fix(hooks): print the --no-track worktree recipe in both guards' remediation text

[huangyiirene]

Fixes#6977

What changed

Both worktree guards print a remediation recipe when they block — the copy an agent reads
at the exact moment it is about to act, and therefore the copy most likely to be run
verbatim. Both printed the recipe with a plain -b. PR #6976 hardened the prescribed form
in CLAUDE.md and AGENTS.md with --no-track; the hooks' copy was left behind.

One word added per file, net ±0 lines in both:

• .claude/hooks/guard-main-checkout.sh:57
• .claude/hooks/guard-main-checkout-bash.sh:544

Both now emit exactly the prescribed form (verbatim from a live invocation of each hook,
with $name interpolated):

 git fetch origin main && git worktree add --no-track ../objectui-TASK -b BRANCH origin/main
cd ../objectui-TASK && pnpm install # then re-run your edits there

Why

The hazard is the one PR #6976 established: plain -b writes the new branch's upstream
keys (branch.NAME.remote, branch.NAME.merge) into the one .git/config that every
linked worktree of a repo shares. That write can fail after the branch is created,
leaving a branch with no worktree — a half state the error text does not name. Read as
"the worktree exists", the agent starts editing the shared primary checkout, which is the
one thing these two guards exist to prevent.

Scope — additive, not a recipe rewrite

The premise was re-verified on origin/main before the first edit and it is narrower than
the sibling card's. This repo's hooks already emitted the fetch-hardened base
(git fetch origin main && … origin/main, landed for ui#6208), so the missing flag was the
only surviving divergence from the prescribed form. Measured on the freshly fetched tree:
grep -rc "no-track" .claude/hooks/ returned 0 across all seven files there (four guards,
three self-test matrices) before this change, and 1 in each of the two edited guards after.

Not in this PR

Each of these was examined and deliberately left alone:

• .changeset/6208-worktree-recipe-fetch-base.md quotes the pre-ui#6208 recipe as the
  record of a past change. Editing a historical changeset rewrites that record —
  adjudicated in PR docs: harden the worktree recipe with --no-track; correct the stale no-mechanical-backstop claim #6976, and restated on the card. check-changeset-overwrite.mjs
  confirms it mechanically: 0 modified, 0 deleted.
• The two -cmp comparison-tree lines in CLAUDE.md / AGENTS.md create no branch and
  take an explicit ref from the caller, so there are no upstream keys to write and the flag
  does not apply. This is the same reading ui#6208 recorded for them.
• .claude/hooks/guard-tree-enum.selftest.sh:99 is a command payload fed to a different
  guard, asserting that ordinary commands pass through it. It is test input, not text any
  agent is told to run, and adding the flag there would change no coverage.

Verification

All at efbfb15, the head of this branch. Nine checks, joined with && so one verdict
covers all of them — exit 0.

Hook self-tests — the only gate that targets .claude/hooks/**
(.github/workflows/hook-selftests.yml), all three matrices it runs:

self-test	result
guard-main-checkout-bash.selftest.sh	121 passed, 0 failed
guard-shared-stash.selftest.sh	41 passed, 0 failed
guard-tree-enum.selftest.sh	36 passed, 0 failed

Identical to the pre-edit baseline, and that is the expected direction rather than a weak
result: no self-test case pins the remediation string, so none needed updating in this
change. Measured, not assumed — a grep -rn for four distinctive fragments of the two block
messages (fetch origin main, dedicated worktree, Blocked: editing on the shared,
WRITES into the shared) across .claude/hooks/*.selftest.sh and scripts/ returns no
match in any self-test.

A second measurement, reported because its absence is the answer rather than an assumption:
this repo ships noguard-main-checkout.selftest.sh. .claude/hooks/ holds four guards
and three matrices; guard-main-checkout.sh is the one guard without one — so for that hook
the direct invocation below is the whole of its coverage.

Behavioural check — both hooks invoked directly with a payload targeting the shared
primary checkout. Each exits 2 (blocked, as designed) and prints the hardened recipe shown
above. The recipe it now prints was then fed back through all three of this repo's Bash
guards (guard-main-checkout-bash, guard-shared-stash, guard-tree-enum): all three exit
0, so the text is runnable and not self-blocked.

Repo gates that reach this diff — check-control-bytes (5809 tracked text files
scanned), check-changeset-presence, check-changeset-no-major, check-changeset-overwrite,
check-shell-escape-residue, check-governed-queue-guard --self-test (132 cases). All ✅.
bash -n clean on both edited files.

Measurement — nothing lints shell files' contents in this repo. Verified on this tree
rather than taken from the workflow header that claims it: eslint.config.js scopes every
block to **/*.{ts,tsx}, and there is no shellcheck step anywhere (the three matches in
the repo are two prose mentions and one # shellcheck source= directive inside
e2e/live/ci/start-backend.sh). check-shell-escape-residue reads fenced blocks in
AGENTS.md, CLAUDE.md, skills/ and content/docs/ — it does not reach .claude/hooks/.
So the self-test matrices plus the direct invocation above are the whole of the coverage
available for this diff, by measurement.

Changeset

check-changeset-presence reports no changeset is owed here (3 files changed, 0 of them
published source of a released package). One is included anyway, with empty frontmatter,
because that is this repo's verified convention for a diff of this shape — the same form as
.changeset/6880-worktree-recipe-no-track.md (PR #6976), 6208-worktree-recipe-fetch-base.md
and 6089-hook-selftests-step-names-drop-counts.md. The empty-frontmatter exemption is a
first-class pass in scripts/check-changeset-presence.mjs.

No skip-changeset label is applied: in this repo that label is a phantom.
scripts/__tests__/ci-cd-pipeline-doc.test.ts pins that nothing under .github/ or
scripts/ may mention it, so applying it would exempt nothing.

Governed surface — draft, human merge

.claude/** is governed. check-governed-queue-guard --test on this diff returns
⛔ GOVERNED — 2 of 3 path(s), and one governed path governs the whole pull request. This PR
stays a draft: not flipped ready, not enqueued, no auto-merge armed. The merge is the
maintainer's, and a human merge is the review record.

Refs: PR #6976 · ui#6880 · ui#6208 · objectstack#13052. The objectstack twin
(objectstack#13663) covers that repo's hooks and is not addressed here.

