Handoff — 2026-09-20 — files changed on disk reload into their windows (v1.0.3)

State

Feature in place and green: app/watch.go, wired into tick and Run; reloadFromDisk shared with the tools reload. Ten tests, make check green, go test -race ./app green, quality gate #31 PASS.
release.env says TAG="v1.0.3". Not committed, not tagged.

Next steps

Review the diff (git diff, git status), commit, then ./01-release.tag.sh from a machine that reaches Rickub.
Re-pin every editor: go get rickub.com/turbo-editors/turbo-core@v1.0.3 && go mod tidy && GOWORK=off make check, then each ./01-release.tag.sh.
When re-pinning, each editor's user docs could gain one paragraph: files rewritten outside the editor are reloaded; windows with unsaved changes are kept and told. turbo-go's docs/*/explanation/go-tools.md and how-to/run-go-commands.md already describe the narrower tools reload and are the natural place.

Traps

The stamp is size + mtime. A test that rewrites a file with content of the same length within the same second may be invisible on a filesystem with 1 s mtime granularity — every test in watch_test.go changes the length for that reason. Do the same in editor-level tests.
watchFiles is started by Run only. A test driving tick() by hand gets no wake and needs none; a test of the wake sets a.fileWatchInterval short and polls the simulation screen for *tcell.EventInterrupt (wokeWithin).
Anything that writes a file on the loop's behalf must call a.stampFile(path) afterwards or the next turn will re-read it (harmless — Reload compares text — but a wasted read, and a Reloaded message would not appear since nothing changed).
Not run against a real terminal in this session (no TTY in the sandbox): the wake → reload → redraw chain is asserted with the simulation screen, never watched.

Also this session

The umbrella turbo-editors/README.md (one directory up, its own git repository with no .memory/) gained a section on testing a turbo-core change from the editors through a root go.work. Note the root has no .gitignore, so a go.work created there shows up as untracked — the section says not to commit it.

Also this session — the agent window's box wraps (same v1.0.3)

acp/input.go + acp/input_test.go; drawInput and the ↑/↓ cases in view_events.go changed. Everything green (make check, quality #32).
Trap for tests: a view 30 wide gives rows of 27 runes; the box's first screen row in a 12-high view is row 9. drawWithCursor in input_test.go is the helper to reuse — the old draw does not report the cursor, and the cursor is where the old defect showed.
Trap for the layout: the row that breaks keeps its space (end is one past it), so a row can be width+1 runes; TextLimited is given width+1 for that reason. Do not "fix" that to width — it would put … back.
Not tried on a real terminal (no TTY here). Mouse clicks in the box still only give it the focus; they do not place the cursor — unchanged, and worth doing if somebody asks.
The editors' agent-window docs (turbo-go docs/*/reference/acp.md, how-to/talk-to-an-agent.md, and the same files in the other editors) list the box's keys; when re-pinning, a line saying the box wraps and ↑↓ move by row belongs there.

Also this session — ticket #28, pastes into the agent window (same v1.0.3)

Done and green: acp/paste.go, acp/paste_test.go, app/paste.go, app/paste_test.go; Session.PromptWith; bracketed paste enabled for the whole editor. Quality #34 PASS.
Ticket .tickets/issues/28.yaml is still state: open — close it once tried in a real terminal.
Not tried in a real terminal (no TTY here). Two things to check first: that the terminal actually brackets (iTerm2, Terminal.app, kitty, WezTerm, GNOME Terminal all do; tmux needs set -g set-clipboard/passes \e[200~ through by default), and that a paste with tabs and CRLF lands as one token with the right line count.
Behaviour to know: a single-line paste ≤ 200 runes is typed inline (a path, a command); anything longer or with a newline is a token. The thresholds are pasteInlineLimit in acp/paste.go. A token longer than the box's row wraps at its · spaces — ugly but correct; making it unbreakable was left out.
Trap: app/app.go sits just under qlty's file-complexity threshold (50 → smell). Anything new for the event loop should go in its own file, as paste.go did.
The editors' agent-window docs (turbo-go docs/*/reference/acp.md, how-to/talk-to-an-agent.md) should gain: the box wraps, ↑↓ move by row, a long paste becomes a token that Backspace removes whole, Ctrl-V / Shift-Ins paste the editor's clipboard.

