24 KiB
scripts/gate — per-worktree build isolation, and one gate suite
Status: revision 5. Approved at revision 4 and IMPLEMENTED; revision 5 records two safety defects review found in the implementation.
Neither was a design gap — both were the implementation failing to honour this document, which is worth stating because it is the case a framing doc cannot prevent by itself:
--prunecould delete every managed directory. §2.6 requires liveness to be established; the code maskedgit worktree listfailure with|| true, so running from outside any repository produced an empty live set — under which every marked directory is an orphan and--prune --forcedeletes all of them, live lanes included. Now: two refusals (not in a worktree; the enumeration failed), and an outside-repo survivor test.--acceptancewas shell-injectable. §2.3 hands the name into a command the runner evaluates; nothing validated it, so--acceptance 'x; rm -rf ~'would have executed. Now: an allowlist of what a cargo target can be named, refused at parse time, with a metacharacter rejection test and a canary.
Also fixed: log directories now carry the PID as well as a
whole-second timestamp (two runs in one second shared a directory and
could overwrite each other's evidence — reintroducing U2/U3 through a
naming choice); the ownership marker is enforced as exactly one line
rather than read head-first; and the prunable test fails loudly
instead of returning green when git worktree add fails, with cleanup
through a Drop guard that survives a panicking assertion.
Developer tooling. Not coherence-affecting — it touches no journey
step, adds no interaction island, adopts no config-registry setting, and
creates no background work (COHERENCE.md §20). It is named here so the
absence is a statement rather than an omission.
Revision 2 answered five review findings: the managed root and ownership marker Q#GS1 needed, the narrowing of Q#GS2 (the script cannot infer touched acceptance suites), Q#GS4's dry-run and eligibility rules, the corrected — conditional — CRDT workspace sweep, and the test-root override.
Revision 4 makes the log runner set -e-safe (the earlier
rc=$? form never executes under set -eu, which scripts/bite
already uses), corrects a stale justification — ambient-root isolation
merged as #206, so the five variables are belt-and-braces for
external and integration paths rather than cover for a missing lane —
and extends the canonical-path rule to directory derivation, which
§2.5 had omitted while §4's symlink test depended on it.
Revision 3 answered three findings, two of which were regressions I
introduced. Rewriting §2 for revision 2 silently dropped two of the
original core responsibilities: ambient-root isolation and durable
sweep logs. The second of those is the U2/U3 remedy, so losing it
would have left this document proposing a fix for a problem it had
stopped solving. Both are restored as §2.2, with the exit-status
hazard in log capture specified rather than left to the implementation.
Prune liveness is made precise in §2.6 — prunable entries and one
canonical path representation — and §4's marker claim is reconciled
with the interface by --init.
1. Why this exists
1.1 The measured problem
Parallel worktrees are about to become the working method. They do not work on this machine today, and the reason is one line of shell config:
set -gx CARGO_TARGET_DIR $HOME/build/cargo-target # fish config.fish:55
Every worktree therefore builds into one directory, and cargo takes an exclusive lock on it. Two lanes building concurrently do not run in parallel — the second blocks — and they invalidate each other's artifacts, so alternating between lanes recompiles from scratch each time. Parallel worktree development under this arrangement is slower than serial.
1.2 The measurement that corrected the first plan
The shared directory was 285G (debug/deps 222G across 12,084
files; debug/incremental 52G). That number drove an initial proposal
to add sccache so per-worktree directories would not lose artifact
sharing.
The number was misleading and the proposal was wrong. 285G is years of accumulation across two projects — pmacs and levcs share that directory. Measured directly, on a cleaned tree:
| wall | size | |
|---|---|---|
cargo build --workspace, cold |
55s | 3.4G |
cargo test --workspace --no-run, cold |
80s | 19G |
And sccache, measured across two target directories:
| hit rate | |
|---|---|
| C/C++ | 50.00% |
| Rust | 0.00% |
Rust rlibs embed their target-dir path in metadata, so dependency
artifacts are not bit-identical between directories; --extern content
hashes differ and misses cascade. sccache does not deliver
cross-worktree Rust reuse without --remap-path-prefix, which would
degrade backtrace paths for every project on the machine.
So the sharing that per-worktree directories cost is worth 80 seconds and 19G per lane. Four lanes is 76G against 604G free. There is nothing to buy back. sccache stays configured — it is cheap and earns its keep on the C/C++ dependencies — but it is not what makes parallel lanes work, and this document exists partly so that claim is not repeated.
1.3 The second problem, which shares a solution
The gate suite in docs/agent-handoff.md §3 is retyped by hand every
time. It has been gotten wrong repeatedly, including twice in the
session that motivated this script:
- A remediation sweep used
--testsinstead of--workspace, silently droppingpmacs_protocolandpmacs_gpu— including protocol tests that same lane had just written. §3 now carries a warning about exactly this. - A full sweep was piped through
grep, discarding the failure output. An intermittent red was then unmatchable againstdocs/ci-red-signatures.md, because a row needs its fragments. This produced registry note U2, and then U3 when it happened again a second time in the same session.
Both are the same failure: a procedure that lives only in prose gets executed differently each time. A script that sets the target directory must already know how to run the suite, so it should own the part of it that is fixed.
2. Design
scripts/gate — POSIX sh, matching scripts/bite's shape: heavy
header comment explaining the reasoning, distinct exit codes, failure
output that says what to do next.
2.1 The managed target root (Q#GS1, Q#GS4)
Everything the script creates lives under one root it owns:
$HOME/build/pmacs-gate-targets/<basename>-<8 hex of CANONICAL worktree path>/
Both parts derive from the canonical physical path of §2.5 — not
from $PWD, which preserves whatever symlinked spelling the caller
happened to use. Two spellings of one worktree must produce one
directory, or the same lane silently builds into two and the isolation
buys nothing while costing double.
Deliberately not $HOME/build/ directly: that holds
cargo-target, go, and cargo, none of which this script may reason
about. A dedicated root means --prune never has to decide whether an
unfamiliar sibling is fair game.
Each managed directory carries an ownership marker at its top level:
.pmacs-gate-target # one line: the absolute worktree path it serves
The marker is what makes deletion safe. It is written at creation, and
a directory without a well-formed marker is never touched, whatever
its name. Discovery is: direct children of the managed root, that are
directories, that contain a readable .pmacs-gate-target whose single
line is an absolute path. Nothing recursive, nothing outside the root,
no name-pattern matching.
PMACS_GATE_TARGET_ROOT overrides the root. It exists for the test
harness (§4) and is documented as test-only; the behavior tests must
never be able to reach the real root.
2.2 Ambient roots and durable logs
Two responsibilities that revision 2's rewrite dropped. They are core, not incidental.
All five ambient roots, fresh per invocation. Every gate command runs with
XDG_CONFIG_HOME XDG_DATA_HOME XDG_STATE_HOME XDG_CACHE_HOME PMACS_STATE_HOME
all pointed at one directory created fresh for that run. The fifth is
not redundant: PMACS_STATE_HOME outranks XDG_STATE_HOME
(src/state.rs), so redirecting only the four XDG variables leaves the
real state root live on a machine that exports it.
This is belt-and-braces, not a workaround for a missing lane. The ambient-root isolation implementation merged as #206; the in-crate paths resolve roots explicitly and no longer depend on the caller's environment. What the five variables cover is everything outside that guarantee — integration suites that spawn the real binary, PTY and daemon fixtures, and anything reaching a production resolution path — where the process under test reads the environment it was handed. A gate runner that scribbles in the developer's real config or data root is a bad failure mode whether or not the crate promises not to, and setting five variables is cheap insurance against it.
HOME is deliberately left alone, matching the existing guidance.
The directory is created under the managed target dir and removed on exit, including on failure. Diagnosis comes from the logs below, not from a retained config tree, and un-reaped ambient directories would accumulate silently.
Every gate's full output goes to a durable log. Per invocation:
<target-dir>/gate-logs/<timestamp>/<NN>-<gate-name>.log
retained after the run, and the workspace sweep log paths are printed
prominently whether the run passes or fails. This is the direct U2/U3
remedy: both notes exist because a sweep's output was filtered through
grep before anyone read it, so an intermittent red could not be
matched against docs/ci-red-signatures.md — a row needs its exact
fragments, and they were gone. A log on disk cannot be filtered away by
the person reading it.
The capture must preserve the gate's own exit status, and the obvious
way does not. In POSIX sh, cmd | tee log reports tee's status,
not cmd's — so a failing gate whose output was teed exits 0 and the
suite reports green. set -o pipefail is not POSIX (dash lacks it)
and this script targets sh.
The specified form is therefore redirection, not a pipeline — and
it must also survive set -e, which scripts/bite and
scripts/feature-census both use (set -eu) and this script will too.
The naive form is doubly wrong: cmd > "$log" 2>&1; rc=$? never
reaches the rc= assignment under set -e, because the shell exits on
the failing command — so the runner dies without printing which gate
failed or where its log is, which is the entire point of capturing it.
A failing command is only exempt from set -e when it is the condition
of an if, so the runner is written as one:
if cargo test --workspace --no-fail-fast -- --skip basedpyright > "$log" 2>&1
then rc=0
else rc=$?
fi
On a non-zero rc the runner prints the failed gate's name and its
log path and then exits non-zero itself. No pipeline exists, so no
status is lost to tee; no bare failing command exists, so no status is
lost to set -e.
The cost is that output is not live; that is the right trade for a gate suite whose failures are read afterwards, and it is exactly how the sweeps were run successfully in the session that motivated this script.
2.3 Interface
scripts/gate [--acceptance SUITE]... [--protocol] [--print-plan]
scripts/gate --print-target-dir
scripts/gate --init
scripts/gate --prune [--force]
-
--acceptance SUITE(repeatable) — the touched acceptance suites. This is the Q#GS2 seam: a script cannot infer from a working tree which suites a change touches, and guessing would be worse than asking, because a wrong guess reads as coverage. §3 stays authoritative for choosing them; the script only runs what it is handed, and prints the list it ran so a PR can quote it. -
--protocol— the change touchesPROTOCOL_VERSION. Adds the CRDT workspace sweep (§2.4). -
--print-plan— print the exact commands and exit without running them. This is what makes drift testable (§4). -
--print-target-dir— print the derived directory and exit. Pure: it creates nothing, so it can be called freely. -
--init— create the managed directory and write its ownership marker, print the path, exit. Runs no gates. The gate path calls the same internal routine, so this is not a second implementation.It exists because §4's verification needs a mutating path it can drive safely: "the marker is written" cannot be tested through a pure printer, and testing it through a real gate run would execute the whole suite inside the suite. It is also genuinely useful — a worktree can be prepared before any work starts.
-
--prune [--force]— see §2.6.
2.4 The gate policy the script encodes
Revision 1 said "both feature configurations", which misstated §3. The corrected encoding, fixed portion first:
cargo fmt --check
cargo clippy --workspace --all-targets -- -D warnings # its own step
cargo test --lib
cargo test --lib --features crdt
cargo test --test <each --acceptance suite>
cargo test --test m4_acceptance -- --skip basedpyright
PMACS_REQUIRE_GPU=1 cargo test -p pmacs-gpu
cargo test --workspace --no-fail-fast -- --skip basedpyright
git diff --check
and only with --protocol, per §3's "touching PROTOCOL_VERSION
STRENGTHENS the sweep line; it does not replace it":
cargo test --workspace --features crdt --no-fail-fast -- --skip basedpyright
One strengthening is proposed rather than assumed. §3 writes the
normal sweep without --no-fail-fast and adds it only in the
protocol-bump form. I propose --no-fail-fast always, and the
justification is that §3's own stated reason is not specific to
bumps — "cargo stops at the first failing target: without it a bump
that breaks eight assertions reports one." A lane that breaks three
unrelated targets has the same problem.
The cost is bounded and paid only on red. A green sweep is byte-for-byte the same work; a red one continues instead of stopping, so it costs the remaining targets' runtime exactly when the complete picture is most wanted. That is the same diagnosability argument that produced U2 and U3, and it is cheap here. If the added red-run time is judged not worth it, dropping this reverts one flag and nothing else.
The CRDT library tests (--lib --features crdt) stay
unconditional — they always were.
2.5 One canonical path representation
Marker creation and prune comparison must use the same spelling of
a path, or a live worktree can be pruned. /home/jeans/Repos/... and a
symlinked route to the same directory are the same worktree and
different strings; git reports the path as it was registered, which
need not match how the marker was written.
The canonical form is the physical absolute path — symlinks
resolved, pwd -P semantics:
canon() { ( cd "$1" 2>/dev/null && pwd -P ) }
Applied at all three points, which is the correction revision 3
needed: directory derivation (§2.1) hashes canon of the worktree,
the marker stores canon of the worktree at creation, and every path
from git worktree list is passed through canon before comparison.
Naming only the last two, as an earlier draft did, leaves the hash
computed from an uncanonicalized $PWD — and then a symlinked
invocation derives a different directory whose marker records the
canonical path. The result is a second target directory for a live
worktree, indistinguishable from an orphan. The symlink test in §4
exists for exactly this and would fail against that draft. A path that cannot be cd-ed into yields
empty, which never matches a marker — correct, because a directory that
is gone is not a live worktree.
This matters concretely here: ~/.config/fish/config.fish and the
cargo config are already symlinks into a dotfiles repo on this machine,
so symlinked spellings are the norm, not a hypothetical.
2.6 Pruning (Q#GS4)
--prune is dry-run by default and prints what it would delete.
Deleting requires --force as a second, explicit step. There is no
automatic pruning, and pruning never happens on the gate path.
A managed directory is eligible when all of:
- it is a direct child of the managed root;
- it contains a readable
.pmacs-gate-targetwhose content is a single absolute path; - that path, canonicalized per §2.5, is not a live worktree.
"Live" is narrower than "listed", and the difference is the whole
correctness of this rule. git worktree list --porcelain keeps
reporting a worktree that was administratively registered but whose
directory was manually deleted — it simply adds a prunable <reason>
line to that entry. Treating every listed path as live would therefore
make exactly the directories most worth reclaiming permanently
ineligible.
So: an entry counts as live only when its record carries no
prunable line. Records are the blank-line-separated blocks of
--porcelain output, read from the primary checkout.
Anything failing any condition is listed as skipped, with the reason rather than passed over silently — a prune that quietly ignores things is how one learns too late that the marker was never written.
2.7 What it cannot do, stated plainly
This is a convention, not an enforcement. An agent that runs bare
cargo test still gets the shared directory and still takes the lock.
Nothing in-repo can prevent that while CARGO_TARGET_DIR is exported
globally, because the environment variable overrides
build.target-dir in .cargo/config.toml — a per-worktree config
file is silently ineffective, which is worse than absent, because it
looks like it worked.
Real enforcement would need one of: dropping the global export (affects
levcs and whatever else uses it), or direnv per worktree. Both are
machine-config changes outside this repo. The proposal is the script
plus a handoff §3 rewrite that points at it; if the convention proves leaky
under real parallel load, direnv is the escalation.
3. Resolved questions
Q#GS1 — directory naming — RESOLVED
<basename>-<8 hex of absolute path> under the managed root of §2.1.
Branch-name derivation was rejected: the name would change on every
branch switch inside one worktree, discarding artifacts precisely when
they are most reusable. Basename alone collides — the recovery
convention in docs/active-work.md actively produces sibling
worktrees with predictable names.
Q#GS2 — gate authority — RESOLVED, NARROWED
The script is authoritative for the fixed executable gates. §3
remains authoritative for policy (why --workspace not --tests,
why --skip basedpyright, when --protocol applies) and for
selecting the touched acceptance suites, which reach the script
only through --acceptance.
Revision 1 claimed the script should own the suite outright. That was wrong for a concrete reason: no script can infer which acceptance suites a change touches, and one that guessed would report coverage it did not have.
Q#GS3 — does the primary checkout get an isolated directory — RESOLVED
Yes. Uniform, no special case. The primary abandons the warm shared directory and pays 80s once. A rule with an exception gets applied wrongly.
Q#GS4 — pruning — RESOLVED
Explicit only, dry-run by default, --force to delete, eligibility and
skip-reporting per §2.6, liveness narrowed to non-prunable records
and paths canonicalized per §2.5.
Q#GS5 — CI — RESOLVED, UNCHANGED
CI runs in fresh containers with their own target directory and no
shared lock, so the isolation is a no-op there. CI's matrix splits the
suite across jobs deliberately (Test (crdt), M4 Perf Gates, …);
collapsing that into one script would serialize the matrix and lose
per-job signal. Worth revisiting; not free, and not this change.
4. Verification
Revision 1 proposed behavior tests with no mechanism to run them.
PMACS_GATE_TARGET_ROOT (§2.1) is that mechanism: every behavior
test points it at a tempfile::tempdir(), and the real managed root is
unreachable from the suite.
The tests live in tests/gate_script_acceptance.rs, so they ride the
existing workspace sweep — a new test, not a CI change (Q#GS5).
The recursion constraint shapes what is testable. A test that ran
scripts/gate for real would run the full suite inside the suite. So
the tests exercise only paths that run no gates — which is a
stronger constraint than "non-mutating", and revision 2 conflated the
two. Asserting that the marker is written requires something that
writes it; a pure printer cannot, and a real gate run must not. That
is what --init is for (§2.3): it mutates, it runs no gates, and the
gate path calls the same routine, so the test is not exercising a
second implementation.
The paths under test:
--print-planmatches §3. Asserts the plan contains--workspaceand never--tests; carries--skip basedpyrighton both suite-wide lines; includes both--libconfigurations; and that--protocoladds the CRDT workspace sweep while the default does not. This is the direct test of Q#GS2's named drift risk, and it is why--print-planexists.--acceptanceis reflected in the plan, once per suite, in order — the seam §3 keeps authority over.- Derivation is stable across a branch switch and differs
between two worktrees — the two properties Q#GS1 turns on.
--print-target-dir, which mutates nothing. --initwrites the marker, and its content is the worktree's path in the §2.5 canonical form. Also that--initis idempotent: running it twice leaves one directory and one marker.- A symlinked spelling of the same worktree derives the same directory and does not prune. The fixture reaches one worktree through a symlink; without §2.5's canonicalization at both ends this is precisely the case that deletes a live lane's artifacts.
--pruneis dry-run by default: against a fixture root holding one eligible directory, one unmarked look-alike, and one marker pointing at a live worktree, it names the eligible one and deletes nothing. Re-run with--force, it deletes exactly that one; the look-alike and the live one survive. This is the test that matters — a prune bug is unrecoverable.- A
prunableworktree record counts as dead. The fixture registers a worktree and deletes its directory withoutgit worktree remove, sogit worktree list --porcelainstill reports it, carryingprunable. Its target directory must be eligible. Treating "listed" as "live" would make exactly the directories most worth reclaiming permanently un-prunable, and nothing else in the suite would notice. - Skip reasons are reported, not silent.
Verified by observation, not automated — and named so the gap is explicit: that a failing gate exits non-zero and says which gate. A deliberately broken tree is a real run of the full suite, so it is a one-off recorded in the lane's PR rather than a test.
Also verified by observation: that the ambient roots are actually redirected and the sweep logs actually appear. Both are properties of a real gate run, so they are confirmed in the same one-off as the exit status above, and named here so the confirmation is not skipped. If the log file is absent, the U2/U3 fix is not real — that is the one outcome of the observed run that would block the lane.
What none of this proves: that agents will use the script. That is §2.7's admission, and no test in this repo can close it.
Gates for the lane itself: the standard suite (via the script, once it
exists — the lane gates itself by construction), plus shellcheck if
available and git diff --check.
5. Not in scope
Dropping the global CARGO_TARGET_DIR export. direnv. sccache
tuning or --remap-path-prefix. CI restructuring (Q#GS5). Pruning the
233G shared directory beyond the 51G of debug/incremental already
reclaimed — it has a co-tenant, and cargo clean there would destroy
levcs's artifacts.