## Problem
Provider tool-call helpers always used the globally injected direct
`Tools.execute` function. When a model received tools from
`session.tools()`, calling `handleToolCalls` or `handle_tool_calls`
therefore discarded the Tool Router session context and caused session
meta-tools such as `COMPOSIO_SEARCH_TOOLS` to fail.
Calling `session.execute()` manually preserved the session, but bypassed
provider behavior such as Anthropic input normalization and schema-alias
restoration.
## Root fix
- Add an explicit execution target to the non-agentic provider helpers:
- TypeScript: `handleToolCalls(session, response)` and
`executeToolCall(session, call)`
- Python: `handle_tool_calls(response=response, session=session)` and
`execute_tool_call(tool_call=call, session=session)`
- Route normalized provider arguments through the supplied Tool Router
session.
- Map session responses back to each helper's existing result shape.
- Keep provider-specific normalization before execution, including
Anthropic schema-alias restoration.
- Reject direct-only options and modifiers when the selected target is a
session, including plain JavaScript calls that bypass the TypeScript
overloads.
- Update OpenAI and Anthropic examples to use the session-aware helpers.
- Harden the docs policy test so setup and execution split across fences
in one sample are still detected.
## Docs review follow-ups
- Reword the concepts-page prohibition so it forbids user-ID-bound
helper calls, not the helpers themselves, matching the provider pages in
this PR.
- Add minimum-version callouts to the OpenAI and Anthropic provider
pages (Python `composio` newer than 0.19.0; TypeScript `@composio/core`
≥ 0.17.0 with `@composio/openai` ≥ 0.12.0 / `@composio/anthropic` ≥
0.11.0), pointing older versions at `session.execute()`.
- Bump `docs/package.json` to `@composio/core` `^0.15.0` and
`@composio/openai` `^0.11.0` (the published majors at the time of the
bump; `@composio/core` 0.16.0 and `composio` 0.19.0 have since released
from `next` without this PR, so its changeset will publish core 0.17.0
and the next Python minor) and annotate each `@errors: 2345` Twoslash
marker with a TODO naming the minor version that retires it; since this
changeset releases minors, all three pins need a manual range bump to
retire the markers. This version of twoslash only throws on *unlisted*
errors, so a stale marker cannot break the build — it would only mask
future TS2345s, which the TODOs now track.
- Update `SESSION_GUARDRAILS` (the block appended to `.md` responses for
agents): add a session-execution bullet (scoped to the OpenAI and
Anthropic helpers, with `session.execute()` for every other provider)
and qualify the direct-execution list with "with a user ID". The
session-execution static test now scans the guardrail blocks like the
execute-version test already did.
- Tighten the docs detector: the Python branch is bounded to the helper
call's argument list (tolerating one level of nested calls) instead of
running past the closing paren, and the TypeScript branch catches whole
user-ID identifiers (`userId`, `user_id`, `uid`) without flagging
session variables like `userSession` — each edge has a regression test.
- Note on the Google provider page that its `executeToolCall` is not
session-aware yet.
## Compatibility and release
Existing user-ID calls remain unchanged and continue to use direct tool
execution. The new session call forms are additive.
The changeset applies minor releases to `@composio/core`,
`@composio/openai`, and `@composio/anthropic` — the new session
overloads are a type-level break for provider subclasses, so patch was
too small. The configured fixed group also includes `@composio/slim`.
The docs site intentionally checks examples against currently published
SDK declarations. The three new TypeScript calls therefore carry exact
Twoslash `TS2345` release-skew annotations; remove them (per the inline
TODOs) once `docs/package.json` picks up `@composio/core` ≥ 0.17.0,
`@composio/openai` ≥ 0.12.0, and `@composio/anthropic` ≥ 0.11.0.
