fix: strip chatgpt transcript tags from agent messages

[ousamabenyounes]

Summary

Fix #118

Strip ChatGPT-style transcript tags from messages before appending them to the SDK session, so agent-originated content cannot be resent to OpenAI-compatible providers with rejected tags like <analysis>, <channel>, or <final>.

Follow-up review fixes:

• Preserve literal transcript-like tags in user prompts instead of sanitizing user-provided content.
• Strip attributed and self-closing transcript tags such as <final reason="done"> and <analysis/> from non-user messages.

Also annotates the descendant task list in the same module so the touched file passes pyright.

Test verification (RED -> GREEN)

RED with the follow-up production fix reverted, while keeping the new regression tests:

uv run pytest tests/test_agent_messages.py -q
FAILED tests/test_agent_messages.py::test_message_to_session_item_preserves_user_literal_tags
FAILED tests/test_agent_messages.py::test_message_to_session_item_strips_attributed_chatgpt_transcript_tags
2 failed, 1 passed

GREEN with the follow-up fix applied:

uv run pytest tests/test_agent_messages.py -q
3 passed

Full suite:

uv run pytest -q
145 passed, 2 warnings

Additional targeted validation:

uv run ruff check strix/core/agents.py tests/test_agent_messages.py
All checks passed!

uv run ruff format --check strix/core/agents.py tests/test_agent_messages.py
2 files already formatted

uv run mypy strix/core/agents.py
Success: no issues found in 1 source file

uv run pyright strix/core/agents.py
0 errors, 0 warnings, 0 informations

Global quality baseline notes, not introduced by this PR:

uv run ruff format --check .
Would reformat: strix/core/hooks.py, strix/interface/utils.py, tests/test_cli_target_list.py, tests/test_local_sources.py

uv run ruff check .
4 existing errors in strix/core/hooks.py, strix/runtime/docker_client.py, strix/telemetry/logging.py

uv run mypy strix/
69 existing errors in TUI/Pygments-related files

[greptile-apps]

Greptile Summary

This PR strips ChatGPT-style transcript tags before messages are appended to the SDK session. The main changes are:

• Added a compiled regex and helper for transcript-tag stripping.
• Applied the helper in agent message conversion.
• Added a focused regression test for simple transcript tags.
• Added a task-list type annotation in descendant cancellation.

Confidence Score: 4/5

The changed message conversion path needs a fix before merging.

• User-provided prompts can lose literal XML-like tags before they reach the SDK session.
• The new pattern handles the tested bare tags, but not attributed or self-closing forms.
• The import and task annotation changes look compatible with the project target.

strix/core/agents.py

Important Files Changed

Filename	Overview
strix/core/agents.py	Adds transcript-tag stripping in session message conversion and a compatible task-list annotation.
tests/test_agent_messages.py	Adds coverage for stripping simple <analysis>, <channel>, and <final> tags while preserving inner text.

Prompt To Fix All With AI

Fix the following 2 code review issues. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 2
strix/core/agents.py:268-271
**User Prompts Lose Literal Tags**

When a user sends text through the CLI, TUI, or resume instruction path, this line strips `<user>`, `<final>`, `<analysis>`, and the other transcript tags before the sender check. A prompt that asks the agent to inspect XML/HTML or preserve a snippet containing those literal tags is changed before it reaches the SDK session, so the model receives different input than the user provided.

```suggestion
        sender = str(message.get("from", "unknown"))
        content = str(message.get("content", ""))
        if sender == "user":
            return cast("TResponseInputItem", {"role": "user", "content": content})
        content = _strip_chatgpt_transcript_tags(content)
```

### Issue 2 of 2
strix/core/agents.py:31-34
**Attributed Transcript Tags Survive**

This pattern only removes bare tags such as `<final>` and `</final>`. If an agent message contains the same transcript tags with attributes or self-closing syntax, such as `<final reason="done">` or `<analysis/>`, those tags stay in the session item and can still be resent to an OpenAI-compatible provider that rejects transcript markup.

```suggestion
_CHATGPT_TRANSCRIPT_TAG_RE = re.compile(
    rf"</?(?:{'|'.join(_CHATGPT_TRANSCRIPT_TAGS)})(?:\s+[^>]*)?\s*/?>",
    flags=re.IGNORECASE,
)
```

_{Reviews (1): Last reviewed commit: "fix: strip chatgpt transcript tags from ..." | Re-trigger Greptile}

[ousamabenyounes]

Addressed the two Greptile review comments in a23be85.

• User prompts now bypass transcript-tag stripping so literal XML/transcript-like snippets are preserved. Covered by test_message_to_session_item_preserves_user_literal_tags.
• Non-user transcript stripping now handles optional attributes and self-closing syntax. Covered by test_message_to_session_item_strips_attributed_chatgpt_transcript_tags.

Verification:

uv run pytest tests/test_agent_messages.py -q
3 passed

uv run pytest -q
145 passed, 2 warnings

Revert proof: with the production fix reverted and the new tests kept, the targeted file fails with 2 failed / 1 passed.