refactor: align codex provider and local runtime

coding-agent/.opencode/prompts/tasks/codex-provider-responses-task-packet.md

@@ -0,0 +1,39 @@
+Goal:
+Replace the current `codex` provider's Chat Completions wire path with the official Codex Responses HTTP SSE path for ChatGPT-authenticated Codex traffic.
+
+Scope:
+- Add a Codex-specific Responses provider that posts to `/responses`.
+- Convert existing agent messages/tools into Responses API request payloads.
+- Parse Responses SSE events into `TextEvent`, `ThinkingEvent`, `ToolCallEvent`, `UsageEvent`, and `DoneEvent`.
+- Wire `provider="codex"` to the new provider and update focused tests.
+
+Out of scope:
+- Responses WebSocket transport and warmup/reuse.
+- Browser-cookie reuse beyond the existing ChatGPT bearer token flow.
+- Changes to non-Codex OpenAI-compatible providers.
+
+Context:
+- ADRs:
+  - `docs/adr/0050-codex-oauth-provider.md`
+  - `docs/adr/0057-codex-provider-responses-wire-protocol.md`
+- Relevant files:
+  - `src/coding_agent/providers/codex_responses.py`
+  - `src/coding_agent/plugins/llm_provider.py`
+  - `tests/providers/test_codex_responses.py`
+  - `tests/coding_agent/plugins/test_llm_provider.py`
+
+Target tests:
+- `uv run pytest tests/providers/test_codex_responses.py tests/coding_agent/plugins/test_llm_provider.py -k "codex" -v`
+- `uv run ruff check src/coding_agent/providers/codex_responses.py src/coding_agent/plugins/llm_provider.py tests/providers/test_codex_responses.py tests/coding_agent/plugins/test_llm_provider.py`
+
+Loop policy:
+- Engineer implements the smallest correct change and runs the target tests.
+- Reviewer reviews only the resulting diff and affected tests.
+- Reviewer reports only P1/P2 findings.
+- Engineer fixes only accepted P1/P2 findings and reruns the same target tests.
+- Verifier reruns the exact target tests and reports pass/fail only.
+
+Stop conditions:
+- At most one review/fix/retest cycle.
+- Escalate architectural redirection or scope expansion to the human.
+- Ignore non-blocking optimization suggestions.

coding-agent/.opencode/prompts/tasks/local-cli-runtime-boundary-task-packet.md

@@ -0,0 +1,48 @@
+Goal:
+Separate the local interactive CLI runtime from the HTTP/control-plane session
+manager while preserving current REPL behavior.
+
+Scope:
+- Identify the smallest behavior-preserving boundary between
+  `coding_agent.cli.repl.InteractiveSession` and server session management.
+- Introduce or prepare a local CLI runtime/session abstraction for pipeline,
+  adapter, tape id, approval memory, model switching, and checkpoint restore.
+- Remove direct local REPL dependence on remote/control-plane-only concepts from
+  `server.SessionManager` where practical in the first slice.
+
+Out of scope:
+- Renaming HTTP endpoints or remote CLI commands.
+- Moving executor, Bee, Docker, or workspace provider modules.
+- Changing checkpoint restore semantics or tape storage formats.
+
+Context:
+- ADRs:
+  - `docs/adr/0048-application-structure-refactor-boundaries.md`
+  - `docs/adr/0054-executor-runtime-terminology-and-resume-first-direction.md`
+  - `docs/adr/0055-session-resume-and-interrupted-run-semantics.md`
+  - `docs/adr/0056-local-cli-control-plane-and-workspace-product-boundaries.md`
+- Relevant files:
+  - `src/coding_agent/cli/repl.py`
+  - `src/coding_agent/cli/commands.py`
+  - `src/coding_agent/app.py`
+  - `src/coding_agent/adapter.py`
+  - `src/coding_agent/server/session_manager.py`
+  - `tests/cli/test_repl.py`
+  - `tests/cli/test_commands.py`
+
+Target tests:
+- `uv run pytest tests/cli/test_repl.py tests/cli/test_commands.py -k "managed_session or model or checkpoint or process_message" -v`
+- `uv run pytest tests/ui/test_session_manager_public_api.py tests/ui/test_session_manager_runtime.py -k "file_session_store or checkpoint_restore or runtime" -v`
+
+Loop policy:
+- Engineer implements the smallest correct change and runs the target tests.
+- Reviewer reviews only the resulting diff and affected tests.
+- Reviewer reports only P1/P2 findings.
+- Engineer fixes only accepted P1/P2 findings and reruns the same target tests.
+- Verifier reruns the exact target tests and reports pass/fail only.
+
+Stop conditions:
+- At most one review/fix/retest cycle.
+- Escalate if local CLI behavior requires changing HTTP server semantics.
+- Escalate if the slice would require migrating persisted session data.
+- Ignore non-blocking optimization suggestions.

coding-agent/.opencode/prompts/tasks/workspace-sandbox-product-boundary-task-packet.md

@@ -0,0 +1,54 @@
+Goal:
+Clarify Workspace as the execution boundary across local repositories, Docker
+sandbox workspaces, future OS-native sandboxes, cloud workspaces, and external
+executor workspace references.
+
+Scope:
+- Document and enforce product terminology around `ExecutionBinding`,
+  `Environment`, and `WorkspaceProvider`.
+- Keep session metadata responsible for workspace binding, not for filesystem or
+  shell implementation details.
+- Prepare executor-related code to sit behind an execution-plane boundary while
+  preserving existing compatibility paths.
+
+Out of scope:
+- Implementing OS-native sandbox support.
+- Rewriting Docker provider behavior.
+- Renaming public remote HTTP protocol fields in this slice.
+- Changing action-safety approval policy.
+
+Context:
+- ADRs:
+  - `docs/adr/0038-workspace-provider-and-sandbox-mvp-boundaries.md`
+  - `docs/adr/0051-external-worker-execution-control-plane.md`
+  - `docs/adr/0054-executor-runtime-terminology-and-resume-first-direction.md`
+  - `docs/adr/0056-local-cli-control-plane-and-workspace-product-boundaries.md`
+- Relevant files:
+  - `src/coding_agent/server/execution_binding.py`
+  - `src/coding_agent/environment/local.py`
+  - `src/coding_agent/environment/cloud.py`
+  - `src/coding_agent/environment/workspace_provider.py`
+  - `src/coding_agent/environment/docker_workspace_provider.py`
+  - `src/coding_agent/server/session_manager.py`
+  - `src/coding_agent/external_executor.py`
+  - `tests/ui/test_execution_binding.py`
+  - `tests/coding_agent/environment/`
+
+Target tests:
+- `uv run pytest tests/ui/test_execution_binding.py -v`
+- `uv run pytest tests/coding_agent/environment/test_local_environment.py tests/coding_agent/environment/test_cloud_environment.py -v`
+- `uv run pytest tests/coding_agent/environment/test_docker_workspace_provider.py -k "binding or unavailable or provider_instance" -v`
+- `uv run pytest tests/coding_agent/test_external_executor.py -k "capability or authorized or workspace" -v`
+
+Loop policy:
+- Engineer implements the smallest correct change and runs the target tests.
+- Reviewer reviews only the resulting diff and affected tests.
+- Reviewer reports only P1/P2 findings.
+- Engineer fixes only accepted P1/P2 findings and reruns the same target tests.
+- Verifier reruns the exact target tests and reports pass/fail only.
+
+Stop conditions:
+- At most one review/fix/retest cycle.
+- Escalate if the slice would require a protocol migration.
+- Escalate if Docker or hosted infrastructure becomes required for deterministic tests.
+- Ignore non-blocking optimization suggestions.

coding-agent/docs/adr/0056-local-cli-control-plane-and-workspace-product-boundaries.md

@@ -0,0 +1,94 @@
+# ADR-0056: Local CLI, Control Plane, and Workspace Product Boundaries
+
+**Status**: Proposed
+**Date**: 2026-06-01
+
+## Context
+
+`coding_agent` now contains three product shapes that grew from the same
+runtime: a local interactive CLI, a remote HTTP control plane, and an executor
+plane for sandboxed or attached execution. The current implementation shares
+large session/runtime objects across those shapes. For example, the local REPL
+uses the server `SessionManager`, and the server session model stores both
+durable metadata and in-process runtime handles.
+
+This makes the product meaning of `local`, `remote`, `session`, `workspace`,
+and `executor` unclear. A local CLI session should behave like a Codex or Claude
+Code session: one user, one local workspace, an interactive REPL, local approval
+prompts, tape persistence, model switching, and checkpoint restore. The remote
+server should be a control plane for sessions, runs, workspaces, approvals, and
+events. Executors should be an execution plane that consumes authorized plans
+and reports status and results.
+
+The existing runtime direction remains valid: AgentKit owns the generic
+Pipeline, Hook Runtime, Tape, Plugin, Tool, Provider, Directive, Approval, and
+Checkpoint mechanisms. Coding Agent owns product policy and workspace semantics.
+
+## Decision
+
+Separate product meanings before moving large amounts of code.
+
+The local CLI product must depend on a local runtime/session abstraction rather
+than on the HTTP control-plane `SessionManager`. That abstraction may reuse
+`app.create_agent`, `PipelineAdapter`, tape stores, checkpoint stores, and
+approval components, but it should not expose remote session, workspace
+inventory, owner lease, executor claim, or HTTP event-queue concerns to the REPL.
+
+The server/control-plane session service remains responsible for HTTP sessions,
+runs, workspace lifecycle, approval interactions, event streaming, resume
+metadata, and executor placement. It should evolve toward a split between
+durable records and in-process runtime handles.
+
+Workspace is the execution boundary for file, patch, shell, archive, diff,
+publish, cleanup, and validation operations. Sessions should record their
+workspace through `ExecutionBinding`/workspace metadata. Local repositories,
+Docker sandbox workspaces, future OS-native sandboxes, cloud workspaces, and
+external executor workspace references should be represented as explicit
+workspace bindings or provider-backed environments instead of being inferred
+from unrelated session fields.
+
+Dogfood this refactor in two bounded tasks:
+
+1. Local CLI runtime boundary: introduce or prepare a local CLI runtime/session
+   abstraction and remove the local REPL's direct dependency on the server
+   `SessionManager` in the smallest behavior-preserving slice.
+2. Workspace/sandbox/control-plane boundary: clarify workspace binding/provider
+   terminology and prepare executor-related code to live behind an execution
+   plane boundary without changing existing remote protocols.
+
+## Alternatives Rejected
+
+- Keep sharing `server.SessionManager` with local CLI — rejected because it
+  makes the simplest local product path depend on remote/control-plane and
+  executor concepts.
+- Move all session, workspace, remote, and executor files in one refactor —
+  rejected because it would create high-risk import churn before the product
+  terms are stable.
+- Put workspace provider semantics into AgentKit — rejected because repository
+  paths, Docker, cloud clients, sandbox policy, and workspace lifecycle are
+  Coding Agent product concerns.
+- Treat Docker sandbox as the only workspace abstraction — rejected because the
+  intended model also includes local repositories, future OS-native sandboxing,
+  cloud workspaces, and external executor workspace references.
+
+## Acceptance Criteria
+
+- [ ] Local CLI task packet exists and names target tests for the REPL/session boundary.
+- [ ] Workspace/sandbox task packet exists and names target tests for workspace binding/provider semantics.
+- [ ] `test_repl_does_not_import_server_session_manager_for_local_runtime`
+- [ ] `test_local_cli_runtime_preserves_model_switch_and_checkpoint_restore`
+- [ ] `test_workspace_binding_product_terms_round_trip`
+- [ ] `uv run pytest tests/cli/test_repl.py tests/cli/test_commands.py -k "managed_session or model or checkpoint or local_runtime" -v`
+- [ ] `uv run pytest tests/ui/test_execution_binding.py tests/coding_agent/environment/test_local_environment.py tests/coding_agent/environment/test_docker_workspace_provider.py -k "binding or workspace_provider or unavailable" -v`
+
+## References
+
+- `docs/adr/0038-workspace-provider-and-sandbox-mvp-boundaries.md`
+- `docs/adr/0048-application-structure-refactor-boundaries.md`
+- `docs/adr/0054-executor-runtime-terminology-and-resume-first-direction.md`
+- `docs/adr/0055-session-resume-and-interrupted-run-semantics.md`
+- `src/coding_agent/cli/repl.py`
+- `src/coding_agent/server/session_manager.py`
+- `src/coding_agent/server/execution_binding.py`
+- `src/coding_agent/environment/`
+- `src/coding_agent/external_executor.py`

