Scaffold fleet Skills: agents/skills source, dist build, installer

[ptr727]

What

Phase 0 and Phase 1 of a multi-phase plan to convert the hub's instruction-heavy docs (comment
style, PR-review conduct, resync safety) into Claude Code / opencode / Codex Skills, so the rules
surface automatically instead of needing to be re-read every session. Opening now, early, to get a
read on the foundation (the distribution mechanism, the installer) before building the remaining
~6 phases on top of it. Expect further commits landing on this same PR as later phases ship.

Phase 0 (scaffolding, no skill content yet):

• .agents/skills/ — canonical, hand-authored home for fleet skills. Codex and opencode read this
  path natively, zero config.
• scripts/build_dist.py — generates a Claude-plugin-compatible copy at
  .claude-plugin/fleet-skills/ (Claude Code does not scan .agents/skills/ directly), with
  --check for CI staleness gating.
• scripts/skills_install.py (+ skills_install.sh/.ps1 thin wrappers, mirroring
  host-setup/agent-safety/install.py's shape) — materializes ~/.agents/skills for Codex/opencode
  and registers the marketplace with Claude Code via the real claude plugin marketplace add/
  install CLI, with --report for staleness checks.
• .claude-plugin/marketplace.json — the Claude Code entry point.
• AGENTS.md pointer (in "Where the Rules Live", verbatim/carried) naming the dependency.

Phase 1: carried-instruction-file-guard skill. Encodes spec/section-model.md's
"Migrating a repo onto the split" procedure (the distinctive-phrase probe) as an active pre-flight
check before any AGENTS.md/GOVERNANCE.md/CODESTYLE.md/WORKFLOW.md overwrite, rather than
the prose-only advisory it was before. This is the highest-leverage skill in the plan: it directly
targets the bug class that once silently deleted a downstream repo's local AGENTS.md additions
during a resync.

Why nested under .claude-plugin/ instead of a top-level dist/

First draft generated output at dist/claude/, which collided with this repo's existing
Python-build-artifact dist/ gitignore entry. Moved to .claude-plugin/fleet-skills/ instead,
avoiding the name clash entirely rather than punching a hole in a blanket ignore.

Verification

• Ran the real install against a scratch AGENTS_HOME, confirmed via claude plugin marketplace list / plugin list that registration and install actually worked, then cleaned up.
• claude plugin validate passes clean on both manifests.
• 13 new unit tests, all passing.
• repo_gate.py (eol, eol-coverage), prose_lint.py (diff), spec/validate.py all clean.
• Live triggering test for the guard skill: loaded it via claude --plugin-dir, prompted a
  session to blindly copy a canonical AGENTS.md over a scratch repo's AGENTS.md carrying a
  local addition. It triggered, identified the local section, refused the overwrite, and proposed
  relocating the content before carrying the rest — the exact procedure it specifies.

Also logs a TODO.md cluster (unrelated to this PR's diff otherwise) for defaulting .py to LF
fleet-wide, found while pinning the two new scripts individually.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Scaffolds a fleet Skills distribution and install mechanism so agent governance rules can be delivered as Skills across Claude Code, Codex, and opencode, starting with a carried-instruction-file guard skill.

Changes:

• Adds a canonical Skills source tree under .agents/skills/ and a generated Claude plugin distribution under .claude-plugin/fleet-skills/, with scripts/build_dist.py to regenerate and --check staleness detection.
• Adds a cross-platform installer (scripts/skills_install.py plus .sh/.ps1 wrappers) with --report staleness reporting and Claude marketplace registration via the claude CLI.
• Updates documentation and pointers (AGENTS + scripts/README) and pins LF line endings for the new Python scripts.

