tavus-cli 0.10.2__tar.gz → 0.10.3__tar.gz

tavus_cli-0.10.3/.release-please-manifest.json

@@ -0,0 +1,3 @@
+{
+  ".": "0.10.3"
+}

{tavus_cli-0.10.2 → tavus_cli-0.10.3}/CHANGELOG.md

@@ -1,5 +1,12 @@
 # Changelog
 
+## [0.10.3](https://github.com/Tavus-Engineering/tavus-mcp/compare/v0.10.2...v0.10.3) (2026-10-16)
+
+
+### Bug Fixes
+
+* preserve every received Builder reply in MCP history ([#78](https://github.com/Tavus-Engineering/tavus-mcp/issues/78)) ([7dab660](https://github.com/Tavus-Engineering/tavus-mcp/commit/7dab660ddaa3110f208f343e90981217c10cdb20))
+
 ## [0.10.2](https://github.com/Tavus-Engineering/tavus-mcp/compare/v0.10.1...v0.10.2) (2026-10-16)
 
 

{tavus_cli-0.10.2 → tavus_cli-0.10.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: tavus-cli
-Version: 0.10.2
+Version: 0.10.3
 Summary: Tavus CLI and MCP server for agentic PAL creation and editing.
 Requires-Python: >=3.11
 Requires-Dist: fastmcp>=2.0.0

{tavus_cli-0.10.2 → tavus_cli-0.10.3}/packages/tavus/pyproject.toml

@@ -1,11 +1,11 @@
 [project]
 name = "tavus"
-version = "0.10.2" # x-release-please-version
+version = "0.10.3" # x-release-please-version
 description = "Official alias for tavus-cli, the Tavus CLI and MCP server."
 readme = "README.md"
 requires-python = ">=3.11"
 dependencies = [
-  "tavus-cli==0.10.2", # x-release-please-version
+  "tavus-cli==0.10.3", # x-release-please-version
 ]
 
 [project.urls]

{tavus_cli-0.10.2 → tavus_cli-0.10.3}/packages/tavus-mcp/pyproject.toml

@@ -1,11 +1,11 @@
 [project]
 name = "tavus-mcp"
-version = "0.10.2" # x-release-please-version
+version = "0.10.3" # x-release-please-version
 description = "Official alias for tavus-cli, which provides the tavus-mcp MCP server."
 readme = "README.md"
 requires-python = ">=3.11"
 dependencies = [
-  "tavus-cli==0.10.2", # x-release-please-version
+  "tavus-cli==0.10.3", # x-release-please-version
 ]
 
 [project.urls]

{tavus_cli-0.10.2 → tavus_cli-0.10.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "tavus-cli"
-version = "0.10.2"
+version = "0.10.3"
 description = "Tavus CLI and MCP server for agentic PAL creation and editing."
 readme = "README.md"
 requires-python = ">=3.11"

{tavus_cli-0.10.2 → tavus_cli-0.10.3}/skills/tavus-agentic-pals/references/chat-mode.md

@@ -88,6 +88,24 @@ timeout covers the join, but budget the first `collect_turn` generously.
 Always `close()` the session and end the conversation in `finally:` —
 orphaned chat sessions still consume quota.
 
+## Conversation history
+
+Reuse the returned `conversation_id` for each turn of the same test. CVI owns
+that conversation's user and assistant history; starting another conversation
+does not restore the previous test. In chat mode, CVI must commit each reply
+before delivering it, without waiting for audio playback. The MCP client does
+not replay earlier turns into the prompt. If a PAL repeats answered questions,
+check that its delivered replies appear in the next CVI model input; seeing
+them in the tool result alone does not prove they were retained.
+
+Builder history is separate. `tavus_pal_builder_history` reads the creator's
+saved dialogue with Charlie, not the PAL's test conversation. The local
+Builder recipe saves each received reply before executing tools or requesting
+a continuation, so a later timeout does not erase earlier dialogue. Automatic
+drafting instructions and test/refinement prompts are not saved as creator
+messages. A resumed Builder receives a bounded snapshot from request-handler;
+the full saved dialogue remains available through the history tools.
+
 For fully offline unit tests, monkeypatch
 `daily_transport.open_conversation_session` with a scripted fake session (see
 `tests/unit/test_build_and_verify.py::FakeSession`).

{tavus_cli-0.10.2 → tavus_cli-0.10.3}/src/tavus_mcp/__init__.py

@@ -1,4 +1,4 @@
 """Tavus CLI and MCP server package."""
 