Generated by Claude Code

---

Generated by Claude Code

[huangyiirene]

Fixes#6977

What changed

Both worktree guards print a remediation recipe when they block — the copy an agent reads
at the exact moment it is about to act, and therefore the copy most likely to be run
verbatim. Both printed the recipe with a plain -b. PR #6976 hardened the prescribed form
in CLAUDE.md and AGENTS.md with --no-track; the hooks' copy was left behind.

One word added per file, net ±0 lines in both:

• .claude/hooks/guard-main-checkout.sh:57
• .claude/hooks/guard-main-checkout-bash.sh:544

Both now emit exactly the prescribed form (verbatim from a live invocation of each hook,
with $name interpolated):

 git fetch origin main && git worktree add --no-track ../objectui-TASK -b BRANCH origin/main
cd ../objectui-TASK && pnpm install # then re-run your edits there

Why

The hazard is the one PR #6976 established: plain -b writes the new branch's upstream
keys (branch.NAME.remote, branch.NAME.merge) into the one .git/config that every
linked worktree of a repo shares. That write can fail after the branch is created,
leaving a branch with no worktree — a half state the error text does not name. Read as
"the worktree exists", the agent starts editing the shared primary checkout, which is the
one thing these two guards exist to prevent.

Scope — additive, not a recipe rewrite

The premise was re-verified on origin/main before the first edit and it is narrower than
the sibling card's. This repo's hooks already emitted the fetch-hardened base
(git fetch origin main && … origin/main, landed for ui#6208), so the missing flag was the
only surviving divergence from the prescribed form. Measured on the freshly fetched tree:
grep -rc "no-track" .claude/hooks/ returned 0 across all seven files there (four guards,
three self-test matrices) before this change, and 1 in each of the two edited guards after.

Not in this PR

Each of these was examined and deliberately left alone:

• .changeset/6208-worktree-recipe-fetch-base.md quotes the pre-ui#6208 recipe as the
  record of a past change. Editing a historical changeset rewrites that record —
  adjudicated in PR docs: harden the worktree recipe with --no-track; correct the stale no-mechanical-backstop claim #6976, and restated on the card. check-changeset-overwrite.mjs
  confirms it mechanically: 0 modified, 0 deleted.
• The two -cmp comparison-tree lines in CLAUDE.md / AGENTS.md create no branch and
  take an explicit ref from the caller, so there are no upstream keys to write and the flag
  does not apply. This is the same reading ui#6208 recorded for them.
• .claude/hooks/guard-tree-enum.selftest.sh:99 is a command payload fed to a different
  guard, asserting that ordinary commands pass through it. It is test input, not text any
  agent is told to run, and adding the flag there would change no coverage.

Verification

All at efbfb15, the head of this branch. Nine checks, joined with && so one verdict
covers all of them — exit 0.

Hook self-tests — the only gate that targets .claude/hooks/**
(.github/workflows/hook-selftests.yml), all three matrices it runs:

self-test	result
guard-main-checkout-bash.selftest.sh	121 passed, 0 failed
guard-shared-stash.selftest.sh	41 passed, 0 failed
guard-tree-enum.selftest.sh	36 passed, 0 failed

Identical to the pre-edit baseline, and that is the expected direction rather than a weak
result: no self-test case pins the remediation string, so none needed updating in this
change. Measured, not assumed — a grep -rn for four distinctive fragments of the two block
messages (fetch origin main, dedicated worktree, Blocked: editing on the shared,
WRITES into the shared) across .claude/hooks/*.selftest.sh and scripts/ returns no
match in any self-test.

A second measurement, reported because its absence is the answer rather than an assumption:
this repo ships noguard-main-checkout.selftest.sh. .claude/hooks/ holds four guards
and three matrices; guard-main-checkout.sh is the one guard without one — so for that hook
the direct invocation below is the whole of its coverage.

Behavioural check — both hooks invoked directly with a payload targeting the shared
primary checkout. Each exits 2 (blocked, as designed) and prints the hardened recipe shown
above. The recipe it now prints was then fed back through all three of this repo's Bash
guards (guard-main-checkout-bash, guard-shared-stash, guard-tree-enum): all three exit
0, so the text is runnable and not self-blocked.

Repo gates that reach this diff — check-control-bytes (5809 tracked text files
scanned), check-changeset-presence, check-changeset-no-major, check-changeset-overwrite,
check-shell-escape-residue, check-governed-queue-guard --self-test (132 cases). All ✅.
bash -n clean on both edited files.

Measurement — nothing lints shell files' contents in this repo. Verified on this tree
rather than taken from the workflow header that claims it: eslint.config.js scopes every
block to **/*.{ts,tsx}, and there is no shellcheck step anywhere (the three matches in
the repo are two prose mentions and one # shellcheck source= directive inside
e2e/live/ci/start-backend.sh). check-shell-escape-residue reads fenced blocks in
AGENTS.md, CLAUDE.md, skills/ and content/docs/ — it does not reach .claude/hooks/.
So the self-test matrices plus the direct invocation above are the whole of the coverage
available for this diff, by measurement.

Changeset

check-changeset-presence reports no changeset is owed here (3 files changed, 0 of them
published source of a released package). One is included anyway, with empty frontmatter,
because that is this repo's verified convention for a diff of this shape — the same form as
.changeset/6880-worktree-recipe-no-track.md (PR #6976), 6208-worktree-recipe-fetch-base.md
and 6089-hook-selftests-step-names-drop-counts.md. The empty-frontmatter exemption is a
first-class pass in scripts/check-changeset-presence.mjs.

No skip-changeset label is applied: in this repo that label is a phantom.
scripts/__tests__/ci-cd-pipeline-doc.test.ts pins that nothing under .github/ or
scripts/ may mention it, so applying it would exempt nothing.

Governed surface — draft, human merge

.claude/** is governed. check-governed-queue-guard --test on this diff returns
⛔ GOVERNED — 2 of 3 path(s), and one governed path governs the whole pull request. This PR
stays a draft: not flipped ready, not enqueued, no auto-merge armed. The merge is the
maintainer's, and a human merge is the review record.

Refs: PR #6976 · ui#6880 · ui#6208 · objectstack#13052. The objectstack twin
(objectstack#13663) covers that repo's hooks and is not addressed here.

