chore(release): prepare v4.0.0 stability release - #46
Conversation
- Settle merged task notifications at native terminal boundaries without blocking new prompts. - Recover original working directories and consumed human prompts from native transcripts. - Preserve continuation order and omit empty process disclosures.
- Match live and historical compactions by native receipts and completed boundary identities. - Heal cached duplicates while retaining canonical details and distinct executions.
- Yield streamed scroll corrections to touch and scrollbar ownership, including keyboard resize. - Suppress immediate native paste replays without blocking intentional repeat pastes. - Cover conversation, new-chat, and side-chat composers with regression tests.
- Merge session fixes for Claude background continuations, history and compaction. - Include iOS scrolling and keyboard paste corrections. - Keep browser, Electron and MCP computer-use feature development outside dev.
- Align backend, Web and bootstrap versions while retaining wire protocol v72. - Document release scope, stability fixes and coordinated upgrade requirements. - Publish versioned bilingual release notes when the release workflow creates a draft.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 6db1ebd5f1
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| limit = min(_CWD_RECORD_BYTES, remaining) | ||
| line = stream.readline(limit) |
There was a problem hiding this comment.
Raise the per-record limit or extract cwd without discarding oversized rows. An allowed image above roughly 3 MiB expands past this 4 MiB JSONL cap after base64 encoding (MAX_SINGLE_ATTACHMENT_BYTES is 6 MiB), so both the leading queue record and the actual user record containing cwd can be skipped. If the session has no later message in the original directory—or later messages reflect a changed directory—recover_session_cwd() returns no cwd and the otherwise valid session is rejected on resume.
AGENTS.md reference: AGENTS.md:L66-L68
Useful? React with 👍 / 👎.
- Size bounded transcript reads for expanded attachments and preceding queue copies. - Cover accepted image size and count limits across both account catalog paths.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 05972c766d
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| cwd = row.get("cwd") | ||
| if (isinstance(cwd, str) and os.path.isabs(cwd) | ||
| and "\x00" not in cwd | ||
| and _sanitize_path(cwd) == source.parent.name): |
There was a problem hiding this comment.
Handle cwd values whose sanitized form exceeds MAX_SANITIZED_LENGTH. Claude stores these transcripts in a prefix-plus-hash project directory (the same case _directory_candidates() handles via startswith(prefix)), so this exact equality can never succeed for a long cwd. If a leading image also causes the SDK lite read to omit cwd, recover_session_cwd() leaves the cwd empty and the otherwise resumable session is rejected. Validate the long-path bucket form while still returning the full cwd recorded in the transcript.
AGENTS.md reference: AGENTS.md:L66-L68
Useful? React with 👍 / 👎.
- Match native UTF-16 project keys alongside the Python SDK catalog keys. - Recover legacy hashes only from explicit native roots to reject later cwd collisions. - Cover native and legacy long paths, malformed buckets and source preservation.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9a91c6999d
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| ) | ||
| if parse_raw and isinstance(data, dict): | ||
| data = self._steers.annotate(data) | ||
| data = self._steers.annotate(data, managed_active=self._turn_active) |
There was a problem hiding this comment.
When a browser query sets _turn_active before the pump consumes an already-buffered autonomous assistant/message_start whose internal user echo was omitted, this passes managed_active=True, so the new _main_activity fallback refuses to establish a background boundary. Although that assistant is routed as background while the human echo is pending, its following unattributed Result then takes the self._turn_active managed branch and terminates the newly submitted browser turn before its exact human echo; the accepted prompt's later output is consequently attached to the wrong lifecycle. The persistent service makes the equivalent assumption in claude_service/server.py:201-203; background detection should be gated by an observed managed human input, not merely query submission.
AGENTS.md reference: AGENTS.md:L60-L65
Useful? React with 👍 / 👎.
- Classify autonomous activity by consumed native input instead of query submission. - Keep pre-echo background terminals and replay ownership separate from the accepted human turn. - Cover pending queries, steering, service recovery and command completion without model calls.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 15d171f048
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| if (!this.recent(receipt) || batch.errors.length | ||
| || (!batch.images.length && !batch.files.length)) return true; |
There was a problem hiding this comment.
When an iOS clipboard import contains both a valid attachment and an invalid or oversized one, pickFiles returns the successfully decoded attachment together with a nonempty errors array. This branch accepts every replay without recording or comparing those successful bytes, so a duplicate paste delivery that arrives after the first import completes but within the replay window appends the valid attachment twice. Keep the successful subset eligible for replay deduplication even when other items report errors; a later explicit paste gesture already clears the receipt.
Useful? React with 👍 / 👎.
- Retain successful attachment receipts when other clipboard items fail. - Apply replay checks to main, side-chat and new-session composers. - Cover partial imports, failed retries, deliberate repeats and replay expiry.
|
LGTM. no P1. Thanks. |
3 participants
Claude background task returns could keep a finished session running, move replies into the wrong timeline, or hide consumed queued prompts. This PR brings the remaining session stability fixes onto master and prepares product v4.0.0, using wire protocol v72.
Changes
The release notes cover accumulated changes since v3.0.0. Most of those features are already on master; this PR adds the remaining session stability fixes and the release metadata. DSH, Electron and managed-browser/MCP computer-use experiments remain outside this release.
Validation
fzfexecutable.git diff --checkpassed.ghscenarios: curated notes, generated-note fallback, an existing draft, and refusal to overwrite a published release.Release sequence
This PR does not create a release tag, publish artifacts, merge itself, or change a running deployment.