## Verification
- `@composio/core`: 1,061 tests passed; typecheck passed
- `@composio/openai`: 34 tests passed; typecheck passed
- `@composio/anthropic`: 53 tests passed; typecheck passed
- Python provider and aliasing suites: 40 passed, 4 skipped
- Focused Python mypy and Ruff checks passed
- Docs static suite: 208 tests passed (including the new guardrail-scan
and detector cases)
- Docs production build passed with the bumped `@composio/core` 0.15.0 /
`@composio/openai` 0.11.0, including Twoslash, TypeScript, and all
generated pages
- Docs lint passed; lint reports only existing warnings
- Changeset status reports the expected minor packages
---------
Co-authored-by: Soumya Medapati <soumyamedapati@soumyas-air.local.meter>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: jkomyno <alberto@composio.dev>
`json_schema_to_model` and `json_schema_to_pydantic_type` now share one
compiled object policy, so provider-facing Pydantic models stop silently
discarding valid free-form arguments and stop ignoring `patternProperties` and
explicit `additionalProperties` controls. Accepted dynamic keys are written back
into `__pydantic_extra__`, which is what `LangchainProvider.wrap_tool` re-reads
with `getattr`.
Acceptance and rejection are asserted from the shared cross-SDK corpus, so the
Python entry points cannot diverge from the TypeScript converters.
## Summary
Agentic providers honor `schema_config={"skip_defaults": True}` for
`args_schema` (defaults stripped, fields become required) but were
**not** applying the same `skip_default` to the function
`__signature__`, so the wrapped tool's signature and `args_schema`
disagreed about which defaulted params were optional. The sibling
`llamaindex` provider already passed it to both.
**Supersedes #3727.** This is the polished version, now that the ag2
migration (#3729) landed and `composio-autogen` imports again — so the
autogen case uses a real import instead of the earlier source-load stub,
and the parity tests run in CI.
## Changes
- **langchain / langgraph** (original commit by @anxkhn): pass
`skip_default=self.skip_default` to
`get_signature_format_from_schema_params`.
- **autogen**: same one-line alignment. autogen builds **only**
`__signature__` (no `args_schema`), so it never had the *mismatch* — but
its signature still ignored `skip_defaults`. Aligned for consistency;
verified on the signature alone.
- **tests**: `TestAgenticSkipDefaultsParity` guards each provider with
`pytest.importorskip`; adds real-import autogen regression tests (no
stub).
- **CI + nox**: install `providers/autogen` in the unit-test job
(`py.test.yml`) and the nox `tst` session so the autogen parity tests
actually run, plus a **hard `import composio_autogen` smoke check** in
CI.
## Why the CI import guard (not just the tests)
The `importorskip`-guarded unit tests **skip** when a provider can't be
imported — which is the correct behavior for optional providers, but
means they'd *silently skip* if autogen broke on a fresh install
(exactly the failure that #3729 just fixed). The extra `python -c
"import composio_autogen"` in the import step is a **hard** check that
fails loudly, so a future dependency breakage can't slip through as a
green build.
`providers/langchain` + `providers/autogen` co-install cleanly
(verified: 59 packages, ag2 0.14.0, no conflict). langgraph continues to
skip in CI — its code path is identical to langchain's, which runs.
## Type of change
- [x] Bug fix
## How Has This Been Tested?
- Base venv (langchain only): langchain parity **passes**, langgraph +
autogen **skip** (no error).
- CI-mirror (langchain + autogen, editable): **4 passed, 2 skipped
(langgraph)** — autogen tests actually run and pass.
- Regression proof, each provider: reverting only the provider line
fails the skip-defaults case with `assert 5 is Parameter.empty`;
restoring passes. Verified for langchain **and** autogen.
- `ruff check`, `ruff format --check`, `mypy`, and full
`tests/test_provider.py` (29 passed, 4 skipped) → clean. `py.test.yml`
parses.
## Checklist
- [x] I ran linters/tests locally and they passed
- [x] I added tests: langchain/langgraph parity + regression, and
autogen parity that runs in CI
- [ ] Changeset: not applicable — Python-only change.
## Credits
Original fix by @anxkhn (#3713 → #3727), preserved as the first commit
and co-authored on the follow-up.
