textlog 0.1.9

macOS clipboard + OCR daemon exposed to Claude Code as an MCP server
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
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184

# Changelog

All notable changes to textlog are documented here.

The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
and the project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).

## [Unreleased]

## [0.1.9] - 2026-05-06

### Changed

- **MCP server `instructions` rewritten as recipe + veto.** Old text
  only described the response shape, so Claude Code often shelled out
  to `tl --help` + `tl logs` to "discover" features it already had
  via MCP — wasting ~500 tokens per session before the first useful
  call. New instructions explicitly list call recipes (`latest →
  textlog__get_recent(n=1)`, `today → textlog__list_today`, etc.) and
  veto the CLI path for reads. Adds ~120 tokens once at handshake,
  recovers ~5× on every clipboard lookup. No schema or behavior
  changes — restart Claude Code to pick up the new handshake string.

## [0.1.8] - 2026-05-05

### Added

- **Body pagination on `textlog__get_capture`.** A 49 KB capture
  overflowed the MCP per-tool token budget when fetched in one shot.
  `get_capture` now accepts optional `offset` (default 0) and `limit`
  (default 8000 chars, capped at 32000) and returns `text_offset`,
  `text_total_chars`, and `truncated` so callers can page through any
  size body. Bodies smaller than the default window return in one
  shot exactly as before.

## [0.1.7] - 2026-05-04

### Changed

- **Breaking MCP shape.** `CaptureSummary.text` is now `text_preview` —
  the first 200 characters of the capture body — accompanied by a new
  `truncated: bool` flag. List-style responses (`textlog__get_recent`,
  `textlog__list_today`, `textlog__search`) no longer dump full
  clipboard bodies into Claude's context. A 5-row `get_recent` that
  used to spend ~10 k tokens on large copies now returns ~250 tokens.
- The `id` field on every summary doubles as a handle for the new
  full-body fetch.

### Added

- `textlog__get_capture(id)` — returns the full untruncated body for a
  single capture. Use this when `text_preview` was marked
  `truncated: true`. Errors with `INVALID_PARAMS` if the row was
  trimmed by the ring buffer or the id never existed.
- `Storage::get_by_id(id)` on the SQLite layer powering the new tool.
- Server-instructions string updated so Claude is told to expand
  truncated previews via `get_capture` instead of guessing at body
  length.

### Why

Before this release the MCP server eagerly serialized the entire
clipboard body for every recent capture, so a single large copy could
blow past 10 k tokens on a default `n=5` request. The fix is at the
response shape, not the on-disk layout — splitting the daily Markdown
file into per-capture chunks would have moved bytes around without
changing what MCP returns. Truncating at the wire boundary plus a
lazy-fetch tool gives the same coverage at ~10× lower token cost.

## [0.1.6] - 2026-04-19

### Changed

- **Breaking default.** `notifications.copy_log_path_on_complete` now
  defaults to `false`. Previously the daemon wrote the daily-MD path
  back to the clipboard after every capture, which cascaded in
  clipboard managers and surprised users who only expected textlog to
  *read* the clipboard. Claude already discovers the daily-MD via
  `md_path` in MCP responses since v0.1.1, so the write-back hasn't
  been necessary for a while.
- Existing configs are not rewritten. If you have
  `copy_log_path_on_complete = true` in `~/textlog/config.toml` and
  want the new behavior, set it to `false` (or delete the line).

## [0.1.5] - 2026-04-19

### Performance

- Move the `NSPasteboard.changeCount` check out of `spawn_blocking` —
  it's a microsecond i64 property read, not a blocking call. The
  blocking-pool handoff was paying a task alloc + context switch every
  tick for nothing.
- Only enter `spawn_blocking` when the counter actually advances and
  string / PNG content needs to be read.
- Exponential idle backoff: active rate = `poll_interval_ms`, doubles
  to a 2 s ceiling after 20 unchanged ticks; any real change snaps
  back to active.
- Default `poll_interval_ms` raised from 250 to 500 ms (matches
  Maccy). Idle wakeups drop by roughly an order of magnitude.

### Added

- `tl perf` command: samples the running daemon's CPU% and RSS via
  `ps` and reports min/avg/max, with config-driven context (poll
  interval, backoff ceiling) and a verdict line. Flags:
  `--duration <secs>` (default 10), `--interval-ms <ms>` (default
  1000).

### Docs

- README notes the launchd respawn-throttle workaround after a binary
  upgrade (`tl uninstall && tl install` clears the throttle).

## [0.1.4] - 2026-04-17

### Fixed

- MCP stdio deadlock: no longer hold `std::io::stdout().lock()`
  across `run_mcp` dispatch. The rmcp stdio transport writes from a
  `spawn_blocking` worker thread that needs the same reentrant lock,
  which caused `initialize` to hang silently. Each subcommand now
  locks stdout on its own.

## [0.1.3] - 2026-04-17

### Added

- Lowercase `-v` short flag alongside `-V` / `--version` / `tl
  version` — all four paths resolve to the same handler.
- `tl update` command: self-updater that shells out to `cargo install
  textlog --force` and prints the `tl uninstall && tl install` hint
  to re-bootstrap the LaunchAgent.

## [0.1.2] - 2026-04-17

### Added

- Storage benchmark (`bench_storage_at_scale`): inserts 10k rows,
  measures FTS5 search latency and daily MD archive size to set a
  performance baseline.

### Docs

- README expanded with real scenarios, full MCP tool reference,
  tuning recipes, and tilde-path policy.

## [0.1.1] - 2026-04-17

### Added

- `md_path` field on every `CaptureSummary` so Claude can cite the
  daily MD archive directly. Makes the daily-archive paste trick work
  without `notifications.copy_log_path_on_complete` — recommended for
  users whose clipboard manager (Raycast, Maccy, Paste, Alfred)
  cascades on the path-back-to-clipboard write.

## [0.1.0] - 2026-04-17

### Added

- Initial public release on crates.io.
- NSPasteboard polling pipeline with privacy filter and per-capture
  SHA-256 dedup.
- Apple Vision OCR (`VNRecognizeTextRequest`) for clipboard images
  and ad-hoc files, with `accurate` / `fast` recognition levels and
  configurable language + confidence thresholds.
- SQLite FTS5 ring buffer (bounded query index) plus a permanent
  daily Markdown archive under `~/textlog/logs`.
- MCP stdio server exposing six tools: `textlog__get_recent`,
  `textlog__list_today`, `textlog__search`, `textlog__ocr_latest`,
  `textlog__ocr_image`, `textlog__clear_since`.
- `tl doctor` with eight health checks: config file, log dir, SQLite
  + FTS5, pasteboard access, notifications, LaunchAgent, MCP
  registration, Vision smoke test.
- LaunchAgent lifecycle: `tl install` / `uninstall` / `start` /
  `stop` / `status`.

[Unreleased]: https://github.com/xicv/textlog/compare/v0.1.5...HEAD
[0.1.5]: https://github.com/xicv/textlog/releases/tag/v0.1.5
[0.1.4]: https://github.com/xicv/textlog/releases/tag/v0.1.4
[0.1.3]: https://github.com/xicv/textlog/releases/tag/v0.1.3
[0.1.2]: https://github.com/xicv/textlog/releases/tag/v0.1.2
[0.1.1]: https://github.com/xicv/textlog/releases/tag/v0.1.1
[0.1.0]: https://github.com/xicv/textlog/releases/tag/v0.1.0