Bug: Imported Codex tasks use <recommended_plugins> as their titles

[maurorozco]

Version: T3 Code 0.0.39, macOS.

Steps: Import existing Codex sessions through onboarding when their transcripts begin with an injected <recommended_plugins> block.

Actual: Multiple tasks are titled <recommended_plugins>.

Expected: Use Codex’s saved task title, or derive a title from the actual user request after excluding injected context.

Cause: AgentSessionScanner.ts derives the title from the first user message’s first line.

The same logic remains in nightly v0.0.40-nightly.20260907.1346. PR #10446 addresses JSON-wrapped generated titles, not this import issue.

[juliusmarminge]

Triage

Confirmed bug. Reproduced in code on current main (same logic as 0.0.39 and v0.0.40-nightly.20260907.1346).

Onboarding import titles Codex threads from the first line of the first user message. When Codex prepends an injected <recommended_plugins> block to that message, every such thread is named <recommended_plugins>.

This is the same first-line fallback already used for other Codex preambles. Existing tests expect title === "<environment_context>" when the prompt starts with that markup.

Where it breaks

• apps/server/src/project/AgentSessionScanner.ts — parseAgentSessionRecords():
  • Claude reads aiTitle from the transcript.
  • Codex never sets title. Fallback is firstUserMessage.text.trim().split("\n")[0].
  • Separate Codex instruction response_items (e.g. <user_instructions>) are dropped when a canonical event_msg exists. Injected XML inside that first user message is not stripped for titles.
  • session_meta is only used for the session id. No name / threadName is parsed.
• apps/server/src/project/AgentSessionImporter.ts — thread.create uses that derived title as-is.
• apps/server/src/project/AgentSessionScanner.test.ts — locks in the current first-line behavior.

Codex’s own thread list already distinguishes name (“Optional user-facing thread title”) from preview (“Usually the first user message”). Import only reads JSONL rollouts and never consults that saved name.

Not this

• fix(server): handle JSON-wrapped titles and verbose Claude output #10446 — unwraps JSON-wrapped generated titles. Different path, already merged.
• feat(web): import recent Codex and Claude threads during onboarding #8066 / feat(web): first-run welcome wizard with agent setup and project import #5362 — introduced this import + first-line fallback.
• feat(web): group onboarding project import by repository #10493 / fix(server): import transcripts with oversized tool records #10430 — later import work; they did not change title derivation.

No duplicate issue found.

Suggested fix

In parseAgentSessionRecords, when title is still null:

1. Prefer a Codex saved title if the transcript/session metadata has name / threadName.
2. Otherwise skip leading injected markup (<recommended_plugins>, <environment_context>, similar instruction/heading-only preambles) and take the first remaining non-empty line of the actual user request (still cap at 100 chars; fall back to "Imported thread").

Keep imported message text as-is so resume/history stay complete. Update the tests that currently expect <environment_context> as the title.

Severity

Minor–medium UX. History and resume still work; imported Codex threads become hard to tell apart. Accepting as an import title bug.

[yashas-salankimatt]

Additional reproduction on T3 Code 0.0.42: after the built-in onboarding import on macOS, 54 of 63 existing threads had malformed titles. Codex titles included <recommended_plugins> and # AGENTS.md instructions for [redacted project path]. Both are injected setup text ahead of the actual request.

The same title behavior affected Codex imports on Linux with T3 Code 0.0.42. Conversation content was present. We corrected local titles through the supported thread.meta.update API; this was a title-only workaround, not a fix for future imports.

Current main still derives a Codex import title from the first line of the first user message (AgentSessionScanner.ts, parseAgentSessionRecords). Codex's saved title, when available, should take precedence. Otherwise the fallback should skip leading setup blocks, including XML-like markup and the # AGENTS.md instructions heading, while leaving imported message text intact.

[wellorbetter]

Additional Windows reproduction with both injected-context titles and Codex subagent imports.

Environment:

• Windows 11 Pro x64, version 10.0.26200 (build 26200)
• T3 Code Desktop 0.0.42
• Codex Desktop
• Current Codex CLI: 0.158.0-alpha.2.1
• Some affected rollout files were created by Codex CLI 0.155.0-alpha.16.4
• Codex sessions include multi-agent runs that spawn workers, reviewers, architects, and other child agents

Reproduction steps:

1. Use Codex Desktop in a local project and run conversations that spawn subagents.
2. Open T3 Code 0.0.42.
3. Import the existing Codex project and conversation history during onboarding.
4. Inspect the imported conversations in the project sidebar.

Observed result:

• T3 imported 137 Codex rollout files as top-level conversations.
• Only 22 were actual top-level sessions.
• 115 were subagent sessions belonging to 9 parent conversations.
• Six imported conversations were titled exactly:
  <in-app-browser-context source="ambient-ui-state">
• Other obviously incorrect titles included:
  • 52 <recommended_plugins>
  • 14 <environment_context>
  • 13 # AGENTS.md instructions
  • 6 other injected-context markers

The original Codex conversations were not corrupted. The problem is that T3 treats child rollout files as independent top-level conversations and derives titles from injected context recorded as user-role content.

The affected session_meta records contain enough information to distinguish them:

• payload.source.subagent
• payload.parent_thread_id

Expected result:

• Codex subagent rollouts should not be imported as unrelated top-level conversations.
• Imported top-level conversations should use the saved thread name or the structured user.text content, rather than injected app/browser/plugin/environment context.

I also evaluated the current logic in PR #11807 against this local dataset. Its subagent predicate identified all 115 child sessions while retaining all 22 top-level sessions, and its structured title selection produced non-empty user-request titles for all 22 retained sessions.

[Yuruzuu]

Still reproducing on 0.0.46-nightly.20261004.2648 on macOS. I imported 54 Codex/Claude chats through the setup wizard, and several titles became # AGENTS.md instructions, <recommended_plugins>, or <system-reminder> instead of their saved names. It makes the imported history pretty hard to navigate.

I fixed the imported titles locally using Codex, but the built-in importer still picked up setup text and tags instead of the actual chat names.