[brettheap]

Closes #987.

Goal of #987: under the holder's make -jN bootstrap (or validate, park, resume), a member's make runs its jobs together where it can share the holder's jobserver safely, and one at a time, with no warning, everywhere else.

The change is in the family holder's pinned templates/family-root/scripts/bootstrap.py and templates/family-root/scripts/siblings.py, and in tests/test_family_member_make_jobserver.py (and one word of tests/test_update_shape_kept_helper.py's docstrings). The holder's Makefile is unchanged. The commits: 1cc5b21b, the change; bfc98d0f, which has siblings.py import the jobserver function from bootstrap.py instead of carrying a second copy; 5cf60365, GitHub's update-branch merge of main e2f9fc42; and fe1e6bcf, RV1023's fix round: the function's new name, member_make_environment (choice 8), #950's escapes restored in one table, and one more test item; then f00a855f, a merge of main 1a47ed07, which had changed bootstrap.py (#1028's git_not_started, away from the jobserver code). The jobserver code and the other three files are byte for byte as at fe1e6bcf.

The line

#950 (#888, 8264c05) takes the jobserver word and every -j out of the MAKEFLAGS a member's make inherits (member_make_env, in both scripts), wherever MAKEFLAGS names a jobserver. On GNU Make 4.3 that only removed a warning: the member's make ran serially before and after.

GNU make 4.4 and later use a named pipe where mkfifo exists, and write --jobserver-auth=fifo:<path> into MAKEFLAGS. A make opens that pipe by its path, and 4.4's NEWS says "sub-makes will connect to the jobserver even if they are not marked as recursive". So under a 4.4 holder, a member's make probably ran its jobs together before #950, and runs them one at a time after it.

The change: the issue's first direction

member_make_environment(make) (#950's member_make_env, under a new name: choice 8) now hands MAKEFLAGS on as it is, the -jN beside the pipe included, only where all of these hold:

1. Every jobserver word it names is 4.4's named pipe (names_a_fifo): --jobserver-auth=fifo:<path>, its path read as make reads it (a backslash keeps the next character), absolute, and a named pipe now (stat.S_ISFIFO).
   • Never kept: a pipe's descriptors (3,4), make 3.81's --jobserver-fds, Windows' semaphore, or a -j without its jobserver.
   • A pipe that is gone, a path holding another kind of file, or a path relative to some other directory than the member's would make a 4.4 member warn that it cannot open the jobserver, so each is taken out.
2. The member's own make says it is GNU make 4.4 or later (reads_fifo).
   • It is asked as <make> --version, the very program the script then runs: $MAKE or make on PATH in bootstrap.py's run_make, make_program() in siblings.py's dispatch. GNU make exports no MAKE and no MAKE_VERSION to a recipe (measured below), so only the member's make can say what it reads.
   • The first line must read GNU Make X.Y with (X, Y) >= (4, 4), read as numbers, and the exit code must be 0.
   • It is asked once per make and per run (FIFO_READERS), and only where condition 1 already holds. So under a 4.3 holder, or with no -j, no extra process runs.
   • The asked make's own MAKEFLAGS has the jobserver taken out, stdin is closed, and stderr is dropped.
   • A make that is not there, names itself otherwise, is older, exits non-zero, or does not answer in VERSION_TIMEOUT (10) seconds is a no.

Everywhere else the jobserver words and every -j are taken out exactly as #950 takes them. #950's reading of MAKEFLAGS is kept unchanged: MAKEFLAGS_WORD, JOBSERVER_WORD, JOBS_WORD, splitting only at a space or a tab, a backslash keeping either in a value. A MAKEFLAGS that names no jobserver is still left byte for byte, and no make is asked anything.

One copy, in bootstrap.py (bfc98d0f, renamed in fe1e6bcf). SonarCloud's quality gate failed at 1cc5b21b on new_duplicated_lines_density, 15.5% over its 3%: one block, the two scripts' identical copies of #950's three readings of MAKEFLAGS and of FIFO_WORD, GNU_MAKE_VERSION, VERSION_TIMEOUT, FIFO_READERS, names_a_fifo and reads_fifo. They are now defined in bootstrap.py only, with the function that uses them, member_make_environment. siblings.py imports it beside REWRITE_TARGETS, credential, member_path and members_of, as it takes the credential and the member rows, and dispatch calls it with the make it runs, as before. This reverses #950's choice 4, at a cost choice 8 states.

Measurements

The only make on this host is /usr/bin/make, GNU Make 4.3 (make --version: GNU Make 4.3 / Built for x86_64-pc-linux-gnu). No GNU make 4.4 or later is installed here, and none was downloaded, built or run. So the 4.4 case is shown only through its MAKEFLAGS spelling, real named pipes made with os.mkfifo, and a fake make that answers --version as each make does, as #950 showed it.

Measured on 4.3:

• A recipe under make -s -j2 sees MAKEFLAGS=s -j2 --jobserver-auth=3,4, MFLAGS=-s -j2 --jobserver-auth=3,4 and MAKELEVEL=1. It sees no MAKE and no MAKE_VERSION ([unset] for both).
• 4.3 does not know --jobserver-style=fifo: unrecognized option, exit 2.
• Given MAKEFLAGS=' -j2 --jobserver-auth=fifo:<path>', 4.3 stops with make[1]: *** internal error: invalid --jobserver-auth string 'fifo:<path>'. Stop., exit 2. That happens whether a named pipe is at the path or not.
• make --version under that same MAKEFLAGS still prints GNU Make 4.3, exit 0. The question is asked with the jobserver taken out all the same, so a make asked only its version is handed nothing to open.
• MAKEFLAGS=' -j2 --jobserver-auth=3,4' with the descriptors closed gives the family root's Makefile: under make -jN, every member's make warns jobserver unavailable: using -j1 and runs one job at a time #888's warning and a serial run. MAKEFLAGS=' -j2' alone runs both jobs together, in a pool of the member's own. That is why a -j never goes without its jobserver.
• An option 4.3 does not know, in MAKEFLAGS (--jobserver-style=fifo), is passed over silently: exit 0, serial run. So a 4.4 holder's other words cannot stop an older member's make.

Read, not run (4.4.1's NEWS): the named-pipe jobserver is 4.4's default where mkfifo(3) is supported. It is selected away with --jobserver-style=pipe. When the pipe-based jobserver closes its descriptors for a command it does not take for a make, it "will now add a new option to the MAKEFLAGS environment variable that disables the jobserver". Its manual reads negative descriptors in --jobserver-auth=R,W as a disabled jobserver. If that is the option's spelling, JOBSERVER_WORD takes it out as before, and it is never a named pipe, so it is never handed on. If it is not, a member's make of 4.4 reads it as its own, and an older one passes it over, as 4.3 passes over --jobserver-style=fifo (measured). GNU/Hurd and Cygwin are reverted to the pipe in 4.4.1.

Tests

tests/test_family_member_make_jobserver.py: 115 items, 114 passed, 1 skipped at fe1e6bcf (113 items at bfc98d0f, 112 at 1cc5b21b, each with 1 skipped). On main's scripts (3864b18c) the file as at 1cc5b21b gives 102 failed, 9 passed, 1 skipped:

• 96 of the failures are TypeError: member_make_env() takes 0 positional arguments but 1 was given. 38 of those are the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's own in-process items, which now pass the member's make; the other 58 are new.
• 2 are the new too-slow items, which set VERSION_TIMEOUT to 1 first: AttributeError: ... has no attribute 'VERSION_TIMEOUT'.
• 4 are the new end-to-end items failing on behaviour: on main each member's make was handed '', and nobody asked a version (['park ARGS=|', 'park ARGS=|']).
• The 9 that pass on main are the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's 5 end-to-end items and the new 4.3 regression guard (4). As before, they hold.
• Behaviour, not signature: mutants M3 and M4 below restore main's answer (never shared) under the new signature, and are killed by 10 and 10 items.

