alysis-code 0.13.6.2__tar.gz → 0.13.8__tar.gz

{alysis_code-0.13.6.2 → alysis_code-0.13.8}/PKG-INFO

@@ -1,7 +1,7 @@
 Metadata-Version: 2.4
 Name: alysis-code
-Version: 0.13.6.2
-Summary: Local CLI coding agent powered by any API.
+Version: 0.13.8
+Summary: Open-source coding agent for hosted and local models.
 Project-URL: Changelog, https://github.com/AlysisAi/alysis-code/blob/main/docs/CHANGELOG.md
 Project-URL: Documentation, https://github.com/AlysisAi/alysis-code/tree/main/docs
 Project-URL: Funding, https://github.com/sponsors/AlysisAi
@@ -37,6 +37,7 @@ Requires-Dist: pillow>=12.3.0
 Requires-Dist: platformdirs>=4.2.0
 Requires-Dist: prompt-toolkit>=3.0.43
 Requires-Dist: pydantic>=2.6.0
+Requires-Dist: pygments<3.0.0,>=2.17.0
 Requires-Dist: rich>=13.7.1
 Requires-Dist: tiktoken>=0.7.0
 Requires-Dist: typer>=0.12.0
@@ -151,8 +152,10 @@ sandbox only when explicitly configured with `verify_sandbox.mode="off"` or
 
 Two layers control what the agent may do — one enforces, one focuses.
 
-**Execution modes** are the enforcement layer. Choose one per command with `--mode`, switch in chat
-with `/mode`, or set a default with `alysis-code config set default_mode <mode>`:
+**Execution modes** are the enforcement layer. Choose one per command with `--mode`, select the
+next message's permissions with `/permissions` or **Shift+Tab**, or set a default with
+`alysis-code config set default_mode <mode>`. A selection leaves the active turn's permissions
+unchanged until the next message starts:
 
 | Mode | Behavior |
 | --- | --- |
@@ -175,6 +178,10 @@ A persona is a convention, never a permission: it can keep or lower your executi
 never raise it. Define your own in `.alysis_personas/*.md`.
 [Learn more](https://alysiscode.com/docs/concepts/personas)
 
+Use `/persona architect` for planning in chat, or `/forge` for a plan-driven execution workflow.
+The legacy chat Plan Mode and its `/plan` commands have been removed; Forge retains its plan views
+and editor. Persona changes made during a running turn apply after that turn finishes.
+
 ## Project Links
 
 - [Website](https://alysiscode.com/)

{alysis_code-0.13.6.2 → alysis_code-0.13.8}/README.md

@@ -93,8 +93,10 @@ sandbox only when explicitly configured with `verify_sandbox.mode="off"` or
 
 Two layers control what the agent may do — one enforces, one focuses.
 
-**Execution modes** are the enforcement layer. Choose one per command with `--mode`, switch in chat
-with `/mode`, or set a default with `alysis-code config set default_mode <mode>`:
+**Execution modes** are the enforcement layer. Choose one per command with `--mode`, select the
+next message's permissions with `/permissions` or **Shift+Tab**, or set a default with
+`alysis-code config set default_mode <mode>`. A selection leaves the active turn's permissions
+unchanged until the next message starts:
 
 | Mode | Behavior |
 | --- | --- |
@@ -117,6 +119,10 @@ A persona is a convention, never a permission: it can keep or lower your executi
 never raise it. Define your own in `.alysis_personas/*.md`.
 [Learn more](https://alysiscode.com/docs/concepts/personas)
 
+Use `/persona architect` for planning in chat, or `/forge` for a plan-driven execution workflow.
+The legacy chat Plan Mode and its `/plan` commands have been removed; Forge retains its plan views
+and editor. Persona changes made during a running turn apply after that turn finishes.
+
 ## Project Links
 