Generated by Claude Code

---

Generated by Claude Code

[huangyiirene]

Fixes#6977

What changed

Both worktree guards print a remediation recipe when they block — the copy an agent reads
at the exact moment it is about to act, and therefore the copy most likely to be run
verbatim. Both printed the recipe with a plain -b. PR #6976 hardened the prescribed form
in CLAUDE.md and AGENTS.md with --no-track; the hooks' copy was left behind.

One word added per file, net ±0 lines in both:

• .claude/hooks/guard-main-checkout.sh:57
• .claude/hooks/guard-main-checkout-bash.sh:544

Both now emit exactly the prescribed form (verbatim from a live invocation of each hook,
with $name interpolated):

 git fetch origin main && git worktree add --no-track ../objectui-TASK -b BRANCH origin/main
cd ../objectui-TASK && pnpm install # then re-run your edits there

Why

The hazard is the one PR #6976 established: plain -b writes the new branch's upstream
keys (branch.NAME.remote, branch.NAME.merge) into the one .git/config that every
linked worktree of a repo shares. That write can fail after the branch is created,
leaving a branch with no worktree — a half state the error text does not name. Read as
"the worktree exists", the agent starts editing the shared primary checkout, which is the
one thing these two guards exist to prevent.

Scope — additive, not a recipe rewrite

The premise was re-verified on origin/main before the first edit and it is narrower than
the sibling card's. This repo's hooks already emitted the fetch-hardened base
(git fetch origin main && … origin/main, landed for ui#6208), so the missing flag was the
only surviving divergence from the prescribed form. Measured on the freshly fetched tree:
grep -rc "no-track" .claude/hooks/ returned 0 across all seven files there (four guards,
three self-test matrices) before this change, and 1 in each of the two edited guards after.

Not in this PR

Each of these was examined and deliberately left alone:

• .changeset/6208-worktree-recipe-fetch-base.md quotes the pre-ui#6208 recipe as the
  record of a past change. Editing a historical changeset rewrites that record —
  adjudicated in PR docs: harden the worktree recipe with --no-track; correct the stale no-mechanical-backstop claim #6976, and restated on the card. check-changeset-overwrite.mjs
  confirms it mechanically: 0 modified, 0 deleted.
• The two -cmp comparison-tree lines in CLAUDE.md / AGENTS.md create no branch and
  take an explicit ref from the caller, so there are no upstream keys to write and the flag
  does not apply. This is the same reading ui#6208 recorded for them.
• .claude/hooks/guard-tree-enum.selftest.sh:99 is a command payload fed to a different
  guard, asserting that ordinary commands pass through it. It is test input, not text any
  agent is told to run, and adding the flag there would change no coverage.

Verification

All at efbfb15, the head of this branch. Nine checks, joined with && so one verdict
covers all of them — exit 0.

Hook self-tests — the only gate that targets .claude/hooks/**
(.github/workflows/hook-selftests.yml), all three matrices it runs:

self-test	result
guard-main-checkout-bash.selftest.sh	121 passed, 0 failed
guard-shared-stash.selftest.sh	41 passed, 0 failed
guard-tree-enum.selftest.sh	36 passed, 0 failed

Identical to the pre-edit baseline, and that is the expected direction rather than a weak
result: no self-test case pins the remediation string, so none needed updating in this
change. Measured, not assumed — a grep -rn for four distinctive fragments of the two block
messages (fetch origin main, dedicated worktree, Blocked: editing on the shared,
WRITES into the shared) across .claude/hooks/*.selftest.sh and scripts/ returns no
match in any self-test.

A second measurement, reported because its absence is the answer rather than an assumption:
this repo ships noguard-main-checkout.selftest.sh. .claude/hooks/ holds four guards
and three matrices; guard-main-checkout.sh is the one guard without one — so for that hook
the direct invocation below is the whole of its coverage.

Behavioural check — both hooks invoked directly with a payload targeting the shared
primary checkout. Each exits 2 (blocked, as designed) and prints the hardened recipe shown
above. The recipe it now prints was then fed back through all three of this repo's Bash
guards (guard-main-checkout-bash, guard-shared-stash, guard-tree-enum): all three exit
0, so the text is runnable and not self-blocked.

Repo gates that reach this diff — check-control-bytes (5809 tracked text files
scanned), check-changeset-presence, check-changeset-no-major, check-changeset-overwrite,
check-shell-escape-residue, check-governed-queue-guard --self-test (132 cases). All ✅.
bash -n clean on both edited files.

Measurement — nothing lints shell files' contents in this repo. Verified on this tree
rather than taken from the workflow header that claims it: eslint.config.js scopes every
block to **/*.{ts,tsx}, and there is no shellcheck step anywhere (the three matches in
the repo are two prose mentions and one # shellcheck source= directive inside
e2e/live/ci/start-backend.sh). check-shell-escape-residue reads fenced blocks in
AGENTS.md, CLAUDE.md, skills/ and content/docs/ — it does not reach .claude/hooks/.
So the self-test matrices plus the direct invocation above are the whole of the coverage
available for this diff, by measurement.

Changeset

check-changeset-presence reports no changeset is owed here (3 files changed, 0 of them
published source of a released package). One is included anyway, with empty frontmatter,
because that is this repo's verified convention for a diff of this shape — the same form as
.changeset/6880-worktree-recipe-no-track.md (PR #6976), 6208-worktree-recipe-fetch-base.md
and 6089-hook-selftests-step-names-drop-counts.md. The empty-frontmatter exemption is a
first-class pass in scripts/check-changeset-presence.mjs.

No skip-changeset label is applied: in this repo that label is a phantom.
scripts/__tests__/ci-cd-pipeline-doc.test.ts pins that nothing under .github/ or
scripts/ may mention it, so applying it would exempt nothing.

Governed surface — draft, human merge

.claude/** is governed. check-governed-queue-guard --test on this diff returns
⛔ GOVERNED — 2 of 3 path(s), and one governed path governs the whole pull request. This PR
stays a draft: not flipped ready, not enqueued, no auto-merge armed. The merge is the
maintainer's, and a human merge is the review record.

Refs: PR #6976 · ui#6880 · ui#6208 · objectstack#13052. The objectstack twin
(objectstack#13663) covers that repo's hooks and is not addressed here.

