[anirudhkumar-nanonets]

Issue: a teammate who opens Trail-Code in Claude Code without trail is never shown the team's learnings, and 14 more learnings sat uncommitted on one laptop.
Why: .gitignore kept .claude/settings.json out of git and .trail/hook.cjs was never committed, so nothing runs at session start for someone without trail.
Fix: commit trail's shared setup (the hook, its one SessionStart entry, .trail's README and ignore file) plus the 6 learnings safe for a public repo, and drop the graft helpers trail init already deletes.

Moves: learnings_committed with first_look = 'false', last 7 days (Metabase db 11, posthog_event_full_view): 4,284 events from 551 installs on 29 Oct (22–29 Oct). Each Trail-Code clone with trail fires one more after pulling this (7 more learnings: 28 on main on 3 Nov, 35 with this PR).

What's in it

• .claude/settings.json (new, tracked): one SessionStart entry for .trail/hook.cjs. Byte-identical to mergeTrailShared({}) on main, so a teammate's trail init writes no diff.
• .trail/hook.cjs, .trail/README.md, .trail/.gitignore: generated from origin/main (hooksShim(), REPO_TRAIL_README). Regenerated on 2 Nov after Install trail without sudo #734 added the no-sudo line to the shim, so the hook matches what trail init writes on main today. A later PR that changes one of these generators must commit the regenerated file with it.
• .gitignore: stops ignoring .claude/settings.json; settings.local.json stays personal. The old reason ("regenerated per machine") was true of the pre-0.23 layout, where settings.json held the status line and every hook. Since 0.23 it holds only the shared entry.
• Removed .claude/helpers/graft-hooks.cjs, graft-statusline.cjs, .claude/skills/graft/SKILL.md: the pre-0.23 layout. trail init deletes them (clearOldLayout), so everyone who ran it had three deletions leaking into git status. Nothing in src, tests or CI reads these copies.
• 6 learnings: overlap warnings on squash-merged branches; the teammate offer needing committed hook wiring; release path and CodeQL; testing the sign-in nudge headless; the TRAIL_CLOUD flag and what uploads; the publish check failing on conflicts. Plus one new learning from this PR: what's tracked now, and how to check the no-trail session start.

Left out, on purpose

• 8 learnings, for a public repo: 4 about the private assign repo and its rollout, the 16 Oct user-test results, one with internal usage numbers and a detail about a private repo's files, one with adoption rates by version, and "Trail is decoupled from assign; no web app yet", a standing learning that's out of date since Trail Cloud (Sign in from your coding agent with a one-time code; release 0.27.0 #706, Connect to Trail Cloud at the moments trail is used #708) and would tell every teammate's Claude not to add sign-in links. They stay untracked in the main checkout; they belong in trail note --personal.
• AGENTS.md, GEMINI.md, .github/copilot-instructions.md: the same 7 KB graft:brain house-rules block from the old brain feature. No current command writes them, links point at the old NanoNets/Graft name, and Claude Code doesn't read them. Learnings carry this now.
• data/.web-token, data/share-link.txt: not part of trail's setup; not read.

Checked

• Real claude -p runs in fresh clones, no trail on PATH, empty TRAIL_HOME, --setting-sources project: on main no SessionStart hook runs and Claude says no learnings are in its context; on this branch the hook hands it the 12 newest learnings and Claude lists their paths.
• trail check (2 Nov, after merging main): no conflicts, 7 worth a look. The Oct 15 one ("every trail init also edits the tracked .claude/settings.json… and those edits leak into unrelated PRs") was about the old layout: the committed file is exactly what init writes now. The rest are about how to run sandboxes or one-off Claude runs, or are this PR's own new learning.
• npm test with a scratch HOME (2 Nov, after merging main): CI passes on Ubuntu and Windows. Locally, the full run on a heavily loaded laptop had hook timeouts. Re-running the failing files alone left 3 failures (graft-to-trail, hosts-trail, upkeep-hooks), and those same 3 fail on clean origin/main on this machine.

Watch out after merge

git overwrites an ignored file when a pull brings a tracked one to the same path (tried in a scratch repo). On trail ≥ 0.23 a personal .claude/settings.json already holds exactly this entry, so nothing changes. Anyone who kept their own hooks there should move them to settings.local.json first.

Follow-ups

• Move the 8 left-out learnings to trail note --personal in the main checkout, and replace the stale "decoupled from assign" one.
• Re-save the init_completed telemetry-timing finding without the adoption numbers, so teammates get the technical part.

Heads-up: once this merges, a git pull replaces any personal (ignored) .claude/settings.json with the shared one. Personal settings belong in .claude/settings.local.json.

🤖 Generated with Claude Code

[trailhq-graft]

🌱 graft blast radius

Nothing outside this diff depends on it. 0 areas changed; no indexed dependents at depth 2.

⚠️ 3 deleted files (.claude/helpers/graft-hooks.cjs, .claude/helpers/graft-statusline.cjs, .claude/skills/graft/SKILL.md) — their dependents cannot be computed from a graph built at this commit, since the files are gone from it.
⚠️ 12 changed files not in the graph (.claude/settings.json, .gitignore, .trail/.gitignore, .trail/README.md, .trail/hook.cjs) — no parser claims the extension, or the index predates the file.

_{graft blast · origin/main...HEAD · depth 2 · 15 changed files}

Open the interactive graph → — click an area to see its dependent symbols at file:line.

[shhdwi]

The committed shared hook file, .claude/settings.json and .trail/README.md match byte for byte what trail init writes on main, so a teammate's init leaves no diff.

Before merging:

1. The three Oct 15 learnings all carry the same ten touches, which have nothing to do with what they say. Six of those paths don't exist (data/.web-token, data/share-link.txt, src/ask/jev-ask.ts, AGENTS.md, GEMINI.md, .github/copilot-instructions.md). With Flag learnings that name code that's gone #731 coming, they'd be flagged stale on every show. Please trim each one to the files it's really about.
2. These learnings name the private NanoNets/assign repo and an internal endpoint (/api/cli/handoffs/claim). This repo is public, so those details probably need to come out.
3. Once merged, a pull overwrites anyone's ignored personal .claude/settings.json. That's worth a line in the description.

[shhdwi]

Merged after one fix commit (e1426d4):

• The three Oct 15 learnings now list only the files each is about (overlap: src/notes/branches.ts, src/notes/edit-overlap.ts; install offer: src/claude/shim-template.ts, src/claude/init.ts; release path: .github/workflows/codeql.yml, scripts/pack-trail.mjs, src/notes/pr-reviews.ts). The six paths that don't exist are gone, so Flag learnings that name code that's gone #731 won't flag them stale.
• Took out the private repo's name and the internal claim endpoint, since this repo is public. The facts in each learning stay.
• Added a line to the description: once this lands, a pull replaces a personal (ignored) .claude/settings.json.

Checked every committed file for tokens, keys, internal URLs and personal data: nothing else found. trail check still reads all 33 learnings. CI's first ubuntu runs failed on a GitHub 504 downloading tree-sitter-cli; the rerun passed.