* fix(#2866): Codex installer strips legacy hooks at end-of-file without trailing newline
The four shape-strip regexes in `bin/install.js` (Codex install path)
required `\r?\n` at end. A stale GSD hook block sitting at end-of-file
without a trailing newline (common — many editors strip them, and the
legacy installer never wrote one) failed every shape, the installer
saw `gsd-check-update` already present, skipped writing the new
Nested-AoT block, and Codex 0.125+ refused to load with
invalid type: map, expected a sequence in `hooks`
Root cause + fix
================
Each shape's terminator changed from `\r?\n` to `(?:\r?\n|$)`, so
end-of-file is also a valid terminator.
Strip logic was lifted into a new pure helper
`stripStaleGsdHookBlocks(configContent)` that the install pipeline now
calls in place of the inline replace chain. The helper is exported via
the GSD_TEST_MODE module.exports for direct unit-test coverage.
Regression test
===============
`tests/bug-2866-codex-strip-no-trailing-newline.test.cjs` exercises
all four historical shapes (Shape 1 — pre-#1755 gsd-update-check;
Shape 2 — flat [[hooks]]+gsd-check-update; Shape 3 — single
[[hooks.SessionStart]] without nested .hooks; Shape 4 — correct
two-block nested) twice each: once with a trailing newline (regression
guard against the existing behavior) and once at end-of-file without a
trailing newline (the reporter's exact repro).
It also asserts:
- the helper is a no-op when no GSD reference is present, and
- Shape 4 strip does not leave an orphaned [[hooks.SessionStart]]
header behind (the same ordering invariant the inline code relied on).
The helper is loaded via `package.json` `bin` field, not a hardcoded
path — `tests/bug-2866-codex-strip-no-trailing-newline.test.cjs`
parses package.json and resolves `pkg.bin['get-shit-done-cc']` to
require the installer.
Closes#2866
* test(#2866): assert TOML structure, not raw-text substrings
CodeRabbit caught the strip assertions using `.includes()` against
raw TOML output. Added a small line-structural parseTomlShape() helper
(table headers + dotted-path key/value map, comments stripped) and
rewrote the assertions to:
- Verify no [[hooks.* table header survives the strip
- Verify no key carries a stale gsd-(update|check)-(check|update) value
- Verify history.persistence is preserved as the parsed string "save-all"
Behaviour is unchanged (the strip function under test is not modified).
The assertions now check structural shape rather than substring presence,
which catches re-shaping regressions that text matching would miss.
No new dependencies — the parser is local to the test and handles only
the small well-formed TOML these tests construct.
* refactor(#2866): replace regex hook strip with TOML AST removal
Per CR feedback on PR #2870: the regex-driven `stripStaleGsdHookBlocks`
implementation was fragile to whitespace, indentation, and key-ordering
variations the regression test never exercised. Variations the regex
silently leaked (verified before the rewrite):
- Shape 4 with an extra blank line between parent/child tables
- Shape 2/3 with `command` ordered before `event`
- Shape 3 with an extra `timeout = 5000` key — worse than a leak: the
regex matched only the command line, leaving `timeout = 5000`
orphaned outside any TOML table (invalid TOML)
- Tight whitespace `event="SessionStart"` (no spaces around `=`)
The structural rewrite uses the TOML parser already present in this file
(`getTomlTableSections` + `getTomlLineRecords` + `parseTomlValue` +
`removeContentRanges` + `collapseTomlBlankLines`):
1. Find every section whose path is `hooks` or starts with `hooks.`.
2. For each, walk the section's line records and parse `command` values
structurally — match by basename equality (`gsd-update-check.js` or
`gsd-check-update.js`), never by regex on raw bytes.
3. Detect orphaned `[[hooks.SessionStart]]` parents: empty body and a
stale child immediately follows → mark for removal.
4. Extend each removal range backward through any preceding
`# GSD Hooks` marker line (detected via line records, not text scan).
5. Remove ranges atomically and collapse resulting blank-line runs.
Legacy hook basenames are hoisted to template-literal constants so the
existing `install-hooks-copy.test.cjs` quoted-literal guard continues to
catch accidental *registration* of the inverted filename, while strip
detection (which legitimately needs both names) bypasses it.
Test coverage added: 8 new sub-tests exercising the four whitespace/
ordering variations (with and without trailing newline) plus a
`[[hooks.UserPromptSubmit]]` user-authored hook to guarantee the strip
only touches GSD-managed sections. 20/20 in the file, 5867/5867 in the
full suite.
ef08a89241