@feigi/fleet-ctl 3.24.21 → 3.24.22
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/package.json +1 -1
- package/scripts/derive-testcmd.sh +9 -9
- package/scripts/worktree.sh +24 -24
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
#!/bin/sh
|
|
2
2
|
# Read one command of a repository's proven Recipe out of its Recipe cache:
|
|
3
|
-
# `install` (the Install step) or `test` (the Test entrypoint).
|
|
3
|
+
# `install` (the Install step) or `test` (the Test entrypoint).
|
|
4
4
|
#
|
|
5
5
|
# The fleet keeps no table of technologies. A Recipe is DERIVED by an agent
|
|
6
6
|
# reasoning over the repository and PROVEN before it is written — the Install
|
|
@@ -8,12 +8,12 @@
|
|
|
8
8
|
# and showed real tests — so this script infers nothing. A missing cache is a
|
|
9
9
|
# refusal naming the step that derives one, never a guess: a guessed command
|
|
10
10
|
# that passes vacuously reads as a green suite to every consumer, which is the
|
|
11
|
-
#
|
|
11
|
+
# hazard the old inference already refused over.
|
|
12
12
|
#
|
|
13
13
|
# The ONE reader of the cache — reused, not reimplemented, by claim-ticket.sh
|
|
14
14
|
# (the Install step it runs and the Test entrypoint it bakes into the runner)
|
|
15
15
|
# and by review-core.mjs's snapshot agent (the Test entrypoint handed to the
|
|
16
|
-
# specialists). Two independent readers is what drifts
|
|
16
|
+
# specialists). Two independent readers is what drifts.
|
|
17
17
|
#
|
|
18
18
|
# THE CACHE is `<workspace>/.fleet/recipe.json`, where <workspace> is the
|
|
19
19
|
# directory holding the repository's COMMON git dir — the main checkout, the
|
|
@@ -65,7 +65,7 @@ set -eu
|
|
|
65
65
|
export LC_ALL=C
|
|
66
66
|
|
|
67
67
|
# Below the locale pin, not above it with `set -eu`: `unset` touches no
|
|
68
|
-
# byte-sensitive tool, but locale-pin
|
|
68
|
+
# byte-sensitive tool, but the locale-pin prose test treats ANY line here that
|
|
69
69
|
# is not a comment, a blank, or `set -[eux]+` as work the pin must sit above,
|
|
70
70
|
# and refuses on principle rather than on this line's own behaviour. Same
|
|
71
71
|
# placement, same reason, as release-ticket.sh's copy.
|
|
@@ -81,7 +81,7 @@ export LC_ALL=C
|
|
|
81
81
|
# and `--git-common-dir` consult no work tree. It stays on the line because the
|
|
82
82
|
# pair is one hazard with one remedy, and "inert today" is a measurement of the
|
|
83
83
|
# current call set, not a property of the script.
|
|
84
|
-
#
|
|
84
|
+
# A prose test pins the line itself.
|
|
85
85
|
unset GIT_DIR GIT_WORK_TREE
|
|
86
86
|
|
|
87
87
|
# shellcheck disable=SC2100 # literal name "derive-testcmd", not arithmetic — the unrelated $derive var assigned below is what the heuristic collides on, not this line
|
|
@@ -124,7 +124,7 @@ cache="${common%/*}/.fleet/recipe.json"
|
|
|
124
124
|
# that ends it rather than anything about a parser.
|
|
125
125
|
[ -e "$cache" ] || die "no Recipe cache at $cache — refusing to infer an Install step or a Test entrypoint; $derive"
|
|
126
126
|
|
|
127
|
-
# An INVOCATION, not a name lookup
|
|
127
|
+
# An INVOCATION, not a name lookup: a version-manager shim satisfies
|
|
128
128
|
# `command -v node` and then fails, and its stderr would arrive below under the
|
|
129
129
|
# cache's name — a corrupt-cache refusal for a cache nobody read.
|
|
130
130
|
nodeerr=$(node -e 0 </dev/null 2>&1) || unrunnable "node is unusable, refusing to read the Recipe cache without the interpreter — $nodeerr"
|
|
@@ -137,10 +137,10 @@ nodeerr=$(node -e 0 </dev/null 2>&1) || unrunnable "node is unusable, refusing t