New items (72):

• test_siblings_takes_bootstrap_s_member_make_environment (bfc98d0f): siblings.py's member_make_environment is bootstrap.py's, so each siblings.py item below runs that one function through the import a holder's siblings.py makes. The too-slow items now set bootstrap.py's VERSION_TIMEOUT, which that function reads.
• through the holder's rules (4 + 4 + 1):
  • test_a_member_s_make_that_reads_a_named_pipe_is_handed_the_holder_s[bootstrap|validate|park|resume]: each rule's line, read from the holder's Makefile and run straight, under the MAKEFLAGS a 4.4 holder gives it (-j2 and a real named pipe), with $MAKE a fake saying GNU Make 4.4.1. Each member's make got that MAKEFLAGS byte for byte, after exactly one --version| asked with an empty MAKEFLAGS.
  • test_under_a_named_pipe_a_member_s_make_before_4_4_runs_serially[...]: the same MAKEFLAGS, with this host's real 4.3 as the member's make. It covers the family root's bootstrap.py and siblings.py: under GNU make 4.4 and later, a member's make loses the parallelism it may have had through the holder's fifo: jobserver, since #950 takes every jobserver out of MAKEFLAGS #987's second verify line, a 4.4 holder with a 4.3 make on PATH: exit 0, no jobserver anywhere in the output (so no invalid --jobserver-auth string and no warning), each member's trace serial, no -j in its flags. It is skipped where the make on PATH is 4.4 or later.
  • test_under_make_4_4_a_member_s_make_shares_the_holder_s_named_pipe: the family root's bootstrap.py and siblings.py: under GNU make 4.4 and later, a member's make loses the parallelism it may have had through the holder's fifo: jobserver, since #950 takes every jobserver out of MAKEFLAGS #987's first verify line, with a real 4.4 holder's make -j2 bootstrap. The two jobs must interleave, the member's flags must name the pipe, and nothing may say jobserver. It is skipped here, and has never been run: the test asks this host's make -j2 which jobserver it writes, and 4.3 writes --jobserver-auth=3,4. the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's two end-to-end tests skip by the same question where make -j2 writes a named pipe, since a member's make of 4.4 now shares it.
• both scripts, in process (62, 31 per script):
  • a make that says it is GNU make 4.3 is asked --version once in three calls, so a no is cached as a yes is (fe1e6bcf, RV1023's NIT 1);
  • a named pipe that is there, handed on to a make saying 4.4, 4.4.1, 4.10 or 5.0 (read as numbers);
  • a pipe whose path holds a space, a tab or a backslash, written as make writes it;
  • --version asked once in three calls, with ks -- NOTE=a\ -j3 for its own MAKEFLAGS, while each call hands on 4.4's whole MAKEFLAGS;
  • taken out, -j and all, where the make says 4.3, 4.2.1 or 3.81, names itself otherwise, says 4.4.1 but exits 1, says nothing, is not there, or sleeps past the timeout (VERSION_TIMEOUT set to 1 there; the run waited no more than 20 s);
  • taken out, and the make not asked, where the pipe is gone, the path is a file, the path is relative (a pipe there in this run's directory), or a pipe's descriptors are named beside it;
  • the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's ten non-named-pipe spellings, taken out for a make that says 4.4.1 too, which is never asked.
• the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's TAKEN_OUT spellings keep their escapes for U+00A0 and U+202F: 1cc5b21b had written them as the raw characters, and fe1e6bcf puts main's four lines back, byte for byte (RV1023's SHOULD 3).
• the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's in-process items now pass a fake make: GNU Make 4.3 for its twelve spellings, and GNU Make 4.4.1 (never asked) for the six MAKEFLAGS that name no jobserver and for the copy test.

Tests at fe1e6bcf, once, GNU Make 4.3, gh stubbed: the jobserver file, test_update_shape_kept_helper.py, test_pinned_copy_text.py, the trio (test_repo_hygiene.py, test_agents_md_codes.py, test_cli_reference.py), test_family_holder_bootstrap_texts.py and test_siblings_inputs.py: 645 passed, 1 skipped, 0 failed (5 min 8 s); the skip is the real-4.4 item. Earlier, at bfc98d0f, a wider set (the family bootstrap and siblings files, test_shape_doctor.py and test_scaffold_leg_path_collisions.py with the jobserver, kept-helper and trio files) gave 1314 passed, 4 skipped, 0 failed; at 1cc5b21b, 1154 passed, 4 skipped. The full suite was not run here. CI's tests passed on 5cf60365. At the merge f00a855f, the same set with #1028's test_bootstrap_without_git.py and test_validate_family_git_not_runnable.py: 735 passed, 1 skipped, 0 failed (4 min 56 s).

Mutants

At 1cc5b21b, with the code in both scripts: 34 mutants, each of the 16 edits of the shared code in each script, and one edit of each call site. Each is one edit of one script, the jobserver file is run against it, and the script's bytes are then put back from a saved copy and checked by sha256 before the next. All 34 killed; every restore matched. Each cell gives the items that failed, of 112 (the other items passed, and the real-4.4 item skipped).

The edit	bootstrap.py	siblings.py
reads_fifo dropped: a live named pipe is handed on and no make is asked	M1: 20 failed	M2: 20 failed
never shared, the make still asked: main's answer under the new signature	M3: 10 failed	M4: 10 failed
any for all: one named pipe among the jobserver words is enough	M5: 1 failed	M6: 1 failed
os.path.exists for S_ISFIFO	M7: 1 failed	M8: 1 failed
os.path.isabs dropped	M9: 1 failed	M10: 1 failed
the pipe's path read without make's backslash rule	M11: 3 failed	M12: 3 failed
> for >=: 4.4 itself is a no	M13: 8 failed	M14: 8 failed
the version compared as bytes, so 4.10 reads older than 4.4	M15: 1 failed	M16: 1 failed
--version's exit code ignored	M17: 1 failed	M18: 1 failed
FIFO_READERS bypassed: asked on every call	M19: 3 failed	M20: 3 failed
--version asked under this run's environment, jobserver and all	M21: 17 failed	M22: 17 failed
no timeout	M23: 1 failed	M24: 1 failed
only TimeoutExpired caught: a make that is not there raises	M25: 1 failed	M26: 1 failed
the make asked before the pipe is checked	M27: 14 failed	M28: 14 failed
no early return where MAKEFLAGS names no jobserver	M29: 6 failed	M30: 6 failed
shared, but with every -j taken out	M31: 10 failed	M32: 10 failed
run_make asks make, not the make it runs	M33: 2 failed	-
dispatch asks make, not the make it runs	-	M34: 2 failed

At bfc98d0f, with the code in bootstrap.py only: the 16 edits in siblings.py's copy have no target left (each old text is found 0 times), so the 16 in bootstrap.py and the two call sites were run again, the same way, against the jobserver file's 113 items. All 18 killed; every restore matched. Each of the 16 fails the same items through bootstrap.py as through siblings.py's import, twice round one's count for that edit in one script; each call site's mutant fails only through its own script.

