Posting here per the maintainer’s direction in google-antigravity/antigravity-cli#997 —
this is about the standalone Antigravity ACP server, not the CLI and not the IDE,
so the CLI repo redirected this report to this forum.
Environment
- Antigravity ACP server 1.1.1 (
agy_acp_server.par+localharness_external),
Linux x86_64, driven over stdio ACP (self-hosted control plane, no IDE involved) - Model:
gemini-3.8-flash-high - Conversation DB:
~/.gemini/antigravity-acp/conversations/<session>.db
Symptom
Every prompt to the session fails before any model call with:
Agent execution error: could not find doneCh for checkpoint
Harness stderr on every attempt:
E ... trajectory_chat_converter.go:1115] no checkpoint found, truncating conversation to step index 3. current tokens: 126983, token limit: 128000
E ... executor.go:489] error in pre-invocation hook: could not find doneCh for checkpoint
E ... errorreport.go:223] agent executor error: could not find doneCh for checkpoint
Agent run failed: executor run failed: could not find doneCh for checkpoint
Restarting the ACP server does not help — the failure is 100% reproducible for this
conversation, while other conversations on the same install compact and run normally
(e.g. one healthy session accumulated 16 completed checkpoints).
Root cause (from the trajectory DB)
The conversation contains exactly one CHECKPOINT step (step type 23) persisted with
status 2 (in-progress) and a 197-byte payload. Its creation timestamp matches a window
where the previous server process died on an upstream model 503 — checkpoint generation
itself makes a model call, so the checkpoint was started, the process died, and the step
was persisted as in-progress forever.
After restart + resume, the truncation path needs a usable checkpoint; the pre-invocation
hook then waits on a doneCh that was only registered in the dead process’s memory, and
the run aborts. Since the stuck step never transitions, every subsequent prompt repeats
the same failure.
Suggested fix direction: on session load, treat in-progress checkpoints from a previous
process as failed/ignorable (and optionally trigger a fresh one) instead of awaiting an
in-memory channel that cannot exist after a restart.
Repro
- Use a long conversation until it approaches the 128k checkpoint token limit.
- Kill the ACP server (or let the model call fail, e.g. 503) while a checkpoint step
is being generated, leaving it persisted with in-progress status. - Restart, resume the session, send a prompt once tokens are near the limit.
Workaround
None found. Rotating/deleting the conversation DB is the only way out (loses the
trajectory). Restarting the server does not help.
Same root cause independently analyzed in google-antigravity/antigravity-cli#997 —
at least two separate users hit this on ACP server 1.1.1.