 - [Website](https://alysiscode.com/)

alysis_code-0.13.8/docs/CHANGELOG.md

@@ -0,0 +1,328 @@
+# Changelog
+
+Notable user-facing changes to Alysis Code are recorded here. This project follows
+[Keep a Changelog](https://keepachangelog.com/en/1.1.0/) and uses
+[Semantic Versioning](https://semver.org/spec/v2.0.0.html) while the public API remains pre-1.0.
+
+## [Unreleased]
+
+No user-facing changes yet.
+
+## [0.13.8] - 2026-09-16
+
+### Changed
+
+- Remove legacy chat Plan Mode, its draft/approve flow, and its prompts and toolbar
+  state. Use the Architect persona for planning in chat; Forge retains its plan views
+  and editor.
+- Replace `/mode` with `/permissions`. Permission selections and Shift+Tab apply to
+  the next user message, with the pending choice visible beside the active permissions.
+- Apply persona changes and approved persona-switch proposals only after the active
+  turn finishes. Preserve pending permission selections when resuming a conversation.
+- Show pending steering messages and queued follow-ups. Shift+Left recalls the newest
+  queued message for editing. Interruptions preserve undelivered work and wait for the
+  current worker to finish before another turn starts.
+- Read large Git diffs in complete pages with a content identifier that detects
+  changes between pages. The agent can select a path or inspect staged changes.
+- Allow a background-process wait of up to 900 seconds, bounded by the remaining
+  run budget and returning early when its requested condition is met.
+- Explain how to correct rejected verification commands, including how to use the
+  session's configured verification contract.
+
+### Fixed
+
+- Keep plain assistant replies in conversation history so follow-up requests retain
+  the previous answer even when the provider returns no extra metadata.
+- Redact credential fields inside nested JSON, tool arguments, feedback text, and
+  support bundles. Omit data that exceeds the safe redaction depth and non-UTF-8
+  diagnostic artifacts that cannot be sanitized reliably.
+- Clean up incomplete support bundles after cancellation or I/O errors, support long
+  Windows export paths, and open report drafts without disrupting terminal output.
+- Inspect Git changes without running configured diff programs, text converters,
+  filesystem monitors, or content filters. Worktree reads that require content
+  filters or submodule helpers return guidance for an approved shell command;
+  staged diffs and unaffected scoped paths remain available.
+- Preserve hosted-account explanations for rolling usage limits, concurrent requests,
+  reserved credits, and temporary routing or capacity problems.
+
+## [0.13.7] - 2026-09-11
+
+### Added
+
+- Rotating usage tips above the chat input while the agent works, with clickable links.
+  Tips hide during dialogs and stay out of conversation history.
+
+### Changed
+
+- The context footer shows remaining conversation space after startup prompts and tools.
+  Provider measurements and calibrated estimates inform the gauge, while capacity warnings
+  continue to use the full request budget.
+- Ctrl+J is the multiline input shortcut; Enter submits. The Alt+Enter alias is removed.
+- Updated public documentation and contributor guidance.
+
+### Fixed
+
+- Setup continuation hints stay visible above the footer on small terminals and long screens.
+- Benchmark wheel builds preserve local build metadata and existing build artifacts, including
+  when a build fails or is interrupted.
+
+## [0.13.6.2] - 2026-09-08
+
+### Fixed
+
+- **Alysis Code account setup and configuration use subscription controls.**
+  Setup identifies the Alysis Code endpoint without the retired MiMo label.
+  Hosted profiles show Alysis Code subscription access and account sign-in,
+  reconnect, and disconnect controls instead of an API-key editor.
+- **Hosted model changes no longer ask for a base URL.** The managed endpoint
+  stays in place through model selection and back navigation. Custom and BYOK
+  providers keep their endpoint settings.
+- **Switching to an Alysis Code account keeps the correct connection.** Signing
+  in refreshes the full active profile before configuration is saved, preventing
+  the previous provider's URL from overwriting the hosted endpoint. Login also
+  switches back to the native agent from a selected delegated runtime. Account
+  status reflects the Alysis login, including after logout when an unrelated
+  provider key is still saved.
+
+## [0.13.6.1] - 2026-09-08
+
+### Fixed
+
+- **V4.1 Flash beta on Alysis Code Pro.** The Pro model picker now includes
+  `deepseek-v4.1-flash-expires-on-0910` and `deepseek-v4-flash-vision-exp`
+  with descriptions. Newly discovered models are captioned "available on your
+  Alysis Code Pro plan" instead of the retired MiMo "trial" wording.
+
+## [0.13.6] - 2026-09-08
+
+### Added
+
