0.5.4

What's Changed

• fix: treat max_retries as retry count by @Andy963 in #26
• feat: add query filters for text matching and date range in TapeQuery by @frostming in #24
• OpenAI codex oauth integration by @niyue in #23
• fix(responses): restore parity with chat completions and refactor parsing by @PsiACE in #21

New Contributors

• @Andy963 made their first contribution in #26
• @niyue made their first contribution in #23

Full Changelog: 0.5.3...0.5.4

0.5.3

What's Changed

• chore(meta): migrate links to bubbuild/republic and align attribution by @PsiACE in #20

Full Changelog: 0.5.2...0.5.3

0.5.2

Full Changelog: 0.5.1...0.5.2

0.5.1

Full Changelog: 0.5.0...0.5.1

Release v0.5.0: Async tape storage

Key Changes

1. Added opt-in Responses API mode

LLM now supports use_responses: bool = False for explicit Responses mode.

• When enabled (and supported by the provider), non-stream requests use the Responses API.
• Added internal conversion from chat/tool history to Responses input items.
• Tool-call extraction now supports both completion-style and responses-style payloads.

Files involved:

• src/republic/llm.py
• src/republic/core/execution.py
• src/republic/clients/chat.py
• tests/test_responses_handling.py

2. Tape subsystem refactored for async support

Tape management now supports both sync and async paths, including async storage and context lifecycle.

• Added AsyncTapeStore protocol.
• Added AsyncTapeStoreAdapter to wrap sync stores for async usage.
• Added AsyncTapeManager for async read/write/handoff/record flows.
• TapeQuery is now generic and supports await query.all() for async stores.
• Tape session now provides async tape methods: read_messages_async, append_async, reset_async, handoff_async.

Files involved:

• src/republic/tape/store.py
• src/republic/tape/manager.py
• src/republic/tape/query.py
• src/republic/tape/session.py
• src/republic/tape/__init__.py

3. Expanded top-level exports for async tape types

republic top-level exports now include:

• AsyncTapeManager
• AsyncTapeStore

Files involved:

• src/republic/__init__.py

4. Reliability and release pipeline fixes

• Fixed TapeEntry.copy() to preserve timestamp during copies.
• Fixed docs deployment workflow permissions by adding contents: write to deploy-docs and updating actions/checkout version.

Files involved:

• src/republic/tape/entries.py
• .github/workflows/on-release-main.yml

Behavior Changes and Upgrade Notes

AsyncTapeStore with sync APIs

When tape_store is an AsyncTapeStore, sync APIs with tape=... raise ErrorPayload.

Use async APIs instead:

• chat_async
• tool_calls_async
• run_tools_async
• stream_async
• stream_events_async

LLM direct tape helpers moved behind session objects

LLM no longer directly exposes:

• tapes()
• handoff(...)
• read_messages(...)
• query(...)
• reset_tape(...)

Migration: use operations on the session returned by llm.tape("<name>").

Documentation Updates

• docs/guides/tape.md now includes Async Tape Store usage and sync/async behavior rules.

Commit List

• 2d330aa fix: add permissions for contents write in deploy-docs job
• edd9400 chore: bump version to 0.5.0 in pyproject.toml, about.py, and uv.lock
• 4ffea2d feat: Refactor tape management to support asynchronous operations (#19)
• 04525fb fix: import time and include timestamp in TapeEntry copy method
• b9069a5 feat: add opt-in Responses API mode (#18)

Upgrade Checklist

• If you use an async tape backend, migrate tape-related calls to async APIs.
• If you relied on removed LLM tape helpers, migrate to llm.tape(name).
• Enable Responses behavior explicitly with LLM(..., use_responses=True) when needed.

0.4.1

What's Changed

• refactor tape query API and document read_entries deprecation by @frostming in #17

Full Changelog: 0.4.0...0.4.1

Release 0.4.0: Async tools

Highlights

• Introduced async tool execution support.
• Simplified public error handling by removing StructuredOutput.
• Added provider-specific max token handling in LLMCore.
• Improved tool error responses by preserving tool arguments on failure.

