Mistral Workflows integration + runtime_bridge extraction + MISTRAL_NATIVE router

[lchoquel]

What this brings to dev

Lands the Mistral Workflows integration and its supporting runtime refactors, and reconciles the latest dev.

Headline features

• Mistral Workflows integration extracted into a standalone package (pipelex-mistralai-workflows); the framework-agnostic core (boundary types, run_pipe_via_bridge, execution modes, ensure_pipelex_booted) is promoted to pipelex.runtime_bridge.* so any host runtime — not just Mistral Workflows — can embed Pipelex.
• runtime_bridge.primitives extraction — activity logic (delivery, hydration, trace flush, graph assembly, submitter hydration, pipe-run-arg) lifted out of pipelex.temporal so both the Temporal track and the Mistral-native track share one body.
• MISTRAL_NATIVE execution mode with a contextvar-backed pipe_router override on the hub, so a host runtime can turn sub-pipe calls into its own child workflows/activities.
• Temporal IDs + observability redesign — workflow_id derived from pipeline_run_id, TypedSearchAttributes, per-activity/per-handle task-queue routing, named worker-runtime profiles, and per-call summaries in the dashboard.

dev reconciliation

Merges current dev (error-handling overhaul, dataclass/pathlib cleanups, the v0.29–v0.30.3 releases). The shared Temporal/PipeStructure work dev already shipped in v0.29.0 is deduplicated, with this branch's net-new work layered on top. Woven points worth noting:

• Error-report propagation threads through the runtime_bridge delivery primitive (execute_delivery now forwards error_report / request_id); Temporal activities keep the @convert_pipelex_errors boundary and recover_error_report recovery.
• Temporal config adopts dev's reviewed defaults (search_attributes off, empty queue_options, default worker-runtime profile, frozen DispatchOptions).
• Mistral workers stay on mistralai 2.x with dev's InferenceErrorCategory / UserAction error-classification enrichment ported onto the 2.x calls.

Verification

make tb, make agent-check (ruff / plxt / pyright / mypy all clean), pipelex validate --all, make agent-test (full suite), and the Temporal error-handling integration tests all pass.

🤖 Generated with Claude Code

---

Summary by cubic

Extracts Mistral Workflows into pipelex-mistralai-workflows and introduces a framework‑agnostic pipelex.runtime_bridge, enabling DIRECT, Temporal, and Mistral‑native execution of Pipelex pipes with shared primitives and updated deps. Also fixes bridge scoping by preserving the caller’s library context and isolating per‑call class registries, and updates the vendored Mistral Workflows docs with corrected examples and links.

• New Features

  • New pipelex.runtime_bridge with run_pipe_via_bridge and execution modes: direct, temporal_blocking, temporal_fire_and_forget, and mistral_native.
  • Shared primitives extracted (hydration, submitter_hydration, delivery, trace_flush, graph_assembly, pipe_classification); Temporal wrappers now thin adapters.
  • Contextvar-backed pipe_router override in the hub to route sub-pipes through a host runtime (used by mistral_native).
  • Error reports and request IDs flow through the delivery primitive for consistent webhook/storage payloads.
  • Bridge now restores the caller’s current library and scopes a per-call ClassRegistry when loading crates to prevent global Kajson leaks; tests cover library-context restore and registry isolation. mistral_list_available_models now raises when no models remain after type filtering.

• Migration

  • Replace the removed pipelex[mistralai-workflows] extra with pip install pipelex-mistralai-workflows and import from pipelex_mistralai_workflows.*.
  • Update to mistralai>=2.4.4 (new import path mistralai.client.*) and temporalio==1.24.0.
  • Requires Python >=3.12.

^{Written for commit d5c90e3. Summary will update on new commits.}

[greptile-apps]

Greptile Summary

This PR adds runtime bridge support and a Mistral-native execution path. The main changes are:

• Extracted shared runtime bridge types, bootstrap helpers, and execution modes.
• Moved Temporal pipe-run primitives into pipelex.runtime_bridge.primitives.
• Added context-scoped pipe router overrides for host runtimes.
• Updated Temporal pipe workflows for shared delivery, hydration, tracing, and error-report handling.
• Ported Mistral plugin code to the Mistral SDK 2.x import surface and reasoning controls.

Confidence Score: 2/5

These issues should be fixed before merging.

