[Bug]: Editing and resending a message shows a context handoff from a model to itself

Aberta
#39,790 0 comentários 0 reações 0 responsáveis Ver no GitHub

Mantenedores costumam responder em até 1 dia

@smirea já está trabalhando nisso.

Desde 20/11/2026.

  • #39791 de @smirea — aberto

Avaliação

Dificuldade
4/5
Tempo estimado
3-5 dias
Facilidade para iniciantes
25/100
Tipo de issue
Bug
Clareza
Claramente especificada
Status de atividade
Estagnada
Stack de tecnologia
typescript
Domínio
frontend, full-stack

Direção de pesquisa

Start with packages/client-runtime/src/handoff.ts and trace how resolveHandoffEndpoints handles the recovery handoff described in the issue. Then inspect the web and mobile handoff rows and the existing behavior requested in #28178. Done means recovery is represented as a fresh session with kept history, without displaying a model-to-itself handoff; issue #39791 already has an open linked pull request.

Escrita pelo modelo de indexação a partir do texto da issue.

Descrição

Before submitting
  • I searched existing issues and did not find a duplicate.
  • I included enough detail to reproduce or investigate the problem.
Area

apps/web

Steps to reproduce

With the dev replay agent (Replay (dev)), in the web client:

  1. Open a thread that has one finished turn.
  2. Send a second message, for example "Second question", and let the turn finish.
  3. On that message choose Edit from here, then Revert and keep changes. The message returns to the composer.
  4. Send it again without changing the model.

Any ACP agent does the same (Antigravity, Grok, Kiro), since none of them can truncate its own conversation. It was first seen on Antigravity with Gemini 4 Neon.

Expected behavior

The timeline says what happened: the provider conversation was restarted in a fresh session and given the kept history. The server already names it that way, since the turn item it writes is titled "Recovered context". No model or provider changed, so the row should not name a source and a target.

#28178 asked for this row so that the thread "shows that the provider conversation couldn't be resumed and a new one was started". docs/user/providers-antigravity.md describes the same event for Edit from here: "the next turn starts a fresh Antigravity session and T3 Code sends it the kept conversation".

Actual behavior

A divider appears above the resent message that reads Context handoff · Multi turn → Multi turn (on Antigravity: Context handoff · 4 Neon → 4 Neon): a handoff from a model to itself.

Root cause: the resume fallback in ProviderTurnStartService writes a handoff turn item whose source and target are the same provider conversation (fromProviderThreadIds is [toProviderThreadId], title "Recovered context"). resolveHandoffEndpoints in packages/client-runtime/src/handoff.ts and the web and mobile rows treat every handoff item alike and draw <from> → <to> under the fixed label "Context handoff", so a recovery reads as a switch that never happened. Since #30188 this path runs on every Edit from here on an ACP agent, not only when a resume fails.

Impact

Cosmetic issue

Version or commit

Nightly v0.1.12-nightly.20261120.4994, and main at 31e31c608d.

Environment

Reproduced with the dev replay agent in the web client (headless Chromium, 1280×800). First seen in the macOS desktop app on an Antigravity thread, Gemini 4 Neon.

Logs or stack traces

Server log for the resent turn on the Antigravity thread:

Recovering provider context in a fresh native session
  driver: antigravity
  reason: native_session_lacks_kept_turns
  errorTag: ProviderAdapterTurnStartError
Screenshots, recordings, or supporting files

After step 4, on main:

Context handoff from Multi turn to Multi turn above the resent message

Workaround

None needed to keep working: the conversation continues with the kept history. The row cannot be hidden or corrected.

Linguagem predominante
TypeScript
Estrelas
24.8k
Forks
6.4k
Merge médio
4h 29min
PRs com merge (30d)
488

Preparar o ambiente

Abrir no Codespaces

Inicia o contêiner de desenvolvimento do projeto no navegador, com a sua própria conta do GitHub.

Primeiros passos

  1. Leia a issue inteira e depois o guia de contribuição do projeto.
  2. Comente na issue dizendo que vai assumir — evita que duas pessoas façam o mesmo trabalho.
  3. Faça um fork do repositório e trabalhe em uma branch.
  4. Abra um pull request que referencie o número da issue.

Mais de pingdotgg/t3code

Todas as issues de pingdotgg/t3code

Issues semelhantes

Mais issues de TypeScript

Receba novas issues na sua caixa de entrada

Um resumo curto de issues do GitHub para quem está começando.