Promote Develop: autoMode Defaults First and Manager-Owned Agent Tool Upgrades - #4096

…tput (#4094)

## Summary

Fixes what a native Windows 11 test of the one-step agent setup
(released as 2.0.1414) showed to be wrong or misleading.

- **`"$defaults"` first in every `autoMode` list.**
- **The problem:** Claude Code's offer to learn rules had written a
24-entry `autoMode.environment` with no `$defaults`, silently replacing
Claude Code's built-in rules. The kit treated such a list as the
machine's choice and left it alone.
- **Install:** for each of `environment`, `allow`, `soft_deny` and
`hard_deny` that exists, the kit adds `$defaults` at index 0 if absent,
moves it there if misplaced, and collapses duplicates. Every other entry
keeps its value and order, and a missing list stays absent. The deny
lists matter most, since one without `$defaults` drops Claude Code's
built-in deny rules. The decision is the maintainer's.
- **`--report`:** a list without `$defaults` first, exactly once, makes
the machine not current. Each case is worded for what it is: only an
absent `$defaults` replaces the built-ins.
- **No opt-out:** a machine can no longer keep a list that replaces the
built-ins, since every install puts `$defaults` back. The README says
so.
- **Archive note.** The kit report said "no checkout to compare against"
even when the archive stamp recorded the commit and the report had
compared it. It now names the commit, and says "no commit" only when the
stamp holds none.
- **Codex presence.** The rules step counts Codex as present when
`~/.codex` exists, which is right, since the Codex IDE extension reads
that file too. The agent-tools table says "not installed" when no
`codex` CLI is on PATH. The rules lines now say "(config directory
present, no codex CLI on PATH)" in that case, on both platforms.
- **`merge-and-release` step 7.** After a hub release, step 7 now also
refreshes the Codex and opencode global rules, but only where the kit in
that same step reads current. That keeps it from rolling the rules back
while the kit declines to.

The logic lives in the shared Python installers, so the `.sh` and `.ps1`
wrappers behave the same. The `install-global-rules` pair's help text
changed identically on both.

Refs #2553

## Verification

The kit's `test_install.py` (215), `tests/test_global_rules.py` (54),
`tests/test_bootstrap.py` (186, with the PowerShell cases under `pwsh`),
`tests/test_cli_usage.py` and `tests/test_build_dist.py` pass.
`build_dist.py --check` is current. ruff, mypy and the prose gate are
clean. `docker_lint.py` passes shellcheck, shfmt, PSScriptAnalyzer and
markdownlint.

Two recorded local passes ran. The first raised two findings, the
per-case wording and step 7's missing rollback guard, both fixed in
7859290. The second found nothing open.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>

## Summary

Answers Copilot's review of promotion PR #4096. That PR's head is
develop, so it takes no fixes directly.

- `tool_manager` in `host-setup/linux/install-agents.sh` returned `bun`
for any `.bun` path before asking npm. So an npm global under an npm
root inside a bun directory was misread, unlike Windows. npm is now
asked first, as on Windows.
- A Nix-packaged agent under `/nix/store/.../node_modules` hit the
generic `node_modules` npm arm first. It was then reported with an npm
reason instead of the Nix one. `tool_manager_by_path` now reads mise,
Homebrew Cellar/Caskroom, `.bun`, and `/nix/store` before the
`node_modules` fallback.

No existing classification changes: mise npm-backend, npm globals under
a Cellar node, linuxbrew, and apt read as before.

## Verification

- New `classify` helper sources the real `tool_manager` with `npm_root`
fixed. Three new tests cover:
  - an npm root under a bun directory, read as npm
  - a bun global npm does not own, read as bun
  - a Nix store `node_modules`, read as nix
- The first and third fail on the old order.
- `tests/test_install_agents.py` passes, 64 tests. shellcheck, shfmt,
and the prose gate are clean.
- A recorded local pass found nothing open.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

… TFM (#4109)

## Summary

Implements the hub side of #4103. Every fleet .NET repo selects its SDK
line in `global.json`, and the C# version follows the TFM rather than
the installed compiler.

- **`dotnet-codestyle` Skill**, "Central build and package
configuration":
- `global.json` is required, with an `sdk` block (`10.0.100`,
`rollForward: latestFeature`, `allowPrerelease: false`). The `test`
runner key appears only where the test project is MTP-based.
- `LangVersion` is left unset in props files and in every project
targeting a .NET TFM. A `netstandard` analyzer or source-generator
project may set a fixed number, since unset it compiles at C# 7.3, which
`Nullable` `enable` refuses.
- `latest`, `latestMajor`, and `preview` are banned everywhere.
`AnalysisLevel` stays `latest-all`.
- The skill states the host requirement and the drift conditions, and
says a major move starts in the hub's `setup-dotnet` steps.
- **`references/testing.md` and `references/project-config.md`** point
at that rule.
- **`CODESTYLE.md` .NET summary** names the rule, so the new check's
`intentRef` lands on it.
- **Audit:** a new `csharp.sdk.globaljson` letter check in
`spec/project-types.json`, and `global.json` in the csharp type's
`requiredFiles`.
- **Baseline:** a hub snippet `catalog/snippets/configs/global.json`
(the `sdk` block only) and its `spec/files.json` entry, so standup and
the presence audit carry the file.
- The generated `.github/skills/` and `.claude-plugin/` copies are
regenerated.