 __all__ = ["__version__"]
-__version__ = "0.10.2"  # x-release-please-version
+__version__ = "0.10.3"  # x-release-please-version

{tavus_cli-0.10.2 → tavus_cli-0.10.3}/src/tavus_mcp/sdk/recipes/build_and_verify.py

@@ -114,6 +114,8 @@ async def build_and_verify(
             answer_provider=None,
             on_builder_reply=None,
             transcript=build_transcript,
+            record_creator_prompt=False,
+            creator_request=prompt,
         )
         await client.builders.publish(builder_id)
         spec = await _assemble_spec(client, pal_id)

{tavus_cli-0.10.2 → tavus_cli-0.10.3}/src/tavus_mcp/sdk/recipes/carrier_builder.py

@@ -95,15 +95,19 @@ async def run_carrier_build(
     answer_provider: AnswerProvider | None,
     on_builder_reply: BuildProgress | None,
     transcript: list[dict[str, Any]],
+    record_creator_prompt: bool = True,
+    creator_request: str | None = None,
 ) -> str:
     """Build or refine one PAL through a real, headless Charlie conversation."""
+    initial_request = creator_request if creator_request is not None else prompt
     runtime: dict[str, Any] = {
         "pal_id": pal_id,
         "builder_id": builder_id,
         "read_documents": {},
         "drafted": False,
+        "creator_request": initial_request,
     }
-    initial_context = await build_carrier_context(client, runtime, initial_request=prompt)
+    initial_context = await build_carrier_context(client, runtime, initial_request=initial_request)
     created = await client.conversations.create_headless_builder(
         {
             "builder_id": builder_id,
@@ -134,6 +138,7 @@ async def run_carrier_build(
     human_driven = answer_provider is not None or answers is not None
     answer_index = 0
     next_message = _complete_brief(prompt) if not human_driven else prompt
+    creator_message = prompt if record_creator_prompt else None
     max_creator_turns = max_rounds + 1 if human_driven else max_rounds
 
     try:
@@ -157,7 +162,8 @@ async def run_carrier_build(
                 client,
                 runtime,
                 next_message,
-                initial_request=prompt,
+                creator_message=creator_message,
+                initial_request=initial_request,
                 transcript=transcript,
             )
             if on_builder_reply is not None:
@@ -176,6 +182,7 @@ async def run_carrier_build(
                     answer_index += 1
                 else:
                     break
+                creator_message = next_message
                 continue
 
             if runtime["drafted"]:
@@ -185,6 +192,7 @@ async def run_carrier_build(
                 "Do not ask me another setup question: name the PAL and dispatch "
                 "a complete first draft now."
             )
+            creator_message = None
 
         if not runtime["drafted"]:
             await _run_creator_turn(
@@ -192,7 +200,8 @@ async def run_carrier_build(
                 runtime,
                 "The build must complete now. Infer any remaining choices from the "
                 "original brief and call draft_pal for a complete first draft.",
-                initial_request=prompt,
+                creator_message=None,
+                initial_request=initial_request,
                 transcript=transcript,
             )
         if not runtime["drafted"]:
@@ -221,27 +230,43 @@ async def _run_creator_turn(
     runtime: dict[str, Any],
     message: str,
     *,
+    creator_message: str | None,
     initial_request: str,
     transcript: list[dict[str, Any]],
 ) -> dict[str, Any]:
     builder_id = str(runtime["builder_id"])
     session = runtime["session"]
-    transcript.append({"role": "user", "text": message})
-    await client.builders.append_messages(
-        builder_id, [{"role": "user", "content": message}]
-    )
+    if creator_message:
+        runtime["creator_request"] = creator_message
+        await client.builders.append_messages(
+            builder_id, [{"role": "user", "content": creator_message}]
+        )
+        transcript.append({"role": "user", "text": creator_message})
 