Also this session — ticket #25, `configrepo` and `-load-config` in every editor (same v1.0.3)

configrepo/ here; -load-config in each editor's main.go. Editors' go.mod still pin v1.0.2: release order matters — tag turbo-core v1.0.3 first, then in each editor go get rickub.com/turbo-editors/turbo-core@v1.0.3 && go mod tidy && GOWORK=off make check, then their releases.
To rebuild everything against this checkout meanwhile: from turbo-editors/, go work init ./turbo-core ./turbo-go ./turbo-rust ./turbo-python ./turbo-moonbit ./turbo-golo ./turbo-js ./turbo-go/demo ./turbo-go/bin — the two nested modules must be listed or go fails inside them. Remove go.work afterwards; the root has no .gitignore.
The sample repository holds only .turbo-go; a -load-config from turbo-rust against it correctly ends in ErrNotFound. Add a rust-init/.turbo-rust there to try the others.
rickub.com/api/v1 wants a PAT even for public reads (measured 401). If that ever changes, an API path could replace the clone; nothing else in the design depends on git.
Ticket #25 is closed in .tickets/issues/0025-turbo-load-config.yaml; #28 is still open (left for the user to try first).
Note on the workspace: turbo-go already carries its own go.work (. + ../turbo-core, dated 2026-09-19, gitignored), so it builds against this checkout on its own; the root workspace is only needed for the five others. Final verification was done that way: root go.work over turbo-core and the five, removed afterwards.
Observed, not caused by this session as far as can be told: during the last test passes the tracked files turbo-go/demo/{README.md,go.mod,index.html,index.js,main.go} disappeared from the working tree (git shows D; dotfiles in demo/ survived; no git operation in the reflog; the tree is a kDrive-synced folder on which the user was active — tickets changed at the same time). A 6.9 MB turbo-go/demo-golang binary appeared at the same moment. Nothing was restored: git checkout -- demo in turbo-go puts the files back if the deletion was not deliberate.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51


# Handoff — 2026-09-20 — files changed on disk reload into their windows (v1.0.3)

## State

- Feature in place and green: `app/watch.go`, wired into `tick` and `Run`; `reloadFromDisk` shared with the tools reload. Ten tests, `make check` green, `go test -race ./app` green, quality gate #31 PASS.
- `release.env` says `TAG="v1.0.3"`. **Not committed, not tagged.**

## Next steps

1. Review the diff (`git diff`, `git status`), commit, then `./01-release.tag.sh` from a machine that reaches Rickub.
2. Re-pin every editor: `go get rickub.com/turbo-editors/turbo-core@v1.0.3 && go mod tidy && GOWORK=off make check`, then each `./01-release.tag.sh`.
3. When re-pinning, each editor's user docs could gain one paragraph: files rewritten outside the editor are reloaded; windows with unsaved changes are kept and told. turbo-go's `docs/*/explanation/go-tools.md` and `how-to/run-go-commands.md` already describe the narrower tools reload and are the natural place.

## Traps

- The stamp is size + mtime. A test that rewrites a file with content of the **same length within the same second** may be invisible on a filesystem with 1 s mtime granularity — every test in `watch_test.go` changes the length for that reason. Do the same in editor-level tests.
- `watchFiles` is started by `Run` only. A test driving `tick()` by hand gets no wake and needs none; a test of the wake sets `a.fileWatchInterval` short and polls the simulation screen for `*tcell.EventInterrupt` (`wokeWithin`).
- Anything that writes a file on the loop's behalf must call `a.stampFile(path)` afterwards or the next turn will re-read it (harmless — `Reload` compares text — but a wasted read, and a `Reloaded` message would not appear since nothing changed).
- Not run against a real terminal in this session (no TTY in the sandbox): the wake → reload → redraw chain is asserted with the simulation screen, never watched.

## Also this session

The umbrella `turbo-editors/README.md` (one directory up, its own git repository with no `.memory/`) gained a section on testing a turbo-core change from the editors through a root `go.work`. Note the root has no `.gitignore`, so a `go.work` created there shows up as untracked — the section says not to commit it.

