Skip to content

Commit 88f7204

Browse files
StanMarekStanMarek
committed
chore: release v0.11.0
1 parent b1239c8 commit 88f7204

4 files changed

Lines changed: 230 additions & 9 deletions

File tree

‎CHANGELOG.md‎

Lines changed: 54 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -7,6 +7,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
77

88
## [Unreleased]
99

10+
## [0.11.0] - 2026-05-03
11+
1012
### Added
1113

1214
- Restored AWS CLI completion spec. 409 service subcommands (`s3`,
@@ -40,6 +42,52 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
4042
of bundled specs; e.g. `git archive --format=` now suggests `tar`/`zip`,
4143
`tar --atime-preserve` suggests `replace`/`system`. See
4244
[ADR 0004](docs/adr/0004-static-arg-suggestions.md).
45+
- Async suggestion feedback in the popup. While script generators are
46+
in-flight, the popup reserves an indicator row that surfaces a spinner
47+
(Loading), a per-provider partial-error label (`PartialError`, gated by
48+
the new `suggest.show_provider_errors` config), or stays Idle. Hard
49+
errors and empty results now expire on a `feedback_dismiss_ms` window
50+
rather than dismissing the popup. New config keys: `suggest.spinner`,
51+
`suggest.show_provider_errors`, `suggest.feedback_dismiss_ms`.
52+
- OSC 7772 percent-encoded buffer framing between the shell and the
53+
proxy. Replaces the legacy OSC 7770 escape-sensitive framing (now
54+
logged once as deprecated) so embedded `\x1b\\` / `\x07` / NUL bytes
55+
in the command line buffer are reported losslessly. The shell hook
56+
emits OSC 7772 with a percent-encoded payload and an explicit cursor
57+
index. OSC 7770 stays parsed for one release for bash/fish parity and
58+
is scheduled for removal in v0.12.0. See
59+
[ADR 0003](docs/adr/0003-osc7772-buffer-framing.md).
60+
- Post-install next-steps summary. The installer now prints a structured
61+
five-step summary (`doctor`, demo command, `Ctrl+/`, `config edit`)
62+
plus the resolved config + spec paths. The headline branches between a
63+
green "installed" and a yellow "partially installed" line based on
64+
whether `~/.zshrc` was actually written; both paths render through
65+
`sanitize_path` so a hostile `$HOME` cannot smuggle terminal escapes.
66+
67+
### Fixed
68+
69+
- Multi-word alias expansion for spec resolution. Aliases are now
70+
tokenised with `shlex` and stored as `Vec<String>`, so
71+
`alias gco='git checkout'` resolves the git spec, walks the
72+
`checkout` subtree, and dispatches `git_branches` / `git_tags` on
73+
the next positional. Chained aliases (`alias gcb='gco -b'`) recurse
74+
through a 16-hop guard with cycle detection. The on-disk alias
75+
cache schema bumps to `format_version: 2` and rejects v1 caches on
76+
load. Flag-prefix completion (`gco -<TAB>`) and the async
77+
`suggest_dynamic` path now share the same alias-target spec walk
78+
so all three entry points stay consistent.
79+
- Terminal popup rendering robustness. Multiple races and edge cases
80+
hardened across the parser/overlay boundary: stale-screen-size
81+
recovery so the popup no longer pins to row 1 when CPR coordinates
82+
fall outside cached dimensions; cumulative `overlay_scroll_deficit`
83+
is reset on shell-side viewport scrolls and discarded outright when
84+
it grows past `cursor_row`; overlay write epoch ordering keeps
85+
ownership only after acknowledged writes; display-mutating CSI
86+
sequences mark the screen dirty so newer popup state survives stale
87+
render bytes losing races; bordered-popup `PartialError` demotion
88+
now repaints the displaced bottom border instead of leaving a
89+
stray `╰──╯`; `pending_failed` / `pending_empty_count` reset
90+
symmetrically on every Idle path.
4391

4492
### Changed
4593

@@ -50,6 +98,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
5098
size-reclaim work and would drop the binary back near the original
5199
ceiling — tracked as a separate spec, not bundled here so neither
52100
change becomes unreviewable.
101+
- ADRs are now tracked in git under `docs/adr/`. The first three —
102+
`0001-pty-proxy-vs-plugin.md`, `0002-vte-vs-vt100.md`, and
103+
`0003-osc7772-buffer-framing.md` — capture historical and current
104+
architectural decisions; `0004` and `0005` ship alongside the
105+
features they describe in this release.
53106

54107
## [0.10.0] - 2026-04-26
55108

@@ -585,6 +638,7 @@ silently changed behaviour.
585638
- **Shell integration** for zsh (full), bash (Ctrl+/), and fish (Ctrl+/)
586639
- **`validate-specs` subcommand** with colored output and item counts
587640

641+
[0.11.0]: https://github.com/StanMarek/ghost-complete/releases/tag/v0.11.0
588642
[0.10.0]: https://github.com/StanMarek/ghost-complete/releases/tag/v0.10.0
589643
[0.9.1]: https://github.com/StanMarek/ghost-complete/releases/tag/v0.9.1
590644
[0.9.0]: https://github.com/StanMarek/ghost-complete/releases/tag/v0.9.0

‎Cargo.lock‎

Lines changed: 8 additions & 8 deletions
Some generated files are not rendered by default. Learn more about customizing how changed files appear on GitHub.