     reply = await _send_and_collect(session, message)
     tool_names: list[str] = []
-    for _ in range(MAX_TOOL_HOPS):
+    reply_texts: list[str] = []
+    for hop in range(MAX_TOOL_HOPS + 1):
+        text = str(reply.get("text") or "").strip()
+        if text:
+            # Save before tool work or another inference can fail and lose this reply.
+            await client.builders.append_messages(
+                builder_id, [{"role": "assistant", "content": text}]
+            )
+            transcript.append({"role": "assistant", "text": text})
+            reply_texts.append(text)
         calls = list(reply.pop("tool_calls", []))
-        if not calls:
+        if not calls or hop == MAX_TOOL_HOPS:
             break
         outcomes: list[str] = []
         for call in calls:
             name = str(call.get("name") or "")
             tool_names.append(name)
-            outcome = await _execute_carrier_tool(client, runtime, call, message)
+            # RQH persists the drafting message too. Keep transport instructions
+            # and automated nudges out of the creator's saved dialogue.
+            outcome = await _execute_carrier_tool(
+                client, runtime, call, str(runtime["creator_request"])
+            )
             outcomes.append(outcome)
             tool_call_id = str(call.get("tool_call_id") or "")
             if tool_call_id:
@@ -270,21 +295,13 @@ async def _run_creator_turn(
         )
         reply = await _send_and_collect(session, followup)
 
-    text = str(reply.get("text") or "").strip()
-    if not text:
-        text = _tool_narration(reply, tool_names)
     structured_reply = {
-        "text": text,
+        "text": "\n\n".join(reply_texts) or _tool_narration(reply, tool_names),
         "tool_names": tool_names,
         "draft_ready": bool(runtime["drafted"]),
         "targets": [],
         "suggestions": [],
     }
-    transcript.append({"role": "assistant", **structured_reply})
-    if text:
-        await client.builders.append_messages(
-            builder_id, [{"role": "assistant", "content": text}]
-        )
     return structured_reply
 
 

{tavus_cli-0.10.2 → tavus_cli-0.10.3}/tests/unit/test_build_and_verify.py

@@ -398,6 +398,10 @@ async def test_build_uses_headless_carrier_and_pal_maker_model(
     assert len(builder_session.sent_texts) == 2
     assert routes["fanout"].call_count == 1
     fanout_body = json.loads(routes["fanout"].calls.last.request.read())
+    assert fanout_body["message"] == (
+        "Build a warm office greeter for delivery drivers. Route packages to "
+        "the lobby shelf and never request access codes."
+    )
     assert fanout_body["briefs"] == {
         "personality": "Warm office greeter for delivery drivers.",
         "objectives": "Route each delivery to the lobby shelf.",
@@ -490,6 +494,153 @@ def test_map_draft_briefs_merges_facts_into_personality() -> None:
     }
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("closeout", ["The greeting is updated.", None])
+@respx.mock
+async def test_carrier_preserves_replies_before_a_tool_followup(
+    clean_env: Path, monkeypatch: pytest.MonkeyPatch, closeout: str | None
+) -> None:
+    """Regression: tool continuations lost earlier replies, especially on timeout."""
+    _write_env(clean_env)
+    routes = _register_build_routes()
+    saved: list[dict[str, str]] = []
+
+    def append(request):
+        saved.extend(json.loads(request.content)["messages"])
+        return Response(200, json=saved)
+
+    routes["append"].mock(side_effect=append)
+    respx.get(f"{PUBLIC}/builder/b1").mock(
+        side_effect=lambda _: Response(200, json={"session": {"chat_history": saved}})
+    )
+    first = "I will keep the no-Memory decision and update the greeting."
+    turns = [
+        {
+            "text": first,
+            "tool_calls": [
+                {
+                    "name": "draft_pal",
+                    "tool_call_id": "greeting",
+                    "arguments": json.dumps({"greeting_brief": "Ask about deliveries."}),
+                }
+            ],
+        }
+    ]
+    if closeout:
+        turns.append({"text": closeout, "tool_calls": []})
+    session = FakeSession("c_builder", turns)
+
+    async def open_session(*_args: Any, **_kwargs: Any) -> FakeSession:
+        return session
+
+    monkeypatch.setattr(daily_transport, "open_conversation_session", open_session)
+    transcript: list[dict[str, Any]] = []
+    async with await _make_client() as client:
+
+        async def run():
+            return await carrier_builder.run_carrier_build(
+                client,
+                builder_id="b1",
+                pal_id="p1",
+                prompt="Change the greeting.",
+                max_rounds=1,
+                answers=[],
+                answer_provider=None,
+                on_builder_reply=None,
+                transcript=transcript,
+            )
+
+        if closeout:
+            await run()
+        else:
+            with pytest.raises(TavusError, match="No CVI reply"):
+                await run()
+        history = await client.builders.chat_history("b1")
+
+    expected = [
+        {"role": "user", "content": "Change the greeting."},
+        {"role": "assistant", "content": first},
+    ]
+    if closeout:
+        expected.append({"role": "assistant", "content": closeout})
+    assert history["messages"] == expected
+    assert [{"role": row["role"], "content": row["text"]} for row in transcript] == expected
+    assert session.closed
+    assert routes["builder_end"].called
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("max_rounds", [1, 2])
+@respx.mock
+async def test_carrier_does_not_save_automatic_nudges_as_creator_messages(
+    clean_env: Path, monkeypatch: pytest.MonkeyPatch, max_rounds: int
+) -> None:
+    """Regression: resuming a build must not attribute client instructions to its creator."""
+    _write_env(clean_env)
+    routes = _register_build_routes()
+    session = FakeSession(
+        "c_builder",
+        [
+            {"text": "Who will use this PAL?", "tool_calls": []},
+            {
+                "text": "I will draft the greeter.",
+                "tool_calls": [
+                    {
+                        "name": "draft_pal",
+                        "tool_call_id": "draft",
+                        "arguments": json.dumps({"sections_brief": "An office greeter."}),
+                    }
+                ],
+            },
+            {"text": "The greeter is ready.", "tool_calls": []},
+        ],
+    )
+
+    refinement = FakeSession("c_builder", [
+        {"text": "I will refine the greeting.", "tool_calls": [{
+            "name": "draft_pal", "tool_call_id": "refine",
+            "arguments": json.dumps({"greeting_brief": "Ask about deliveries."}),
+        }]},
+        {"text": "The greeting is ready.", "tool_calls": []},
+    ])
+    builders = iter([session, refinement])
+
+    async def open_session(conversation_id: str, *_args: Any, **_kwargs: Any) -> FakeSession:
+        if conversation_id == "c_builder":
+            return next(builders)
+        return FakeSession("c_smoke", [{"text": "Hello!", "tool_calls": []}])
+
+    routes["judge"].mock(side_effect=[
+        Response(200, json={"overall": "partial", "summary": "Improve the greeting."}),
+        Response(200, json={"overall": "pass"}),
+    ])
+    monkeypatch.setattr(daily_transport, "open_conversation_session", open_session)
+    async with await _make_client() as client:
+        result = await build_and_verify(
+            client, prompt="Build an office greeter.", face_id="rabc",
+            max_rounds=max_rounds, probes=["Hello"], auto_refine=True, max_refine_rounds=1,
+        )
+    transcript = result["build_transcript"]
+
+    saved = [
+        row
+        for call in routes["append"].calls
+        for row in json.loads(call.request.content)["messages"]
+    ]
+    assert [row for row in saved if row["role"] == "user"] == [
+        {"role": "user", "content": "Build an office greeter."},
+    ]
+    assert [row for row in transcript if row["role"] == "user"] == [
+        {"role": "user", "text": "Build an office greeter."},
+    ]
+    assert len(session.sent_texts) == 3
+    # RQH also saves this field; automated drafting must not add a fake user turn.
+    assert all(
+        json.loads(call.request.content)["message"] == "Build an office greeter."
+        for call in routes["fanout"].calls
+    )
+
+
 @pytest.mark.parametrize(
     "args",
     [