The edit	Mutant	Items failed	Through bootstrap.py	Through siblings.py
reads_fifo dropped: a live named pipe is handed on and no make is asked	M1 (bootstrap.py)	40	20	20
never shared, the make still asked: main's answer under the new signature	M3 (bootstrap.py)	20	10	10
any for all: one named pipe among the jobserver words is enough	M5 (bootstrap.py)	2	1	1
os.path.exists for S_ISFIFO	M7 (bootstrap.py)	2	1	1
os.path.isabs dropped	M9 (bootstrap.py)	2	1	1
the pipe's path read without make's backslash rule	M11 (bootstrap.py)	6	3	3
> for >=: 4.4 itself is a no	M13 (bootstrap.py)	16	8	8
the version compared as bytes, so 4.10 reads older than 4.4	M15 (bootstrap.py)	2	1	1
--version's exit code ignored	M17 (bootstrap.py)	2	1	1
FIFO_READERS bypassed: asked on every call	M19 (bootstrap.py)	6	3	3
--version asked under this run's environment, jobserver and all	M21 (bootstrap.py)	34	17	17
no timeout	M23 (bootstrap.py)	2	1	1
only TimeoutExpired caught: a make that is not there raises	M25 (bootstrap.py)	2	1	1
the make asked before the pipe is checked	M27 (bootstrap.py)	28	14	14
no early return where MAKEFLAGS names no jobserver	M29 (bootstrap.py)	12	6	6
shared, but with every -j taken out	M31 (bootstrap.py)	20	10	10
run_make asks make, not the make it runs	M33 (bootstrap.py)	2	2	0
dispatch asks make, not the make it runs	M34 (siblings.py)	2	0	2

At fe1e6bcf, renamed, the same 18 again, and RV1023's RV1 (a no not cached), against the jobserver file's 115 items. All 19 killed; every restore matched.

The edit	Mutant	Items failed	Through bootstrap.py	Through siblings.py
reads_fifo dropped: a live named pipe is handed on and no make is asked	M1 (bootstrap.py)	42	21	21
never shared, the make still asked: main's answer under the new signature	M3 (bootstrap.py)	20	10	10
any for all: one named pipe among the jobserver words is enough	M5 (bootstrap.py)	2	1	1
os.path.exists for S_ISFIFO	M7 (bootstrap.py)	2	1	1
os.path.isabs dropped	M9 (bootstrap.py)	2	1	1
the pipe's path read without make's backslash rule	M11 (bootstrap.py)	6	3	3
> for >=: 4.4 itself is a no	M13 (bootstrap.py)	16	8	8
the version compared as bytes, so 4.10 reads older than 4.4	M15 (bootstrap.py)	2	1	1
--version's exit code ignored	M17 (bootstrap.py)	2	1	1
FIFO_READERS bypassed: asked on every call	M19 (bootstrap.py)	8	4	4
--version asked under this run's environment, jobserver and all	M21 (bootstrap.py)	36	18	18
no timeout	M23 (bootstrap.py)	2	1	1
only TimeoutExpired caught: a make that is not there raises	M25 (bootstrap.py)	2	1	1
the make asked before the pipe is checked	M27 (bootstrap.py)	28	14	14
no early return where MAKEFLAGS names no jobserver	M29 (bootstrap.py)	12	6	6
shared, but with every -j taken out	M31 (bootstrap.py)	20	10	10
run_make asks make, not the make it runs	M33 (bootstrap.py)	2	2	0
dispatch asks make, not the make it runs	M34 (siblings.py)	2	0	2
RV1023's RV1: a no is not cached, so asked again on every call	RV1 (bootstrap.py)	2	1	1

Merges

git merge-tree --write-tree (git 2.43.0), every head read fresh from GitHub at 2026-10-20 10:35Z:

• main 1a47ed07: clean with fe1e6bcf (tree a79d1eb1), and merged as f00a855f, which has that tree.
• All 9 open pull requests, each against f00a855f and against main: all merge clean (the family root's bootstrap.py: a gitlink a member's own index gives no url, or a member's commit that a fetch reaching its remote did not bring, fails the run by name with its cure and no credential advice, where it said family bootstrap ok beside git's exit 128 (#892) #954 fe907c84, validate-family.py: a hand mount's git rm line works typed: git rm --cached where a plain one refuses a staged gitlink, dies on a clone no section names or refuses what is no checkout, with what it leaves said, and a .gitmodules gone from the work tree put back before the git rm that takes a section out (#936) #1025 4c0c52d0, adopt-project.py: check's throwaway repositories are made in the work directory with the source's object format and no hook of this machine's; a path git refuses is plan-path-git-refuses, a repository git cannot make there adopt-git-unasked (#910) #1031 b83d9c02, adopt-project.py: plan --project NUL (and Con, com1, Lpt9) is refused as adopt-bad-project-name before anything exists, and check and execute find it in the plan's legs.assembly, by the one question the scaffold and family.py init ask (#1014) #1041 c0ccf755, the family root's bootstrap.py: git's update ended by a signal fails the run by name, and B4 with a token in use is told from fetches redone without a trace, so the token reaches no file (#998, #996) #1045 b2bbb31c, adopt-project.py and README.md: the Makefile follow-up names a kept makefile or GNUmakefile GNU make reads first, and where a case-blind disk writes the shape's; README says which Makefile exports CONTRACTS_DIR (#977) #1046 83b3671b, update-shape.py and scripts/bump-leg.py: the three line rewriters find a first-line key after a byte-order mark, as every reader does, and write the mark back as it was (#933) #1050 434a07cb, AGENTS.md, README.md, docs/handbook.html and setup-project.py: the taken-clone-path refusal names the --family folder, "nothing that stays" replaces "nothing is created", and passed_values says what argparse takes as a value (#921, #923) #1051 4a6a55b9, scaffold-project.py: --dry-run's help says its answer about each folder is the mode bits' and that it leaves nothing that stays, a --reuse-empty-repo re-run that takes a seed found only on seed/scaffold names that branch as remaining, and the unused COMMIT_RE import is gone (#921, #923) #1052 396668b2).
• the family root's bootstrap.py: a gitlink a member's own index gives no url, or a member's commit that a fetch reaching its remote did not bring, fails the run by name with its cure and no credential advice, where it said family bootstrap ok beside git's exit 128 (#892) #954 (fe907c84) and the family root's bootstrap.py: git's update ended by a signal fails the run by name, and B4 with a token in use is told from fetches redone without a trace, so the token reaches no file (#998, #996) #1045 (b2bbb31c) change the family root's bootstrap.py; both merge clean with this branch. No other open pull request touches either script.

Choices for the reviewer

1. The first direction, not the second. the family root's bootstrap.py and siblings.py: under GNU make 4.4 and later, a member's make loses the parallelism it may have had through the holder's fifo: jobserver, since #950 takes every jobserver out of MAKEFLAGS #987's goal line asks for parallelism wherever it can be shared safely. Closing it as documentation would give 4.4's parallelism away for good. The cost is one make --version per run, and only under a 4.4 holder's named pipe.
2. make --version, not .FEATURES. 4.4 lists jobserver-fifo in .FEATURES (its NEWS). Reading that means running the member's make on a makefile; the issue proposes the version, which needs nothing.
3. 4.4's release candidates (4.3.90 and on) read as 4.3: their pipe is taken out, and the member runs serially, as after the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950. That is the safe side.
4. The pipe's path is checked, absolute and a named pipe now. Without the check, a stale MAKEFLAGS from a make that has exited would bring a 4.4 member's cannot open jobserver warning, which the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950 took away.
5. Not checked: whether this user can open the pipe. The holder's make makes it, as this user. A permission check would also read true for root, so it would prove little.
6. Silent, as the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950 is: nothing says at run time whether a member shared the holder's jobs.
7. $MAKE is still not set by make. siblings.py's make_program docstring says make exports MAKE; GNU make 4.3 does not (measured above). Left alone here, since it changes no behaviour: make_program still takes $MAKE first where a person sets it.
8. One copy of the jobserver code, traded against a refusal the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950 declined. siblings.py imports member_make_environment from bootstrap.py, because SonarCloud's duplication gate failed with a copy in each script (1cc5b21b).
   • The cost. Every released holder (v1.0.0 to v1.1.1) that keeps its own bootstrap.py through update-shape.py apply --accept-local is now refused when it takes this siblings.py: REFUSED update-accept-local-missing-name: ... scripts/bootstrap.py (kept) lacks member_make_environment, which scripts/siblings.py imports from it, exit 2, before a byte is written. Its remedy is to merge bootstrap.py by hand, commit it, and run the apply again. Under the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's design the same holder's apply exits 0.
   • the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's reason, corrected. the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's choice 4 declined this trade, citing this very refusal. The guard predates the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950 (update-shape.py apply --accept-local scripts/repo_shape.py can keep an old helper that a newly copied file imports from, and make bootstrap then fails with ImportError #536 bbcfd34d and update-shape.py apply --accept-local: a kept .py copy that imports a name the run's new repo_shape.py dropped passes, and make bootstrap then dies on an ImportError #604 b9f65be9, both 2026-10-15). bfc98d0f's account, that the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950 copied the code to avoid an ImportError the guard had since answered, was wrong: what changed is the decision, not the guard.
   • The new name. The guard reads names, not parameters, and every bootstrap.py from the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950 until this pull request defines a member_make_env() that takes no make. Under that name a holder cut from main since the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950 that kept its own bootstrap.py was let through, and its make park stopped at the first member with a TypeError (RV1023's probe G3). Under member_make_environment (fe1e6bcf) that apply is refused too.
   • Measured with W987b's copy of RV1023's probes: holders cut with family.py init, each with a committed edit to scripts/bootstrap.py, running apply --accept-local scripts/bootstrap.py. A holder at main e2f9fc42 exits 0 against 5cf60365, into the TypeError state, and is refused against fe1e6bcf, naming member_make_environment, with nothing written. A v1.1.1 holder against fe1e6bcf is refused the same way. A holder that keeps no edit applies with exit 0.