Reviewed changes

Copilot reviewed 16 out of 16 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
TODO.md	Adds a backlog cluster entry about potentially pinning *.py to LF fleet-wide.
scripts/build_dist.py	Implements generator for .claude-plugin/fleet-skills/ plus --check staleness detection.
scripts/test_build_dist.py	Unit tests for build_dist.py regeneration and staleness behavior.
scripts/skills_install.py	Implements cross-platform installer and --report staleness reporting.
scripts/test_skills_install.py	Unit tests for skills_install.py materialization and report behavior.
scripts/skills_install.sh	Bash wrapper to run the installer with Python 3.
scripts/skills_install.ps1	PowerShell wrapper to run the installer with Python 3 on Windows.
scripts/README.md	Documents the new generator and installer scripts and their read-only modes.
AGENTS.md	Adds a pointer explaining that some rules are packaged as Skills and how to install/report.
.gitattributes	Pins LF for the newly added Python scripts and their tests.
.claude-plugin/marketplace.json	Adds Claude marketplace manifest for the local plugin source.
.claude-plugin/fleet-skills/.claude-plugin/plugin.json	Adds generated Claude plugin manifest listing the packaged skills.
.claude-plugin/fleet-skills/.source-digest	Adds generated source digest stamp used for staleness detection.
.claude-plugin/fleet-skills/skills/carried-instruction-file-guard/SKILL.md	Adds generated copy of the carried instruction file guard skill.
.agents/skills/README.md	Documents .agents/skills/ as the canonical, hand-authored Skills source.
.agents/skills/carried-instruction-file-guard/SKILL.md	Adds the canonical carried instruction file guard skill content.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Pull request overview

Copilot reviewed 16 out of 16 changed files in this pull request and generated no new comments.

Suppressed comments (6)

.agents/skills/README.md:10

• This paragraph is hard-wrapped. Per GOVERNANCE.md "Documentation Style Conventions" (Markdown) use one logical paragraph per line with no hard-wrap line-length limit.

Codex and opencode read this directory directly (`.agents/skills/<name>/SKILL.md`), no install
step required, walking from a downstream repo's working directory up to its own repository root.
Claude Code does not scan this path. `scripts/build_dist.py` generates a Claude-plugin-compatible
copy at `.claude-plugin/fleet-skills/`, published through `.claude-plugin/marketplace.json`.

.agents/skills/README.md:13

• This paragraph is hard-wrapped. Per GOVERNANCE.md "Documentation Style Conventions" (Markdown) use one logical paragraph per line with no hard-wrap line-length limit.

Empty today: this is Phase 0 scaffolding. See `AGENTS.md` for how a repo depends on these skills
and `scripts/README.md` for `build_dist.py` and the installer.

.agents/skills/README.md:5

• This paragraph is hard-wrapped. Per GOVERNANCE.md "Documentation Style Conventions" (Markdown) use one logical paragraph per line with no hard-wrap line-length limit.

This issue also appears in the following locations of the same file:

• line 7
• line 12

Canonical source for the fleet's Claude Code / opencode / Codex Skills, one directory per skill:
`<name>/SKILL.md` plus optional `scripts/` and `references/`. This is the only place a skill's
content is hand-authored. Everything else derived from it is generated, never hand-edited.

AGENTS.md:96

• The new paragraph is hard-wrapped across multiple lines. Per GOVERNANCE.md "Documentation Style Conventions" (Markdown) use one logical paragraph per line with no hard-wrap line-length limit.

Some of the rules above are also packaged as Claude Code / opencode / Codex Skills, hand-authored
at [`.agents/skills/`](./.agents/skills/), so they surface automatically instead of needing to be
re-read every session. Run `scripts/skills_install.sh` (or `.ps1` on Windows) once per machine to
install them. `scripts/skills_install.py --report` says whether this machine is current. A rule
that keeps needing to be restated is a sign the install is missing or stale, not that the rule
does not exist.

scripts/skills_install.py:167

• If claude is available but marketplace registration / plugin install fails, the script still exits 0, which makes automation think the full install succeeded. Treat a claude install failure as a non-zero exit (while still writing the stamp) so callers can detect and handle it.

    claude_registered = False
    if claude_available():
        claude_registered = register_claude_marketplace()
    else:
        print("`claude` not found on PATH, skipping Claude Code marketplace registration "
              "(Codex/opencode global skills were still installed).", file=sys.stderr)

    home.mkdir(parents=True, exist_ok=True)
    stamp_path.write_text(json.dumps(build_stamp(claude_registered), indent=2) + "\n", encoding="utf-8")
    print(f"Installed to {home / 'skills'}. Claude Code marketplace registered: {claude_registered}.")
    return 0

.agents/skills/carried-instruction-file-guard/SKILL.md:23

• This Markdown file uses hard-wrapped prose (sentences/paragraphs split across multiple lines). Per GOVERNANCE.md "Documentation Style Conventions" (Markdown), use one logical paragraph per line with no hard-wrap line-length limit. Since this is the canonical skill source, please reflow the whole document and then regenerate the .claude-plugin/fleet-skills/ copy from it.

