release: v0.5.2 — hardened no-tools recovery + release-workflow guard

PyPI 0.5.1 was published from the pre-review commit (the first v0.5.1 tag push),
and PyPI versions are immutable, so the corrected build can't reuse 0.5.1. Ship
the review-hardened code from main as 0.5.2.

- openai.py: recovery stream that errors after emitting text now preserves the
  streamed text in its terminal DONE result (review finding #2).
- RELEASE_NOTES.md: the agent-layer nudge is a user-role message, not a system
  message (review finding #1); add v0.5.2 notes.
- version 0.5.2 across pyproject.toml, src/hive/__init__.py, uv.lock; CHANGELOG.md
  and docs/changelog.md 0.5.2 entries (folds in Phase 3, already on main).
- .github/workflows/release.yml: skip publishing when the version already exists
  on PyPI, so a moved/re-pushed tag can't trigger a conflicting re-upload.
- tests: recovery-stream-error-preserves-streamed-text.

Author: chiruu12

.github/workflows/release.yml

@@ -21,7 +21,20 @@ jobs:
       - name: Build package
         run: uv build
 
+      - name: Check whether this version is already on PyPI
+        id: pypi_check
+        run: |
+          version="$(grep -m1 -E '^version = ' pyproject.toml | sed -E 's/version = "([^"]+)"/\1/')"
+          echo "version=$version"
+          if curl -sf -o /dev/null "https://pypi.org/pypi/hive-agent/$version/json"; then
+            echo "hive-agent $version is already on PyPI -- skipping publish (a moved/re-pushed tag can't re-upload an immutable version)."
+            echo "exists=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "exists=false" >> "$GITHUB_OUTPUT"
+          fi
+
       - name: Publish to PyPI
+        if: steps.pypi_check.outputs.exists == 'false'
         run: uv publish --token "$PYPI_TOKEN" --check-url https://pypi.org/simple/
         env:
           PYPI_TOKEN: ${{ secrets.PYPI_TOKEN }}

CHANGELOG.md

@@ -1,5 +1,16 @@
 # Changelog
 
+## [0.5.2] — 2026-05-31
+
+### Fixed
+- **Hardened no-tools recovery** (review follow-up to 0.5.1) — streaming recovery now fires only when the `tool_use_failed` occurs *before* any text reaches the caller (no duplicated output), and a recovery stream that errors after emitting text preserves that text in its terminal result. A `tool_use_failed` on a request that *did* offer tools is no longer swallowed. The agent-layer wrap-up nudge and the adapter's text-only recovery nudge are now `user`-role messages (some strict providers reject mid-thread `system` messages); the agent nudge is sent only for the wrap-up call and isn't written to the logged conversation.
+
+### Added
+- **Phase 3 simulation core** — registry-driven world catalogs (events & jobs) and wired simulation feedback loops.
+
+### CI
+- Release workflow skips publishing when the version already exists on PyPI, so moving or re-pushing a tag can't trigger a conflicting re-upload.
+
 ## [0.5.1] — 2026-05-31
 
 ### Fixed

RELEASE_NOTES.md

@@ -1,3 +1,41 @@
+# Hive v0.5.2 Release Notes
+
+## Headline: Hardened No-Tools Recovery
+
+A follow-up to v0.5.1 that incorporates review feedback on the no-tools recovery
+path, plus the Phase 3 simulation core that landed on `main`.
+
+## What's Fixed
+
+- **Streaming recovery is scoped to pre-content failures**: the OpenAI-compatible
+  adapter now recovers a `tool_use_failed` only when it happens *before* any text has
+  reached the caller, so a mid-stream error can never produce duplicated output. If a
+  recovery stream errors after emitting text, the terminal result preserves the text
+  already streamed.
+- **Recovery no longer masks real errors**: a `tool_use_failed` on a request that
+  *did* offer tools (e.g. a malformed tool schema) is surfaced instead of being
+  swallowed by the recovery retry.
+- **No mid-thread system messages**: both the agent-layer wrap-up nudge and the
+  adapter's text-only recovery nudge are sent as `user`-role messages, which strict
+  providers accept mid-conversation (some reject mid-thread `system` messages). The
+  agent nudge is sent only for the wrap-up call and is not written to the logged
+  conversation.
+
+## Also Included
+
+- **Phase 3 simulation core**: registry-driven world catalogs (events & jobs) and
+  wired simulation feedback loops.
+
+## Upgrade
+
+```bash
+pip install --upgrade hive-agent
+```
+
+No code changes required. The happy path (tools provided, or no error) is unchanged.
+
+---
+
 # Hive v0.5.1 Release Notes
 
 ## Headline: Resilient No-Tools Wrap-Up Across Strict Providers
@@ -17,8 +55,9 @@ even though the tools that ran during the loop already persisted.
   of raising -- the tools already ran. Covers both `generate_with_metadata` and
   `generate_stream`.
 - **Belt-and-suspenders nudge** at the agent layer: `Agent.run_once` now appends a
-  "tool budget exhausted, reply in plain text" system message before the final
-  wrap-up call, reducing the chance of hitting the error at all.
+  "tool budget exhausted, reply in plain text" user message before the final
+  wrap-up call, reducing the chance of hitting the error at all. (A user-role message
+  is used deliberately -- some strict providers reject mid-thread system messages.)
 