Breaking Changes

StructuredOutput removed from public API

StructuredOutput(value, error) is no longer used for non-streaming APIs.
Success now returns plain values, and failures raise ErrorPayload.

Affected API families include:

• LLM.chat*, LLM.tool_calls*, LLM.if_*, LLM.classify*, LLM.embed*
• Tape.chat*, Tape.tool_calls*
• ToolExecutor.execute(...) now raises ErrorPayload for validation/input/tool lookup failures

Also removed/simplified:

• ContextSelection removed (read_messages(...) returns list[dict[str, Any]])
• QueryResult removed (TapeQuery.all() returns list[TapeEntry])

Migration pattern:

• Replace .value / .error checks with try/except ErrorPayload.
• Search old usage quickly with:

rg -n "StructuredOutput|\\.value\\b|\\.error\\b" src tests

For detailed migration examples, see docs/guides/breaking-changes.md.

Features

• feat: implement async tool execution (86a5f43)

  • Added async execution path for tool calls.
  • Updated chat client and tool executor integration.
  • Expanded contract and UX test coverage around async behavior.

• feat: add provider-specific max tokens handling in LLMCore (a79be07)

  • LLMCore now applies max token behavior with provider-specific handling.

Fixes

• fix: return tool arguments on error in tool response handling (279a36e)
  • Tool execution error responses now preserve original tool arguments for better debugging and recovery.

Refactors and Tooling

• refactor: Refactor error handling and remove StructuredOutput (025b045)
  • Aligned client/core error semantics to Python-style exceptions.
  • Updated tape query/context/session return types.
  • Refreshed tests for new behavior.
• CI/dev tooling updates:
  • Switched workflow hooks from pre-commit to prek.
  • Updated related config/docs and lockfile.

Diff Summary

From 0.3.2 to 0.4.0:

• 5 commits
• 25 files changed
• 998 insertions, 696 deletions

Recommended Upgrade Checklist

• Update call sites that rely on .value / .error.
• Add/adjust exception handling for ErrorPayload.
• Re-run test suites covering chat/tool-call/tape flows.
• Validate provider-specific token limits in your production configs.

0.3.2

What's Changed

• refactor: decouple tape from chat client by @frostming in #10

Full Changelog: 0.3.1...0.3.2

0.3.1

What's Changed

• feat: add model parameter to tool decorator for enhanced functionality by @frostming in #8
• fix: include tool error as model context by @frostming in #9

New Contributors

• @frostming made their first contribution in #8

Full Changelog: 0.3.0...0.3.1

Release 0.3.0: Tape-First UX Refresh

This release refines Republic around a tape-first, practical developer workflow. The API surface is smaller and more intentional, docs/examples were fully rewritten, and reliability improved in streaming and fallback behavior.

Highlights

• Tape-first flow with explicit anchors (handoff) and queryable execution history.
• Unified LLM facade for chat, tools, streaming, text decisions, and embeddings.
• Structured return pattern (value + error) across core interfaces.
• Improved event streaming, including more robust tool-call delta merging.
• Better error classification and fallback handling across provider exception shapes.
• New docs and runnable examples with a simpler, usage-first style.
• Contract-style test suite focused on user-facing behavior.

Example

import os

from republic import LLM

llm = LLM(model="openrouter:openrouter/free", api_key=os.environ["LLM_API_KEY"])

out = llm.chat("Describe Republic in one sentence.", max_tokens=48)
print(out.value if out.error is None else out.error.message)

tape = llm.tape("release-notes")
tape.handoff("draft_v1", state={"owner": "assistant"})

reply = tape.chat("Summarize the latest changes in three bullets.", max_tokens=96)
print(reply.value if reply.error is None else reply.error.message)

Upgrade Notes

• Main usage is now centered on LLM + structured outputs.
• Docs were reorganized to a Quickstart + focused Guides layout.
• Streaming/tool behavior is aligned around explicit inspectable results.
• Error/retry/fallback logic was updated for mixed provider environments.
• Core dependency surface now includes pydantic.

Thanks

Derived from lightning-ai/litai, inspired by pydantic/pydantic-ai, and adapted into Republic’s tape-first execution style.