gpt-5.5 reports 2432-token context window in 0.137.0, causing repeated auto-compaction and usage drain

Summary

After updating to Codex CLI/Desktop 0.137.0 / Codex Desktop app 26.602.40724, gpt-5.5 sessions report model_context_window: 2432. That appears to trigger auto-compaction almost immediately and repeatedly, including in codex exec CLI sessions. Each resumed trivial CLI turn auto-compacted before answering and total logged usage increased by roughly 26k tokens per turn.

This is affecting both Codex Desktop-originated threads and codex exec CLI threads. Since this started, usage/rate-limit drain also seems noticeably faster than expected for a 5x plan.

Expected behavior

• gpt-5.5 should report/use its real context window, not 2432 tokens.
• Auto-compaction should not trigger immediately for tiny conversations.
• After a compaction, the next tiny resumed turn should not immediately auto-compact again.
• Repeated auto-compactions caused by incorrect context-window accounting should not burn subscription usage/rate limit.

Actual behavior

• A trivial codex exec thread (Reply with OK only) reported model_context_window: 2432.
• First turn used/logged ~25k tokens, despite the prompt requiring only a tiny response.
• Each subsequent codex exec resume auto-compacted before answering.
• Total logged usage rose by roughly 26k tokens per resumed turn:
• after turn 1: 25473
• after turn 2: 51553
• after turn 3: 77826
• Desktop-originated threads show the same model_context_window: 2432 and repeated compacted / context_compacted entries.

Environment

OS: CachyOS
Kernel: Linux 7.0.10-2-cachyos x86_64 GNU/Linux
Desktop: Hyprland
Session: Wayland
Codex CLI: codex-cli 0.137.0
Codex Desktop app version: 26.602.40724
Electron version: 42.1.0
Codex Desktop Linux wrapper/build: ilysenko/codex-desktop-linux 0.8.2, commit 44798b3400f68a0080deaad2eb81288140617ae2
Codex Desktop build generatedAt: 2026-06-06T13:05:53.523Z
Model: gpt-5.5
Plan type shown in token_count events: prolite

Desktop build metadata came from:

~/Documents/code/codex-desktop-linux/codex-new-app/resources/codex-linux-build-info.json

CLI repro

Thread:

codex://threads/019e9d19-6d62-7e83-8651-2a48f14d0f78

Commands:

codex exec --json -m gpt-5.5 --skip-git-repo-check "Reply with OK only."

codex exec resume 019e9d19-6d62-7e83-8651-2a48f14d0f78 \
  --json --skip-git-repo-check "Reply with OK2 only."

codex exec resume 019e9d19-6d62-7e83-8651-2a48f14d0f78 \
  --json --skip-git-repo-check "Reply with OK3 only."

Rollout path:

~/.codex/sessions/2026/06/06/rollout-2026-06-06T14-22-29-019e9d19-6d62-7e83-8651-2a48f14d0f78.jsonl

Observed rollout/token-count sequence:

2026-06-06T13:22:30.651Z task_started model_context_window=2432
2026-06-06T13:22:33.942Z token_count total_tokens=25473 model_context_window=2432

2026-06-06T13:22:59.082Z task_started model_context_window=2432
2026-06-06T13:23:13.537Z type=compacted
2026-06-06T13:23:13.539Z token_count total_tokens=25473 last_token_usage.total_tokens=6022 model_context_window=2432 rate_limits=null
2026-06-06T13:23:13.541Z event_msg type=context_compacted
2026-06-06T13:23:15.784Z token_count total_tokens=51553 last_token_usage.total_tokens=26080 model_context_window=2432

2026-06-06T13:23:57.656Z type=compacted / compaction token_count total_tokens=51553 model_context_window=2432
2026-06-06T13:23:59.985Z token_count total_tokens=77826 model_context_window=2432

