The non-git fallback used a substring match, so a negation line such as
!AGENTS/foo counted as the pattern AGENTS/ already being ignored and the
real rule was never appended.
Resolving the root to (pwd) meant running an agent CLI anywhere created an
AGENTS/ git repo, an AGENTS.md, two root symlinks, and a docs/ tree in that
directory. Scaffolding now requires a git repo or a pre-existing agent file
(AGENTS.md, CLAUDE.md, or AGENTS/); elsewhere it is a no-op.
Two related defects on the `$GIT_SERVER` auto-create path added in #150.
The test suite was creating repositories on a live forge. mkrep resolves a
server from `$GIT_SERVER` plus `$GITEA_URL`/`$GITEA_HOST`, and this repo
doubles as a real ~/.config/fish where all of them are exported, so the six
sections that call a bare `mkrep <dir>` took the auto-create branch and
contacted the server. That is how an empty `rootiest/repo` came to exist on
git.rootiest.dev on 2026-09-14 (since deleted). It also explains why those
cases looked flaky rather than broken: standalone they passed, because the
repo existed and mkrep linked instead of creating, while under
run-tests.fish they failed with `Error: no available login` — a throwaway
XDG_CONFIG_HOME leaves `tea` with no credentials. Passing for that reason is
worse than failing.
The suite now neutralizes all five variables once at setup, the same
shadow-to-empty pattern already used per-section for --check-existing, whose
comment flagged this exact hazard. Nothing depended on the ambient value:
every section wanting a server sets its own `set -lx GIT_SERVER`, and one
section asserts the opposite outright.
mkrep itself now confirms before an implicit create. Creating a repository
on a forge is the only outward-facing thing mkrep does, and on this path an
exported variable is all it takes to reach it, so `mkrep foo` — which reads
as purely local — would make a repo on a server without saying so. It asks
first, defaulting to no; declining leaves the local repo with no remote and
still exits 0. Provenance is what gates the prompt, not the resolved value:
--server, --remote and --new-remote all state outright what they will do, so
none of them prompts, and --yes skips the question. Where it cannot be asked
(a script, a pipe, any non-interactive shell) creation is skipped rather than
assumed, with a note on stderr naming the flags that would allow it.
Nine new assertions cover the skip, its stderr note, --yes, and --server not
prompting. The interactive read is verified by hand under a PTY, both
answers, but is not in the suite: that needs a pseudo-terminal, and the
answer parsing it guards is a single `string match`.
Both back the new `md` wrapper and nothing else, so both land in the
Optional tier, skipped by `fish-deps install`/`sync` unless `--optional`
(or `--all`) is passed.
firejail is a plain system package everywhere, so it needs no special
handling. marktext is not: upstream ships an AUR package and its own
GitHub release assets, and no distro carries it under a common name. Its
`_fdc_pm` entry is therefore deliberately empty, and a `marktext-release`
special offers the AUR package via paru/yay where one is present and
otherwise installs upstream's AppImage to ~/.local/bin/marktext.
The release assets embed their version in the filename, so there is no
stable /releases/latest/download URL to fetch -- `_fish_deps_marktext_appimage`
reads the download URL from the GitHub API instead. Upstream builds the
Linux AppImage for x86_64 only, and the helper says so rather than
downloading an unusable binary.
`fish-deps update` refreshes marktext through the AUR where available, and
otherwise only when ~/.local/bin/marktext exists -- a distro-packaged
marktext belongs to that package manager, and dropping an AppImage into
~/.local/bin would shadow it.
`md` forwards every argument to marktext untouched except two flags of its
own: `--read-only`/`-r` and `--foreground`. By default it detaches via
`bkg`, so the shell stays usable and the editor outlives the window that
launched it.
MarkText has no read-only mode, so `-r` sandboxes it with firejail, binding
each named file read-only. The subtlety is that MarkText is single-instance:
a plain launch hands the file to an already-running, unsandboxed, writable
window and exits, silently defeating the sandbox. `-r` therefore also passes
a private `--user-data-dir`, which forces an independent instance the
read-only bind actually covers, plus `--no-sandbox`, since Electron's own
sandbox needs the user namespaces firejail has already taken away.
Flags whose entire purpose is terminal output (`--version`, `-v`/`--verbose`,
`--debug`) imply `--foreground`; backgrounding them would send the output
you asked for to /dev/null.