Generated by Claude Code

---

Generated by Claude Code

[huangyiirene]

Fixes#6977

What changed

Both worktree guards print a remediation recipe when they block — the copy an agent reads
at the exact moment it is about to act, and therefore the copy most likely to be run
verbatim. Both printed the recipe with a plain -b. PR #6976 hardened the prescribed form
in CLAUDE.md and AGENTS.md with --no-track; the hooks' copy was left behind.

One word added per file, net ±0 lines in both:

• .claude/hooks/guard-main-checkout.sh:57
• .claude/hooks/guard-main-checkout-bash.sh:544

Both now emit exactly the prescribed form (verbatim from a live invocation of each hook,
with $name interpolated):

 git fetch origin main && git worktree add --no-track ../objectui-TASK -b BRANCH origin/main
cd ../objectui-TASK && pnpm install # then re-run your edits there

Why

The hazard is the one PR #6976 established: plain -b writes the new branch's upstream
keys (branch.NAME.remote, branch.NAME.merge) into the one .git/config that every
linked worktree of a repo shares. That write can fail after the branch is created,
leaving a branch with no worktree — a half state the error text does not name. Read as
"the worktree exists", the agent starts editing the shared primary checkout, which is the
one thing these two guards exist to prevent.

Scope — additive, not a recipe rewrite

The premise was re-verified on origin/main before the first edit and it is narrower than
the sibling card's. This repo's hooks already emitted the fetch-hardened base
(git fetch origin main && … origin/main, landed for ui#6208), so the missing flag was the
only surviving divergence from the prescribed form. Measured on the freshly fetched tree:
grep -rc "no-track" .claude/hooks/ returned 0 across all seven files there (four guards,
three self-test matrices) before this change, and 1 in each of the two edited guards after.

Not in this PR

Each of these was examined and deliberately left alone:

• .changeset/6208-worktree-recipe-fetch-base.md quotes the pre-ui#6208 recipe as the
  record of a past change. Editing a historical changeset rewrites that record —
  adjudicated in PR docs: harden the worktree recipe with --no-track; correct the stale no-mechanical-backstop claim #6976, and restated on the card. check-changeset-overwrite.mjs
  confirms it mechanically: 0 modified, 0 deleted.
• The two -cmp comparison-tree lines in CLAUDE.md / AGENTS.md create no branch and
  take an explicit ref from the caller, so there are no upstream keys to write and the flag
  does not apply. This is the same reading ui#6208 recorded for them.
• .claude/hooks/guard-tree-enum.selftest.sh:99 is a command payload fed to a different
  guard, asserting that ordinary commands pass through it. It is test input, not text any
  agent is told to run, and adding the flag there would change no coverage.

Verification

All at efbfb15, the head of this branch. Nine checks, joined with && so one verdict
covers all of them — exit 0.

Hook self-tests — the only gate that targets .claude/hooks/**
(.github/workflows/hook-selftests.yml), all three matrices it runs:

self-test	result
guard-main-checkout-bash.selftest.sh	121 passed, 0 failed
guard-shared-stash.selftest.sh	41 passed, 0 failed
guard-tree-enum.selftest.sh	36 passed, 0 failed

Identical to the pre-edit baseline, and that is the expected direction rather than a weak
result: no self-test case pins the remediation string, so none needed updating in this
change. Measured, not assumed — a grep -rn for four distinctive fragments of the two block
messages (fetch origin main, dedicated worktree, Blocked: editing on the shared,
WRITES into the shared) across .claude/hooks/*.selftest.sh and scripts/ returns no
match in any self-test.

A second measurement, reported because its absence is the answer rather than an assumption:
this repo ships noguard-main-checkout.selftest.sh. .claude/hooks/ holds four guards
and three matrices; guard-main-checkout.sh is the one guard without one — so for that hook
the direct invocation below is the whole of its coverage.

Behavioural check — both hooks invoked directly with a payload targeting the shared
primary checkout. Each exits 2 (blocked, as designed) and prints the hardened recipe shown
above. The recipe it now prints was then fed back through all three of this repo's Bash
guards (guard-main-checkout-bash, guard-shared-stash, guard-tree-enum): all three exit
0, so the text is runnable and not self-blocked.

Repo gates that reach this diff — check-control-bytes (5809 tracked text files
scanned), check-changeset-presence, check-changeset-no-major, check-changeset-overwrite,
check-shell-escape-residue, check-governed-queue-guard --self-test (132 cases). All ✅.
bash -n clean on both edited files.

Measurement — nothing lints shell files' contents in this repo. Verified on this tree
rather than taken from the workflow header that claims it: eslint.config.js scopes every
block to **/*.{ts,tsx}, and there is no shellcheck step anywhere (the three matches in
the repo are two prose mentions and one # shellcheck source= directive inside
e2e/live/ci/start-backend.sh). check-shell-escape-residue reads fenced blocks in
AGENTS.md, CLAUDE.md, skills/ and content/docs/ — it does not reach .claude/hooks/.
So the self-test matrices plus the direct invocation above are the whole of the coverage
available for this diff, by measurement.

Changeset

check-changeset-presence reports no changeset is owed here (3 files changed, 0 of them
published source of a released package). One is included anyway, with empty frontmatter,
because that is this repo's verified convention for a diff of this shape — the same form as
.changeset/6880-worktree-recipe-no-track.md (PR #6976), 6208-worktree-recipe-fetch-base.md
and 6089-hook-selftests-step-names-drop-counts.md. The empty-frontmatter exemption is a
first-class pass in scripts/check-changeset-presence.mjs.

No skip-changeset label is applied: in this repo that label is a phantom.
scripts/__tests__/ci-cd-pipeline-doc.test.ts pins that nothing under .github/ or
scripts/ may mention it, so applying it would exempt nothing.

Governed surface — draft, human merge

.claude/** is governed. check-governed-queue-guard --test on this diff returns
⛔ GOVERNED — 2 of 3 path(s), and one governed path governs the whole pull request. This PR
stays a draft: not flipped ready, not enqueued, no auto-merge armed. The merge is the
maintainer's, and a human merge is the review record.

Refs: PR #6976 · ui#6880 · ui#6208 · objectstack#13052. The objectstack twin
(objectstack#13663) covers that repo's hooks and is not addressed here.