 A multi-action turn (e.g. "make three notes") now completes on Groq without surfacing
 a 400.

docs/changelog.md

@@ -1,5 +1,16 @@
 # Changelog
 
+## [0.5.2] -- 2026-05-31
+
+### Fixed
+- **Hardened no-tools recovery** (review follow-up to 0.5.1): streaming recovery now fires only when the `tool_use_failed` occurs *before* any text reaches the caller (no duplicated output), and a recovery stream that errors after emitting text preserves that text in its terminal result. A `tool_use_failed` on a request that *did* offer tools is no longer swallowed. The agent-layer wrap-up nudge and the adapter's text-only recovery nudge are now `user`-role messages (some strict providers reject mid-thread `system` messages); the agent nudge is sent only for the wrap-up call and isn't written to the logged conversation.
+
+### Added
+- **Phase 3 simulation core**: registry-driven world catalogs (events & jobs) and wired simulation feedback loops.
+
+### CI
+- Release workflow skips publishing when the version already exists on PyPI, so moving or re-pushing a tag can't trigger a conflicting re-upload.
+
 ## [0.5.1] -- 2026-05-31
 
 ### Fixed

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "hive-agent"
-version = "0.5.1"
+version = "0.5.2"
 description = "Local-first agent OS. Spawn persistent AI agents that collaborate, write code, and use tools autonomously."
 authors = [{ name = "chiruu12" }]
 license = { text = "MIT" }

src/hive/__init__.py

@@ -1,6 +1,6 @@
 """Hive - Autonomous agent OS."""
 
-__version__ = "0.5.1"
+__version__ = "0.5.2"
 
 from hive.agents.existence import ExistenceLoop
 from hive.agents.goal_strategy import Goal, GoalContext, GoalStrategy

src/hive/models/openai.py

@@ -299,19 +299,25 @@ async def generate_stream(
             "stream": True,
             "stream_options": {"include_usage": True},
         }
+        recovered_text: list[str] = []
         try:
             stream = await self._client.chat.completions.create(**recovery_kwargs)
             async for event in self._consume_stream(stream, t0):
+                if event.type == StreamEventType.TEXT and event.text:
+                    recovered_text.append(event.text)
                 yield event
         except Exception:
             logger.warning(
-                "No-tools recovery stream for model %s failed; returning empty text",
+                "No-tools recovery stream for model %s failed; returning streamed text so far",
                 self._model,
             )
+            # If the recovery stream errored mid-flight, _consume_stream never emitted its
+            # terminal DONE -- synthesize one carrying whatever text already reached the
+            # caller, so consumers reconstructing from DONE.result don't lose it.
             yield StreamEvent(
                 type=StreamEventType.DONE,
                 result=GenerateResult(
-                    message=Message.assistant(""),
+                    message=Message.assistant("".join(recovered_text)),
                     model=self._model,
                     duration_ms=int((time.time() - t0) * 1000),
                 ),

tests/models/test_tool_use_recovery.py

@@ -175,6 +175,33 @@ async def fake_create(**kwargs: Any) -> Any:
         assert done.type == StreamEventType.DONE
         assert done.result.message.content == ""
 
+    @pytest.mark.asyncio
+    async def test_recovery_stream_error_preserves_streamed_text(self) -> None:
+        # Double failure: initial no-tools call rejects, then the recovery stream emits
+        # some text and errors mid-flight. The terminal DONE must carry the text already
+        # streamed (not empty), so consumers reconstructing from DONE.result don't lose it.
+        p = OpenAI(api_key="sk-test")
+
+        async def fake_create(**kwargs: Any) -> Any:
+            if not _has_nudge(kwargs["messages"]):
+                raise _ToolUseFailedError()
+
+            async def gen() -> Any:
+                yield _stream_chunk(content="Saved the ")
+                yield _stream_chunk(content="notes")
+                raise _ToolUseFailedError()
+
+            return gen()
+
+        p._client.chat.completions.create = fake_create
+        events = await _collect(p.generate_stream([Message.user("make three notes")], None))
+
+        texts = [e.text for e in events if e.type == StreamEventType.TEXT]
+        assert texts == ["Saved the ", "notes"]
+        done = events[-1]
+        assert done.type == StreamEventType.DONE
+        assert done.result.message.content == "Saved the notes"
+
     @pytest.mark.asyncio
     async def test_error_after_text_propagates_without_duplicate(self) -> None:
         # If the stream fails *after* text already reached the caller, recovery would