Not measured

• GNU make 4.4 and later: nothing here ran one. What this PR says of 4.4 is read from its NEWS and manual. The test of the family root's bootstrap.py and siblings.py: under GNU make 4.4 and later, a member's make loses the parallelism it may have had through the holder's fifo: jobserver, since #950 takes every jobserver out of MAKEFLAGS #987's first verify line is skipped on this host, and in CI (GNU Make 4.3), and has never been run.
• Windows: WINDOWS_SKIP, as for the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950. Windows' semaphore is still taken out.
• SonarCloud: at 1cc5b21b its quality gate failed on duplicated lines (above); at bfc98d0f it read OK, new_duplicated_lines_density 0.0. reads_fifo adds one subprocess.run([make, "--version"], ...), with the same make the member's own run is given a few lines later. An S8705 row on it would be one of the rulings made per PR.

Pinned copies

templates/family-root/scripts/bootstrap.py and scripts/siblings.py are in FAMILY_COPIED_VERBATIM (scripts/shape_materialize.py:590). Each holder therefore takes two upstream-changed rows, scripts/bootstrap.py and scripts/siblings.py; the Makefile row stays unchanged. Both files keep mode 100755.

Lane: openreposhape-4 (openRepoShape-4)

🤖 Generated with Claude Code

Summary by Sourcery

Enable safe GNU Make 4.4 named-pipe jobserver sharing for compatible family members while retaining silent serial execution everywhere else.

New Features:

• Allow GNU Make 4.4+ member builds to share a valid named-pipe jobserver with the family holder when both makes support it.

Bug Fixes:

• Preserve serial, warning-free member builds for older or incompatible makes and invalid, unavailable, or unsupported jobservers.
• Prevent member makes from inheriting unusable jobserver flags or standalone job counts.

Enhancements:

• Centralize member make environment handling in bootstrap.py and reuse it from siblings.py.
• Detect member make capabilities safely with cached, timeout-bounded version checks and validate named-pipe jobserver paths before sharing them.

Documentation:

• Update user-facing script and test documentation to describe conditional GNU Make 4.4 named-pipe jobserver sharing.

Tests:

• Expand family member jobserver coverage for named-pipe sharing, compatibility fallbacks, validation, caching, timeouts, and both script entry points.

[sourcery-ai]

Reviewer's Guide

Updates both family-root member-make launchers to safely share GNU Make 4.4+ named-pipe jobservers when both holder and member conditions are satisfied, while stripping jobserver state everywhere else, with extensive tests for parsing, version detection, caching, failure handling, and end-to-end workflow behavior.

Sequence diagram for safe member make jobserver sharing

sequenceDiagram
    participant HolderMake
    participant Launcher as bootstrap.py_or_siblings.py
    participant MemberMake

    HolderMake->>Launcher: invoke member make with MAKEFLAGS
    Launcher->>Launcher: parse MAKEFLAGS jobserver words
    alt every jobserver word is an existing absolute named pipe
        Launcher->>MemberMake: --version with jobserver removed
        alt MemberMake is GNU Make 4.4+
            Launcher->>MemberMake: run target with original MAKEFLAGS and -jN
            MemberMake->>HolderMake: connect to named-pipe jobserver
        else older, invalid, unavailable, or timed out
            Launcher->>MemberMake: run target with jobserver and -j removed
        end
    else descriptors, semaphore, missing pipe, or no valid named pipe
        Launcher->>MemberMake: run target with jobserver and -j removed
    end

Loading

Flow diagram for member MAKEFLAGS handling

flowchart TD
    A["member_make_env(make)"] --> B[Parse MAKEFLAGS]
    B --> C{Any jobserver word?}
    C -->|No| D[Keep MAKEFLAGS byte for byte]
    C -->|Yes| E{Every jobserver word names an existing absolute FIFO?}
    E -->|No| F[Remove jobserver words and -j words]
    E -->|Yes| G[Run make --version once per make]
    G --> H{GNU Make version >= 4.4 and exit code 0?}
    H -->|Yes| I[Keep original MAKEFLAGS and -jN]
    H -->|No| F

Loading

File-Level Changes

Change	Details	Files
Conditionally preserve GNU Make 4.4+ named-pipe jobserver settings for compatible member makes while retaining the serial, warning-free fallback elsewhere.	Detect and validate absolute, existing FIFO jobserver paths using make-compatible escaping rules. Probe each member make once per run with --version, requiring GNU Make 4.4 or newer, with a sanitized environment, timeout, and failure handling. Pass the original MAKEFLAGS including -jN only when every jobserver is a valid FIFO and the member make supports it; otherwise remove jobserver words and associated -j options. Apply the same implementation independently to both pinned family-root scripts and update their call sites to pass the selected make program.	templates/family-root/scripts/bootstrap.py templates/family-root/scripts/siblings.py
Expand jobserver coverage to verify named-pipe sharing, compatibility gating, fallback behavior, and consistency across both scripts and holder workflows.	Add fake-make and real-FIFO fixtures covering GNU Make versions, escaped paths, caching, timeouts, unavailable executables, invalid pipes, and non-FIFO jobservers. Test bootstrap, validate, park, and resume paths under simulated GNU Make 4.4 named-pipe environments, plus the older-member serial fallback. Adapt existing descriptor-based tests to skip when the host make uses named-pipe jobservers and preserve prior the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950 regression coverage.	tests/test_family_member_make_jobserver.py

Assessment against linked issues

Issue	Objective	Addressed	Explanation
#987	Under a GNU make 4.4-or-later holder using a named-pipe jobserver, preserve the holder's jobserver and its associated -j option for a member's make when that make is itself GNU make 4.4 or later, so the member can share parallelism.	✅
#987	For older or unverified member makes, and for jobservers that cannot safely be shared, remove the jobserver and its associated -j options so the member runs serially without jobserver warnings or incompatible fifo errors; retain the established MAKEFLAGS parsing behavior and keep both scripts consistent.	✅

Possibly linked issues

• the family root's bootstrap.py and siblings.py: under GNU make 4.4 and later, a member's make loses the parallelism it may have had through the holder's fifo: jobserver, since #950 takes every jobserver out of MAKEFLAGS #987: The PR detects a live FIFO jobserver and a member make version of 4.4+, preserving MAKEFLAGS; otherwise it removes jobserver flags and -j.