Generated by Claude Code

---

Generated by Claude Code

[huangyiirene]

Fixes#6977

What changed

Both worktree guards print a remediation recipe when they block — the copy an agent reads
at the exact moment it is about to act, and therefore the copy most likely to be run
verbatim. Both printed the recipe with a plain -b. PR #6976 hardened the prescribed form
in CLAUDE.md and AGENTS.md with --no-track; the hooks' copy was left behind.

One word added per file, net ±0 lines in both:

• .claude/hooks/guard-main-checkout.sh:57
• .claude/hooks/guard-main-checkout-bash.sh:544

Both now emit exactly the prescribed form (verbatim from a live invocation of each hook,
with $name interpolated):

 git fetch origin main && git worktree add --no-track ../objectui-TASK -b BRANCH origin/main
cd ../objectui-TASK && pnpm install # then re-run your edits there

Why

The hazard is the one PR #6976 established: plain -b writes the new branch's upstream
keys (branch.NAME.remote, branch.NAME.merge) into the one .git/config that every
linked worktree of a repo shares. That write can fail after the branch is created,
leaving a branch with no worktree — a half state the error text does not name. Read as
"the worktree exists", the agent starts editing the shared primary checkout, which is the
one thing these two guards exist to prevent.

Scope — additive, not a recipe rewrite

The premise was re-verified on origin/main before the first edit and it is narrower than
the sibling card's. This repo's hooks already emitted the fetch-hardened base
(git fetch origin main && … origin/main, landed for ui#6208), so the missing flag was the
only surviving divergence from the prescribed form. Measured on the freshly fetched tree:
grep -rc "no-track" .claude/hooks/ returned 0 across all seven files there (four guards,
three self-test matrices) before this change, and 1 in each of the two edited guards after.

Not in this PR

Each of these was examined and deliberately left alone:

• .changeset/6208-worktree-recipe-fetch-base.md quotes the pre-ui#6208 recipe as the
  record of a past change. Editing a historical changeset rewrites that record —
  adjudicated in PR docs: harden the worktree recipe with --no-track; correct the stale no-mechanical-backstop claim #6976, and restated on the card. check-changeset-overwrite.mjs
  confirms it mechanically: 0 modified, 0 deleted.
• The two -cmp comparison-tree lines in CLAUDE.md / AGENTS.md create no branch and
  take an explicit ref from the caller, so there are no upstream keys to write and the flag
  does not apply. This is the same reading ui#6208 recorded for them.
• .claude/hooks/guard-tree-enum.selftest.sh:99 is a command payload fed to a different
  guard, asserting that ordinary commands pass through it. It is test input, not text any
  agent is told to run, and adding the flag there would change no coverage.

Verification

All at efbfb15, the head of this branch. Nine checks, joined with && so one verdict
covers all of them — exit 0.

Hook self-tests — the only gate that targets .claude/hooks/**
(.github/workflows/hook-selftests.yml), all three matrices it runs:

self-test	result
guard-main-checkout-bash.selftest.sh	121 passed, 0 failed
guard-shared-stash.selftest.sh	41 passed, 0 failed
guard-tree-enum.selftest.sh	36 passed, 0 failed

Identical to the pre-edit baseline, and that is the expected direction rather than a weak
result: no self-test case pins the remediation string, so none needed updating in this
change. Measured, not assumed — a grep -rn for four distinctive fragments of the two block
messages (fetch origin main, dedicated worktree, Blocked: editing on the shared,
WRITES into the shared) across .claude/hooks/*.selftest.sh and scripts/ returns no
match in any self-test.

A second measurement, reported because its absence is the answer rather than an assumption:
this repo ships noguard-main-checkout.selftest.sh. .claude/hooks/ holds four guards
and three matrices; guard-main-checkout.sh is the one guard without one — so for that hook
the direct invocation below is the whole of its coverage.

Behavioural check — both hooks invoked directly with a payload targeting the shared
primary checkout. Each exits 2 (blocked, as designed) and prints the hardened recipe shown
above. The recipe it now prints was then fed back through all three of this repo's Bash
guards (guard-main-checkout-bash, guard-shared-stash, guard-tree-enum): all three exit
0, so the text is runnable and not self-blocked.

Repo gates that reach this diff — check-control-bytes (5809 tracked text files
scanned), check-changeset-presence, check-changeset-no-major, check-changeset-overwrite,
check-shell-escape-residue, check-governed-queue-guard --self-test (132 cases). All ✅.
bash -n clean on both edited files.

Measurement — nothing lints shell files' contents in this repo. Verified on this tree
rather than taken from the workflow header that claims it: eslint.config.js scopes every
block to **/*.{ts,tsx}, and there is no shellcheck step anywhere (the three matches in
the repo are two prose mentions and one # shellcheck source= directive inside
e2e/live/ci/start-backend.sh). check-shell-escape-residue reads fenced blocks in
AGENTS.md, CLAUDE.md, skills/ and content/docs/ — it does not reach .claude/hooks/.
So the self-test matrices plus the direct invocation above are the whole of the coverage
available for this diff, by measurement.

Changeset

check-changeset-presence reports no changeset is owed here (3 files changed, 0 of them
published source of a released package). One is included anyway, with empty frontmatter,
because that is this repo's verified convention for a diff of this shape — the same form as
.changeset/6880-worktree-recipe-no-track.md (PR #6976), 6208-worktree-recipe-fetch-base.md
and 6089-hook-selftests-step-names-drop-counts.md. The empty-frontmatter exemption is a
first-class pass in scripts/check-changeset-presence.mjs.

No skip-changeset label is applied: in this repo that label is a phantom.
scripts/__tests__/ci-cd-pipeline-doc.test.ts pins that nothing under .github/ or
scripts/ may mention it, so applying it would exempt nothing.

Governed surface — draft, human merge

.claude/** is governed. check-governed-queue-guard --test on this diff returns
⛔ GOVERNED — 2 of 3 path(s), and one governed path governs the whole pull request. This PR
stays a draft: not flipped ready, not enqueued, no auto-merge armed. The merge is the
maintainer's, and a human merge is the review record.

Refs: PR #6976 · ui#6880 · ui#6208 · objectstack#13052. The objectstack twin
(objectstack#13663) covers that repo's hooks and is not addressed here.