+- **DeepSeek V4.1 Flash internal beta.** `deepseek-v4.1-flash-expires-on-0910`
+  (announced 2026-09-08) is now on the DeepSeek preset menu. It is a
+  new-architecture model with native multimodal input, served on the unchanged
+  `https://api.deepseek.com` base URL under a temporary id that stops resolving
+  on 2026-09-10; DeepSeek bills it at the `deepseek-v4-flash` rate and caps
+  accounts at 20 concurrent requests. Registry metadata assumes the V4 Flash
+  shape (1M context, 384K output, vision, thinking `low|high|max`) until
+  DeepSeek publishes the GA id and limits. `deepseek-v4-pro` stays the default.
+
+### Changed
+
+- **DeepSeek footer labels follow vendor spelling.** Names now appear as
+  `DeepSeek-V4-Pro` and `DeepSeek-V4.1-Flash`; dated beta ids omit their
+  `-expires-on-<date>` suffix from the displayed label.
+
+## [0.13.5] - 2026-09-07
+
+### Added
+
+- Ship eight bundled workflow skills: `address-pr-comments`, `code-review`,
+  `commit`, `debug`, `fix-ci`, `release-notes`, `security-review`, and `skill-creator`.
+- Select a relevant skill automatically with a bounded, reasoning-off model call;
+  read its instructions before other task tools, with ordinary routing as fallback.
+- Invoke skills with `$<name> [task]` at an idle chat prompt, with live completion
+  and one-turn attachment. `/skill` remains supported.
+- Add `bundled_skills_enabled` configuration, project and user override precedence,
+  and model-in-loop bundled workflow evaluations with launch-readiness checks.
+
+### Fixed
+
+- Harden skill instruction boundaries, selection ordering, and per-turn cleanup.
+- Preserve resolved skill settings in child agents and restore slash completion.
+- Recognize silent and inline verification results and preserve authoritative
+  verification evidence when supplemental checks run.
+- Bound stalled OpenAI-compatible and Responses streams, including traffic that
+  contains no meaningful progress.
+
+## [0.13.4] - 2026-09-05
+
+Model-catalog and TUI-access refresh.
+
+### Added
+
+- Added GPT-6 Astra as the default OpenAI model, with aliases, pricing, context-window,
+  reasoning, OpenRouter, and ChatGPT-subscription metadata.
+- Added newly available first-party and gateway routes for Claude Fable 5.1, Gemini 3.8
+  Flash, Qwen 3.8, GLM-5.3, Kimi K3, Doubao Seed 2.1, DeepSeek V4 snapshots, Grok 4.6,
+  and related hosted variants.
+
+### Changed
+
+- Refreshed every hosted preset's model picker, aliases, validation model, retirement
+  mappings, pricing, context limits, and reasoning contracts against provider catalogs.
+- Updated TUI model labels to preserve vendor naming and made profile pickers show each
+  profile's model, host, active state, retired-model state, and duplicates.
+- Migrated the Perplexity preset from the retiring Sonar Chat Completions endpoint to the
+  Agent API, including hosted web search and saved-profile migration.
+- Replaced the OpenAI economy and validation model with GPT-5.6 Luna and kept
+  GPT-5.3 Codex limited to the Responses preset.
+
+### Fixed
+
+- OpenAI Chat Completions now uses `max_completion_tokens` for GPT-5 and newer models.
+- Native Gemini tool and response schemas are projected onto Gemini's supported OpenAPI
+  subset before requests are sent.
+
+## [0.13.3] - 2026-08-28
+
+### Fixed
+
+- Restored the Ubuntu Python 3.11 and 3.12 CI jobs after the repository cleanup by aligning Ruff
+  formatting, updating moved sandbox contract paths, preserving the security-documentation
+  contract, making immediate shell-output coverage independent of thread scheduling, and allowing
+  harmless floating-point clock precision in deadline coverage.
+
+## [0.13.2] - 2026-08-28
+
+Public launch and repository-layout refresh.
+
+### Added
+
+- A theme-adaptive README banner with dedicated light and dark artwork.
+- `alysis-code` as the canonical executable, while retaining `alysis` as the short alias.
+
+### Changed
+
+- Simplified the public README around installation, core capabilities, modes, and documentation.
+- Moved governance files under `.github/`, release and changelog documents under `docs/`, and
+  sandbox and benchmark sources under `scripts/` for a cleaner repository root.
+- Legacy Sylliptor environment and command compatibility now forwards silently while retaining
+  diagnostic logging for support.
+
+### Fixed
+
+- Device-login approval URLs now remain pinned to the configured Alysis site instead of trusting
+  a backend-provided origin.
+
+## [0.13.1] - 2026-08-27
+
+### Fixed
+
+- Kept terminal input, selection, scrolling, paste hints, and theme behavior consistent across
+  chat and configuration screens.
+- Made intentional clean stops exit successfully while preserving non-zero exits for genuine
+  failures and unknown stop reasons.
+- Added bounded retries for connections interrupted during a response, with consistent deadline
+  and diagnostic reporting.
+