A downstream repo's `AGENTS.md`/`GOVERNANCE.md`/`CODESTYLE.md`/`WORKFLOW.md` can hold two different
kinds of content mixed in one file: sections that are stale copies of the hub's fleet-wide rules,
and local rules the repo wrote for a fault the fleet has never seen elsewhere. Re-vendoring the
hub's canonical version over the whole file deletes the second kind silently, because nothing
about the diff looks wrong. This has actually happened: a resync replaced a repo's `AGENTS.md`

[ptr727]

Addressing the 6 suppressed findings from the round-2 review, fixed in dfdc428:

This paragraph is hard-wrapped. Per GOVERNANCE.md "Documentation Style Conventions" (Markdown) use one logical paragraph per line with no hard-wrap line-length limit.

Correct on all 4 locations (.agents/skills/README.md x3, AGENTS.md:96) plus the whole of carried-instruction-file-guard/SKILL.md you flagged separately below. All four files reflowed to single-line paragraphs; the .claude-plugin/fleet-skills/ copy regenerated from the reflowed source.

If claude is available but marketplace registration / plugin install fails, the script still exits 0, which makes automation think the full install succeeded. Treat a claude install failure as a non-zero exit (while still writing the stamp) so callers can detect and handle it.

Correct. main() now distinguishes claude simply not being on PATH (a partial-but-expected install for a Codex/opencode-only machine, still exits 0) from claude present but registration actually failing (a real failure, now exits 1, stamp still written either way). Added 3 tests covering all three paths.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Copilot reviewed 16 out of 16 changed files in this pull request and generated no new comments.

Suppressed comments (4)

scripts/skills_install.py:24

• The Usage examples omit the scripts/ prefix even though this file lives under scripts/. This is inconsistent with AGENTS.md (which points to scripts/skills_install.py --report) and can confuse users running from repo root.

Usage: python3 skills_install.py            (installs)
       python3 skills_install.py --report   (read-only: is this machine current?)
       AGENTS_HOME=/x python3 skills_install.py   (override the global skills target, for testing)

scripts/skills_install.py:15

• The docstring refers to wrapper names install.sh / install.ps1, but the wrappers added in this PR are skills_install.sh / skills_install.ps1. Keeping these names accurate avoids sending users to non-existent files.

This issue also appears on line 22 of the same file.

Both wrappers (install.sh, install.ps1) call this, so every OS runs one tested code path.

scripts/skills_install.sh:6

• The wrapper's invocation examples assume running from within scripts/ (e.g., ./skills_install.sh), but AGENTS.md instructs running scripts/skills_install.sh from repo root. Aligning these examples avoids confusion.

#   ./skills_install.sh              installs the fleet skills for this user
#   AGENTS_HOME=/x ./skills_install.sh   overrides the target (testing)

scripts/skills_install.ps1:5

• The PowerShell wrapper's usage examples assume the current directory is scripts/ (e.g., .\skills_install.ps1), but AGENTS.md instructs running it via scripts/skills_install.ps1 from repo root. Aligning the examples avoids confusion.

#   .\skills_install.ps1
#   $env:AGENTS_HOME = "C:\path"; .\skills_install.ps1   # override the target (testing)

[ptr727]

Addressing the 4 new suppressed findings from round 3, fixed in 3c0a66e:

The Usage examples omit the scripts/ prefix even though this file lives under scripts/. This is inconsistent with AGENTS.md (which points to scripts/skills_install.py --report) and can confuse users running from repo root.

Correct, fixed in the module docstring.

The docstring refers to wrapper names install.sh / install.ps1, but the wrappers added in this PR are skills_install.sh / skills_install.ps1.

Correct, a leftover from copying the host-setup/agent-safety/install.py shape without renaming the reference. Fixed.

The wrapper's invocation examples assume running from within scripts/ ... but AGENTS.md instructs running scripts/skills_install.sh from repo root.