|
|
|
137
137
|
# is guarded on its own:
|
|
138
138
|
#
|
|
139
139
|
# stderr is kept APART — to a file, never merged into the capture — because
|
|
140
|
-
# NODE_OPTIONS or NODE_DEBUG chatter lands there
|
|
140
|
+
# NODE_OPTIONS or NODE_DEBUG chatter lands there.
|
|
141
141
|
#
|
|
142
142
|
# stdout carries the value FRAMED, `recipe<` before it and `>recipe` after,
|
|
143
|
-
# and the WHOLE capture must be exactly that frame, byte for byte
|
|
143
|
+
# and the WHOLE capture must be exactly that frame, byte for byte:
|
|
144
144
|
# the anchored match alone only proves the ends line up, so a shim whose
|
|
145
145
|
# own chatter happens to spell out `recipe<`/`>recipe` around the real
|
|
146
146
|
# frame would still pass it and splice that chatter into the value — node
|
|
@@ -237,7 +237,7 @@ value=${value%"$close"}
|
|
|
237
237
|
# caller's "$@" as an unrelated top-level command instead of args reaching
|
|
238
238
|
# the Test entrypoint, and one containing a `#`-led word swallows everything
|
|
239
239
|
# after it, "$@" included, as a comment — measured on both. A trailing
|
|
240
|
-
# newline is checked here too
|
|
240
|
+
# newline is checked here too: the frame above lets a value keep one
|
|
241
241
|
# where an unframed `$(…)` capture always dropped it silently, so a value
|
|
242
242
|
# that used to read as `true;` now reads as `true;\n` and would otherwise
|
|
243
243
|
# slip past a check written for the no-newline case. Checked on the raw
|
package/scripts/worktree.sh
CHANGED
|
@@ -19,7 +19,7 @@
|
|
|
19
19
|
# and this file is sourced by the same scripts.
|
|
20
20
|
#
|
|
21
21
|
# Not folded into json.sh, which is scoped to JSON escaping: a filesystem
|
|
22
|
-
# predicate and a `git` reader are neither.
|
|
22
|
+
# predicate and a `git` reader are neither.
|
|
23
23
|
|
|
24
24
|
# The byte `wt_listing` substitutes for a newline found INSIDE a worktree path.
|
|
25
25
|
#
|
|
@@ -71,7 +71,7 @@ nl_path() {
|
|
|
71
71
|
# split into two and every consumer's `substr($0,10)` truncated the path at the
|
|
72
72
|
# newline. Measured on a real linked worktree at `.../wt/fix-33<LF>slug`, git
|
|
73
73
|
# 2.50.1: three scripts each handed a downstream consumer a path not on disk,
|
|
74
|
-
# and none of them refused
|
|
74
|
+
# and none of them refused — the defect already fixed in inflight.sh.
|
|
75
75
|
#
|
|
76
76
|
# `--porcelain -z` terminates every attribute with NUL instead, which needs git
|
|
77
77
|
# 2.36.0 — a floor this repo already exceeds through no-undo-audit.sh's
|
|
@@ -80,7 +80,7 @@ nl_path() {
|
|
|
80
80
|
# NOT `awk -v RS='\0'`. That is a gawk/BWK extension, and the awk this fleet
|
|
81
81
|
# actually runs on macOS — /usr/bin/awk, BWK awk 20200816 — does not merely
|
|
82
82
|
# ignore it: it stops dead at the first NUL and reports ONE record for a listing
|
|
83
|
-
# of any length. Measured
|
|
83
|
+
# of any length. Measured, all three spellings, `-v RS='\0'`,
|
|
84
84
|
# `-v RS='\000'` and `BEGIN{RS="\0"}`, every one of them `count=1`. No awk
|
|
85
85
|
# program can hold a NUL byte either, so the swap has to happen before awk sees
|
|
86
86
|
# the stream at all.
|
|
@@ -153,14 +153,14 @@ wt_listing() {
|
|
|
153
153
|
# alone reads its absence as unknown, which is the permanent refusal the callers
|
|
154
154
|
# exist to stop producing.
|
|
155
155
|
#
|
|
156
|
-
# `[ ! -L "$look" ]` in the loop condition, and it is the whole
|
|
156
|
+
# `[ ! -L "$look" ]` in the loop condition, and it is the whole fix. Every
|
|
157
157
|
# `test` primary except `-L` STATS, so it follows a symlink: on a dangling one
|
|
158
158
|
# `-e` is false while `-L` is true, and the predicate answered established-absent
|
|
159
159
|
# for a path `git worktree add` treats as occupied. Measured — worktree-audit.sh
|
|
160
160
|
# reported such a path as `MISSING on disk` with `ahead=0, dirty=0`, which is
|
|
161