+## [0.13.0] - 2026-08-26
+
+### Changed
+
+- Renamed Sylliptor to Alysis Code. The package is now `alysis-code`, the command is `alysis`, and
+  the Python module is `alysis_code`.
+- Preserved the deprecated `sylliptor` command and migration support for existing configuration,
+  credentials, plugins, and project directories. See the [migration guide](migration-alysis-code.md).
+- Improved run-budget handling, deadline-aware shell waits, persistent services, edit diagnostics,
+  sampling controls, and build provenance.
+
+### Security
+
+- Added redaction at log and telemetry write boundaries for credential-shaped values.
+
+## [0.12.0] - 2026-08-18
+
+### Added
+
+- Added Code, Architect, Ask, and Debug personas.
+- Added provider-neutral Terminal-Bench and Harbor adapters.
+- Added deterministic release builds, dependency auditing, SBOM generation, and provenance
+  attestations.
+
+### Changed
+
+- Improved evidence-based finalization, blocker reporting, provider authentication, model
+  discovery, context management, and verification repair.
+- Strengthened boundaries around credentials, protected paths, filesystem access, and sensitive
+  tool output.
+
+## [0.11.2] - 2026-07-27
+
+### Added
+
+- Added route arbitration, turn-contract enforcement, regression baselines, and evidence-backed
+  completion checks.
+- Added tracked process cleanup, workspace provisioning, incomplete-subagent containment, and new
+  Moonshot/Kimi model presets.
+- Added a master switch for web tools and search backends.
+
+### Changed
+
+- Refreshed the ChatGPT subscription model catalog and provider-reported final-answer handling.
+
+## [0.11.1] - 2026-07-20
+
+### Changed
+
+- Redesigned the interactive configuration screen with grouped settings, clearer summaries, and
+  improved provider and model selection.
+- Added provider-specific reasoning controls and clearer per-role model overrides.
+
+### Fixed
+
+- Applied configuration changes made before the first message without restarting the terminal UI.
+- Corrected input-budget calculations for shared-window models.
+
+## [0.11.0] - 2026-07-18
+
+### Added
+
+- Added specialist subagent roles and capability-aware delegation.
+- Added live subagent progress and result attribution to the terminal UI.
+- Added an opt-in visual designer for image generation.
+
+### Security
+
+- Constrained child tools, execution modes, deadlines, and result handling so delegated output
+  cannot change parent authority or permissions.
+
+## [0.10.0] - 2026-07-13
+
+### Added
+
+- Added ChatGPT Codex subscription login, encrypted refreshable credentials, and live model
+  discovery.
+- Added DDGS web search as a keyless fallback and expanded hosted search support.
+- Added repository mapping, focused test discovery, cached update prompts, and safer workspace
+  previews.
+
+### Changed
+
+- Improved provider-aware prompt caching, usage accounting, terminal navigation, Forge planning,
+  and OpenAI Responses streaming recovery.
+
+## [0.9.8] - 2026-06-29
+
+### Added
+
+- Added stronger completion contracts and verification evidence for one-shot and managed tasks.
+- Added shell waiting, durable service tools, and stricter web-fetch provenance.
+
+### Changed
+
+- Improved Forge verification, runtime deadlines, provider adapters, usage tracking, and recovery
+  from unknown tool calls.
+
+## Earlier releases
+
+Release notes for versions before 0.9.8 remain available in the
+[GitHub Releases archive](https://github.com/AlysisAi/alysis-code/releases).
+
+[Unreleased]: https://github.com/AlysisAi/alysis-code/compare/v0.13.8...HEAD
+[0.13.8]: https://github.com/AlysisAi/alysis-code/compare/v0.13.7...v0.13.8
+[0.13.7]: https://github.com/AlysisAi/alysis-code/compare/v0.13.6.2...v0.13.7
+[0.13.6.2]: https://github.com/AlysisAi/alysis-code/compare/v0.13.6.1...v0.13.6.2
+[0.13.6.1]: https://github.com/AlysisAi/alysis-code/compare/v0.13.6...v0.13.6.1
+[0.13.6]: https://github.com/AlysisAi/alysis-code/compare/v0.13.5...v0.13.6
+[0.13.5]: https://github.com/AlysisAi/alysis-code/compare/v0.13.4...v0.13.5
+[0.13.4]: https://github.com/AlysisAi/alysis-code/compare/v0.13.3...v0.13.4
+[0.13.3]: https://github.com/AlysisAi/alysis-code/compare/v0.13.2...v0.13.3
+[0.13.2]: https://github.com/AlysisAi/alysis-code/compare/v0.13.1...v0.13.2
+[0.13.1]: https://github.com/AlysisAi/alysis-code/compare/v0.13.0...v0.13.1
+[0.13.0]: https://github.com/AlysisAi/alysis-code/compare/v0.12.0...v0.13.0
+[0.12.0]: https://github.com/AlysisAi/alysis-code/compare/v0.11.2...v0.12.0
+[0.11.2]: https://github.com/AlysisAi/alysis-code/compare/v0.11.1...v0.11.2
+[0.11.1]: https://github.com/AlysisAi/alysis-code/compare/v0.11.0...v0.11.1
+[0.11.0]: https://github.com/AlysisAi/alysis-code/compare/v0.10.0...v0.11.0
+[0.10.0]: https://github.com/AlysisAi/alysis-code/compare/v0.9.8...v0.10.0
+[0.9.8]: https://github.com/AlysisAi/alysis-code/releases/tag/v0.9.8