Summary for that single CLI thread:

compacted entries: 2
context_compacted events: 2

Desktop examples

Desktop thread 1

codex://threads/019e9d06-8e5e-7822-8524-1124c725fe0a
originator: Codex Desktop
cli_version: 0.137.0
model: gpt-5.5
reasoning_effort: xhigh
tokens_used in state_5.sqlite: 54950
rollout: ~/.codex/sessions/2026/06/06/rollout-2026-06-06T14-01-53-019e9d06-8e5e-7822-8524-1124c725fe0a.jsonl

Observed:

2026-06-06T13:02:05.355Z token_count total_tokens=26965 model_context_window=2432
2026-06-06T13:02:39.458Z type=compacted
2026-06-06T13:02:39.482Z event_msg type=context_compacted
2026-06-06T13:02:43.890Z token_count total_tokens=54950 model_context_window=2432
2026-06-06T13:03:37.195Z type=compacted
2026-06-06T13:03:37.218Z event_msg type=context_compacted

Desktop thread 2

codex://threads/019e9d0e-88ff-7242-a090-351eaaf652d9
originator: Codex Desktop
cli_version: 0.137.0
model: gpt-5.5
reasoning_effort: low
tokens_used in state_5.sqlite: 134426
rollout: ~/.codex/sessions/2026/06/06/rollout-2026-06-06T14-10-36-019e9d0e-88ff-7242-a090-351eaaf652d9.jsonl

Observed in this short session:

compacted rollout entries: 6
context_compacted event messages: 6

Early sequence:

2026-06-06T13:10:58.488Z token_count total_tokens=26236 model_context_window=2432
2026-06-06T13:11:27.363Z type=compacted
2026-06-06T13:11:27.371Z event_msg type=context_compacted
2026-06-06T13:11:36.559Z token_count total_tokens=53348 model_context_window=2432
2026-06-06T13:11:58.157Z type=compacted
2026-06-06T13:11:58.165Z event_msg type=context_compacted

Comparison with earlier thread

An older 0.136.0 Desktop thread with the same model reported a much larger context window:

codex://threads/019e93cd-fbcc-75b0-850f-c61fd4ae48a0
cli_version: 0.136.0
model: gpt-5.5
model_context_window observed in token_count events: 258400

The affected 0.137.0 sessions above report:

model_context_window: 2432

That looks like the likely trigger for the immediate/repeated compaction behavior.

Local inspection commands

codex --version
# codex-cli 0.137.0

sqlite3 ~/.codex/state_5.sqlite \
  "select id, datetime(created_at,'unixepoch'), datetime(updated_at,'unixepoch'), tokens_used, model, reasoning_effort, cli_version, source, title, rollout_path from threads order by created_at desc limit 5;"

python3 - <<'PY'
import json, sqlite3
ids = [
  '019e9d19-6d62-7e83-8651-2a48f14d0f78',
  '019e9d06-8e5e-7822-8524-1124c725fe0a',
  '019e9d0e-88ff-7242-a090-351eaaf652d9',
]
con = sqlite3.connect('/home/jay/.codex/state_5.sqlite')
for id in ids:
    path = con.execute('select rollout_path from threads where id=?', (id,)).fetchone()[0]
    compacted = context = 0
    windows = []
    totals = []
    with open(path) as f:
        for line in f:
            o = json.loads(line)
            if o.get('type') == 'compacted':
                compacted += 1
            if o.get('type') == 'event_msg':
                payload = o.get('payload', {})
                if payload.get('type') == 'context_compacted':
                    context += 1
                if payload.get('type') == 'token_count':
                    info = payload.get('info', {})
                    windows.append(info.get('model_context_window'))
                    totals.append(info.get('total_token_usage', {}).get('total_tokens'))
    print(id, 'compacted=', compacted, 'context_compacted=', context, 'windows=', sorted(set(windows)), 'totals=', totals)
PY

Impact

This makes Codex difficult to use because short threads can repeatedly auto-compact instead of progressing normally, and usage/rate-limit drain appears to be much faster than expected since the bug started, even on a 5x subscription plan.