‎Cargo.toml‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -12,7 +12,7 @@ members = [
1212
]
1313

1414
[workspace.package]
15-
version = "0.10.0"
15+
version = "0.11.0"
1616
edition = "2021"
1717
rust-version = "1.86"
1818
license = "MIT"

‎benchmarks/v0.11.0.md‎

Lines changed: 167 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,167 @@
1+
# Benchmarks — v0.11.0
2+
3+
Date: 2026-05-03
4+
Platform: macOS (Apple Silicon)
5+
6+
Criterion's reported `change` column compares against the previous saved local
7+
run, which on this machine is v0.10.0 stable. Deltas here measure the work that
8+
landed between v0.10.0 and v0.11.0: AWS spec restoration (#101), local-project
9+
providers (#97), OSC 7772 buffer framing (#94), async suggestion feedback (#98),
10+
multi-word alias expansion (#95), static `args.suggestions` (#96), popup page
11+
navigation (#99), and the popup rendering hardening pass (#100). Three new
12+
benchmark groups (`osc7772_decode`, `feedback_render`, plus `alias_hit_*` cases
13+
in `engine_suggest_sync` and `with_static_suggestions_tar` in `spec_resolution`)
14+
ship as new baselines.
15+
16+
## vt_parse_throughput
17+
18+
| Benchmark | Time | vs v0.10.0 |
19+
|-----------|------|------------|
20+
| plain_text | 197.36 µs | +10.6% |
21+
| ansi_colored | 246.11 µs | +10.5% |
22+
| cursor_heavy | 266.70 µs | −2.9% improvement |
23+
24+
The plain_text and ansi_colored regressions trace to the OSC dispatch fan-out
25+
in `gc-parser`: the OSC handler now matches OSC 7770 (legacy), 7771 (cursor),
26+
and 7772 (percent-encoded buffer) sequences, plus the per-frame deprecation
27+
warn for 7770. The plain_text input contains no OSCs, so the cost is purely
28+
the dispatch table grow; cursor_heavy is dominated by CUP/CUF/CUB sequences,
29+
which sit on a different code path and improve marginally as side benefit of
30+
the parser refactor.
31+
32+
## osc7772_decode (new in v0.11.0)
33+
34+
| Benchmark | Time | Notes |
35+
|-----------|------|-------|
36+
| 100 | 574.03 ns | 100 byte payload (typical command line) |
37+
| 1024 | 4.0154 µs | 1 KiB payload (long pasted line) |
38+
| 8192 | 28.623 µs | 8 KiB payload (worst-case heredoc body) |
39+
40+
Per-byte percent-decode cost lands at ~5.7 ns/byte, well under the keystroke
41+
budget. The shell-side encoder benchmark is left out — emission happens once
42+
per prompt redraw at human typing cadence, never on the hot path.
43+
44+
## feedback_render (new in v0.11.0)
45+
46+
| Benchmark | Time | Notes |
47+
|-----------|------|-------|
48+
| indicator_row_width_60 | 121.96 ns | Steady-state spinner repaint |
49+
| indicator_row_width_60_varying_frame | 123.15 ns | Frame-varying input (defeats const-folding) |
50+
51+
The async-feedback indicator row repaints at ≤16 fps in the worst case
52+
(spinner cadence). Sub-microsecond per-frame cost means the feature is free
53+
on the render path.
54+
55+
## fuzzy_ranking
56+
57+
| Benchmark | Time | vs v0.10.0 |
58+
|-----------|------|------------|
59+
| 1k_3char | 101.44 µs | +1.5% (flat) |
60+
| 10k_3char | 1.0743 ms | +4.5% within noise |
61+
| 10k_empty | 350.17 µs | +2.0% (flat) |
62+
63+
Within run-to-run noise. Nucleo's hot path is unchanged; the small drift is
64+
attributable to the `EnumValue` `SuggestionKind` variant added for static
65+
`args.suggestions` priority routing.
66+
67+
## spec_loading
68+
69+
| Benchmark | Time | vs v0.10.0 |
70+
|-----------|------|------------|
71+
| load_717_specs | 288.14 ms | +241% expected |
72+
73+
**Expected.** This is the AWS spec restoration showing up at startup-load
74+
time. The bench loads the embedded spec corpus once, and the corpus grew
75+
from ~47 MB minified pre-AWS to ~83 MB minified with AWS — 17 139 added
76+
subcommands and 99 537 added options. The load happens once per process at
77+
init, so the increase is on cold-start, not on the keystroke path. Embedded
78+
specs heap walk (`memory/embedded_specs_heap_walk` = 4.23 ms) confirms the
79+
runtime walk over the full corpus is still well-bounded.