161
|
# what the fleet controller reads to decide whether a replacement member would
|
|
162
|
-
# redo work or destroy it. Same `-e`/lstat split
|
|
163
|
-
#
|
|
162
|
+
# redo work or destroy it. Same `-e`/lstat split claim-ticket.sh guards one
|
|
163
|
+
# layer up.
|
|
164
164
|
#
|
|
165
165
|
# In the LOOP condition rather than as a fourth clause on the result, because
|
|
166
166
|
# that placement answers a second shape for free: a dangling symlink standing in
|
|
@@ -175,19 +175,19 @@ wt_listing() {
|
|
|
175
175
|
# remove` that deleted a symlink's target, with no live claim; and
|
|
176
176
|
# release-ticket.sh's rc-255 halt path, which leaves the link with the branch and
|
|
177
177
|
# the `in-progress` label still alive — a live claim, mid-release, that failed
|
|
178
|
-
# partway. "Not established-absent" is the only answer true of both.
|
|
178
|
+
# partway. "Not established-absent" is the only answer true of both.
|
|
179
179
|
#
|
|
180
|
-
# `look=${look:-/}` INSIDE the loop, and that placement is the whole
|
|
180
|
+
# `look=${look:-/}` INSIDE the loop, and that placement is the whole fix:
|
|
181
181
|
# `${p%/*}` on `/x` yields the empty string, not `/`, so a path whose every
|
|
182
182
|
# ancestor below the root is gone used to fall out on "" and answer unknown about
|
|
183
183
|
# an absence the searchable root proves. The same restore written AFTER the loop
|
|
184
184
|
# reads identically and is wrong — nothing enters the loop on an empty `$1`, so
|
|
185
185
|
# it would rewrite that to `/` too and turn `gone ""` into established-absent.
|
|
186
|
-
# Inside, it only ever rewrites what the loop just truncated.
|
|
187
|
-
# holds that matrix, `gone ""` included, because no caller can
|
|
188
|
-
# callers test the path non-empty first, and worktree-audit.sh
|
|
189
|
-
# `git worktree list`, which never emits an empty one — so a
|
|
190
|
-
# alone cannot tell the two placements apart.
|
|
186
|
+
# Inside, it only ever rewrites what the loop just truncated. The suite's direct
|
|
187
|
+
# test of `gone` holds that matrix, `gone ""` included, because no caller can
|
|
188
|
+
# reach it: the callers test the path non-empty first, and worktree-audit.sh
|
|
189
|
+
# reads its own off `git worktree list`, which never emits an empty one — so a
|
|
190
|
+
# caller-level suite alone cannot tell the two placements apart.
|
|
191
191
|
#
|
|
192
192
|
# `!=`, not a non-empty test: `${p%/*}` returns p unchanged when p holds no
|
|
193
193
|
# slash, so the emptiness form spins forever on one. git emits absolute paths to
|
|
@@ -197,8 +197,8 @@ wt_listing() {
|
|
|
197
197
|
# One definition, in one file, because every caller asks one question. Answered
|
|
198
198
|
# separately they drift, and the halves of this fleet that protect a member's
|
|
199
199
|
# work stop agreeing about whether there is any work there to protect — which is
|
|
200
|
-
# exactly what
|
|
201
|
-
# site rather than in the predicate.
|
|
200
|
+
# exactly what separate copies of it did: three, and only one compensating, at
|
|
201
|
+
# a call site rather than in the predicate.
|
|
202
202
|
#
|
|
203
203
|
# 0 ONLY for established absent; 1 covers present AND cannot-stat, so a caller
|
|
204
204
|
# needing those apart pairs this with its own `[ ! -e ]`, as release-ticket.sh's
|
|
@@ -217,7 +217,7 @@ gone() {
|
|
|
217
217
|
# worktree on disk holding an uncommitted file and for which prune is a no-op
|
|
218
218
|
# — the permanent-refusal shape the surrounding guards exist to prevent.
|
|
219
219
|
# Refused in the predicate rather than at each call site for the reason the
|
|
220
|
-
# rest of this comment gives: answered separately, the callers drift.
|
|
220
|
+
# rest of this comment gives: answered separately, the callers drift.
|
|
221
221
|
nl_path "$1" && return 1
|
|
222
222
|
look=$1
|
|
223
223
|
