My Work puts what needs you first
My Work, the page that shows what your agents are doing, was rebuilt from a full UI review (#6210, #6304). What needs you sits at the top, long lists fold, and every page works on a phone, by keyboard and with a screen reader.
- What needs you comes first. On the home page the first request waiting on you is at the top, with the one orange button on the screen. Requests are one-line rows; Idle and Shipped fold to one row each, with their counts.
- Shorter pages. On a real install: home is about 2,600px tall instead of 5,138; the List opens with 5 rows instead of 195; My Rules is about 4,400px instead of over 17,000.
- Reply where you read. A request's drawer puts the reply box right under its progress line, in every state. While a reply is on its way, "Reply here" waits, so a message can never go to a second conversation as a second paid turn.
- A focus area's page reads top to bottom: what needs you, the asks not started yet, then the board. A failed save says which change failed and offers Try again on that screen only. Past focus shows the day an area was marked done.
- My Rules: one rule to decide at a time, with one orange Approve. On a phone, swipe to decide.
- Fits a phone. Every page is 390px wide on a 390px phone, with a bottom bar for My Work, Focus and My Rules.
- Keyboard and screen readers. Every control has a name; tooltips open from Tab; focus comes back to where you were after a dialog, the drawer, Dismiss or Try again; controls are 44px on touch.
- Theme before load. Your light or dark choice is drawn before the first frame, loading and error screens included.
- Plain errors. A failure is one sentence in your language, with the raw reason folded under Details. Nothing retries by itself.
- Your language and clock. Times show in your language from the recorded time (French reads a 24-hour clock), and a day from another year shows its year. New strings are translated into French, Korean, Turkish and Chinese; corrections from native speakers are welcome.
- Less motion. A live update moves at most two things, and nothing moves with reduced motion on.
Observation TV is removed
- The TV page and its read-only remote access are gone from the open-source worker (#6295):
/tv,tv.htmland theCLAUDE_MEM_TV_TOKENsetting. The classic viewer and My Work are unchanged. - If you watched TV from another machine: set
CLAUDE_MEM_LOCAL_API_TOKEN, which protects network access now, and deleteCLAUDE_MEM_TV_TOKEN. Until you do, a worker that still finds the TV token with a non-loopbackCLAUDE_MEM_WORKER_HOSTand no local API token binds to127.0.0.1(::1for an IPv6 host) and logs aSECURITYerror, so the full API is never open to the network without a token. - Framing protection now covers every page the worker serves.
Already shipped in 15.20.0 and 15.21.0
The rest of the same review reached you in 15.20.0:
- Setup: a step bar, Back on every screen, plain errors, and Stop import (#6233).
- Classic viewer: one header row, a searchable project picker, "Find in this session", and no React error 185 at the end of a long timeline (#6229).
- Import from Mem0 and Supermemory: acceptance fixes (#6206).
- Fixes: Codex Desktop sessions are read live on macOS (#6285, thanks @MaxWang007), and the worker starts when
settings.jsonis not valid JSON (#6259).
15.21.0 followed: Code Mode starts with claude-mem enabled (#6278), auto-clear resumes with a continuation prompt (#6294), Hermes Desktop and GitHub Copilot in VS Code connect over MCP (#6284, #6286), and setup finishes when your prompt history is too large for one upload (#6280).
Full notes: 15.20.0 and 15.21.0.
Also new
- The SessionStart banner is a card in Claude Code terminals (#6274). On Claude Code 2.1.326 or later, the Rules + Tools line, "Previously" and the timeline are drawn as one framed card. A session mirrored to the Claude app shows one line instead of the hook's name in front of every line of the banner.
Fixes
- The worker keeps answering while it works through a backlog. Health checks, hooks and the viewer no longer wait behind:
- a hook spool backlog: the drain reads files in parallel batches of 64 (#6260, thanks @Zhong-fan) and removes and claims them without blocking (#6296, from #6254 by @Afloat16); reported in #6247 by @soarkit;
- the project list at worker start, which read every row and now reads the project indexes (#6270; reported in #6246 by @soarkit);
- the Claude CLI check that corpus prime and query run on a cache miss (#6256, thanks @Afloat16; reported in #6250 by @soarkit);
- Chroma sync's checkpoint, which was rewritten for every row (about 28 MB of writes for 2,000 rows) and is now written once per 250 ms (#6204, thanks @Zhong-fan; reported in #6203 by @soarkit).
- Folder CLAUDE.md files: updates fetch up to 8 folders at a time instead of one by one (#6264; reported in #6248 by @soarkit),
claude-mem generatereads a project's observations once instead of once per folder (#6265; reported in #6249 by @soarkit), and a failed folder update is logged with its folder and tried again by the next observation for that folder (#6208, thanks @0xABCD01). - Long Bash output keeps its end (#6292; reported in #6282 by @micronox). When Claude Code cut an output over about 30,000 characters and saved the full text to a file, the memory came from the start only, so the last lines (test summaries, final errors) were lost. claude-mem now reads the end from that file.
- Native memory notes: a note with
type:at the top of its frontmatter keeps its category (#6244), and a note's project is found when the transcript's first prompt is long (#6272), thanks @Afloat16. - Recent sessions include adopted worktrees (#6268, thanks @Afloat16). After a worktree was merged into its parent, OpenClaw's
claude-mem-recentshowed no previous sessions. - Search: Chinese and Japanese terms inside longer unspaced words are found in corpus evidence (#6213, thanks @liuyinchu).
smart_outline,smart_searchandsmart_unfoldfind JavaScript functions defined by assignment, such asexports.read = function …(#6245), and each Java method overload unfolds on its own (#6257), thanks @Afloat16. - Quieter error reports: a context read that passes the 120-second reader deadline names the slow read and logs a warning, not an error (#6255), and on Windows a spool entry that Windows keeps locked gets one warning instead of an error report on every drain (#6269).
- Two failures that were dropped are logged now (#6302): auto-clear says so in the transcript when it could not hand Claude its request to save the work state, and the server logs why
/v1/infocould not read the vector search status. - Turkish: clearer My Work wording and an up-to-date Turkish README (#6287, thanks @alkank).
For contributors
- The commit-message test that needs bash 4 is skipped on macOS's bash 3.2, and still runs on Linux CI (#6297).