Generated by Claude Code

---

Generated by Claude Code

[huangyiirene]

Fixes#6977

What changed

Both worktree guards print a remediation recipe when they block — the copy an agent reads
at the exact moment it is about to act, and therefore the copy most likely to be run
verbatim. Both printed the recipe with a plain -b. PR #6976 hardened the prescribed form
in CLAUDE.md and AGENTS.md with --no-track; the hooks' copy was left behind.

One word added per file, net ±0 lines in both:

• .claude/hooks/guard-main-checkout.sh:57
• .claude/hooks/guard-main-checkout-bash.sh:544

Both now emit exactly the prescribed form (verbatim from a live invocation of each hook,
with $name interpolated):

 git fetch origin main && git worktree add --no-track ../objectui-TASK -b BRANCH origin/main
cd ../objectui-TASK && pnpm install # then re-run your edits there

Why

The hazard is the one PR #6976 established: plain -b writes the new branch's upstream
keys (branch.NAME.remote, branch.NAME.merge) into the one .git/config that every
linked worktree of a repo shares. That write can fail after the branch is created,
leaving a branch with no worktree — a half state the error text does not name. Read as
"the worktree exists", the agent starts editing the shared primary checkout, which is the
one thing these two guards exist to prevent.

Scope — additive, not a recipe rewrite

The premise was re-verified on origin/main before the first edit and it is narrower than
the sibling card's. This repo's hooks already emitted the fetch-hardened base
(git fetch origin main && … origin/main, landed for ui#6208), so the missing flag was the
only surviving divergence from the prescribed form. Measured on the freshly fetched tree:
grep -rc "no-track" .claude/hooks/ returned 0 across all seven files there (four guards,
three self-test matrices) before this change, and 1 in each of the two edited guards after.

Not in this PR

Each of these was examined and deliberately left alone:

• .changeset/6208-worktree-recipe-fetch-base.md quotes the pre-ui#6208 recipe as the
  record of a past change. Editing a historical changeset rewrites that record —
  adjudicated in PR docs: harden the worktree recipe with --no-track; correct the stale no-mechanical-backstop claim #6976, and restated on the card. check-changeset-overwrite.mjs
  confirms it mechanically: 0 modified, 0 deleted.
• The two -cmp comparison-tree lines in CLAUDE.md / AGENTS.md create no branch and
  take an explicit ref from the caller, so there are no upstream keys to write and the flag
  does not apply. This is the same reading ui#6208 recorded for them.
• .claude/hooks/guard-tree-enum.selftest.sh:99 is a command payload fed to a different
  guard, asserting that ordinary commands pass through it. It is test input, not text any
  agent is told to run, and adding the flag there would change no coverage.

Verification

All at efbfb15, the head of this branch. Nine checks, joined with && so one verdict
covers all of them — exit 0.

Hook self-tests — the only gate that targets .claude/hooks/**
(.github/workflows/hook-selftests.yml), all three matrices it runs:

self-test	result
guard-main-checkout-bash.selftest.sh	121 passed, 0 failed
guard-shared-stash.selftest.sh	41 passed, 0 failed
guard-tree-enum.selftest.sh	36 passed, 0 failed

Identical to the pre-edit baseline, and that is the expected direction rather than a weak
result: no self-test case pins the remediation string, so none needed updating in this
change. Measured, not assumed — a grep -rn for four distinctive fragments of the two block
messages (fetch origin main, dedicated worktree, Blocked: editing on the shared,
WRITES into the shared) across .claude/hooks/*.selftest.sh and scripts/ returns no
match in any self-test.

A second measurement, reported because its absence is the answer rather than an assumption:
this repo ships noguard-main-checkout.selftest.sh. .claude/hooks/ holds four guards
and three matrices; guard-main-checkout.sh is the one guard without one — so for that hook
the direct invocation below is the whole of its coverage.

Behavioural check — both hooks invoked directly with a payload targeting the shared
primary checkout. Each exits 2 (blocked, as designed) and prints the hardened recipe shown
above. The recipe it now prints was then fed back through all three of this repo's Bash
guards (guard-main-checkout-bash, guard-shared-stash, guard-tree-enum): all three exit
0, so the text is runnable and not self-blocked.

Repo gates that reach this diff — check-control-bytes (5809 tracked text files
scanned), check-changeset-presence, check-changeset-no-major, check-changeset-overwrite,
check-shell-escape-residue, check-governed-queue-guard --self-test (132 cases). All ✅.
bash -n clean on both edited files.

Measurement — nothing lints shell files' contents in this repo. Verified on this tree
rather than taken from the workflow header that claims it: eslint.config.js scopes every
block to **/*.{ts,tsx}, and there is no shellcheck step anywhere (the three matches in
the repo are two prose mentions and one # shellcheck source= directive inside
e2e/live/ci/start-backend.sh). check-shell-escape-residue reads fenced blocks in
AGENTS.md, CLAUDE.md, skills/ and content/docs/ — it does not reach .claude/hooks/.
So the self-test matrices plus the direct invocation above are the whole of the coverage
available for this diff, by measurement.

Changeset

check-changeset-presence reports no changeset is owed here (3 files changed, 0 of them
published source of a released package). One is included anyway, with empty frontmatter,
because that is this repo's verified convention for a diff of this shape — the same form as
.changeset/6880-worktree-recipe-no-track.md (PR #6976), 6208-worktree-recipe-fetch-base.md
and 6089-hook-selftests-step-names-drop-counts.md. The empty-frontmatter exemption is a
first-class pass in scripts/check-changeset-presence.mjs.

No skip-changeset label is applied: in this repo that label is a phantom.
scripts/__tests__/ci-cd-pipeline-doc.test.ts pins that nothing under .github/ or
scripts/ may mention it, so applying it would exempt nothing.

Governed surface — draft, human merge

.claude/** is governed. check-governed-queue-guard --test on this diff returns
⛔ GOVERNED — 2 of 3 path(s), and one governed path governs the whole pull request. This PR
stays a draft: not flipped ready, not enqueued, no auto-merge armed. The merge is the
maintainer's, and a human merge is the review record.

Refs: PR #6976 · ui#6880 · ui#6208 · objectstack#13052. The objectstack twin
(objectstack#13663) covers that repo's hooks and is not addressed here.