{alysis_code-0.13.6.2 → alysis_code-0.13.8}/pyproject.toml

@@ -4,8 +4,8 @@ build-backend = "hatchling.build"
 
 [project]
 name = "alysis-code"
-version = "0.13.6.2"
-description = "Local CLI coding agent powered by any API."
+version = "0.13.8"
+description = "Open-source coding agent for hosted and local models."
 readme = "README.md"
 requires-python = ">=3.11"
 license = { text = "Apache-2.0" }
@@ -44,6 +44,7 @@ dependencies = [
   "platformdirs>=4.2.0",
   "pydantic>=2.6.0",
   "prompt-toolkit>=3.0.43",
+  "pygments>=2.17.0,<3.0.0",
   "rich>=13.7.1",
   "tiktoken>=0.7.0",
   "typer>=0.12.0",

{alysis_code-0.13.6.2 → alysis_code-0.13.8}/src/alysis_code/__init__.py

@@ -1,3 +1,3 @@
 __all__ = ["__version__"]
 
-__version__ = "0.13.6.2"
+__version__ = "0.13.8"

{alysis_code-0.13.6.2 → alysis_code-0.13.8}/src/alysis_code/agent/README.md

@@ -5,14 +5,15 @@ managed execution flows.
 
 ## Contents
 
-- `session.py` builds session state from configuration, workspace context, and
+- [session.py](session.py) builds session state from configuration, workspace context, and
   runtime kind.
-- `tools_assembly.py` builds the concrete tool surface for a session.
-- `turn.py` runs the turn loop and tool-call iteration.
-- `routing.py` handles routing and final response shaping.
-- `prompt_context.py` prepares workspace, convention, skill, plugin, and
+- [tools_assembly.py](tools_assembly.py) builds the concrete tool surface for a session.
+- [turn/core.py](turn/core.py) runs the turn loop and tool-call iteration.
+- [turn_path.py](turn_path.py) defines execution posture and reply-language directives.
+- [prompt_context.py](prompt_context.py) prepares workspace, convention, skill, plugin, and
   verification context.
-- `verification.py` and `verification_commands.py` support verification flows.
+- [verification.py](verification.py) and [verification_commands.py](verification_commands.py)
+  support verification flows.
 
