Antigravity: "Running view_file" never gets a completed status mid-turn, so the UI shows it hanging for minutes and the trail log records fake durations

Open
#22 1 comment 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

  • #31 by @warhammered — closed without merging

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
55/100
Issue type
Bug
Clarity
Mostly clear
Activity status
Active
Tech stack
typescript
Domain
backend, testing

Research direction

Start by tracing the ACP adapter’s handling of session/update messages and how it matches tool call IDs to open view_file calls; the issue does not name source files. Enable DUSK_DEBUG_ACP=1 and reproduce by asking Antigravity to read several workspace files. Done means read calls emit their completion before bot.turn.end, covered by a test for an Antigravity read.

Written by the indexing model from the issue text.

Description

Version

0.2.15 (also seen on 0.2.12 before the update)

Platform

Own server (Windows 11 + WSL2 Ubuntu 24.04)

Engine

Gemini 4 Hydrogen via Antigravity ACP (agy_acp_server.par in bwrap)

What happened

Expected: each view_file tool call is marked done when Antigravity finishes it, like Read on the Claude engine.

Actual: the chat shows "Running view_file" for minutes (users see "Digging around" for 10 to 14 minutes). In logs/trail-server.jsonl (16 Nov 22:32 to 17 Nov 18:33 WIB):

count
Running view_file events 393
logged ok:true, all within 2 s of bot.turn.end 195
logged ok:false mid-turn 189
ok:true logged mid-turn 0
completions sharing the same second with others 358 (bursts up to 20)

Median logged duration 52 s, p90 366 s, max 957 s. Claude engine Read in the same period: 119 calls, median 0.2 s, 97% ok. Running client_view_file: 10 calls, 3 ok.

So a view_file is never closed when it actually finishes. It is closed in a batch, either as failed when the next step starts or as ok when the turn ends. The status the user sees and the durations in the log are both wrong, which makes every Gemini turn look stuck and makes the "slow model" notice fire.

Steps to reproduce
  1. Bot on Antigravity / Gemini 4 Hydrogen, ask it to read 5 files in its workspace and summarize them.
  2. Watch the chat: each "Running view_file" stays until the turn ends.
  3. grep 'view_file' ~/.duskbot/logs/trail-server.jsonl and compare ts with the turn's bot.turn.end.
Likely cause (hypothesis, not confirmed in source)

The ACP adapter does not match Antigravity's tool_call_update with status: completed to the open view_file call (different toolCallId, or the update arrives without the status field the adapter checks), so the call is only settled by the end-of-step or end-of-turn sweep.

Suggested fix

Log the raw session/update messages for one view_file with DUSK_DEBUG_ACP=1 and settle the call on the first update that carries completed/failed or content for that toolCallId. Add a test that bot.tool for an Antigravity read is emitted before bot.turn.end.

Dominant language
TypeScript
Stars
277
Forks
27
Avg merge
2h 47m
Merged PRs (30d)
1

Getting set up

This project ships no dev container, Dockerfile or contributing guide, so setting up is up to you: start from its README, and see our first-contribution guide for the general steps.

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from DuskBotAI/duskbot

All issues in DuskBotAI/duskbot

Similar issues

More TypeScript issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.