insights 0.6.6

What changes for you

• The page calls the instruction file AGENTS.md (or CLAUDE.md) where it said "rule file". .claude/rules/ keeps its own name. Wording only: no number or level changes (#28).
• The browser's find works across tabs. The text in every tab and panel is in the page, hidden until found, and a "Show everything" control opens it all (#30).
• A broken count no longer blanks a tab. A count that is not a number makes its own row read "couldn't check", and the rest of the tab shows (#29).
• Advice on your AGENTS.md, for your busiest projects, as cards on the Review tab (#31): a pointer to a file that does not exist, an AGENTS.md that never loads, no command to check the work, lines that restate the repo, and your weakest line with a rewrite. An AGENTS.md that Claude Code never reads is no longer counted as loaded.

Known limits

• A full scan still takes 20 minutes or more on a laptop with many sessions, and some of its checks run out of time on a first run. This is next.
• The weakest-line card needs the AI reader and shows "couldn't check" when its quote cannot be verified against your file.
• Advice never changes a level.
• One run's AI part costs about 350k tokens on a large setup, against a target of about 200k.
• Folders without git are not read as projects.
• Folders of generated files that each hold a rule file still count as separate rule files.
• Share still sends the older read's lines until the receiving side accepts the new fields.
• Every threshold is a guess until it is checked against levels confirmed on calls.

Pull requests: #27, #28, #29, #30, #31. (#27 changes no behaviour.)

What's Changed

• insights: the date is read in one place by @teocns in #27
• insights: the instruction file is called AGENTS.md by @teocns in #28
• insights: a count that is not a number no longer breaks a tab by @teocns in #29
• insights: the browser's find works across tabs by @teocns in #30
• insights: advice on your AGENTS.md by @teocns in #31
• Release insights 0.6.6 by @teocns in #32

Full Changelog: insights--v0.6.5...insights--v0.6.6

insights 0.6.5

What changes for you

• Repos are found from where a session worked, not only from the folder it started in: files edited there, cd or git -C there, commits there. A repo that was only read from does not count. When most of your sessions start outside a repo, the page says so, and a zero reads "couldn't see", not "didn't build" (#14).
• Every tab shows the same number for the same thing (#15).
  • Programs that start the agent show as scripts, SDK programs and schedules on both tabs.
  • Tools from others give two counts: skills from others, and MCP servers.
  • Evals read "changed in 30 days, no run recorded" when nothing ran.
  • Every project read is listed. Level 3 reads "0 fully shown, 3 partly", and each part names its project.
  • Plurals are fixed, quotes are whole sentences, and schedules count only when they are your own.
  • The Share tab marks every line that comes from the older read.
• One rule-file count, and the same projects on every run (#16). Copies of rule files that Claude Code keeps in its plugin cache are not counted as yours. Two projects with the same activity are always picked in the same order.
• The scan can no longer run forever on an odd path (#19). A session that had written to a path starting with // could make the scan walk up for ever. It now stops, and odd targets (//…, /dev/null, shell variables, URLs) are not treated as folders. One session log that takes more than 5 minutes is skipped and reported as "couldn't check"; the rest of the run carries on.
• A skill you type as /name counts as used (#11). A skill that is yours stays yours when teammates edited some of its files, and the code it names outside its folder is read.
• Sessions in the desktop app or an editor count toward your busiest projects, not only terminal sessions (#13).
• The scan reads MCP servers built into your own repos (#18). They are not counted toward a level yet.

Known limits

• A full scan takes 20 minutes or more on a laptop with many sessions, and many of its checks run out of time on a first run. A fix is in progress for the next version.
• One run's AI part costs about 350k tokens on a large setup, against a target of about 200k.
• Folders without git are not read as projects.
• Folders of generated files that each hold a rule file still count as separate rule files.
• Share still sends the older read's lines until the receiving side accepts the new fields.
• Every threshold is a guess until it is checked against levels confirmed on calls.

Pull requests: #11, #13, #14, #15, #16, #17, #18, #19. (#17 changes no behaviour.)

What's Changed

• insights 0.3.1: VS Code chats count; page opens on Ubuntu by @teocns in #1
• insights 0.4.0: level decided by AI readers, the v14.2 way by @teocns in #2
• insights 0.5.0: no levels, no scores by @teocns in #3
• gateway: claude --prefer X falls back to rotation, never to claude's own login by @teocns in #4
• insights 0.6.0: the v15 setup tool by @teocns in #6
• gateway: the macOS service runs with an app's priority by @teocns in #5
• insights 0.6.1: the page opens on Your level by @teocns in #7
• insights 0.6.2: the scan's cache cleans up after itself by @teocns in #8
• insights 0.6.3: a level is read from what your files show by @teocns in #9
• insights 0.6.4: "What you've built" redesigned by @teocns in #10
• insights: repos are found from where a session worked by @teocns in #14
• insights: every tab prints the same number for the same thing by @teocns in #15
• insights: one rule-file count; the same projects on every run by @teocns in #16
• insights: no shipped line points at a person by @teocns in #17
• insights: the scan cannot loop forever on an odd path by @teocns in #19
• insights: a skill typed as /name counts as used by @AlexJov996 in #11
• insights: the most active project is read from every session the person typed in, in any app, not only the terminal by @AlexJov996 in #13
• Contributing guide, pull request template and a changelog for insights by @teocns in #22
• A tag and a GitHub Release are created when a plugin's version changes by @teocns in #23
• insights: the scan reads MCP servers built into a person's own repos by @AlexJov996 in #18
• Release insights 0.6.5 by @teocns in #26

New Contributors

• @teocns made their first contribution in #1
• @AlexJov996 made their first contribution in #11

Full Changelog: https://github.com/proxify-dev/agentic-kit/commits/insights--v0.6.5