{tavus_cli-0.10.2 → tavus_cli-0.10.3}/uv.lock

@@ -1303,7 +1303,7 @@ wheels = [
 
 [[package]]
 name = "tavus-cli"
-version = "0.10.2"
+version = "0.10.3"
 source = { editable = "." }
 dependencies = [
     { name = "fastmcp" },

{tavus_cli-0.10.2 → tavus_cli-0.10.3}/workers/tavus-mcp/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "@tavus/tavus-mcp-worker",
-  "version": "0.10.2",
+  "version": "0.10.3",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "@tavus/tavus-mcp-worker",
-      "version": "0.10.2",
+      "version": "0.10.3",
       "dependencies": {
         "@modelcontextprotocol/sdk": "^1.29.0",
         "agents": "^0.12.2",

{tavus_cli-0.10.2 → tavus_cli-0.10.3}/workers/tavus-mcp/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tavus/tavus-mcp-worker",
-  "version": "0.10.2",
+  "version": "0.10.3",
   "private": true,
   "type": "module",
   "scripts": {

tavus_cli-0.10.3/workers/tavus-mcp/src/version.ts

@@ -0,0 +1 @@
+export const VERSION = "0.10.3"; // x-release-please-version

tavus_cli-0.10.2/.release-please-manifest.json

@@ -1,3 +0,0 @@
-{
-  ".": "0.10.2"
-}

tavus_cli-0.10.2/workers/tavus-mcp/src/version.ts

@@ -1 +0,0 @@
-export const VERSION = "0.10.2"; // x-release-please-version