Repository navigation
[macOS][Computer History] Recorder creates Skysight segments but 10-minute summaries never start #40386
Description
What version of the Codex App are you using (From “About Codex” dialog)?
ChatGPT desktop 26.818.61809 (build 7019)
Additional bundled versions:
Computer History plugin: 1.0.1000854
Bundled CLI: codex-cli 0.149.0-alpha.4.3
What subscription do you have?
ChatGPT Pro
What platform is your computer?
Darwin 25.6.0 arm64 arm
What issue are you seeing?
Computer History has remained empty after being enabled for several days. The settings page continues to show:
No computer history yet
The recording stage itself is working:
- Computer History reports
state=running. SkyComputerUseServiceis running.- Fresh event-stream segments continue to appear under the Skysight segment directory.
- The effective observation settings use
observeas the default for both applications and URLs, with empty allowlists and blocklists.
However, no memory summary is ever written. At the latest check:
Skysight event-stream segment directories: 55
~/.codex/memories/extensions/skysight/resources files: 0
The current segment started normally at 2026-08-24T11:20:00Z, and the recording session had been running since 2026-08-24T10:43:31Z after the latest restart. Multiple complete 10-minute windows passed without a single 10min summary file.
Fully quitting and reopening ChatGPT did not resolve the problem. Updating the app from 26.818.41509 (build 6962) to 26.818.61809 (build 7019) also did not resolve it.
The updated application contains Contents/Resources/codex_chronicle, but the summarization stage does not produce output. This appears to be a scheduler or helper-launch failure after successful event capture, rather than a screen-recording permission or observation-filter failure.
What steps can reproduce the bug?
-
Enable Computer History in ChatGPT desktop on macOS.
-
Leave the application running and use the computer normally for longer than several 10-minute summary windows.
-
Confirm that Computer History reports
runningand that new directories appear under:~/Library/Group Containers/2DC432GLL2.com.openai.sky.CUAService/Library/Caches/ComputerUse/Skysight/segments/ -
Count the recorded segments and generated summaries:
find "$HOME/Library/Group Containers/2DC432GLL2.com.openai.sky.CUAService/Library/Caches/ComputerUse/Skysight/segments" \ -mindepth 1 -maxdepth 1 -type d | wc -l find "$HOME/.codex/memories/extensions/skysight/resources" \ -type f | wc -l
-
Observe that segment count continues to increase while the resources count remains zero.
-
Open Settings → Computer History and observe that the History panel still says
No computer history yet.
Actual result from the latest check:
55 segments
0 summary resources
The issue persists after a full application restart and update.
What is the expected behavior?
After eligible activity has been captured for a complete summary window:
- the background summarizer should start;
- at least one
10minMarkdown summary should be written under~/.codex/memories/extensions/skysight/resources/; - the Computer History UI should show generated history instead of remaining permanently empty;
- if summary generation cannot start, the UI or diagnostics should expose a specific actionable error rather than reporting the feature as running.
Acceptance Criteria
- When eligible Skysight segments cover a complete 10-minute window, the 10-minute summary job is queued and completes.
- A successful job writes a
10minMarkdown resource under~/.codex/memories/extensions/skysight/resources/and makes the result visible in Computer History. - If the job cannot be queued or launched, diagnostics expose the failing scheduler/helper stage and an actionable error.
- Recording-only operation cannot remain silently reported as a healthy end-to-end Computer History state indefinitely.
Validation
- Add a regression test covering recorder output followed by 10-minute summary scheduling.
- Verify the path on macOS with default-observe settings and empty allowlists/blocklists.
- Verify that a summarizer launch failure is surfaced without stopping or corrupting the raw event stream.
Additional information
- macOS Screen & System Audio Recording permission is enabled for ChatGPT.
- Application and website permission lists load successfully.
- The bundled
codex_chronicleexecutable is present. - A read-only diagnostic run reported zero failed checks. Its warnings concerned unrelated endpoint-protection inspection, Gatekeeper assessment, and local thread-inventory findings.