The function is autoloaded and so never shadows an `md` function or alias
defined elsewhere -- fish only looks in functions/ when nothing named `md`
exists. A real `md` *binary* would be shadowed, so the body hands off to it
verbatim whenever marktext is not installed. No conf.d file and no
opinionated guard: `md` is a novel name rather than a command shadow, the
same as `bkg` and `detach`.
mkrep only ever runs git init, never a commit, so a freshly created
repo has no HEAD yet. The gitea/gitlab default templates chained
`git push -u origin HEAD` unconditionally after linking the remote,
which fails immediately regardless of the remote ("src refspec HEAD
does not match any") -- reproduced by a real user hitting it on the
first mkrep --server call. Guard the push on HEAD actually resolving
to a commit; skipping it is the correct outcome (nothing to push yet),
and a real push failure once a commit exists still propagates.
$GITEA_HOST/$GITLAB_HOST are bare hostnames (git.example.com); only
$GITEA_URL/$GITLAB_URL are expected to carry a scheme. mkrep was using
_HOST values as-is, producing a schemeless clone URL when only the
_HOST var was set.
mkrep can now pick a git host from --server, $GIT_SERVER, and
$GITEA_URL/$GITEA_HOST/$GITLAB_URL/$GITLAB_HOST (base URL only -- type
must come from --server/$GIT_SERVER, so setting those URL vars for an
unrelated tool can't silently turn a plain mkrep call into a
remote-creating one). Before creating anything it checks via
gh/glab/tea whether the repo already exists and links instead of
recreating it; --check-existing runs just that check and reports
without touching remotes.
Function entry pages previously reused the man-page pipeline's single
indented block plus its paragraph-guessing heuristics (_is_prose's
per-line word count, _as_table's 2+ row minimum), so whether a
Description, Arguments, or Exit Status section landed as flowing
prose, a table, or an unhighlighted code block depended on incidental
shape -- a short wrapped line, a single argument, a narrow column --
rather than which section it was. `cat`, `copy`, and `ltr` each ended
up formatted differently for no functional reason.
Add render_entry_site, a site-only renderer that builds each entry
straight from the parsed function header instead of re-deriving
structure from indented text: every present section (Synopsis,
Description, Arguments, Exit Status, Returns, Notes, Example) gets its
own `###` heading, Arguments/Exit Status become a table via the new
_kv_rows column parser (which also fixes narrow single-space columns
and name-only rows with wrapped continuations), and everything else is
unwrapped into normal paragraphs. The man-page/pandoc path
(render_entry, build_concat) is untouched.
Gittyup commits via libgit2 directly and never invokes gpg, silently
ignoring commit.gpgsign — root cause of an unsigned commit reaching
main. Adds a tracked .githooks/pre-push that rejects any push
carrying a commit with no signature or a bad signature, bypassable
with --no-verify. Wiring core.hooksPath to it is a per-machine
concern, done separately in user-dots, not shipped here.
- Removes a formatting error that broke the visual of
the --remote flag in both the documentation and
the function's help output.
-----
Impacted files:
functions/mkrep.fish
Raw multi-line armored key piped through 'echo | gpg --import' came out
CRC-corrupted (Invalid keyring) on the first real run -- something in the
secret/env round-trip mangles embedded newlines. Regenerated the bot key
(old one is unrecoverable, secrets are write-only) and store it
base64-encoded, decoded with base64 -d before import.
agy uses different resume syntax than claude: -c/--continue always
resumes the most-recent session, while --conversation takes a
specific session id. Previously -r/--resume was blanket-mapped to
-c/--continue, breaking `agy --resume <id>` and `agy -r=<id>` style
invocations. Now bare -r/--resume (no id following) maps to
-c/--continue; -r/--resume given an id (via =id or a following
non-flag word) maps to --conversation(=id).
actions@gitea was never a verified email on any Gitea account, so the
CI docs-regen commit (git commit && git push, done client-side in the
runner) could never show as verified regardless of server-side
[repository.signing] config -- Gitea only signs commits it generates
itself (merge button, web editor, wiki), never ones a client pushes.
Import a dedicated passphrase-less key for a new fishconfig-bot
account (verified email, no login) from the CI_GPG_PRIVATE_KEY repo
secret and sign with it instead.
git branch -vv marks column 1 with '+' (not '*') for a branch checked
out in another linked worktree. Only '*' was stripped from $1, so a
gone branch shown with '+' left a bogus "+" entry in $gone_branches
that then failed to delete: error: branch '+' not found.
Add a regression case reproducing the '+'-marked gone-branch line via
the existing git-clean mock handler in tests/test-network-fish.fish.
mkcd's create-and-cd behavior plus git init, remote linking, and
optional remote creation via a user-configured command template
($MKREP_REMOTE_CMD, {name}/{user} placeholders).
The guard added in 208ad95 had no test. This is the failure worth
covering, because it is the one that does not announce itself: with an
empty dump every row renders as DEFAULT, indistinguishable from a config
where nothing is set, so the user toggles from a false baseline.