Correct for both skills_install.sh and skills_install.ps1. Fixed, and labeled the bare commands ("Run:"/"Or:") rather than capitalizing the path they open on, since a path is not something a comment-casing fix can capitalize without breaking it — matching the existing .gitattributes "Configure with: git config ..." precedent.

(The 6 earlier-round suppressed findings shown again in this round's digest were already answered in the previous comment on this PR.)

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Copilot reviewed 16 out of 16 changed files in this pull request and generated 1 comment.

Suppressed comments (3)

scripts/test_build_dist.py:27

• The test uses a dist/claude/... path even though this PR intentionally moved the generated output under .claude-plugin/ to avoid the repo's dist/ meaning. Using a matching temp path makes the test align with the documented layout and avoids reintroducing the old location in future edits.

        self.tmp = Path(self.enterContext(__import__("tempfile").TemporaryDirectory()))
        self.skills_src = self.tmp / ".agents" / "skills"
        self.dist_plugin = self.tmp / "dist" / "claude" / "fleet-skills"
        self.addCleanup(self._restore, build_dist.SKILLS_SRC, build_dist.DIST_PLUGIN,
                         build_dist.PLUGIN_MANIFEST, build_dist.DIGEST_STAMP)
        build_dist.SKILLS_SRC = self.skills_src
        build_dist.DIST_PLUGIN = self.dist_plugin
        build_dist.PLUGIN_MANIFEST = self.dist_plugin / ".claude-plugin" / "plugin.json"
        build_dist.DIGEST_STAMP = self.dist_plugin / ".source-digest"

scripts/build_dist.py:89

• is_stale() only checks .source-digest. If the plugin manifest or skills/ directory is missing/corrupted but the digest stamp remains, --check can incorrectly report the dist as current. Treat missing DIST_PLUGIN, plugin.json, or skills/ as stale.

def is_stale():
    """Whether the generated plugin needs regenerating: missing, or built from different source bytes."""
    if not DIGEST_STAMP.is_file():
        return True
    return DIGEST_STAMP.read_text(encoding="utf-8").strip() != source_digest(skill_names())

scripts/skills_install.py:141

• report() will currently crash on an unreadable/invalid JSON stamp, and it reports "current" purely on commit equality even when either the stamp or this checkout is marked dirty. Handling decode/IO errors and treating dirty sources as stale makes --report reliable for scripting.

    stamp = json.loads(stamp_path.read_text(encoding="utf-8"))
    current = source_ref()
    stale = stamp.get("source", {}).get("commit") != current.get("commit")
    print(json.dumps({"stamp": stamp, "currentCommit": current.get("commit"), "stale": stale}, indent=2))
    return 1 if stale else 0

[ptr727]

Addressing the 3 new suppressed findings from round 4, fixed in a4eb06a:

The test uses a dist/claude/... path even though this PR intentionally moved the generated output under .claude-plugin/ to avoid the repo's dist/ meaning.

Correct, a leftover variable-name mismatch from the rename. Renamed to .claude-plugin.

is_stale() only checks .source-digest. If the plugin manifest or skills/ directory is missing/corrupted but the digest stamp remains, --check can incorrectly report the dist as current.

Correct. Now also checks that plugin.json and each skill's generated SKILL.md actually exist, not only the digest stamp. Added 2 regression tests (deleted manifest, deleted generated skill).

report() will currently crash on an unreadable/invalid JSON stamp, and it reports "current" purely on commit equality even when either the stamp or this checkout is marked dirty.

Correct on both counts. report() now catches a decode/read error and reports stale rather than crashing, and treats a dirty checkout as stale even when the commit matches (trusting "current" there would mean trusting bytes that were never actually installed). Added 2 regression tests.

(The 10 earlier-round suppressed findings shown again in this round's digest were already answered in the previous comments on this PR.)

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Copilot reviewed 16 out of 16 changed files in this pull request and generated 1 comment.

Suppressed comments (4)

.agents/skills/README.md:7

• README claims the directory is "Empty today" but this PR adds at least carried-instruction-file-guard, so the description is now inaccurate.

Empty today: this is Phase 0 scaffolding. See `AGENTS.md` for how a repo depends on these skills and `scripts/README.md` for `build_dist.py` and the installer.