while [ ! -e "$look" ] && [ ! -L "$look" ] && [ "$look" != "${look%/*}" ]; do
|
|
@@ -259,11 +259,11 @@ gone() {
|
|
|
259
259
|
# started from a detached HEAD is different: `BISECT_START` holds that HEAD's
|
|
260
260
|
# full SHA, which the code below still turns into `refs/heads/<sha>` and
|
|
261
261
|
# collects — matching only a branch literally named with that 40-hex string,
|
|
262
|
-
# which nothing in this fleet's naming ever produces.
|
|
262
|
+
# which nothing in this fleet's naming ever produces.
|
|
263
263
|
#
|
|
264
264
|
# `2>/dev/null` on the admin-dir read, NOT the `2>&1` the fleet's message-text
|
|
265
|
-
# captures fold in, because this capture is used as a PATH. Measured
|
|
266
|
-
#
|
|
265
|
+
# captures fold in, because this capture is used as a PATH. Measured: a
|
|
266
|
+
# `~/.gitconfig` with a key outside any section makes every git
|
|
267
267
|
# command print `error: key does not contain a section: …` to stderr AT EXIT 0,
|
|
268
268
|
# so `2>&1` returns that line glued in front of the git dir, every `[ -e ]`
|
|
269
269
|
# below then matches nothing, and a worktree holding an interrupted rebase reads
|
|
@@ -274,7 +274,7 @@ gone() {
|
|
|
274
274
|
# that is not there, and only the second is an answer.
|
|
275
275
|
#
|
|
276
276
|
# `wt_op` is this function's OUTPUT, read by the sourcing script; so is
|
|
277
|
-
# `wt_op_held`, read by `wt_holding` below and by worktree-audit.sh
|
|
277
|
+
# `wt_op_held`, read by `wt_holding` below and by worktree-audit.sh —
|
|
278
278
|
# SC2034, as for `wt_listing`.
|
|
279
279
|
#
|
|
280
280
|
# The admin dir is normally found by asking git FROM $1 (`rev-parse
|
|
@@ -294,7 +294,7 @@ gone() {
|
|
|
294
294
|
# the worktree path it was registered for, and that registry lives under this
|
|
295
295
|
# repo's `.git`, never under the worktree — a directory `rm -rf`'d out from
|
|
296
296
|
# under git leaves its bookkeeping, and whatever rebase or bisect state it
|
|
297
|
-
# held, untouched.
|
|
297
|
+
# held, untouched.
|
|
298
298
|
# shellcheck disable=SC2034
|
|
299
299
|
wt_op_state() {
|
|
300
300
|
wt_op=
|
|
@@ -357,7 +357,7 @@ wt_op_state() {
|
|
|
357
357
|
# there, is marked `prunable` too — the two are not the same fault), is what
|
|
358
358
|
# tells `wt_op_state` it may look for that admin dir a different way; the
|
|
359
359
|
# ordinary route — asking git FROM the worktree's own directory — cannot,
|
|
360
|
-
# because there is no directory left to ask from.
|
|
360
|
+
# because there is no directory left to ask from.
|
|
361
361
|
#
|
|
362
362
|
# A record naming neither a `branch` nor the literal `detached` line is a HEAD
|
|
363
363
|
# git itself could not classify from the worktree's own directory — measured,
|
|
@@ -368,7 +368,7 @@ wt_op_state() {
|
|
|
368
368
|
# is untouched — so this case resolves through the SAME registry route as a
|
|
369
369
|
# gone worktree, unconditionally, rather than answering "cannot tell" for a
|
|
370
370
|
# worktree that may hold nothing of this caller's at all. Only once even that
|
|
371
|
-
# route cannot read an admin dir does this fall back to "cannot tell".
|
|
371
|
+
# route cannot read an admin dir does this fall back to "cannot tell".
|
|
372
372
|
#
|
|
373
373
|
# Each record is read whole before it is judged — `branch`, `detached`, and
|
|
374
374
|
# the boundary that ends it — never decided line by line, because a record's
|
|
@@ -462,7 +462,7 @@ wt_h_settle() {
|
|
|
462
462
|
fi
|
|
463
463
|
# Neither a `branch` line nor the literal `detached` line: git could not
|
|
464
464
|
# classify this record's HEAD from the worktree's own directory at all (a
|
|
465
|
-
# corrupted admin HEAD is
|
|
465
|
+
# corrupted admin HEAD is that shape, measured) — go straight to the
|
|
466
466
|
# registry route, unconditionally, since the ordinary one already failed by
|
|
467
467
|
# definition of being here.
|
|
468
468
|
wt_h_check_op "$1" 1
|