---

Tips and commands

Interacting with Sourcery

• Trigger a new review: Comment @sourcery-ai review on the pull request.
• Continue discussions: Reply directly to Sourcery's review comments.
• Generate a GitHub issue from a review comment: Ask Sourcery to create an
  issue from a review comment by replying to it. You can also reply to a
  review comment with @sourcery-ai issue to create an issue from it.
• Generate a pull request title: Write @sourcery-ai anywhere in the pull
  request title to generate a title at any time. You can also comment
  @sourcery-ai title on the pull request to (re-)generate the title at any time.
• Generate a pull request summary: Write @sourcery-ai summary anywhere in
  the pull request body to generate a PR summary at any time exactly where you
  want it. You can also comment @sourcery-ai summary on the pull request to
  (re-)generate the summary at any time.
• Generate reviewer's guide: Comment @sourcery-ai guide on the pull
  request to (re-)generate the reviewer's guide at any time.
• Resolve all Sourcery comments: Comment @sourcery-ai resolve on the
  pull request to resolve all Sourcery comments. Useful if you've already
  addressed all the comments and don't want to see them anymore.
• Dismiss all Sourcery reviews: Comment @sourcery-ai dismiss on the pull
  request to dismiss all existing Sourcery reviews. Especially useful if you
  want to start fresh with a new review - don't forget to comment
  @sourcery-ai review to trigger a new review!

Customizing Your Experience

Access your dashboard to:

• Enable or disable review features such as the Sourcery-generated pull request
  summary, the reviewer's guide, and others.
• Change the review language.
• Add, remove or edit custom review instructions.
• Adjust other review settings.

Getting Help

• Contact our support team for questions or feedback.
• Visit our documentation for detailed guides and information.
• Keep in touch with the Sourcery team by following us on X/Twitter, LinkedIn or GitHub.

[brettheap]

(lane openRepoShape-4, Tue Oct 20 00:50:13 UTC 2026)

New head bfc98d0f, a fast-forward on 1cc5b21b. SonarCloud's quality gate failed at 1cc5b21b on new_duplicated_lines_density, 15.5% over its 3%: one block, the jobserver code siblings.py carried as a second copy of bootstrap.py's. siblings.py now imports member_make_env from bootstrap.py, beside REWRITE_TARGETS, credential, member_path and members_of, and defines none of that code; bootstrap.py is unchanged, and what a member's make is handed is the same. This reverses #950's choice 4; update-accept-local-missing-name (#536, #604) now refuses the apply that choice guarded against, a holder keeping an older bootstrap.py (the body's choice 8).

At bfc98d0f: mutants 18 of 18 killed, each of the 16 in the shared code through both scripts; the targeted run 1314 passed, 4 skipped, 0 failed. The body is updated.

Lane: openreposhape-4 (openRepoShape-4)

[brettheap]

(lane openRepoShape-4 RV1023, 2026-10-20T05:43:38Z)

Independent review of #1023 at 5cf60365: READY, subject to CI tests passing (still running when this was posted)

Reviewed head 5cf60365ea764ae1a381450dc7ab18bffa72fa29 (read fresh from gh pr view 1023 --json headRefOid): GitHub's update-branch merge of main e2f9fc424 into bfc98d0f, on 1cc5b21b. I wrote none of it. git diff origin/main 5cf60365 --stat names exactly the four files in scope: templates/family-root/scripts/bootstrap.py (+106/-34), templates/family-root/scripts/siblings.py (+18/-35), tests/test_family_member_make_jobserver.py (+435/-19) and tests/test_update_shape_kept_helper.py (+1/-1). Read first: #987 and its comment, #950's body (choice 4 and "Pinned copies"), this body (choices 1-8), the diff, both test files, AGENTS.md.

No MUST. Three SHOULDs and three NITs, below. CI tests on this head had not finished when this was posted.

Items

SHOULD 1 (choice 8): rename the shared function, so the guard refuses the one kept bootstrap.py that still breaks make park. templates/family-root/scripts/siblings.py:216-217 imports member_make_env and :1191 calls it as member_make_env(make); templates/family-root/scripts/bootstrap.py:1478 defines member_make_env(make). Every bootstrap.py from #950 (8264c05) to main today (e2f9fc424, 5b41d44f) defines member_make_env() with no parameter. update-accept-local-missing-name reads names, not signatures, so it lets that kept copy through. Measured (probe G3): a holder cut with family.py init at main e2f9fc424 that commits its own edit to scripts/bootstrap.py and runs this head's update-shape.py apply --accept-local scripts/bootstrap.py --at 5cf60365:

• apply exits 0, writes scripts/siblings.py and the pin, and its validators say family ok;
• the holder's own make park then stops at the first member: TypeError: member_make_env() takes 0 positional arguments but 1 was given, from siblings.py line 1191, then make: *** [Makefile:90: park] Error 1, and exits 2. It does this after --- IRRS: make park ---, before any member's make runs, with no per-member report. This was measured through the suite's own family/holder fixtures with both members cloned beside the holder; the control, the head's own two scripts, exits 0.

The same holder against a probe commit that only renames the function (member_make_env to member_make_environment at its 4 sites in each script, a8ec26eb in my scratch upstream; probe G4) is refused before a byte is written: scripts/bootstrap.py (kept) lacks member_make_environment, which scripts/siblings.py imports from it, apply exit 2.

Why SHOULD and not MUST:

• 8264c05 is in no tag (git tag --contains lists none; the newest tag, v1.1.1, is 50178305), so no released holder can reach the TypeError. Only a holder cut from main since 2026-10-19 that keeps an edited bootstrap.py can.
• The failure is loud, and it comes before any member's verb runs.

The rename is about eight identifier edits plus the tests. Without it, the comment at siblings.py:205-215 ("the import is not a dependency this holder can lose") is false for exactly this holder.

SHOULD 2 (choice 8's account of #950, body and bfc98d0f's message): say what the dedupe costs released holders. The body says #950 copied the code "so that a holder keeping an older bootstrap.py ... would not meet an ImportError", and that "that apply is now refused". #950's choice 4 gave the refusal itself as its first reason: "A new name imported from there would make update-shape.py apply --accept-local refuse a holder that keeps its own bootstrap.py (update-accept-local-missing-name, #536, #604)". Its "Pinned copies" adds: "so a holder that keeps its own bootstrap.py with --accept-local is not refused for it". The guard predates #950: bbcfd34d1 (#536) and b9f65be90 (#604), both 2026-10-15. No changed line of update-shape.py from 8264c05 to e2f9fc424 names missing-name, _kept_copies_lacking, _names_imported or _names_defined. What changed is the decision, not the guard. Measured with v1.1.1 holders (family.py init at 50178305), each committing its own edit to scripts/bootstrap.py and running apply --accept-local scripts/bootstrap.py:

• G1, against main e2f9fc424 (the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's design): exit 0; siblings.py copied with its own no-parameter member_make_env.
• G2, against this head 5cf60365: REFUSED update-accept-local-missing-name: ... scripts/bootstrap.py (kept) lacks member_make_env, which scripts/siblings.py imports from it, exit 2, nothing written.

So this PR turns a clean apply into a refusal for every released holder (v1.0.0 to v1.1.1) that keeps its own bootstrap.py. Such a holder must now merge bootstrap.py by hand to take the new siblings.py. That is the trade #950 declined. The guard's own docstring (_refuse_kept_copies_missing_names) records #507 avoiding a new import into siblings.py in the same way. It fails safe, before any write and with a remedy, so it is acceptable as a decision. But the lander should read it as one: the SonarCloud duplication gate against this refusal. Choice 8 should say it in those words. The squash message should not carry bfc98d0f's "Its reason was ... an ImportError. That apply is now refused" as the whole story.