- No raw event-stream contents are included because they may contain private on-screen activity.
Activity
MilkyWay008 commentedon Aug 24, 2026
55 segment dirs but zero files under
~/.codex/memories/extensions/skysight/resourcesmeans recording is fine and the summarizer helper is the part that never runs. I'd checklog show --last 2h --predicate 'process CONTAINS "codex_chronicle" OR eventMessage CONTAINS "skysight"'to see if the helper even launches and what it errors on, plus~/.codex/logand~/Library/Logsright at a 10-minute boundary. Also worth confirming~/.codexisn't symlinked oddly and runningspctl -a -vvagainst the codex_chronicle binary in case Gatekeeper/TCC is blocking it, toggling Computer History off and on resets the summarizer state if nothing else shows.grevgeny commentedon Aug 24, 2026
Thanks — I ran the suggested checks:
codex_chroniclelaunch or Skysight runtime messages; it only contains the diagnosticlog showqueries themselves.~/.codexis a normal directory, not a symlink.codesign --verify --strictreports thatcodex_chronicleis valid on disk and satisfies its Designated Requirement.ChatGPT.appas a Notarized Developer ID application.runningand producing fresh event-stream segments, while the Skysight resources directory remains empty.This further narrows the failure to the scheduler/queue before the summarizer helper is launched, rather than the recorder or an apparent Gatekeeper/signature failure.
grevgeny commentedon Aug 26, 2026
Update: I now have a deterministic root-cause reproduction. Please disregard my earlier conclusion that the failure occurs before the summarizer launches.
Current versions:
26.820.60940codex-cli 0.150.0-alpha.8The recorder remains healthy and produces fresh Skysight segments. The scheduler also launches the memory writer approximately every ten minutes. However, ten recent writer attempts failed identically with:
The writer subsequently falls back to HTTPS and receives the same 401. Consequently, it produces no output and
~/.codex/memories/extensions/skysight/resourcesremains empty.The generated writer command includes:
I reproduced the failure outside Computer History using the bundled CLI and the same relevant flags:
With
--ignore-user-config:Removing only
--ignore-user-config:Retaining
--ignore-user-configbut adding:Result:
Relevant local state:
codex login statusreportsLogged in using ChatGPT.cli_auth_credentials_store = "auto"is configured.auth.jsonfile.OPENAI_API_KEY,CODEX_API_KEY, or API base-URL override.The causal chain therefore appears to be:
This also conflicts with the CLI flag’s documented contract that authentication still uses
CODEX_HOME:codex-rs/exec/src/cli.rscodex-rs/config/src/config_toml.rsA narrow fix for the internal Computer History caller would be to retain
--ignore-user-configwhile explicitly passing the App-resolved credential-store mode, for example:A regression test should cover an isolated internal writer with credentials available only through a mocked Keychain, no
auth.json, and ignored user configuration. It should verify that the request is authenticated and that a10minsummary resource is written.No settings or application files were changed during this investigation.
GeT-LeFt commentedon Aug 27, 2026
Thanks for the deterministic authentication finding. I reproduced the same user-visible failure on the same current desktop/CLI versions, but my local authentication state differs in an important way.
Environment
26.820.60940(build7119)codex-cli 0.150.0-alpha.81.0.100081626.819.1000816Logged in using ChatGPTDifferent authentication state
In this reproduction,
$CODEX_HOME/auth.jsonis present and has existed since August 2. I inspected structure only, not credential values:0600auth_mode:chatgptOPENAI_API_KEYis nullOPENAI_API_KEY,CODEX_API_KEY, or API base-URL override is present in the environmentcli_auth_credentials_storeis not explicitly configured inconfig.tomlRunning
codex login statusrefreshed the auth file timestamp. The desktop app refreshed it again approximately 19 minutes later, confirming that the ChatGPT credential set was still active. Despite that refresh, summary generation did not recover.Controlled observations
I also tested the previously suspected CLI-path issue. For one controlled run, both the ChatGPT process and
SkyComputerUseServiceinherited:That run lasted about 39 minutes and crossed multiple complete event windows. Representative completed segments contained 96, 124, and 171 eligible events. No summary resource was generated, so the CLI-path hypothesis is ruled out for this reproduction.