• Library context can be cleared after crate-backed bridge and workflow runs.
• Temporal pipe-run options can be silently ignored for both blocking and fire-and-forget starts.
• Mistral text prompts can be sent with system instructions after user content.

Focus on pipelex/runtime_bridge/bridge.py, pipelex/temporal/tprl_pipe/wf_pipe_router.py, pipelex/temporal/tprl_pipe/temporal_pipe_run.py, and pipelex/plugins/mistral/mistral_factory.py.

Important Files Changed

Filename	Overview
pipelex/runtime_bridge/bridge.py	Adds the host-runtime bridge entry point and scoped library handling for crate-backed runs.
pipelex/temporal/tprl_pipe/temporal_pipe_run.py	Wraps Pipelex pipe execution in Temporal workflows for blocking and fire-and-forget modes.
pipelex/temporal/tprl_pipe/wf_pipe_router.py	Runs pipe-router child workflows with per-workflow libraries and buffered graph tracing.
pipelex/plugins/mistral/mistral_factory.py	Builds Mistral SDK clients, chat messages, OCR inputs, and OCR outputs.

Comments Outside Diff (4)

1. pipelex/temporal/tprl_pipe/wf_pipe_router.py, line 173-177 (link)

   Preserve workflow library

   The workflow opens a crate-backed library and then always tears down the current library context. If this child workflow code is entered while another library context is already active, that outer context is erased after the pipe finishes. Any later code in the same workflow task that depends on the previous current library can fail lookups or run without the intended library.

   Prompt To Fix With AI

   This is a comment left during a code review.
   Path: pipelex/temporal/tprl_pipe/wf_pipe_router.py
   Line: 173-177
   
   Comment:
   **Preserve workflow library**
   
   The workflow opens a crate-backed library and then always tears down the current library context. If this child workflow code is entered while another library context is already active, that outer context is erased after the pipe finishes. Any later code in the same workflow task that depends on the previous current library can fail lookups or run without the intended library.
   
   How can I resolve this? If you propose a fix, please make it concise.

2. pipelex/temporal/tprl_pipe/temporal_pipe_run.py, line 71-75 (link)

   Forward run options

   TemporalPipeRun stores workflow_execution_timeout and retry_policy, and the factory passes config defaults into the instance, but this executor creation drops those stored values. A caller that overrides the timeout or retry policy through make_temporal_pipe_run(...) will still start WfPipeRun with the factory defaults, so required deadlines or retry behavior are silently misapplied.

   Prompt To Fix With AI

   This is a comment left during a code review.
   Path: pipelex/temporal/tprl_pipe/temporal_pipe_run.py
   Line: 71-75
   
   Comment:
   **Forward run options**
   
   `TemporalPipeRun` stores `workflow_execution_timeout` and `retry_policy`, and the factory passes config defaults into the instance, but this executor creation drops those stored values. A caller that overrides the timeout or retry policy through `make_temporal_pipe_run(...)` will still start `WfPipeRun` with the factory defaults, so required deadlines or retry behavior are silently misapplied.
   
   How can I resolve this? If you propose a fix, please make it concise.

3. pipelex/temporal/tprl_pipe/temporal_pipe_run.py, line 112-116 (link)

   Forward start options

   The fire-and-forget path has the same option drop as the blocking path. start() builds a new executor without forwarding the timeout and retry settings stored on this TemporalPipeRun, so async pipe runs can ignore the configured workflow deadline or retry policy even though the instance was constructed with those values.

   Prompt To Fix With AI

   This is a comment left during a code review.
   Path: pipelex/temporal/tprl_pipe/temporal_pipe_run.py
   Line: 112-116
   
   Comment:
   **Forward start options**
   
   The fire-and-forget path has the same option drop as the blocking path. `start()` builds a new executor without forwarding the timeout and retry settings stored on this `TemporalPipeRun`, so async pipe runs can ignore the configured workflow deadline or retry policy even though the instance was constructed with those values.
   
   How can I resolve this? If you propose a fix, please make it concise.

4. pipelex/plugins/mistral/mistral_factory.py, line 86-90 (link)

   Send system first

   This builds Mistral chat messages with the user message before the system message. The structured-generation path in this same file sends the system message first, and chat providers commonly validate or prioritize messages by role order. For prompts with both system_text and user content, text generation can receive a different prompt shape than object generation, leading to provider errors or ignored system instructions.

   Prompt To Fix With AI

   This is a comment left during a code review.
   Path: pipelex/plugins/mistral/mistral_factory.py
   Line: 86-90
   
   Comment:
   **Send system first**
   
   This builds Mistral chat messages with the user message before the system message. The structured-generation path in this same file sends the system message first, and chat providers commonly validate or prioritize messages by role order. For prompts with both `system_text` and user content, text generation can receive a different prompt shape than object generation, leading to provider errors or ignored system instructions.
   
   
   
   How can I resolve this? If you propose a fix, please make it concise.

Prompt To Fix All With AI

Fix the following 5 code review issues. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 5
pipelex/runtime_bridge/bridge.py:227-236
**Restore library context**

This scoped helper clears the current library instead of restoring the one that was active before the bridge call. When a host runtime invokes `run_pipe_via_bridge` with a `library_crate_dump` from inside an already scoped library, the bridge returns with no current library set. Later pipe or concept lookups in the same async context can then raise `RuntimeError("No current library set...")` or run against the wrong library state.

```suggestion
    library_manager = get_library_manager()
    library_id = f"runtime_bridge_{uuid4().hex[:8]}"
    prev_library_id = get_current_library_id_or_none()
    library_manager.open_library(library_id=library_id)
    set_current_library(library_id=library_id)
    try:
        library_manager.load_from_crate(library_id=library_id, crate=library_crate)
        yield library_id
    finally:
        library_manager.teardown(library_id=library_id)
        if prev_library_id is not None:
            set_current_library(library_id=prev_library_id)
        else:
            teardown_current_library()
```

### Issue 2 of 5
pipelex/temporal/tprl_pipe/wf_pipe_router.py:173-177
**Preserve workflow library**

The workflow opens a crate-backed library and then always tears down the current library context. If this child workflow code is entered while another library context is already active, that outer context is erased after the pipe finishes. Any later code in the same workflow task that depends on the previous current library can fail lookups or run without the intended library.

### Issue 3 of 5
pipelex/temporal/tprl_pipe/temporal_pipe_run.py:71-75
**Forward run options**

`TemporalPipeRun` stores `workflow_execution_timeout` and `retry_policy`, and the factory passes config defaults into the instance, but this executor creation drops those stored values. A caller that overrides the timeout or retry policy through `make_temporal_pipe_run(...)` will still start `WfPipeRun` with the factory defaults, so required deadlines or retry behavior are silently misapplied.

### Issue 4 of 5
pipelex/temporal/tprl_pipe/temporal_pipe_run.py:112-116
**Forward start options**

The fire-and-forget path has the same option drop as the blocking path. `start()` builds a new executor without forwarding the timeout and retry settings stored on this `TemporalPipeRun`, so async pipe runs can ignore the configured workflow deadline or retry policy even though the instance was constructed with those values.

### Issue 5 of 5
pipelex/plugins/mistral/mistral_factory.py:86-90
**Send system first**

This builds Mistral chat messages with the user message before the system message. The structured-generation path in this same file sends the system message first, and chat providers commonly validate or prioritize messages by role order. For prompts with both `system_text` and user content, text generation can receive a different prompt shape than object generation, leading to provider errors or ignored system instructions.

```suggestion
        if system_text := llm_job.llm_prompt.system_text:
            messages.append(SystemMessage(content=system_text))

        if user_content:
            messages.append(UserMessage(content=user_content))
```

_{Reviews (1): Last reviewed commit: "Merge branch 'dev' into feature/Mistral-..." | Re-trigger Greptile}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1db5ecb987

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[cubic-dev-ai]

24 issues found across 114 files

_{You’re at about 92% of the monthly reviewed-line limit. You may want to disable incremental reviews to conserve quota. Reviews will continue until that limit is exceeded. If you need help avoiding interruptions, please contact contact@cubic.dev.}

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name=".claude/skills/workflows/references/getting-started/value-proposition.mdx">

<violation number="1" location=".claude/skills/workflows/references/getting-started/value-proposition.mdx:111">
P2: Broken link: `../appendices/streaming` points to a non-existent `appendices/` directory. Readers will get a 404.</violation>
</file>

<sub>Note: This PR contains a large number of files. cubic only reviews up to 100 files per PR, so some files may not have been reviewed. cubic prioritizes the most important files to review.
On a pro plan you can use ultrareview for larger PRs.

Re-trigger cubic</sub>