Reaching the guard needs a real terminal -- the isatty check sits in
front of it -- so the case runs fish under a pty via python3's stdlib
pty module, which this suite already depends on. An empty
__config_settings_state is shadowed in to fake the failure.
The pty reader's 15s deadline is load-bearing rather than defensive: if
the guard regresses, config-settings does not fail, it opens the TUI and
blocks on getch(), so an unbounded read would hang the suite instead of
failing it. Verified both ways -- passes with the guard, and with the
guard removed the deadline fires and the case fails with a legible
excerpt rather than a screenful of escape sequences.
An empty dump does not fail loudly: the TUI renders every row as
DEFAULT, which is indistinguishable from a config where nothing is set.
That is a wrong answer rather than a missing one -- the user would be
looking at ON rows reported as DEFAULT -- so the launcher now checks the
dump is non-empty and bails with a message instead. The taxonomy alone
guarantees output on any working checkout. Also guard a failed mktemp,
which would otherwise send the dump to /state and hand an empty path to
rm -rf.
README: mention that / searches sub-categories across every category.
test_concat_section_five_stays_verbatim forbids backticks inside
generated Section 5 entries -- function headers are rendered verbatim, so
inline code markup there breaks the man page.
config-settings is now a launcher for scripts/config-settings-tui.py,
drawn with Python's stdlib curses. The seven fish files that hand-rolled
the ANSI renderer are gone, along with the golden harness that had to pin
their byte-exact output.
The TUI is a child process, so it can neither read the session's global
variables nor write them. State goes in as a dump from the new
__config_settings_state; the edits come back as a fish script that
config-settings sources, which is what lets the Session page's `set -g`
land in the caller's shell instead of in a child that is about to exit.
Every edit is emitted as a call to __config_settings_apply or
__config_settings_set_value, so list splitting, the SCROLLBACK_HISTORY_*
export mirror and the shadow-warning suppression all stay in the fish
layer that already owned them.
The consequence, and the one behaviour change: edits are applied in one
batch on exit rather than on each keypress. The status bar shows a
pending count.
New: `/` filters the current page, and on the Universal and Session pages
it reaches into every category's sub-categories, listing hits as
"Category › Sub" so a sub-category can be toggled without drilling into
its parent first. Also a `?` help overlay, mouse selection, and a
drill-down page that leads with the category's own toggle.
Gone with the renderer: the four width tiers, the wrap-aware erase
arithmetic, the stty/dd/od raw key reader, the panel-height bookkeeping
and the hand-written redraw differ. curses owns all of it, and the alt
screen plus absolute addressing makes the desync class behind 608b022,
4210f3b, 93fc5e0 and 3c4f720 unreachable.
The sub-category taxonomy is NOT duplicated in Python: it travels in the
state dump, still sourced from __config_settings_subcats. The category,
Sponge and Paths row tables move into Python, consolidating the two
copies the fish renderers kept.
Dependency: python3 with curses. Stdlib on Arch, Fedora and a full
Debian/Ubuntu python3; python3-minimal alone lacks _curses. The launcher
checks for both and names what is missing. Called out in the README.
Verified: 416/416 assertions, plus a live end-to-end in a sandbox HOME
confirming the Universal page writes universal scope only (U1/G0) and the
Session page global scope only (U0/G1).
Add scripts/config-settings-tui.py, a stdlib-curses prototype of the
config-settings interface, plus a gate that runs its --self-test.
The backend is stubbed: values live in an in-memory dict and nothing is
read from or written to fish variables. This is here to evaluate the
render engine and the interaction model before committing to a rewrite.
Why curses rather than more ANSI arithmetic:
- No flicker, structurally. curses diffs its virtual screen against the
physical one and emits only the changed cells, which is what
__config_settings_diff_redraw.fish reimplements by hand.
- Alternate screen plus absolute addressing. Stray output cannot desync
the display, so the bug class behind 608b022 (fish's read prompt),
4210f3b (a shadow warning on stderr), 93fc5e0 and 3c4f720 (line wrap
breaking the erase height) cannot occur at all.
- Resize is a repaint rather than wrap-factor arithmetic.
- Overlays, panes, live filtering and mouse input cost a few lines each.
The layout departs from the current single panel: a page sidebar with a
live filter on the left, a scrolling detail pane on the right, a help
overlay on '?', and mouse selection. Sub-category drill-down, the
tri-state badges and the Sponge/Paths value rows all carry over.
Dependency note: python3 with the curses module. That is stdlib on Arch,
Fedora and a full Debian/Ubuntu python3; python3-minimal alone does not
carry _curses, so the test asserts the import.
The prototype is not wired into config-settings and nothing existing
changed.
rootiest
fishconfig-bot
Claude
Gitea Actions