ctx value. (PR #75 by @ilayu-blip)