Generated by Claude Code

---

Generated by Claude Code

[huangyiirene]

Fixes#6977

What changed

Both worktree guards print a remediation recipe when they block — the copy an agent reads
at the exact moment it is about to act, and therefore the copy most likely to be run
verbatim. Both printed the recipe with a plain -b. PR #6976 hardened the prescribed form
in CLAUDE.md and AGENTS.md with --no-track; the hooks' copy was left behind.

One word added per file, net ±0 lines in both:

• .claude/hooks/guard-main-checkout.sh:57
• .claude/hooks/guard-main-checkout-bash.sh:544

Both now emit exactly the prescribed form (verbatim from a live invocation of each hook,
with $name interpolated):

 git fetch origin main && git worktree add --no-track ../objectui-TASK -b BRANCH origin/main
cd ../objectui-TASK && pnpm install # then re-run your edits there

Why

The hazard is the one PR #6976 established: plain -b writes the new branch's upstream
keys (branch.NAME.remote, branch.NAME.merge) into the one .git/config that every
linked worktree of a repo shares. That write can fail after the branch is created,
leaving a branch with no worktree — a half state the error text does not name. Read as
"the worktree exists", the agent starts editing the shared primary checkout, which is the
one thing these two guards exist to prevent.

Scope — additive, not a recipe rewrite

The premise was re-verified on origin/main before the first edit and it is narrower than
the sibling card's. This repo's hooks already emitted the fetch-hardened base
(git fetch origin main && … origin/main, landed for ui#6208), so the missing flag was the
only surviving divergence from the prescribed form. Measured on the freshly fetched tree:
grep -rc "no-track" .claude/hooks/ returned 0 across all seven files there (four guards,
three self-test matrices) before this change, and 1 in each of the two edited guards after.

Not in this PR

Each of these was examined and deliberately left alone:

• .changeset/6208-worktree-recipe-fetch-base.md quotes the pre-ui#6208 recipe as the
  record of a past change. Editing a historical changeset rewrites that record —
  adjudicated in PR docs: harden the worktree recipe with --no-track; correct the stale no-mechanical-backstop claim #6976, and restated on the card. check-changeset-overwrite.mjs
  confirms it mechanically: 0 modified, 0 deleted.
• The two -cmp comparison-tree lines in CLAUDE.md / AGENTS.md create no branch and
  take an explicit ref from the caller, so there are no upstream keys to write and the flag
  does not apply. This is the same reading ui#6208 recorded for them.
• .claude/hooks/guard-tree-enum.selftest.sh:99 is a command payload fed to a different
  guard, asserting that ordinary commands pass through it. It is test input, not text any
  agent is told to run, and adding the flag there would change no coverage.

Verification

All at efbfb15, the head of this branch. Nine checks, joined with && so one verdict
covers all of them — exit 0.

Hook self-tests — the only gate that targets .claude/hooks/**
(.github/workflows/hook-selftests.yml), all three matrices it runs:

self-test	result
guard-main-checkout-bash.selftest.sh	121 passed, 0 failed
guard-shared-stash.selftest.sh	41 passed, 0 failed
guard-tree-enum.selftest.sh	36 passed, 0 failed

Identical to the pre-edit baseline, and that is the expected direction rather than a weak
result: no self-test case pins the remediation string, so none needed updating in this
change. Measured, not assumed — a grep -rn for four distinctive fragments of the two block
messages (fetch origin main, dedicated worktree, Blocked: editing on the shared,
WRITES into the shared) across .claude/hooks/*.selftest.sh and scripts/ returns no
match in any self-test.

A second measurement, reported because its absence is the answer rather than an assumption:
this repo ships noguard-main-checkout.selftest.sh. .claude/hooks/ holds four guards
and three matrices; guard-main-checkout.sh is the one guard without one — so for that hook
the direct invocation below is the whole of its coverage.

Behavioural check — both hooks invoked directly with a payload targeting the shared
primary checkout. Each exits 2 (blocked, as designed) and prints the hardened recipe shown
above. The recipe it now prints was then fed back through all three of this repo's Bash
guards (guard-main-checkout-bash, guard-shared-stash, guard-tree-enum): all three exit
0, so the text is runnable and not self-blocked.

Repo gates that reach this diff — check-control-bytes (5809 tracked text files
scanned), check-changeset-presence, check-changeset-no-major, check-changeset-overwrite,
check-shell-escape-residue, check-governed-queue-guard --self-test (132 cases). All ✅.
bash -n clean on both edited files.

Measurement — nothing lints shell files' contents in this repo. Verified on this tree
rather than taken from the workflow header that claims it: eslint.config.js scopes every
block to **/*.{ts,tsx}, and there is no shellcheck step anywhere (the three matches in
the repo are two prose mentions and one # shellcheck source= directive inside
e2e/live/ci/start-backend.sh). check-shell-escape-residue reads fenced blocks in
AGENTS.md, CLAUDE.md, skills/ and content/docs/ — it does not reach .claude/hooks/.
So the self-test matrices plus the direct invocation above are the whole of the coverage
available for this diff, by measurement.

Changeset

check-changeset-presence reports no changeset is owed here (3 files changed, 0 of them
published source of a released package). One is included anyway, with empty frontmatter,
because that is this repo's verified convention for a diff of this shape — the same form as
.changeset/6880-worktree-recipe-no-track.md (PR #6976), 6208-worktree-recipe-fetch-base.md
and 6089-hook-selftests-step-names-drop-counts.md. The empty-frontmatter exemption is a
first-class pass in scripts/check-changeset-presence.mjs.

No skip-changeset label is applied: in this repo that label is a phantom.
scripts/__tests__/ci-cd-pipeline-doc.test.ts pins that nothing under .github/ or
scripts/ may mention it, so applying it would exempt nothing.

Governed surface — draft, human merge

.claude/** is governed. check-governed-queue-guard --test on this diff returns
⛔ GOVERNED — 2 of 3 path(s), and one governed path governs the whole pull request. This PR
stays a draft: not flipped ready, not enqueued, no auto-merge armed. The merge is the
maintainer's, and a human merge is the review record.

Refs: PR #6976 · ui#6880 · ui#6208 · objectstack#13052. The objectstack twin
(objectstack#13663) covers that repo's hooks and is not addressed here.