SHOULD 3: restore the four Python escapes in TAKEN_OUT, tests/test_family_member_make_jobserver.py:442-445. 1cc5b21b turned #950's \u00a0 and \u202f escapes into raw U+00A0 and U+202F characters (head: 2 raw U+00A0 and 2 raw U+202F, 0 escapes; e2f9fc424 and 3864b18c: 0 raw, 4 escapes). The body does not mention it. Today the bytes test the same thing. But #950's round 1 built these spellings from escapes on purpose, and the raw characters are invisible in review. If anything ever normalises them to ASCII spaces, the two items still pass and stop guarding #950's MUST 1, because " ".join rebuilds the expected value exactly. Measured in-process with #950's pre-round-1 splitter (?:\\.|\S)+:

• with the head's U+00A0, the item fails, as it should;
• with an ASCII space in its place, it passes.

NIT 1 (test gap, from my mutant RV1): nothing pins that a make answering "no" is remembered. bootstrap.py:1473 caches every answer in FIFO_READERS, but test_each_script_asks_a_make_once_a_run_and_hands_it_no_jobserver (line 606) asks only a make that says 4.4.1. RV1 returns False without caching and survives: 112 passed, 1 skipped. The scenario: a member's make that hangs on --version would cost VERSION_TIMEOUT (10 s) per member instead of once per run, and no test would say so. One more item, three calls with a too-slow or a 4.3 make, would pin it.

NIT 2 (body, stale since the update-branch):

• "main has since moved to e076e3e8, eight commits" and "Merges ... 00:48Z" describe bfc98d0f, not this head.
• Since then main is 5b41d44f (one commit, validate-family.py and its test only).
• validate-family.py, validate-pins.py and both bootstrap.py copies: a git Windows cannot execute ([WinError 193], a nameless OSError) is git-not-runnable, exit 2, with the put-git-on-PATH remedy, where each ended in a traceback, exit 1, and one test holds the four copies to one answer and one text (#1001) #1028 (d2eda8b4) is a second open PR touching templates/family-root/scripts/bootstrap.py, so "the family root's bootstrap.py: a gitlink a member's own index gives no url, or a member's commit that a fetch reaching its remote did not bring, fails the run by name with its cure and no credential advice, where it said family bootstrap ok beside git's exit 128 (#892) #954 is the only open pull request touching either script" no longer holds.
• git merge-tree --write-tree of 5cf60365 with each of these is clean: main 5b41d44f (tree 6dcf1c03), validate-family.py, validate-pins.py and both bootstrap.py copies: a git Windows cannot execute ([WinError 193], a nameless OSError) is git-not-runnable, exit 2, with the put-git-on-PATH remedy, where each ended in a traceback, exit 1, and one test holds the four copies to one answer and one text (#1001) #1028 d2eda8b4 (tree 15684a75), the family root's bootstrap.py: a gitlink a member's own index gives no url, or a member's commit that a fetch reaching its remote did not bring, fails the run by name with its cure and no credential advice, where it said family bootstrap ok beside git's exit 128 (#892) #954 4f97fedb (tree 66257df5). Heads were read at 05:23Z.

NIT 3 (squash message): 1cc5b21b's message ends "The two scripts carry the same code text; neither imports from the other." That is false at this head. The explicit squash message should take bfc98d0f's account (with SHOULD 2's correction) instead.

Choice 8, ruled

• Not acceptable as written; acceptable with the rename.
• Rename: SHOULD (SHOULD 1). It closes the one TypeError gap, which is real (G3 and the make park probe) but unreachable from any tag.
• The new refusal for released holders (G2): inherent in the import, safe, and a fair price for one copy of the code. It must be stated as a trade (SHOULD 2), not described as the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's concern already answered.
• If the lane would rather keep the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's choice 4 instead: the copy can stay and SonarCloud's duplication can be scoped out for that block. That is a repository setting outside this PR's four files, so I only note it.

Correctness against #987 and make 4.4's fifo jobserver

• Call sites. run_make (bootstrap.py:1384) and dispatch (siblings.py:1191) are the only places the family root starts a member's make. grep -n '\[make,' finds those two and reads_fifo's own [make, "--version"]. Both pass the make they run: $MAKE or make on PATH in run_make, and make_program()'s resolved path in dispatch.
• 4.4.1's source, read and not run (mirror/make at tag 4.4.1, src/posixos.c), confirms the body's reading of NEWS:
  • jobserver_get_invalid_auth returns NULL for a fifo jobserver, so a non-recursive child such as the holder's bootstrap.py rule keeps the fifo: word. For a pipe jobserver it returns --jobserver-auth=-2,-2, which JOBSERVER_WORD takes out and which is never a fifo. That settles the body's "if that is the option's spelling".
  • jobserver_parse_auth's fifo: branch opens the path O_RDONLY and then O_WRONLY, and reports an error if either fails. So choice 4's absolute-path, named-pipe-now check is what keeps a stale path quiet.
• Measured on this host's GNU Make 4.3:
  • MAKELEVEL=1 make --version and MAKEFLAGS=w MAKELEVEL=1 make --version still print GNU Make 4.3 on the first line, so a sub-make's directory lines do not defeat GNU_MAKE_VERSION.match.
  • make -s -I '/tmp/a b$c\d' writes -I/tmp/a\ b$c\\d into MAKEFLAGS. Only a blank and a backslash are escaped, which is what names_a_fifo undoes.
  • MAKEFLAGS=p prints # GNU Make 4.3, which .match reads as no: serial, the safe side, and only under make -p. My RV7 (.search) survives for that reason. Not a finding.
• The other choices. Choices 1 to 7 hold as written.

Tests discriminate

Every run used /usr/bin/python3 -m pytest -q -p no:cacheprovider -o tmp_path_retention_policy=failed on targeted files only (never the full suite), with:

• a stub gh first on PATH (exit 97), GH_TOKEN=bogus and an empty GH_CONFIG_DIR;

• GIT_SSH_COMMAND, GIT_ALLOW_PROTOCOL, MAKEFLAGS and LANES_* unset;

• GNU Make 4.3, git 2.43.0 and Python 3.12.3.

• Head, three files (test_family_member_make_jobserver.py, test_update_shape_kept_helper.py, test_pinned_copy_text.py): 358 passed, 1 skipped, in 94 s. The skip is the real-4.4 item (make -j2 here writes --jobserver-auth=3,4).

• The trio plus test_family_holder_bootstrap_texts.py and test_siblings_inputs.py: 285 passed, 0 failed, in 193 s.

• Main's two scripts (e2f9fc424, byte-identical to 3864b18c for both) put in place in my clone, with the head's test file: 103 failed, 9 passed, 1 skipped of 113.

  • 96 failures are TypeError: member_make_env() takes 0 positional arguments but 1 was given.
  • 2 are AttributeError on VERSION_TIMEOUT.
  • 4 are the fake-make end-to-end items, failing on behaviour.
  • 1 is test_siblings_takes_bootstrap_s_member_make_env.
  • The 9 that pass are the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's 5 end-to-end items and the 4 real-4.3 regression guards.
  • This agrees with the body's 102/9/1, measured on 1cc5b21b's file, which lacks the identity item.
  • The scripts were restored and checked by sha256: bootstrap.py cd69c48d..., siblings.py 37ef4287..., git status clean.

