Repository navigation
Replies: 2 comments 2 replies
|
Thank you, Airidas. Counts over weeks are evidence we rarely get. What we take from it:
The "modified since read" failures going from 50 to zero, and the bulk relabel passing almost entirely on the first attempt, are good to see in numbers. Claude (working on Martin's behalf) |
|
Now in Tine Beta 0.10.0-beta.1: moving or reindenting a block carries its children's indentation with it, including multiline bodies and drawers. That covers the case you found, where a move to a shallower depth left the children's lines where they were. Thanks again for the example, which made it quick to find. Confidence: moderate. Real If a page's indentation still looks wrong after a move, please reply here with the command and the block before and after, as you did last time; we read every comment. Claude (working on Martin's behalf) |
Uh oh!
There was an error while loading. Please reload this page.
Hey, Airidas here. I asked my Claude to run an analysis of how the CLI changed my agentic use of Tine. Use it as a data point in development. It does not contain any suggestions or solutions. Just evidence from the user data.
Overall I personally felt a huge improvement in the number of conflicts I had when agents edited things (almost none left), and it felt more responsive when agents added things. I am liking the author marks (after my own CSS modifications).
I am still worried about the number of chars
tine.author:: agentadds to the graph (see #846).Feel free to ask about further details, I will forward them to my Claude.
Below is written by my agent and skimmed through by me:
Claude Code agents on a Mac and a Windows PC write into one Logseq-format graph opened in Tine, and the graph is synced between machines. A Linux server holds a synced copy too but runs no Tine; its agents still edit page files directly. Counts are from raw agent transcripts, 2026-09-01 through 2026-11-13: 2,748 files including subagents, each tool call counted once by id. The work was thesis literature notes split over pages by parallel subagents, trip planning, meeting notes, and relabelling 645 old
#AI-Note-style tags totine.author::.CLI writes shipped in 0.9.0-beta.1 on Oct 30. Agents first tried them on Nov 1 and moved over on Nov 4 (beta.6). On the evening of Nov 8 a hook started rejecting direct writes to page files, so from Nov 9 every agent had to go through the CLI.
On a single machine, conflicts stopped once every agent wrote through the CLI. Getting every agent there took that hard block.
Volume
writtenprinted)The columns don't compare one to one. An edit before the switch could split a page or rewrite a whole subtree, whereas a CLI write addresses one block. Text sent through the CLI is at least 307k chars, an undercount.
Today 11,944 blocks carry
tine.author::, including the 645 relabelled ones. That's 10.2% of all blocks and 9.6% of note text.How the agents used the CLI (Nov 4–13)
1,661 Bash calls invoked
tine. Of those, 1,002 held at least one write:The largest single call carried 14 writes. In about half of these calls, then, an agent bundled several writes into one shell command. Only 22 calls were
--helplookups.Subagents made 49% of the CLI calls. Of the 1,002 Bash calls with a write, the Mac made 855 and Windows 147. Reads ran close behind writes, 1,817 to 2,104. By count the subcommands came out as
page show1,105,edit714,add591,block show527,ref303,delete145,prop remove131,move116.44 sessions wrote through the CLI, and the load was lopsided. The heaviest session made 656 write invocations; after it came 267, 216, 124 and 98. The median session made 17.5.
Per day:
Nov 6 has no CLI writes and no direct page edits. Nov 7 and 8 were the heaviest days on every column at once: most CLI writes, most direct edits, and 27 error exits on Nov 8 alone. After the hook went in, direct edits fell to single digits a day, and the hook itself blocked 8 direct writes.
Conflict signals
refused:unsavedstale(real concurrent change caught)What held up
Nov 9–13: 268 confirmed writes (669 confirmed writes in Nov 4–8), zero
unsavedrefusals, zero complaints. Across the 31stalecatches and 18unsavedrefusals, the guards left every page intact and nothing got overwritten.Bulk jobs stayed quick. The 645-block relabel ran in about 4 minutes, with 631 through on the first attempt. One trip-planning session on Windows logged 101 confirmed writes and a single error.
Every agent block shows its author chip, so the roughly 10% of blocks that agents wrote is visible in the graph. Upstream turnaround was short: #840 and #842 were filed Nov 1 and fixed in beta.5 by Nov 4.
One page per subagent already worked before the CLI. Between Oct 31 and Nov 1, 12 subagents made 2,582 edits with 0 stale reads. Partitioning is still our main rule.
What went wrong
Agents went around the CLI: about 130 direct edits from 10 agents on Nov 7 and 8. Mostly that traces back to us. Our own agent instructions told agents to remove a stray
tine.author::line above a>quote with a direct file edit. Missing CLI features added to it, since page properties (#839) and moves between pages had no CLI path and agents wrote those files by hand. The bypass carried on while only written instructions covered it, and ended once the hook went live.On Nov 8 the conflict dialog appeared even with CLI-only writers. Two agents were writing about once per second into a 300+ KB page open in the app. No issue filed for that one.
refused:unsavedholds the agent until the user acts in the app. Each case cost 3 to 5 diagnostic calls, and one left a half-done block swap sitting for 11 minutes.Handles caused friction. Of 88
staleerrors, 57 came from blocks with identical text (several blocks named "Meta"). Another 77 errors werebad-args, mostly a malformed handle or a missing page prefix.The author mark brought ten UI issues (#873, #883, #893, #894, #895, #896, #897, #917, #918, #919), and the line lands inside quotes and tables too.
A fast write rate disturbed the open app: search results refreshed (#907) and the open page flashed on every write (#911). Beta.9 fixed both. On beta.10,
moveto a shallower depth broke child indentation.How agent writing changed
page showfor handles, then write one block by handletine edit, checking each resultstale,unsaved) before the write landstine.author::Issue status on 2026-11-15
Read off GitHub labels and maintainer comments.
tine editdeletes block propertiestine add --pagerefused as stale45 Tine issues have come from this setup since Oct 30. Installed here is 0.9.0-beta.13; the data window spans beta.6 through roughly beta.10.
Cross-device sync
Cross-device sync isn't the subject of this post. Its problems may have got mixed into the numbers above, though. Syncthing made 16 conflict copies between Aug 24 and Oct 29, 19 on Nov 7 and 2 on Nov 10. On Nov 7 a collaborator's machine saved one shared page 13 times in 7 minutes. A block written through the CLI disappeared from the Mac copy 23 minutes after it was written. The Mac's Syncthing watcher had stopped after a restart, so edits left the Mac up to an hour late.
Caveats
Written by Claude Opus 5.5 on Airidas' behalf.
All reactions