 ## Scope
 

{alysis_code-0.13.6.2 → alysis_code-0.13.8}/src/alysis_code/agent/prompt_context.py

@@ -22,7 +22,6 @@ from ..extensions.activation import (
 from ..extensions.models import normalize_extension_id, plugin_slug_from_id
 from ..extensions.state import load_global_state, load_project_state
 from ..personas import DEFAULT_PERSONA, normalize_persona, persona_modes_enabled
-from ..plan_mode import extract_approved_plan_user_message
 from ..repo_scan import (
     _MANIFEST_SPECS,
     _README_NAMES,
@@ -1460,10 +1459,7 @@ def _build_repo_task_brief_message(
     }:
         return None
 
-    clean = (
-        extract_approved_plan_user_message(pending_instruction)
-        or str(pending_instruction or "").strip()
-    )
+    clean = str(pending_instruction or "").strip()
     if not clean or _is_host_managed_user_context_message(clean):
         return None
     if clean[:1] in {"/", ":"} and "\n" not in clean:
@@ -1828,11 +1824,10 @@ def refresh_session_task_brief_from_observed_turn(
         return False
     messages_obj, existing_index, current_content, inserted_placeholder = ensured
 
-    approved_plan_message = extract_approved_plan_user_message(instruction)
-    if not approved_plan_message and material_edit_count <= 0:
+    if material_edit_count <= 0:
         return inserted_placeholder
 
-    clean = approved_plan_message or str(instruction or "").strip()
+    clean = str(instruction or "").strip()
     if not clean or _is_host_managed_user_context_message(clean):
         return inserted_placeholder
     if clean[:1] in {"/", ":"} and "\n" not in clean:
@@ -2090,9 +2085,6 @@ def _build_user_message(
     image_paths: list[str] | None,
 ) -> tuple[dict[str, Any], dict[str, Any]]:
     log_payload: dict[str, Any] = {"content": instruction}
-    display_content = extract_approved_plan_user_message(instruction)
-    if display_content and display_content != instruction:
-        log_payload["display_content"] = display_content
     if not image_paths:
         return {"role": "user", "content": instruction}, log_payload
 

{alysis_code-0.13.6.2 → alysis_code-0.13.8}/src/alysis_code/agent/session.py

@@ -17,7 +17,6 @@ from typing import TYPE_CHECKING, Any
 
 from .. import __version__
 from ..agent import _patchable
-from ..agentbox_integration import AgentBoxTelemetry
 from ..background_runner import (
     DisabledBackgroundRunner,
     LazyBackgroundShellRunner,
@@ -549,11 +548,11 @@ class AgentSession:
     one_shot_execution: bool = False
     # Persona mode (code|architect|ask|debug). A convention layered on the
     # execution-mode gate, never an enforcement layer; "code" is the no-op
-    # persona. See docs/persona_modes_design.md.
+    # persona. See docs/personas.md.
     persona: str = "code"
     # The user's chosen execution mode remembered while a narrowing persona
     # (architect/ask) is active, so switching back to code/debug restores it.
-    # None when the active persona narrows nothing. An explicit /mode <exec>
+    # None when the active persona narrows nothing. An explicit /permissions <exec>
     # always wins and clears the restore point (and restores the write scope).
     persona_restore_mode: str | None = None
     # The user's base allow_write_globs snapshotted alongside
@@ -629,7 +628,6 @@ class AgentSession:
     cache_keepalive: ParentCacheKeepalive | None = None
     crash_diagnostics: CrashDiagnosticLogger | None = None
     crash_diagnostic_log_path: str | None = None
-    agentbox_telemetry: AgentBoxTelemetry | None = None
     process_group_registry: ProcessGroupRegistry | None = None
     # Empty-response handling is budgeted per session, not per turn: two failed
     # recovery cycles mean the endpoint is not answering, and re-spending the
@@ -840,8 +838,6 @@ class AgentSession:
             if self.mcp_manager is not None:
                 self.mcp_manager.close()
         finally:
-            if self.agentbox_telemetry is not None:
-                self.agentbox_telemetry.close(error=reason not in {"session_close", "completed"})
             self.store.close()
 