coding-agent/docs/adr/0057-codex-provider-responses-wire-protocol.md

@@ -0,0 +1,39 @@
+# ADR-0057: Codex Provider Uses Responses Wire Protocol
+
+**Status**: Proposed
+**Date**: 2026-06-02
+
+## Context
+
+The current `codex` provider reuses the OpenAI-compatible Chat Completions provider and sends requests to `https://chatgpt.com/backend-api/codex/chat/completions`. The official Codex CLI does not use that path for ChatGPT-authenticated Codex traffic.
+
+OpenAI Codex CLI uses the Responses wire API against `https://chatgpt.com/backend-api/codex/responses`, with ChatGPT bearer auth plus `ChatGPT-Account-ID`, Codex identity headers, and SSE/WebSocket streaming semantics. Treating the ChatGPT Codex backend as Chat Completions causes Cloudflare/browser-gate failures and protocol-level mismatches.
+
+## Decision
+
+Replace the `codex` provider implementation with a Codex-specific Responses provider. The first implementation will support HTTP SSE against `/responses`; WebSocket prewarm/reuse remains a later enhancement.
+
+The `openai` and `openai_compat` providers continue to use Chat Completions through `OpenAICompatProvider`. Only `provider="codex"` switches to the Codex Responses provider.
+
+## Alternatives Rejected
+
+- Keep `codex` on `OpenAICompatProvider` — rejected because the official Codex CLI no longer uses Chat Completions for Codex traffic.
+- Add a new parallel `codex-responses` provider — rejected because the existing `codex` provider name already means ChatGPT-authenticated Codex and should have correct semantics.
+- Implement WebSocket first — rejected to keep the first repair narrow; the official CLI also has HTTP Responses fallback behavior.
+
+## Acceptance Criteria
+
+- [ ] `test_codex_provider_uses_responses_provider`
+- [ ] `test_stream_posts_to_codex_responses_with_chatgpt_headers`
+- [ ] `test_stream_converts_function_call_output_item_done`
+- [ ] `uv run pytest tests/providers/test_codex_responses.py tests/coding_agent/plugins/test_llm_provider.py -k "codex" -v`
+
+## References
+
+- `src/coding_agent/providers/codex_responses.py`
+- `src/coding_agent/plugins/llm_provider.py`
+- `tests/providers/test_codex_responses.py`
+- `tests/coding_agent/plugins/test_llm_provider.py`
+- `https://github.com/openai/codex/blob/main/codex-rs/model-provider-info/src/lib.rs`
+- `https://github.com/openai/codex/blob/main/codex-rs/codex-api/src/endpoint/responses.rs`
+- `https://github.com/openai/codex/blob/main/codex-rs/model-provider/src/bearer_auth_provider.rs`

coding-agent/src/coding_agent/cli/local_runtime.py

@@ -0,0 +1,55 @@
+"""Local CLI runtime/session boundary.
+
+This module is the CLI-facing seam for local interactive sessions. The first
+slice intentionally delegates to the existing server SessionManager so behavior
+stays unchanged while the REPL stops depending on control-plane symbols.
+"""
+
+from __future__ import annotations
+
+from typing import Protocol, runtime_checkable, Any
+
+from coding_agent.server.session_manager import SessionManager
+
+
+@runtime_checkable
+class LocalCliSessionManager(Protocol):
+    """Subset of session behavior used by the interactive local REPL."""
+
+    async def create_session(self, **kwargs: Any) -> str: ...
+
+    def get_session(self, session_id: str) -> Any: ...
+
+    async def get_session_async(self, session_id: str) -> Any: ...
+
+    async def ensure_session_runtime(self, session_id: str) -> Any: ...
+
+    async def replace_session_runtime_config(
+        self,
+        session_id: str,
+        *,
+        model_name: str,
+    ) -> Any: ...
+
+    async def restore_checkpoint(self, session_id: str, checkpoint_id: str) -> None: ...
+
+    async def run_agent(self, session_id: str, prompt: str) -> None: ...
+
+
+def create_local_cli_session_manager() -> LocalCliSessionManager:
+    """Create the local REPL session manager.
+
+    The concrete implementation remains the existing managed local session
+    service for now. Keeping this construction here prevents REPL code from
+    importing server/control-plane types directly.
+    """
+
+    return SessionManager(
+        storage_config={
+            "http_session_backend": "fs",
+            "runtime_backend": "jsonl",
+        }
+    )
+
+
+__all__ = ["LocalCliSessionManager", "create_local_cli_session_manager"]

coding-agent/src/coding_agent/cli/repl.py

@@ -14,6 +14,10 @@
 from coding_agent.approval import ApprovalPolicy
 from coding_agent.cli.commands import handle_command
 from coding_agent.cli.input_handler import InputHandler, expand_pasted_refs
+from coding_agent.cli.local_runtime import (
+    LocalCliSessionManager,
+    create_local_cli_session_manager,
+)
 from coding_agent.cli.terminal_output import (
     get_prompt_output,
     print_pt,
@@ -25,7 +29,6 @@
 from coding_agent.ui.stream_renderer import StreamingRenderer
 from coding_agent.ui.rich_consumer import RichConsumer
 from coding_agent.ui.status_footer import StatusFooter
-from coding_agent.server.session_manager import SessionManager
 from coding_agent.wire.protocol import TurnEnd, WireMessage
 
 
@@ -469,13 +472,8 @@ def _is_root_turn_end(message: WireMessage) -> bool:
     return isinstance(message, TurnEnd) and not message.agent_id
 
 
-def _local_session_manager() -> SessionManager:
-    return SessionManager(
-        storage_config={
-            "http_session_backend": "fs",
-            "runtime_backend": "jsonl",
-        }
-    )
+def _local_session_manager() -> LocalCliSessionManager:
+    return create_local_cli_session_manager()
 
 
 async def run_repl(config: Config):