AGENTS.md:91

• This instructs running scripts/skills_install.sh from the current repo, but scripts/ is explicitly hub-hosted and "never carried" (scripts/README.md), so downstream repos that carry AGENTS.md typically won't have this path. The instruction should say to run the installer from a ProjectTemplate hub checkout.

Some of the rules above are also packaged as Claude Code / opencode / Codex Skills, hand-authored at [`.agents/skills/`](./.agents/skills/), so they surface automatically instead of needing to be re-read every session. Run `scripts/skills_install.sh` (or `.ps1` on Windows) once per machine to install them. `scripts/skills_install.py --report` says whether this machine is current. A rule that keeps needing to be restated is a sign the install is missing or stale, not that the rule does not exist.

scripts/build_dist.py:75

• regenerate() unconditionally calls shutil.rmtree(DIST_PLUGIN) when the path exists. If a stray file or symlink exists at .claude-plugin/fleet-skills, this will raise (NotADirectoryError) instead of regenerating.

    if DIST_PLUGIN.exists():
        shutil.rmtree(DIST_PLUGIN)

scripts/build_dist.py:89

• is_stale() docstring says it "checks the manifest" beyond the digest stamp, but the implementation only checks that PLUGIN_MANIFEST exists (not its contents). Either validate the manifest contents or adjust the docstring to match what is actually checked.

    """Whether the generated plugin needs regenerating: missing, corrupted, or built from
    different source bytes. Checks the manifest and each skill's directory too, not only the
    digest stamp, since a stamp surviving a partial deletion would otherwise report current over
    a plugin that no longer actually resolves."""

[ptr727]

Addressing the 4 new suppressed findings from round 5, fixed in 2805d08:

README claims the directory is "Empty today" but this PR adds at least carried-instruction-file-guard, so the description is now inaccurate.

Correct, dropped the stale claim.

This instructs running scripts/skills_install.sh from the current repo, but scripts/ is explicitly hub-hosted and "never carried" (scripts/README.md), so downstream repos that carry AGENTS.md typically won't have this path. The instruction should say to run the installer from a ProjectTemplate hub checkout.

Correct catch, and an important one — my pointer text contradicted this repo's own existing "reach, don't carry" convention for scripts/. Reworded to instruct reaching a hub checkout first, matching how the rest of Hub-Hosted Tooling is phrased.

regenerate() unconditionally calls shutil.rmtree(DIST_PLUGIN) when the path exists. If a stray file or symlink exists at .claude-plugin/fleet-skills, this will raise (NotADirectoryError) instead of regenerating.

Correct. Now unlinks a stray file/symlink instead of only handling the directory case.

is_stale() docstring says it "checks the manifest" beyond the digest stamp, but the implementation only checks that PLUGIN_MANIFEST exists (not its contents).

Correct — closed the gap rather than just fixing the docstring: is_stale() now parses the manifest and validates its skills list actually matches the current skill set, so a hand-edited or corrupted manifest is caught too, not only a missing one.

(Earlier-round suppressed findings shown again in this round's digest were already answered in previous comments on this PR.)

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Copilot reviewed 16 out of 16 changed files in this pull request and generated no new comments.

Suppressed comments (2)

scripts/build_dist.py:83

• regenerate() writes .source-digest without a trailing newline. This can violate the repo-wide .editorconfig insert_final_newline = true expectation and can make a freshly-regenerated tree fail style/EOL checks.

    DIGEST_STAMP.write_text(source_digest(names), encoding="utf-8")

scripts/skills_install.py:68

• source_ref() claims to report whether "the tree is dirty", but the current git status --porcelain call is scoped only to .agents/skills. That can report dirty=False even when other repo files affecting the install (e.g. .claude-plugin/marketplace.json / generated plugin content) have local modifications, which makes --report assert "current" over bytes that were never installed from the recorded commit.

    # A repo-relative pathspec, not str(SKILLS_SRC).
    # With `git -C <repo>`, an absolute path can fail to match anything.
    # That silently and permanently reports dirty=False.
    status = git("status", "--porcelain", "--", SKILLS_SRC.relative_to(ROOT).as_posix())
    ref["dirty"] = bool(status)

[ptr727]

Addressing the round-20 finding, fixed in b8dbd03:

tree_digest() sorts Path objects directly (sorted(skill_dir.rglob('*'))). Path ordering depends on the platform-specific string form (e.g., 'a\b' vs 'a/b'), so two OSes can hash the same files in a different order and produce different digests even though you hash relative paths with .as_posix() afterwards. Sort using the same normalized key you hash (relative_to(root).as_posix()) so the digest is stable across platforms.