After the authentication refresh, I observed another complete window:
By 10:01:57Z:
$CODEX_HOME/memories/extensions/skysight/resourcesstill contained zero files;codex exec,codex_chronicle, or other summary process was observed by periodic process sampling;The bundled
codex_chroniclebinary does contain the expected internal flags:Current conclusion
This is the same recorder-healthy / resources-empty symptom, but I cannot confirm the same local
401path becauseauth.jsonexists, is actively refreshed, and no independent401 Unauthorizedevidence is emitted locally.It may be a second silent failure mode, or the internal writer may still be failing to consume the available ChatGPT credentials before it emits observable logs. Please consider exposing the writer launch result and stderr/exit reason in diagnostics so these cases can be distinguished.
No credential values, raw event contents, or private paths are included here.
KennyNoDie commentedon Sep 1, 2026
Independent reproduction on a newer desktop and bundled CLI version.
Environment
26.825.51511(build7377)codex-cli 0.151.0-alpha.7.2codex login status:Logged in using ChatGPTcli_auth_credentials_store = "auto"is configured~/.codex/auth.jsonis absent before and aftercodex login statusDeterministic A/B reproduction
I ran the bundled CLI in an empty temporary directory.
Control:
Result:
Candidate fix, with all other relevant inputs unchanged:
Result:
Real Computer History writer
The automatic writer remains broken. A completed 10-minute segment ending at
2026-09-01T09:50:00Zcontained 270 events. The internal writer then failed with repeated Responses WebSocket 401 errors, and~/.codex/memories/extensions/skysight/resourcesstill contained zero summary files.This independently confirms that the issue persists in
26.825.51511, and that explicitly forwarding the App-resolved credential-store mode fixes the bundled CLI invocation while the internal Computer History caller remains unchanged.No credential values or raw Computer History event contents were inspected or included.
josh-oai commentedon Sep 14, 2026
Thanks for the detailed reproduction. We’ve merged a fix for the Keychain authentication issue preventing Computer History summaries from starting, planned for an upcoming update. Please let us know if summaries still fail after updating.
KennyNoDie commentedon Sep 15, 2026
Retested after updating, including a full application restart.
Environment
26.908.70816(build9275)codex-cli 0.154.0-alpha.6.21.0.1001067/26.913.1001067codex login status:Logged in using ChatGPTcli_auth_credentials_store = "auto"is configured~/.codex/auth.jsonis still absent; authentication remains Keychain-onlyThis is newer than the previously reported
26.825.51511/0.151.0-alpha.7.2reproduction.Recording still works; summaries still do not start
After enabling Computer History on the updated build, a complete 10-minute window closed normally:
2026-09-15T04:00:00Z–04:10:00ZeventCount=113The internal writer then failed at
12:10:34+0800(~26.7s) with:~/.codex/memories/extensions/skysight/resourcesremained empty.A later complete window (
04:10:00Z–04:20:00Z, 193 events) failed the same way.Restart did not change the result
ChatGPT was fully quit and reopened at
14:23:03+0800. The helper was a new process. Recording resumed. The first complete post-restart window:2026-09-15T06:20:00Z–06:30:00ZeventCount=287failed at
14:30:28+0800from the new helper process, again after ~27.0s, with the samecodex exec failed ... HTTP 401 Unauthorizedonwss://api.openai.com/v1/responses.resources/is still empty after that window.Bundled CLI A/B on this updated build
Control, empty temporary directory, bundled CLI:
Result:
Candidate fix, all other relevant inputs unchanged:
Result:
The helper binary still contains
--ignore-user-configand does not containcli_auth_credentials_store.This update and a subsequent restart therefore do not yet include the caller-side fix. No credential values or raw Computer History event contents were inspected or included.