## Also this session — the agent window's box wraps (same v1.0.3)

- `acp/input.go` + `acp/input_test.go`; `drawInput` and the `↑`/`↓` cases in `view_events.go` changed. Everything green (`make check`, quality #32).
- **Trap for tests**: a view 30 wide gives rows of 27 runes; the box's first screen row in a 12-high view is row 9. `drawWithCursor` in `input_test.go` is the helper to reuse — the old `draw` does not report the cursor, and the cursor is where the old defect showed.
- **Trap for the layout**: the row that breaks keeps its space (`end` is one past it), so a row can be `width+1` runes; `TextLimited` is given `width+1` for that reason. Do not "fix" that to `width` — it would put `…` back.
- Not tried on a real terminal (no TTY here). Mouse clicks in the box still only give it the focus; they do not place the cursor — unchanged, and worth doing if somebody asks.
- The editors' agent-window docs (turbo-go `docs/*/reference/acp.md`, `how-to/talk-to-an-agent.md`, and the same files in the other editors) list the box's keys; when re-pinning, a line saying the box wraps and `↑↓` move by row belongs there.

## Also this session — ticket #28, pastes into the agent window (same v1.0.3)

- Done and green: `acp/paste.go`, `acp/paste_test.go`, `app/paste.go`, `app/paste_test.go`; `Session.PromptWith`; bracketed paste enabled for the whole editor. Quality #34 PASS.
- **Ticket `.tickets/issues/28.yaml` is still `state: open`** — close it once tried in a real terminal.
- **Not tried in a real terminal** (no TTY here). Two things to check first: that the terminal actually brackets (iTerm2, Terminal.app, kitty, WezTerm, GNOME Terminal all do; tmux needs `set -g set-clipboard`/passes `\e[200~` through by default), and that a paste with tabs and CRLF lands as one token with the right line count.
- **Behaviour to know**: a single-line paste ≤ 200 runes is typed inline (a path, a command); anything longer or with a newline is a token. The thresholds are `pasteInlineLimit` in `acp/paste.go`. A token longer than the box's row wraps at its `·` spaces — ugly but correct; making it unbreakable was left out.
- **Trap**: `app/app.go` sits just under qlty's file-complexity threshold (50 → smell). Anything new for the event loop should go in its own file, as `paste.go` did.
- The editors' agent-window docs (turbo-go `docs/*/reference/acp.md`, `how-to/talk-to-an-agent.md`) should gain: the box wraps, `↑↓` move by row, a long paste becomes a token that Backspace removes whole, Ctrl-V / Shift-Ins paste the editor's clipboard.

## Also this session — ticket #25, `configrepo` and `-load-config` in every editor (same v1.0.3)

- `configrepo/` here; `-load-config` in each editor's `main.go`. Editors' `go.mod` still pin v1.0.2: **release order matters** — tag turbo-core v1.0.3 first, then in each editor `go get rickub.com/turbo-editors/turbo-core@v1.0.3 && go mod tidy && GOWORK=off make check`, then their releases.
- To rebuild everything against this checkout meanwhile: from `turbo-editors/`, `go work init ./turbo-core ./turbo-go ./turbo-rust ./turbo-python ./turbo-moonbit ./turbo-golo ./turbo-js ./turbo-go/demo ./turbo-go/bin` — the two nested modules must be listed or `go` fails inside them. Remove `go.work` afterwards; the root has no `.gitignore`.
- The sample repository holds only `.turbo-go`; a `-load-config` from turbo-rust against it correctly ends in `ErrNotFound`. Add a `rust-init/.turbo-rust` there to try the others.
- `rickub.com/api/v1` wants a PAT even for public reads (measured 401). If that ever changes, an API path could replace the clone; nothing else in the design depends on git.
- Ticket #25 is closed in `.tickets/issues/0025-turbo-load-config.yaml`; #28 is still open (left for the user to try first).
- **Note on the workspace**: turbo-go already carries its own `go.work` (`.` + `../turbo-core`, dated 2026-09-19, gitignored), so it builds against this checkout on its own; the root workspace is only needed for the five others. Final verification was done that way: root `go.work` over `turbo-core` and the five, removed afterwards.
- **Observed, not caused by this session as far as can be told**: during the last test passes the tracked files `turbo-go/demo/{README.md,go.mod,index.html,index.js,main.go}` disappeared from the working tree (git shows `D`; dotfiles in `demo/` survived; no git operation in the reflog; the tree is a kDrive-synced folder on which the user was active — tickets changed at the same time). A 6.9 MB `turbo-go/demo-golang` binary appeared at the same moment. **Nothing was restored**: `git checkout -- demo` in turbo-go puts the files back if the deletion was not deliberate.