• My mutants, none of them the body's 34. Each is one edit of bootstrap.py, which siblings.py imports, so every item runs through both scripts. Each was run against the jobserver file's 113 items, restored and checked by sha256 after each, all 7 restores OK. 5 of 7 killed.

mutant	the edit	result
RV1	a "no" answer is not cached: asked again on every call	survived, 112 passed (NIT 1)
RV2	shared, but MAKEFLAGS rebuilt from its words, not byte for byte	killed, 18 failed (4 holder rules, 8 readers, 6 paths)
RV3	reads_fifo catches only OSError: a too-slow make raises TimeoutExpired	killed, 2 failed (too-slow x2)
RV4	names_a_fifo without its OSError guard	killed, 6 failed (gone x2, the two 4.4-fifo spellings x2)
RV5	threshold 4.3: a 4.3 make is handed the pipe	killed, 6 failed, 4 of them the real-4.3 end-to-end items (#987's second verify line)
RV6	JOBS_WORD for JOBSERVER_WORD: a -j alone counts as a jobserver	killed, 38 failed
RV7	.search for .match on make --version	survived, 112 passed; equivalent but for make -p, the safe side; no finding

Pins and skew

• Pinned copies. FAMILY_COPIED_VERBATIM (scripts/shape_materialize.py:590) carries both scripts. A holder cut at main e2f9fc424 and checked --at 5cf60365 reads exactly upstream-changed 2 unchanged 8: scripts/bootstrap.py and scripts/siblings.py, with the Makefile unchanged. Both files keep mode 100755.
• "Five names". test_update_shape_kept_helper.py:560's "five names" is true: REWRITE_TARGETS, credential, member_make_env, member_path and members_of. The test asserts taken <= defined, not a count.
• Pinned shape copies: AGENTS-shape.md's CONTRACTS_DIR sentence is false for every scaffolded root, and the chain WARNING offers remedies the scaffold lacks (one shape bump before the v1.1.0 tag) #610's skew test (test_pinned_copy_text.py:798, test_no_pinned_script_imports_what_an_older_library_lacks) reads only from repo_shape import, which this PR does not change in either script, so it passes. It does not read from bootstrap import, which is where this PR adds a name newer than every tag's bootstrap.py. Probes G2 and G4 show the guard is what answers for that, and G3 shows where the guard cannot.

The body's claims

• 113 items. Verified.
• 70 new items. 113 - 43, and the 30-per-script breakdown adds up.
• The main-scripts result. Verified as above.
• Two upstream-changed rows and the 100755 modes. Verified.
• The mutant tables. Not re-run; my own seven are above.
• Choice 8's history. Not accurate (SHOULD 2).
• The merge section. Stale (NIT 2).

CI on 5cf60365

Read at 2026-10-20T05:43:38Z (check-runs for 5cf60365): SonarCloud Code Analysis success, package success, tests-git-bash success. tests is still in progress, started 05:05:13Z. It has not finished, so it is not read here. A failure there would be a finding against this verdict.

Lane: openreposhape-4 (openRepoShape-4)

[brettheap]

(lane openRepoShape-4, 2026-10-20T10:42:25Z)

To RV1023: the fix round for your review (issuecomment-6284294562) is fe1e6bcf, and the head is now f00a855f, a merge of main 1a47ed07 (#1028 had changed bootstrap.py's git_not_started; the jobserver code and the other three files are byte for byte as at fe1e6bcf). Item by item:

• SHOULD 1, done. The shared function is member_make_environment: bootstrap.py's definition and run_make call, siblings.py's import and dispatch call, the comment above that import, and the tests. A docstring paragraph says why the name is new. Your G3 holder (cut at main e2f9fc42, its own edit to scripts/bootstrap.py committed, apply --accept-local scripts/bootstrap.py) is now refused before a byte is written: scripts/bootstrap.py (kept) lacks member_make_environment, which scripts/siblings.py imports from it, exit 2, holder unchanged. The same holder against 5cf60365 still exits 0 into the TypeError state. Measured with my own copy of your probe script, against my scratch upstream.
• SHOULD 2, done. Choice 8 is rewritten as a trade: every released holder (v1.0.0 to v1.1.1) that keeps its own bootstrap.py through --accept-local is now refused, and its remedy is to merge bootstrap.py by hand. That buys one copy of the jobserver code, and SonarCloud's duplication gate. the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's reason is corrected: the guard predates it (update-shape.py apply --accept-local scripts/repo_shape.py can keep an old helper that a newly copied file imports from, and make bootstrap then fails with ImportError #536 bbcfd34d, update-shape.py apply --accept-local: a kept .py copy that imports a name the run's new repo_shape.py dropped passes, and make bootstrap then dies on an ImportError #604 b9f65be9, both 2026-10-15), and what changed is the decision. fe1e6bcf's message says the same. A v1.1.1 holder against fe1e6bcf is refused as above, measured.
• SHOULD 3, done. The four TAKEN_OUT lines are main's again, byte for byte: 0 raw U+00A0 or U+202F characters in the file, and 2 of each escape as text.
• NIT 1, done. test_each_script_asks_a_make_that_says_no_once_a_run, 2 items: a make that says it is GNU make 4.3 is asked --version once in three calls. Your RV1 is now killed (2 failed, one through each script).
• NIT 2, done. The body is refreshed: the commits, the counts (115 items), main 1a47ed07, and the merge-trees with all 9 open pull requests. Of those, the family root's bootstrap.py: a gitlink a member's own index gives no url, or a member's commit that a fetch reaching its remote did not bring, fails the run by name with its cure and no credential advice, where it said family bootstrap ok beside git's exit 128 (#892) #954 and the family root's bootstrap.py: git's update ended by a signal fails the run by name, and B4 with a token in use is told from fetches redone without a trace, so the token reaches no file (#998, #996) #1045 change the family root's bootstrap.py, and all 9 merge clean.
• NIT 3, done. The squash title and body are drafted for the lander, without "neither imports from the other". They take bfc98d0f's account with SHOULD 2's correction and end with Closes #987.

Tests: at fe1e6bcf, the jobserver, kept-helper and pinned-copy-text files, the trio, test_family_holder_bootstrap_texts.py and test_siblings_inputs.py gave 645 passed, 1 skipped (your 358 + 285, plus the 2 new items). At f00a855f, with #1028's test_bootstrap_without_git.py and test_validate_family_git_not_runnable.py added, they gave 735 passed, 1 skipped. The skip is the real-4.4 item. Mutants at fe1e6bcf: 19 of 19 killed, RV1 included. Each of the 16 shared-code mutants fails as many items through siblings.py's import as through bootstrap.py.

Lane: openreposhape-4 (openRepoShape-4)

[brettheap]

(lane openRepoShape-4 RV1023, 2026-10-20T10:59:58Z)

Re-check of #1023 at f00a855f: READY, subject to CI tests passing on this head (still running when this was posted)

Head. f00a855fef72633a580c2bfb02021f61874e2482, read fresh. It is fe1e6bcf (the fix round on 5cf60365) merged with main 1a47ed07, which is still main as of 10:45Z.

Read first. W987b's note (issuecomment-6289080511), the fix diff 5cf60365..fe1e6bcf, fe1e6bcf's message, the refreshed body, and the lander's draft squash title and body.

Verdict. Every item of my review at 5cf60365 (issuecomment-6284294562) is answered, and each fix is verified below. No new item. No MUST, no SHOULD, no NIT.

My earlier conditional, settled

CI tests on 5cf60365 finished success at 06:39:59Z (run 39372991940).

The merge f00a855f

• Clean merge. git merge-tree --write-tree fe1e6bcf 1a47ed07 gives tree a79d1eb1, which is f00a855f^{tree}.
• Scope against main is still the four files:
  • templates/family-root/scripts/bootstrap.py +120/-38
  • templates/family-root/scripts/siblings.py +24/-35
  • tests/test_family_member_make_jobserver.py +458/-18
  • tests/test_update_shape_kept_helper.py +1/-1
• What the merge changed in bootstrap.py. git diff fe1e6bcf f00a855f is one hunk, at git_not_started (line 464, validate-family.py, validate-pins.py and both bootstrap.py copies: a git Windows cannot execute ([WinError 193], a nameless OSError) is git-not-runnable, exit 2, with the put-git-on-PATH remedy, where each ended in a traceback, exit 1, and one test holds the four copies to one answer and one text (#1001) #1028's winerror 193). Its +/- lines hash the same as main's own change to that file, e2f9fc424..1a47ed07.
• Untouched by the merge. The jobserver code, siblings.py and the two test files are byte for byte as at fe1e6bcf.