     def _hook_warning(self, message: str, *, code: str = "hook_warning") -> None:
@@ -1154,8 +1150,6 @@ class AgentSession:
                     persistent_has_media=request_contains_media(self.messages),
                 )
             self.store.append("llm_usage", usage_record.to_payload())
-            if self.agentbox_telemetry is not None:
-                self.agentbox_telemetry.record_usage(usage_record)
             return usage_record
         except Exception as exc:  # noqa: BLE001 -- accounting cannot break the agent turn
             self.store.append(
@@ -1298,15 +1292,21 @@ class AgentSession:
         )
         if rewrite_payload is not None:
             self.store.append("final_summary_rewrite", rewrite_payload)
-        assistant_message = None
+        # Keep ordinary text replies in history so follow-up requests retain the
+        # assistant turn even when the provider supplies no opaque metadata.
+        assistant_message = (
+            None if internal_fallback else {"role": "assistant", "content": emitted_text}
+        )
         if assistant_response is not None:
-            candidate_message = assistant_message_from_response(
+            assistant_message = assistant_message_from_response(
                 assistant_response,
                 content=emitted_text,
             )
-            if PROVIDER_METADATA_KEY in candidate_message:
-                assistant_message = candidate_message
-        extra_payload = {"message": assistant_message} if assistant_message is not None else None
+        extra_payload = (
+            {"message": assistant_message}
+            if assistant_message is not None and PROVIDER_METADATA_KEY in assistant_message
+            else None
+        )
         if internal_fallback and self.subagent_depth > 0:
             # A nested run's locally generated stop report is internal state. The
             # nested surface forwards assistant messages up to the parent's panel,
@@ -1709,29 +1709,16 @@ class AgentSession:
         self._turn_owner_thread_id = turn_thread_id
         self._bind_provider_retry_observer(self.client)
         try:
-            if self.agentbox_telemetry is None:
-                return _run_turn(
-                    self,
-                    instruction,
-                    image_paths=image_paths,
-                    routing_mode_override=routing_mode_override,
-                    ephemeral_system_messages=ephemeral_system_messages,
-                    ephemeral_user_messages=ephemeral_user_messages,
-                    cancellation_token=cancellation_token,
-                    chat_only=chat_only,
-                )
-            self.agentbox_telemetry.task(instruction)
-            with self.agentbox_telemetry.turn():
-                return _run_turn(
-                    self,
-                    instruction,
-                    image_paths=image_paths,
-                    routing_mode_override=routing_mode_override,
-                    ephemeral_system_messages=ephemeral_system_messages,
-                    ephemeral_user_messages=ephemeral_user_messages,
-                    cancellation_token=cancellation_token,
-                    chat_only=chat_only,
-                )
+            return _run_turn(
+                self,
+                instruction,
+                image_paths=image_paths,
+                routing_mode_override=routing_mode_override,
+                ephemeral_system_messages=ephemeral_system_messages,
+                ephemeral_user_messages=ephemeral_user_messages,
+                cancellation_token=cancellation_token,
+                chat_only=chat_only,
+            )
         except CooperativeCancellationError as exc:
             # The budget watchdog cancels through the same cooperative channel a
             # user does, so the two are told apart by reason. A budget stop is a
@@ -3234,10 +3221,6 @@ def create_session(
             execution_deadline=execution_deadline,
             crash_diagnostics=crash_diagnostics,
             crash_diagnostic_log_path=resolved_crash_diagnostic_log_path,
-            agentbox_telemetry=AgentBoxTelemetry.from_env(
-                root=root,
-                runtime_version=f"alysis-{__version__}",
-            ),
             process_group_registry=process_group_registry,
         )
         if subagent_depth == 0 and store.enabled: