fix(agent): run the pre-push gate without pnpm exec

[patroza]

Why

pnpm exec runs a dependency-status check before the command it wraps. When the
lockfile has moved — a rebase does exactly that — pnpm decides node_modules
must be purged and reinstalled, then aborts because a git hook has no TTY:

ERR_PNPM_ABORTED_REMOVE_MODULES_DIR_NO_TTY

The push then fails with a pnpm error that says nothing about the gate, because
the gate never started. It reads as a broken ship gate rather than a stale
lockfile, which is the worst version of this failure: the thing that reports
problems is itself the thing that appears broken.

Hit live while rebasing an effect-app PR onto a moved base.

What

Invoke node directly from .githooks/pre-push instead of pnpm exec node.

How

Nothing is lost by dropping pnpm exec here: the gate puts
node_modules/.bin on PATH itself (withRepoBin), which is what pnpm exec
was providing. The hook needs only a node binary, and the shell that runs a git
hook already has one.

Fixed first in effect-app#838 and ported unchanged to the other repos that share
this gate.

Remarks

Same class as the other headless failures this gate has surfaced: a tool that
silently changes behaviour based on TTY or CI detection breaks inside a hook, and
does it confusingly. Worth remembering when adding anything else to this path.

Flow doc updated: n/a (agent tooling)
E2E coverage: n/a — shell hook, no product behaviour; exercised by every agent push
Design angles: n/a (one-line fix)

[patroza]

Why

pnpm exec runs a dependency-status check before the command it wraps. When the
lockfile has moved — a rebase does exactly that — pnpm decides node_modules
must be purged and reinstalled, then aborts because a git hook has no TTY:

ERR_PNPM_ABORTED_REMOVE_MODULES_DIR_NO_TTY

The push then fails with a pnpm error that says nothing about the gate, because
the gate never started. It reads as a broken ship gate rather than a stale
lockfile, which is the worst version of this failure: the thing that reports
problems is itself the thing that appears broken.

Hit live while rebasing an effect-app PR onto a moved base.

What

Invoke node directly from .githooks/pre-push instead of pnpm exec node.

How

Nothing is lost by dropping pnpm exec here: the gate puts
node_modules/.bin on PATH itself (withRepoBin), which is what pnpm exec
was providing. The hook needs only a node binary, and the shell that runs a git
hook already has one.

Fixed first in effect-app#838 and ported unchanged to the other repos that share
this gate.

Remarks

Same class as the other headless failures this gate has surfaced: a tool that
silently changes behaviour based on TTY or CI detection breaks inside a hook, and
does it confusingly. Worth remembering when adding anything else to this path.

Flow doc updated: n/a (agent tooling)
E2E coverage: n/a — shell hook, no product behaviour; exercised by every agent push
Design angles: n/a (one-line fix)

[patroza]

Why

pnpm exec runs a dependency-status check before the command it wraps. When the
lockfile has moved — a rebase does exactly that — pnpm decides node_modules
must be purged and reinstalled, then aborts because a git hook has no TTY:

ERR_PNPM_ABORTED_REMOVE_MODULES_DIR_NO_TTY

The push then fails with a pnpm error that says nothing about the gate, because
the gate never started. It reads as a broken ship gate rather than a stale
lockfile, which is the worst version of this failure: the thing that reports
problems is itself the thing that appears broken.

Hit live while rebasing an effect-app PR onto a moved base.

What

Invoke node directly from .githooks/pre-push instead of pnpm exec node.

How

Nothing is lost by dropping pnpm exec here: the gate puts
node_modules/.bin on PATH itself (withRepoBin), which is what pnpm exec
was providing. The hook needs only a node binary, and the shell that runs a git
hook already has one.

Fixed first in effect-app#838 and ported unchanged to the other repos that share
this gate.

Remarks

Same class as the other headless failures this gate has surfaced: a tool that
silently changes behaviour based on TTY or CI detection breaks inside a hook, and
does it confusingly. Worth remembering when adding anything else to this path.

Flow doc updated: n/a (agent tooling)
E2E coverage: n/a — shell hook, no product behaviour; exercised by every agent push
Design angles: n/a (one-line fix)

[patroza]

Why

pnpm exec runs a dependency-status check before the command it wraps. When the
lockfile has moved — a rebase does exactly that — pnpm decides node_modules
must be purged and reinstalled, then aborts because a git hook has no TTY:

ERR_PNPM_ABORTED_REMOVE_MODULES_DIR_NO_TTY

The push then fails with a pnpm error that says nothing about the gate, because
the gate never started. It reads as a broken ship gate rather than a stale
lockfile, which is the worst version of this failure: the thing that reports
problems is itself the thing that appears broken.

Hit live while rebasing an effect-app PR onto a moved base.

What

Invoke node directly from .githooks/pre-push instead of pnpm exec node.

How

Nothing is lost by dropping pnpm exec here: the gate puts
node_modules/.bin on PATH itself (withRepoBin), which is what pnpm exec
was providing. The hook needs only a node binary, and the shell that runs a git
hook already has one.

Fixed first in effect-app#838 and ported unchanged to the other repos that share
this gate.

Remarks

Same class as the other headless failures this gate has surfaced: a tool that
silently changes behaviour based on TTY or CI detection breaks inside a hook, and
does it confusingly. Worth remembering when adding anything else to this path.

Flow doc updated: n/a (agent tooling)
E2E coverage: n/a — shell hook, no product behaviour; exercised by every agent push
Design angles: n/a (one-line fix)

[patroza]

Why

pnpm exec runs a dependency-status check before the command it wraps. When the
lockfile has moved — a rebase does exactly that — pnpm decides node_modules
must be purged and reinstalled, then aborts because a git hook has no TTY:

ERR_PNPM_ABORTED_REMOVE_MODULES_DIR_NO_TTY

The push then fails with a pnpm error that says nothing about the gate, because
the gate never started. It reads as a broken ship gate rather than a stale
lockfile, which is the worst version of this failure: the thing that reports
problems is itself the thing that appears broken.

Hit live while rebasing an effect-app PR onto a moved base.

What

Invoke node directly from .githooks/pre-push instead of pnpm exec node.

How

Nothing is lost by dropping pnpm exec here: the gate puts
node_modules/.bin on PATH itself (withRepoBin), which is what pnpm exec
was providing. The hook needs only a node binary, and the shell that runs a git
hook already has one.

Fixed first in effect-app#838 and ported unchanged to the other repos that share
this gate.

Remarks

Same class as the other headless failures this gate has surfaced: a tool that
silently changes behaviour based on TTY or CI detection breaks inside a hook, and
does it confusingly. Worth remembering when adding anything else to this path.

Flow doc updated: n/a (agent tooling)
E2E coverage: n/a — shell hook, no product behaviour; exercised by every agent push
Design angles: n/a (one-line fix)

[patroza]

Why

pnpm exec runs a dependency-status check before the command it wraps. When the
lockfile has moved — a rebase does exactly that — pnpm decides node_modules
must be purged and reinstalled, then aborts because a git hook has no TTY:

ERR_PNPM_ABORTED_REMOVE_MODULES_DIR_NO_TTY

The push then fails with a pnpm error that says nothing about the gate, because
the gate never started. It reads as a broken ship gate rather than a stale
lockfile, which is the worst version of this failure: the thing that reports
problems is itself the thing that appears broken.

Hit live while rebasing an effect-app PR onto a moved base.

What

Invoke node directly from .githooks/pre-push instead of pnpm exec node.

How

Nothing is lost by dropping pnpm exec here: the gate puts
node_modules/.bin on PATH itself (withRepoBin), which is what pnpm exec
was providing. The hook needs only a node binary, and the shell that runs a git
hook already has one.

Fixed first in effect-app#838 and ported unchanged to the other repos that share
this gate.

Remarks

Same class as the other headless failures this gate has surfaced: a tool that
silently changes behaviour based on TTY or CI detection breaks inside a hook, and
does it confusingly. Worth remembering when adding anything else to this path.

Flow doc updated: n/a (agent tooling)
E2E coverage: n/a — shell hook, no product behaviour; exercised by every agent push
Design angles: n/a (one-line fix)

[patroza]

Why

pnpm exec runs a dependency-status check before the command it wraps. When the
lockfile has moved — a rebase does exactly that — pnpm decides node_modules
must be purged and reinstalled, then aborts because a git hook has no TTY:

ERR_PNPM_ABORTED_REMOVE_MODULES_DIR_NO_TTY

The push then fails with a pnpm error that says nothing about the gate, because
the gate never started. It reads as a broken ship gate rather than a stale
lockfile, which is the worst version of this failure: the thing that reports
problems is itself the thing that appears broken.

Hit live while rebasing an effect-app PR onto a moved base.

What

Invoke node directly from .githooks/pre-push instead of pnpm exec node.

How

Nothing is lost by dropping pnpm exec here: the gate puts
node_modules/.bin on PATH itself (withRepoBin), which is what pnpm exec
was providing. The hook needs only a node binary, and the shell that runs a git
hook already has one.

Fixed first in effect-app#838 and ported unchanged to the other repos that share
this gate.

Remarks

Same class as the other headless failures this gate has surfaced: a tool that
silently changes behaviour based on TTY or CI detection breaks inside a hook, and
does it confusingly. Worth remembering when adding anything else to this path.

Flow doc updated: n/a (agent tooling)
E2E coverage: n/a — shell hook, no product behaviour; exercised by every agent push
Design angles: n/a (one-line fix)

[patroza]

Why

pnpm exec runs a dependency-status check before the command it wraps. When the
lockfile has moved — a rebase does exactly that — pnpm decides node_modules
must be purged and reinstalled, then aborts because a git hook has no TTY:

ERR_PNPM_ABORTED_REMOVE_MODULES_DIR_NO_TTY

The push then fails with a pnpm error that says nothing about the gate, because
the gate never started. It reads as a broken ship gate rather than a stale
lockfile, which is the worst version of this failure: the thing that reports
problems is itself the thing that appears broken.

Hit live while rebasing an effect-app PR onto a moved base.

What

Invoke node directly from .githooks/pre-push instead of pnpm exec node.

How

Nothing is lost by dropping pnpm exec here: the gate puts
node_modules/.bin on PATH itself (withRepoBin), which is what pnpm exec
was providing. The hook needs only a node binary, and the shell that runs a git
hook already has one.

Fixed first in effect-app#838 and ported unchanged to the other repos that share
this gate.

Remarks

Same class as the other headless failures this gate has surfaced: a tool that
silently changes behaviour based on TTY or CI detection breaks inside a hook, and
does it confusingly. Worth remembering when adding anything else to this path.

Flow doc updated: n/a (agent tooling)
E2E coverage: n/a — shell hook, no product behaviour; exercised by every agent push
Design angles: n/a (one-line fix)