Correct, fixed exactly as suggested — sorts by relative_to(root).as_posix() explicitly now, the same key that gets hashed.

No regression test added, flagging honestly rather than keeping one that doesn't prove anything: I wrote a test mocking rglob() to return files in reverse order and asserting the digest was unchanged. It passed against both the pre-fix and post-fix code, because the pre-fix code also calls sorted() on whatever rglob() returns — my mock's reordering was erased by that sorted() call in both versions, so the test never actually exercised the platform-specific divergence this finding is about. On closer look, PosixPath's own comparison already coincides with the .as_posix() string form on this Linux host, so there's no way to construct two paths that sort differently by raw Path vs by posix-string here — the bug is real but only reproducible on Windows. Discarded the test rather than keep one giving false confidence. The fix stands on inspection: sorting by the same key you hash by is strictly more correct regardless of whether I can locally prove the failure mode.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Copilot reviewed 16 out of 16 changed files in this pull request and generated no new comments.

Suppressed comments (1)

scripts/build_dist.py:13

• The PR introduces a staleness check mode (--check) and new self-tests, but the existing CI workflow doesn’t invoke them. For example, .github/workflows/validate-task.yml currently runs scripts/test_prose_lint.py, scripts/test_repo_gate.py, and scripts/test_pr_review.py, but not scripts/test_build_dist.py, scripts/test_skills_install.py, or python3 scripts/build_dist.py --check. Without wiring these into CI, the generated .claude-plugin/fleet-skills/ tree can drift undetected and the new installer/generator behavior won’t be continuously validated.

Usage: python3 scripts/build_dist.py           regenerate the plugin from .agents/skills/
       python3 scripts/build_dist.py --check   read-only: exit 1 if the plugin is stale
"""

[ptr727]

Addressing the round-21 finding, fixed in 028fb74:

The PR introduces a staleness check mode (--check) and new self-tests, but the existing CI workflow doesn't invoke them ... Without wiring these into CI, the generated .claude-plugin/fleet-skills/ tree can drift undetected and the new installer/generator behavior won't be continuously validated.

Correct — a real gap, not a nitpick, since it meant every fix earlier in this PR was only ever going to be checked by hand going forward. Added test_build_dist.py and test_skills_install.py to the existing "Run script self-tests step", and a new "Check fleet skills are current step" running build_dist.py --check, matching the shape of the existing "Check repo gates step". Verified locally by running the exact commands CI will run, not just trusting the workflow YAML looks right.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Copilot reviewed 17 out of 17 changed files in this pull request and generated no new comments.

Suppressed comments (1)

scripts/skills_install.py:221

• --report can incorrectly claim a machine is current when the installer was run from a dirty checkout and the working tree later becomes clean at the same commit. The stamp already records source.dirty, but stale only checks the current checkout’s dirty bit, not the stamp’s, so a dirty install can be treated as current solely because the commit SHA matches.

    stale = bool(
        current_commit is None
        or stamp_commit != current_commit
        or current.get("dirty")
    )

[ptr727]

Addressing the round-22 finding, fixed in 81f3c44:

--report can incorrectly claim a machine is current when the installer was run from a dirty checkout and the working tree later becomes clean at the same commit. The stamp already records source.dirty, but stale only checks the current checkout's dirty bit, not the stamp's, so a dirty install can be treated as current solely because the commit SHA matches.

Correct. stale now also reads the stamp's own recorded dirty flag, alongside the current checkout's — the two catch different things: the current flag catches a machine that's gone dirty since install, the stamp's flag catches the install itself having happened from a dirty state, which no later clean state can retroactively verify. Verified the new test actually catches this before committing: reverted to the pre-fix code first, confirmed the test failed (0 != 1), then restored the fix and confirmed it passes — not just assumed.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Copilot reviewed 17 out of 17 changed files in this pull request and generated no new comments.