Codex VS Code follow-up turn fails with context_length_exceeded after needs_follow_up=true

What version of the IDE extension are you using?

0.140.0-alpha.2

What subscription do you have?

ChatGPT Plus

Which IDE are you using?

VS Code

What platform is your computer?

Linux 6.12.74+deb13+1-amd64 x86_64 unknown

What issue are you seeing?

Codex stops with this user-facing message:

Codex ran out of room in the model's context window. Start a new thread or clear earlier history before retrying.

The corresponding websocket failure in the logs is:

event.kind=response.failed
error.code=context_length_exceeded
error.message="Your input exceeds the context window of this model. Please adjust your input and try again."

The confusing part is the state immediately before the failure. The post-sampling token usage logs repeatedly show the turn entering a follow-up/continuation state, while the local context-limit flags are still false:

thread_id=019ed7f3-2509-76d3-8975-a5b51ab0411b
turn_id=019ed7fa-e2a2-7a62-a617-0e780a2b06ce
model=gpt-5.5
reasoning_effort=xhigh
total_usage_tokens=282204
estimated_token_count=Some(304934)
auto_compact_scope_tokens=282204
auto_compact_scope_limit=334800
full_context_window_limit=None
full_context_window_limit_reached=false
token_limit_reached=false
model_needs_follow_up=true
has_pending_input=false
needs_follow_up=true

There are several similar rows leading up to the failure:

total_usage_tokens=237867, estimated_token_count=Some(257051), needs_follow_up=true
total_usage_tokens=242488, estimated_token_count=Some(261711), needs_follow_up=true
total_usage_tokens=257185, estimated_token_count=Some(276508), needs_follow_up=true
total_usage_tokens=270057, estimated_token_count=Some(290746), needs_follow_up=true
total_usage_tokens=274825, estimated_token_count=Some(297507), needs_follow_up=true
total_usage_tokens=282204, estimated_token_count=Some(304934), needs_follow_up=true

All of those rows also report:

full_context_window_limit_reached=false
token_limit_reached=false

Then the same turn fails with:

event.kind=response.failed
error.code=context_length_exceeded
message="Your input exceeds the context window of this model. Please adjust your input and try again."

So the failure pattern looks like this:

post sampling token usage:
  model_needs_follow_up=true
  needs_follow_up=true
  token_limit_reached=false
  full_context_window_limit_reached=false

then:
  event.kind=response.failed
  error.code=context_length_exceeded

This looks like the follow-up / remote compaction continuation request is still too large when it is sent, even though the local turn accounting did not mark the token/context limit as reached.

I’m not sure whether the root cause is remote compaction, follow-up handling, token estimation, or the UI error message, but the logs make the current behavior hard to reason about: local accounting says the limit was not reached, but the next request fails with backend context_length_exceeded.

What steps can reproduce the bug?

I can reproduce this with long-running Codex VS Code sessions using gpt-5.5 with reasoning_effort=xhigh, especially during repo-heavy tasks where Codex reads many files and runs shell commands.

My config:

model = "gpt-5.5"
model_reasoning_effort = "xhigh"
service_tier = "default"

[projects."/srv/repsnord/app"]
trust_level = "trusted"

[projects."/"]
trust_level = "trusted"

[projects."/srv"]
trust_level = "trusted"

Typical reproduction flow:

1. Open Codex in VS Code.
2. Use model gpt-5.5 with reasoning_effort=xhigh.
3. Start a repo-heavy task that reads rules/docs/source files and runs shell commands.
4. Let the turn continue until Codex starts producing automatic follow-up/continuation behavior.
5. Codex eventually stops with: “Codex ran out of room in the model's context window.”
6. Inspect ~/.codex/logs_2.sqlite.
7. The logs show repeated model_needs_follow_up=true / needs_follow_up=true rows followed by response.failed context_length_exceeded.

Example thread:

thread_id=019ed7f3-2509-76d3-8975-a5b51ab0411b
turn_id=019ed7fa-e2a2-7a62-a617-0e780a2b06ce
app.version=0.140.0-alpha.2
originator=codex_vscode
model=gpt-5.5
reasoning_effort=xhigh

I also saw the same failure pattern in another thread:

thread_id=019ed7dd-d86c-7270-a147-404cdf301f0f
model=gpt-5.5
app.version=0.140.0-alpha.2
event.kind=response.failed
error.code=context_length_exceeded

What is the expected behavior?

If Codex logs:

model_needs_follow_up=true
needs_follow_up=true

then I would expect one of these outcomes:

1. Codex successfully performs the follow-up turn after compaction.
2. Codex compacts further before sending the follow-up request.
3. Codex stops earlier with a specific message that the follow-up/compaction request could not fit.
4. Codex logs the local context-limit flags consistently before the backend rejects the request.

Right now the local post-sampling state says:

full_context_window_limit_reached=false
token_limit_reached=false

but the next request fails with:

error.code=context_length_exceeded

That makes it difficult to tell whether the failure is caused by model context, remote compaction, follow-up handling, token estimation, or just the UI message.

Additional information

I generated a local evidence bundle from:

~/.codex/logs_2.sqlite
~/.codex/state_5.sqlite

Safe summary:

Codex VS Code 0.140.0-alpha.2
model=gpt-5.5
reasoning_effort=xhigh
originator=codex_vscode
wire_api=responses
transport=responses_websocket
x-codex-beta-features includes remote_compaction_v2

The most relevant local log pattern is:

model_needs_follow_up=true
needs_follow_up=true
token_limit_reached=false
full_context_window_limit_reached=false

followed by:

event.kind=response.failed
error.code=context_length_exceeded

I can provide more redacted excerpts if needed.

thread_id=019edad6-cd0e-7cb2-ae4e-0905f5521948
turn_id=019edaed-ff97-7f01-b5c0-3bfa00468fb0
model=gpt-5.5
reasoning_effort=xhigh
total_usage_tokens=273243
estimated_token_count=Some(280362)
auto_compact_scope_tokens=273243
auto_compact_scope_limit=334800
auto_compact_limit_scope=Total
auto_compact_window_ordinal=None
auto_compact_window_prefill_tokens=None
full_context_window_limit=None
full_context_window_limit_reached=false
token_limit_reached=false
model_needs_follow_up=true
has_pending_input=false
needs_follow_up=true

total_usage_tokens=311998
estimated_token_count=Some(319123)
auto_compact_scope_tokens=311998
auto_compact_scope_limit=334800
auto_compact_limit_scope=Total
auto_compact_window_ordinal=None
auto_compact_window_prefill_tokens=None
full_context_window_limit=None
full_context_window_limit_reached=false
token_limit_reached=false
model_needs_follow_up=true
has_pending_input=false
needs_follow_up=true

event.kind=response.failed
error.code=context_length_exceeded
error.message="Your input exceeds the context window of this model. Please adjust your input and try again."

type=invalid_request_error
code=context_length_exceeded
param=input

auto_compact_window_ordinal=None
auto_compact_window_prefill_tokens=None