80+
81+
zstd-compressing the embedded JSON corpus (separate plan tracked in
82+
follow-on work) is the principled reclaim path here and would drop both
83+
binary size and load time meaningfully.
84+
85+
## spec_resolution
86+
87+
| Benchmark | Time | vs v0.10.0 |
88+
|-----------|------|------------|
89+
| shallow | 2.0979 µs | +8.4% |
90+
| deep | 1.2872 µs | +1.7% (flat) |
91+
| with_static_suggestions_tar | 6.0840 µs | new in v0.11.0 |
92+
93+
The shallow regression (~163 ns absolute) covers the alias-aware spec walk
94+
(`resolve_ctx_for_spec_walk` helper + cycle guard) added for multi-word
95+
alias expansion. Deep walks dominate descent time, so the alias preamble is
96+
amortised away.
97+
98+
## transform_pipeline
99+
100+
| Benchmark | Time | vs v0.10.0 |
101+
|-----------|------|------------|
102+
| simple | 26.598 µs | +2.5% (flat) |
103+
| regex | 113.72 µs | +2.0% (flat) |
104+
| json | 35.870 µs | +33.6% expected |
105+
106+
The json regression traces to the `json_extract_array` extension landed in
107+
the docs/correctness pass — `json_extract` now branches on whether the
108+
target is a scalar or an array projection, adding a per-row clone for the
109+
array case. The bench exercises the new array path, so the +33% is the
110+
cost of correctness, not avoidable overhead. Previous behaviour silently
111+
emitted wrong completions on array-shaped JSON (caught in v0.10.0
112+
`### Corrected`).
113+
114+
## engine_suggest_sync
115+
116+
| Benchmark | Time | vs v0.10.0 |
117+
|-----------|------|------------|
118+
| command_position | 17.677 µs | +2.4% within noise |
119+
| subcommand_with_spec | 16.863 µs | +0.6% (flat) |
120+
| filesystem_fallback | 1.0955 ms | +2.3% within noise |
121+
| alias_hit_single | 16.034 µs | new in v0.11.0 (apples-to-apples vs subcommand_with_spec, +-2 µs) |
122+
| alias_hit_multi | 1.1587 ms | new in v0.11.0 (multi-word alias path) |
123+
124+
`alias_hit_single` (alias `g=git`, ctx `g ch<TAB>`) lands within ±2 µs of
125+
`subcommand_with_spec` — alias expansion plus the spec walk is in the same
126+
budget as a non-aliased spec walk. `alias_hit_multi` (alias `gco='git
127+
checkout'`, ctx `gco m<TAB>`) is dominated by the filesystem walk over the
128+
bench's 2 k-file tmp dir (git checkout has a `filepaths` template), not by
129+
the alias machinery.
130+
131+
## priority_sort
132+
133+
| Benchmark | Time | vs v0.10.0 |
134+
|-----------|------|------------|
135+
| empty_query_10k | 723.97 µs | +1.4% (flat) |
136+
| fuzzy_query_10k | 650.20 µs | +5.7% within noise |
137+
138+
## memory
139+
140+
| Benchmark | Time | Notes |
141+
|-----------|------|-------|
142+
| embedded_specs_heap_walk | 4.23 ms | Walks every spec on the heap once |
143+
144+
Up from ~3.0 ms in v0.10.0; the raw walk now traverses the AWS subtree.
145+
Still well under the runtime memory budget — the embedded-spec heap test
146+
budget rose from 64 MiB to 128 MiB to admit AWS, with ~104 MiB used.
147+
148+
## Analysis
149+
150+
Three baselines added (`osc7772_decode`, `feedback_render`,
151+
`engine_suggest_sync/alias_hit_*`, `spec_resolution/with_static_suggestions_tar`).
152+
153+
The two material regressions are both **expected and acceptable**:
154+
155+
1. `spec_loading/load_717_specs` +241% — AWS spec restoration. One-time
156+
cold-start cost, not on the keystroke path. Reclaim plan: zstd-compress
157+
embedded specs (separate spec).
158+
2. `transform_pipeline/json` +33.6% — `json_extract_array` branch added
159+
for correctness. Previous v0.10.0 behaviour was silently wrong on
160+
array-shaped JSON; the new cost is structural, not optimisable.
161+
162+
Everything else lands within run-to-run noise (±10%). No keystroke-path
163+
regression worth blocking the release.
164+
165+
The `osc7772_decode` baseline (574 ns at typical buffer size) confirms the
166+
new buffer framing is invisible on the hot path. The `feedback_render`
167+
indicator row at 122 ns/frame confirms the async feedback UI ships free.

0 commit comments

Comments
 (0)