---------
Co-authored-by: Anas Khan <83116240+anxkhn@users.noreply.github.com>
This PR:
- supersedes https://github.com/ComposioHQ/composio/pull/3712,
preserving the original commits by @serhiizghama (kept as
author/co-author)
- fixes `OpenAIProvider.handleToolCalls` to execute **every** tool call
in an assistant message — it previously only read `tool_calls[0]`, so
parallel tool calls (on by default) were dropped and their
`tool_call_id`s went unanswered, failing the next request with the "must
be followed by tool messages responding to each `tool_call_id`" error
- limits execution to the **first choice** — with `n > 1` the previous
loop ran each tool call once per choice and orphaned the `tool_call_id`s
belonging to the alternative completions
- aligns the Python SDK (`OpenAIProvider.handle_tool_calls`) with the
same first-choice behavior so both SDKs stay in parity
- fixes the stale `@composio/openai` test that modeled "multiple tool
calls" as separate `n` choices rather than one message's `tool_calls`
array
- clarifies in the changeset that the calls run **sequentially**, in the
order the model returned them (here "parallel" means the model issued
several calls in one turn, not concurrent execution), so each
`tool_call_id` is answered exactly once and deterministically
- adds regression tests (TS + Python): both calls run for parallel tool
calls in one message, and only the first choice runs when `n > 1`
Full `@composio/core` (996) and `@composio/openai` suites green; Python
`handle_tool_calls` tests, Ruff, and mypy clean; `tsc` clean.
---------
Co-authored-by: serhiizghama <zmrser@gmail.com>
This PR:
- stops the Python SDK from silently retrying `tools.execute` and
`tools.proxy`, which are non-idempotent POST writes. The
Stainless-generated client retries POSTs by default (`max_retries=2`) on
read timeouts, 429, 5xx, and connection errors. A read timeout is unsafe
to retry — the request may already be in flight on the backend — so a
retry can duplicate the side effect, e.g. send an email twice. Reported
in https://github.com/ComposioHQ/composio/issues/3586
- routes both write paths through a new `client.without_retries` — a
cached, retry-disabled (`max_retries=0`) clone of the `HttpClient`.
Reads and lists keep the default retries, so resilience is unchanged for
idempotent calls
- overrides `HttpClient.copy()` to re-inject the required `provider`
keyword and re-aliases `with_options`, since the generated `copy()`
rebuilds via `self.__class__(...)` without `provider` — previously
`with_options(max_retries=0)` raised `TypeError` on the subclass
- caches the no-retry sibling instead of cloning per call: each clone
creates a new `ContextVar`, and dynamically-created `ContextVar`s are
never garbage-collected
- adds `tests/test_no_retry_writes.py`: writes hit the transport exactly
once on a retryable 5xx, reads still retry-then-succeed (proving the
scoping), and a regression guard for the `copy()` override
- interim fix only — the durable solution is idempotency keys, tracked
in https://github.com/ComposioHQ/composio/issues/3654 (gated on backend
support), which supersedes this once available
No changeset (Python-only change; changesets are TypeScript-only).
This PR is **part 1 of 3** splitting
https://github.com/ComposioHQ/composio/pull/3505 to make the removal of
the old 2025 custom tools easier to review. It carries the **Python**
slice.
- removes the legacy `composio.tools.custom_tool` registry path:
`core.models.custom_tools`, `ExecuteRequestFn`, the old fallback
execution wiring, the security tests, and the example
- preserves the 2026 tool-router APIs: `composio.experimental.tool()`,
`composio.experimental.Toolkit`, inline custom-tool execution,
preload/attach/use flows, and `session.custom_tools()`
- bumps the Python workspace and all provider packages `0.13.1` →
`0.14.0`
## Notes
- This slice is a byte-identical subset of #3505 — the three split
branches recombine to that PR's exact tree. See #3505 for the original
local verification logs (`uv run --frozen nox -s chk`, targeted pytest);
CI re-runs per PR.
The `_python_reserved` set was hardcoded to only `{"for", "async"}`,
missing `"from"` and all other Python keywords. Tools from toolkits like
Intercom that have a `from` parameter caused `ValueError: 'from' is not
a valid parameter name` when building function signatures.
Extracted keyword substitution logic into shared utilities
(`substitute_reserved_python_keywords` / `reinstate_reserved_python_keywords`)
using `keyword.iskeyword()` from stdlib for complete coverage, and applied
it to all affected providers: langchain, gemini, langgraph, autogen, llamaindex.
Closes PLEN-1671
jkomyno