Generated by Claude Code

---

Generated by Claude Code

[huangyiirene]

Fixes#6977

What changed

Both worktree guards print a remediation recipe when they block — the copy an agent reads
at the exact moment it is about to act, and therefore the copy most likely to be run
verbatim. Both printed the recipe with a plain -b. PR #6976 hardened the prescribed form
in CLAUDE.md and AGENTS.md with --no-track; the hooks' copy was left behind.

One word added per file, net ±0 lines in both:

• .claude/hooks/guard-main-checkout.sh:57
• .claude/hooks/guard-main-checkout-bash.sh:544

Both now emit exactly the prescribed form (verbatim from a live invocation of each hook,
with $name interpolated):

 git fetch origin main && git worktree add --no-track ../objectui-TASK -b BRANCH origin/main
cd ../objectui-TASK && pnpm install # then re-run your edits there

Why

The hazard is the one PR #6976 established: plain -b writes the new branch's upstream
keys (branch.NAME.remote, branch.NAME.merge) into the one .git/config that every
linked worktree of a repo shares. That write can fail after the branch is created,
leaving a branch with no worktree — a half state the error text does not name. Read as
"the worktree exists", the agent starts editing the shared primary checkout, which is the
one thing these two guards exist to prevent.

Scope — additive, not a recipe rewrite

The premise was re-verified on origin/main before the first edit and it is narrower than
the sibling card's. This repo's hooks already emitted the fetch-hardened base
(git fetch origin main && … origin/main, landed for ui#6208), so the missing flag was the
only surviving divergence from the prescribed form. Measured on the freshly fetched tree:
grep -rc "no-track" .claude/hooks/ returned 0 across all seven files there (four guards,
three self-test matrices) before this change, and 1 in each of the two edited guards after.

Not in this PR

Each of these was examined and deliberately left alone:

• .changeset/6208-worktree-recipe-fetch-base.md quotes the pre-ui#6208 recipe as the
  record of a past change. Editing a historical changeset rewrites that record —
  adjudicated in PR docs: harden the worktree recipe with --no-track; correct the stale no-mechanical-backstop claim #6976, and restated on the card. check-changeset-overwrite.mjs
  confirms it mechanically: 0 modified, 0 deleted.
• The two -cmp comparison-tree lines in CLAUDE.md / AGENTS.md create no branch and
  take an explicit ref from the caller, so there are no upstream keys to write and the flag
  does not apply. This is the same reading ui#6208 recorded for them.
• .claude/hooks/guard-tree-enum.selftest.sh:99 is a command payload fed to a different
  guard, asserting that ordinary commands pass through it. It is test input, not text any
  agent is told to run, and adding the flag there would change no coverage.

Verification

All at efbfb15, the head of this branch. Nine checks, joined with && so one verdict
covers all of them — exit 0.

Hook self-tests — the only gate that targets .claude/hooks/**
(.github/workflows/hook-selftests.yml), all three matrices it runs:

self-test	result
guard-main-checkout-bash.selftest.sh	121 passed, 0 failed
guard-shared-stash.selftest.sh	41 passed, 0 failed
guard-tree-enum.selftest.sh	36 passed, 0 failed

Identical to the pre-edit baseline, and that is the expected direction rather than a weak
result: no self-test case pins the remediation string, so none needed updating in this
change. Measured, not assumed — a grep -rn for four distinctive fragments of the two block
messages (fetch origin main, dedicated worktree, Blocked: editing on the shared,
WRITES into the shared) across .claude/hooks/*.selftest.sh and scripts/ returns no
match in any self-test.

A second measurement, reported because its absence is the answer rather than an assumption:
this repo ships noguard-main-checkout.selftest.sh. .claude/hooks/ holds four guards
and three matrices; guard-main-checkout.sh is the one guard without one — so for that hook
the direct invocation below is the whole of its coverage.

Behavioural check — both hooks invoked directly with a payload targeting the shared
primary checkout. Each exits 2 (blocked, as designed) and prints the hardened recipe shown
above. The recipe it now prints was then fed back through all three of this repo's Bash
guards (guard-main-checkout-bash, guard-shared-stash, guard-tree-enum): all three exit
0, so the text is runnable and not self-blocked.

Repo gates that reach this diff — check-control-bytes (5809 tracked text files
scanned), check-changeset-presence, check-changeset-no-major, check-changeset-overwrite,
check-shell-escape-residue, check-governed-queue-guard --self-test (132 cases). All ✅.
bash -n clean on both edited files.

Measurement — nothing lints shell files' contents in this repo. Verified on this tree
rather than taken from the workflow header that claims it: eslint.config.js scopes every
block to **/*.{ts,tsx}, and there is no shellcheck step anywhere (the three matches in
the repo are two prose mentions and one # shellcheck source= directive inside
e2e/live/ci/start-backend.sh). check-shell-escape-residue reads fenced blocks in
AGENTS.md, CLAUDE.md, skills/ and content/docs/ — it does not reach .claude/hooks/.
So the self-test matrices plus the direct invocation above are the whole of the coverage
available for this diff, by measurement.

Changeset

check-changeset-presence reports no changeset is owed here (3 files changed, 0 of them
published source of a released package). One is included anyway, with empty frontmatter,
because that is this repo's verified convention for a diff of this shape — the same form as
.changeset/6880-worktree-recipe-no-track.md (PR #6976), 6208-worktree-recipe-fetch-base.md
and 6089-hook-selftests-step-names-drop-counts.md. The empty-frontmatter exemption is a
first-class pass in scripts/check-changeset-presence.mjs.

No skip-changeset label is applied: in this repo that label is a phantom.
scripts/__tests__/ci-cd-pipeline-doc.test.ts pins that nothing under .github/ or
scripts/ may mention it, so applying it would exempt nothing.

Governed surface — draft, human merge

.claude/** is governed. check-governed-queue-guard --test on this diff returns
⛔ GOVERNED — 2 of 3 path(s), and one governed path governs the whole pull request. This PR
stays a draft: not flipped ready, not enqueued, no auto-merge armed. The merge is the
maintainer's, and a human merge is the review record.

Refs: PR #6976 · ui#6880 · ui#6208 · objectstack#13052. The objectstack twin
(objectstack#13663) covers that repo's hooks and is not addressed here.

Generated by Claude Code

---

Generated by Claude Code