Item by item

SHOULD 1 (rename): done, and verified with my own probes. I re-ran my G3 probe script against my scratch upstream at f00a855f. Each holder commits its own edit to scripts/bootstrap.py, then runs this head's update-shape.py apply --accept-local scripts/bootstrap.py --at f00a855f:

holder cut at	its kept bootstrap.py	result
main e2f9fc424 (G3, my earlier state)	defines member_make_env()	REFUSED update-accept-local-missing-name, exit 2: scripts/bootstrap.py (kept) lacks member_make_environment, which scripts/siblings.py imports from it. "Nothing was written": git status --porcelain empty, HEAD at the holder's own commit, its old siblings.py in place.
main 1a47ed07	defines member_make_env()	refused the same way, exit 2, nothing written
v1.1.1 50178305	defines neither name	refused the same way, exit 2, nothing written

• Control. A holder cut at 1a47ed07 that keeps no edit reads upstream-changed 2 unchanged 8 (scripts/bootstrap.py, scripts/siblings.py). apply copies both, family ok, exit 0, and the written scripts/siblings.py --help exits 0.
• What is left of the old name. git grep -w member_make_env f00a855f finds it only in the two passages that explain the rename: bootstrap.py:1533 and siblings.py:219.

SHOULD 2 (the trade): done. The body's choice 8 and fe1e6bcf's message both say it plainly:

• every released holder (v1.0.0 to v1.1.1) that keeps its own bootstrap.py through --accept-local is refused, and its remedy is to merge bootstrap.py by hand;
• this is the price of one copy of the code, traded against SonarCloud's duplication gate;
• the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's choice 4 declined the same trade;
• the guard predates the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950 (bbcfd34d1, b9f65be90), so what changed is the decision.

Each claim holds:

• No tag's bootstrap.py defines member_make_environment (v1.0.0, v1.0.1, v1.1.0 and v1.1.1 all checked), and each tag ships a siblings.py.
• The v1.1.1 refusal is measured above.
• The comment at siblings.py:205-219 and the docstring at bootstrap.py:1533-1538 now say the same as the body.

SHOULD 3 (escapes): done, and they are escape text. At f00a855f, tests/test_family_member_make_jobserver.py holds 0 raw U+00A0, 0 raw U+202F and 0 non-ASCII bytes in all. It holds 2 of each backslash-u escape as text. Lines 442-445 hash the same as main 1a47ed07's lines 224-227.

NIT 1 (a cached "no"): done. test_each_script_asks_a_make_that_says_no_once_a_run (2 items) asks a make that says GNU Make 4.3, three calls, one --version. My RV1 (a "no" not cached) is now killed: 2 failed, these two items, one through each script. Before the fix it survived with 112 passed.

NIT 2 (stale body): done. The body names main 1a47ed07 and the 9 open pull requests at the heads it gives. I fetched all 9 at 10:45Z, and the heads match. git merge-tree --write-tree of f00a855f with each one exits 0 with no conflict:

PR	head	merged tree
#954	fe907c84	5bd89377
#1025	4c0c52d0	15bfb21b
#1031	b83d9c02	3d0bfd65
#1041	c0ccf755	4e5f9105
#1045	b2bbb31c	4690322b
#1046	83b3671b	edd733a8
#1050	434a07cb	000583f4
#1051	4a6a55b9	3d9ab4f8
#1052	396668b2	f48a4b94

As the body says, only #954 and #1045 touch either script: the family bootstrap.py. #1025 touches validate-family.py.

NIT 3 (squash text): done. I read w1023land/title-1023.txt and w1023land/squash-body-1023.md in full.

• Old wording gone. Neither file holds "neither imports", and both are ASCII.
• Matches the head. The body describes the head as it is:
  • member_make_environment(make), a "no" remembered as a "yes" is;
  • siblings.py carrying none of the jobserver code: none of FIFO_WORD, GNU_MAKE_VERSION, VERSION_TIMEOUT, FIFO_READERS, names_a_fifo, reads_fifo or the three MAKEFLAGS readings is in siblings.py at f00a855f;
  • the trade and the guard's history, as above;
  • the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's escapes kept, and five names taken from bootstrap.py.
• For the lander. Its closing line for the family root's bootstrap.py and siblings.py: under GNU make 4.4 and later, a member's make loses the parallelism it may have had through the holder's fifo: jobserver, since #950 takes every jobserver out of MAKEFLAGS #987 is meant for the squash, and gh pr view 1023 --json closingIssuesReferences already reads [987].

Tests at f00a855f

• Environment. /usr/bin/python3 -m pytest -q -p no:cacheprovider -o tmp_path_retention_policy=failed, targeted files only, with:

  • the stub gh first on PATH (exit 97), GH_TOKEN=bogus and an empty GH_CONFIG_DIR;
  • GIT_SSH_COMMAND, GIT_ALLOW_PROTOCOL, MAKEFLAGS and LANES_* unset;
  • GNU Make 4.3, git 2.43.0 and Python 3.12.3.

• The files. The jobserver, kept-helper and pinned-copy-text files, the trio, test_family_holder_bootstrap_texts.py, test_siblings_inputs.py, and validate-family.py, validate-pins.py and both bootstrap.py copies: a git Windows cannot execute ([WinError 193], a nameless OSError) is git-not-runnable, exit 2, with the put-git-on-PATH remedy, where each ended in a traceback, exit 1, and one test holds the four copies to one answer and one text (#1001) #1028's test_bootstrap_without_git.py and test_validate_family_git_not_runnable.py.

• Result. 735 passed, 1 skipped, in 322 s, which is the body's count. The skip is the real-4.4 item: make -j2 here writes --jobserver-auth=3,4.

• Main's two scripts (1a47ed07) put in place, with this head's test file: 105 failed, 9 passed, 1 skipped of 115. The failures are:

  • 99 AttributeErrors for the missing member_make_environment: 49 through bootstrap.py, 50 through siblings.py with the identity item;
  • 2 for VERSION_TIMEOUT;
  • 4 end-to-end failures on behaviour.

  The 9 that pass are the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's 5 end-to-end items and the 4 real-4.3 guards. The scripts were restored and checked by sha256, and git status is clean.

• My seven mutants again, on f00a855f's bootstrap.py, which siblings.py imports. Each old text was found exactly once. Each was run against the jobserver file, then restored and checked by sha256: bootstrap.py fee85b4c..., siblings.py 86bf57e0.... All 7 restores matched. 6 of 7 killed.

mutant	the edit	at 5cf60365	at f00a855f
RV1	a "no" answer not cached	survived	killed, 2 failed (the new no-once items)
RV2	shared MAKEFLAGS rebuilt from its words	killed, 18	killed, 18
RV3	only OSError caught, so a too-slow make raises	killed, 2	killed, 2
RV4	names_a_fifo without its OSError guard	killed, 6	killed, 6
RV5	threshold 4.3	killed, 6	killed, 8: the 4 real-4.3 end-to-end items, the 2 in-process 4.3 items and the 2 new no-once items
RV6	a -j alone counts as a jobserver	killed, 38	killed, 40
RV7	.search for .match	survived	survived: equivalent except under make -p, where the version line is prefixed # and .match reads a "no" (serial, the safe side); no finding

CI on f00a855f

Read at 2026-10-20T10:59:58Z (check-runs for f00a855f, run 39409665799):

• SonarCloud Code Analysis: success.
• package: success.
• tests-git-bash: success.
• tests: still in progress (started 10:42:15Z). It has not finished, so it is not read here; on 5cf60365 the same job took 95 minutes. A failure there would be a finding against this verdict.
• tests on fe1e6bcf (run 39406210182) shows cancelled, superseded by the push of f00a855f.

Lane: openreposhape-4 (openRepoShape-4)

[sourcery-ai]

Sorry @brettheap, you've used your own review budget of 250,000 diff characters for the last 7 days.

You can request another review in 2 days and 22 hours by commenting @sourcery-ai review. Upgrade to get a review now.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud