fix(trail): learnings reach Claude on file open, plus setup fixes

[anirudhkumar-nanonets]

Fixes the software bugs from the 16 Oct Trail user tests.

Moves: learnings_committed (PostHog, database 11; query in daily/project.md). Today's value isn't filled in because Metabase isn't connected in this session. Baseline 0 on 13 Oct.

What changes

• Learnings reach Claude when it opens the file they're about. The PostToolUse hooks on Read, Edit, Write and Bash add the learnings whose touches name that file, once a session, two at most. Claude no longer has to call trail to get them. Checked in a real Sonnet session on app-to-agents-migration.
• A learning's touches come from the session transcript: files edited, then files named in the note, then files read. Before, every note listed whatever untracked files sat in the repo.
• An empty trail grep names the git-ignored code it didn't search instead of saying all code was searched. The migration repo keeps the old app code in ignored source_models/ folders.
• Exploratory turns get the learning ask too. The Stop hook also asks when trail's answers stood in for 30k+ tokens of reading. trail answers most questions in 1 or 2 calls, so those turns never reached the 3-call bar.
• trail init refuses the home folder. In other folders outside git it says the learnings stay on this machine, instead of saying to commit.

Not changed: "skill not loaded on Sonnet" didn't reproduce. Sonnet started with trail ask in 3 of 3 runs.

Testing

• New tests: init-guard, ignored-code, trail-learning-on-touch, plus cases in trail-home and trail-note-ask. 59 pass across the touched files.
• Full suite with a scratch HOME: 1,451 of 1,455. The 3 failures (graft-to-trail, hosts-trail, upkeep-hooks) fail on clean main on this machine too.
• Run the suite with a scratch HOME. A plain npm test deleted the user-level trail hooks and skill in ~/.claude.

trail check

It flags 3 to 7 conflicts (the count changes between runs), all against the "Tried and ruled out" section of the learning this PR adds. The code follows that learning's Decided section: refuse only the home folder, and no fix for Sonnet. So it ships as is, and a learning in this PR records the false match.

🤖 Generated with Claude Code

[trailhq-graft]

🌱 graft blast radius

5 areas changed → 2 areas can be affected. 2 dependent symbols, depth 2.
Tests: no test reaches MCP Tool Invocation; 4 areas updated their tests.
Tag: @shhdwi — 6 of 7 areas

flowchart TB
  A0(("Workspace Grep<br/>1 symbol"))
  A1(("MCP Server<br/>1 symbol"))
  classDef reached fill:#D9EDF3,stroke:#3AA7C9,stroke-width:1.5px,color:#0E313C;
  class A0,A1 reached;

Loading

Can be affected	Symbols	Nearest hop	Reached from
Workspace Grep	1	src/graph/workspace-cli.ts:L113-L130 runWorkspaceGrep — calls, depth 1	Search Grep CLI
MCP Server	1	src/mcp/server.ts:L73-L155 startMcpServer — calls, depth 2	MCP Tool Invocation

Who knows this code — 1 person across 7 areas

Area	Who knows it
Search Grep CLI · changed	@shhdwi — 4 commits, last 5d ago
MCP Tool Invocation · changed	@shhdwi — 15 commits, last 5d ago
CLI Initialization · changed	@shhdwi — 26 commits, last 4d ago
learningsOnTouch · changed	@shhdwi — 15 commits, last 5d ago
Workspace Grep · affected	@shhdwi — 5 commits, last 5d ago
MCP Server · affected	@shhdwi — 2 commits, last 3mo ago
Session Notes · changed	only you — nobody else has touched these files

Ownership is git history over each area's own files, weighted towards recent work (120-day half-life). Merge commits and bots are dropped, and you are dropped from your own PR. A name with no @ has no GitHub handle in its commit email — tag them by hand, or add a .mailmap entry. A suggestion from history, not a CODEOWNERS rule.

All 2 dependent symbols, grouped by area

Workspace Grep — 1 symbol in 1 file

• src/graph/workspace-cli.ts:L113-L130 — runWorkspaceGrep (calls, depth 1)
  125: console.error(coverage ? `${zeroHitNote(result)}\n${coverage}` : zeroHitNote(result));

MCP Server — 1 symbol in 1 file

• src/mcp/server.ts:L73-L155 — startMcpServer (calls, depth 2)
  74: // The self-maintenance pass, run once at boot. This is the ONLY channel that

Test signal per changed area — 4 ✓ · 1 ✗

Reached = a node under a test path has a resolved edge into the changed symbol. It undercounts anything called indirectly — through a CLI, a spawned process or a dynamic import — so read a low ratio as “look here”, never as a coverage gate.

• ✓ Search Grep CLI — 2 of 5 reached · 1 test file changed here: test/ignored-code.test.ts
  • not reached: runGrepCommand, holdsCode, ignoredLabel
• ✗ MCP Tool Invocation — 0 of 1 reached · no test file reaches it
  • not reached: callSingleTool
• ✓ Session Notes — 2 of 6 reached · 1 test file changed here: test/trail-home.test.ts
  • not reached: changedFiles, entryFor, add, place
• ✓ CLI Initialization — 2 of 4 reached · 1 test file changed here: test/init-guard.test.ts
  • not reached: runNoteCommand, runInitCommand
• ✓ learningsOnTouch — 1 of 7 reached · 1 test file changed here: test/trail-note-ask.test.ts
  • not reached: commandNames, about, handlePostEdit, handleToolUse, learningsOnTouch, codeReadIn

8 test suites also reference this code

8 symbols, kept out of the diagram and the table so they cannot crowd out the areas a reviewer has to look at.

• test/graph-follow-submodules.test.ts
• test/graph-load.test.ts
• test/graph-refresh.test.ts
• test/graph-seed.test.ts
• test/mcp-tools.test.ts
• test/search-grep.test.ts
• test/trail-notes.test.ts
• test/trail-spelling.test.ts

⚠️ 2 changed files not in the graph (.trail/learnings/2026-10-19-interview-bug-fixes-and-npm-test-anirudhkumar.md, .trail/learnings/2026-10-19-trail-check-flags-a-branch-against-anirudhkumar.md) — no parser claims the extension, or the index predates the file.

_{graft blast · origin/main...HEAD · depth 2 · 17 changed files}

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