Follow-ups filed rather than done here: #4106 (the Linux installer
installs only the newest SDK line by default) and #4107 (the hub's
`setup-dotnet` steps read `global.json` once every .NET repo carries
it).

The shape matches what the PhotoCleaner #262 change verified under SDK
10.0.404 and 11.0.100.

Part of #4103. The closing keyword belongs on the develop -> main
promotion.

## Open Findings From the Local Pass

- `.agents/skills/dotnet-codestyle/SKILL.md:121` (`introduced`): the
netstandard exception sits between "Those float with the installed SDK's
compiler." and the IDE0028/CS8652 sentence that explains it. That
sentence then reads as a result of the fixed number. The edit budget was
spent, so it was fixed in this pull request's review loop, in 9fd565e,
with its own recorded pass (no findings).

- `spec/project-types.json:9` (`introduced`), raised by the local pass
on b953e56: the csharp type's `requiredFiles` lists only
`.editorconfig` and `.gitattributes`, while the new check makes a root
`global.json` required for every C# repo. That push's edit budget was
spent, so it was fixed in the next push (requiredFiles, then the snippet
and baseline entry below).

## Declined Findings From the Local Pass

- A multi-targeting library (`netstandard2.0;net10.0`, or
`net8.0;net10.0`), or a netstandard library that is not an analyzer, has
no LangVersion instruction it can follow. Declined under `GOVERNANCE.md`
"Trust Boundaries and Hardening Effort": no fleet repo multi-targets,
and the only netstandard project in the fleet is an analyzer. A repo
that adds one is the moment to extend the rule.

## Review Follow-Ups

- The Copilot finding on the LangVersion scope was fixed in 0d54f2f.
The local pass on that fix also tightened the check: the version must be
the `100` band, a `global.json` with no `sdk` block is drift, and the
check's sentences were shortened.
- A TFM major move keeps the old SDK line installed in the hub's steps
until every repo has moved (b953e56). That ordering point is also noted
on #4107.

- The local passes on the follow-up pushes also made the version exactly
`<major>.0.100` rather than any 1xx band, scoped the LangVersion rule to
files the repo holds (the SDK's own targets set it), banned
`EnablePreviewFeatures`, named the devcontainer image as a second
SDK-line site, and moved a netstandard project's fixed `LangVersion`
with the TFM. The last pass on the head found nothing.

## Verification

- `scripts/build_dist.py --check`: generated skill distributions are
current.
- `scripts/prose_lint.py --diff origin/develop`: no violations.
- `spec/validate.py`: OK.
- `markdownlint-cli2` on the changed Markdown: 0 issues.
- Full `pytest tests/` on the head: 4009 passed, 11 skipped.
- `scripts/repo_gate.py`: all checks ok.
- Four recorded `local-strict-review` passes: six findings fixed in
round one, one in round two, one listed above and fixed, and a clean
pass on that fix.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

…Step (#4111)

## Summary

Answers CodeRabbit's review of promotion PR #4096 on the
`install-agents` scripts. The guiding rule is unchanged: run only a
command that upgrades exactly the tool's own package.

- **bun's directories.** `BUN_INSTALL` and `BUN_INSTALL_GLOBAL_DIR` are
honored on both platforms through one helper each: `bun_global_dir` on
Linux, `Get-BunDirectory` on Windows. A moved bun carries no `.bun`
segment, so it is recognized by these directories. On Windows that
includes the bin directory, where bun's copied shim sits.
`bunfig.toml`'s `install.globalDir` is deliberately not read.
- **mise (Linux).** mise's own listing is asked before any path shape,
so a tool under a relocated `MISE_DATA_DIR` reads as mise.
- **mise shims (Linux).** A command whose real path is the `mise`
executable is resolved through `mise which` before it is classified. So
a Homebrew-installed mise no longer leads to `brew upgrade mise`. A
failed lookup is reported, and nothing runs.
- **herdr post step.** One step, `herdr_after_upgrade` /
`Invoke-HerdrAfterUpgrade`, reinstalls herdr's integrations and gives
the restart guidance. It now runs after both the self-updater and a
successful manager upgrade.
- **Scoop roots (Windows).** With `SCOOP` or `SCOOP_GLOBAL` unset,
`Get-ScoopRoot` reads `root_path` and `global_path` from scoop's own
`config.json`, then falls back to the defaults.
- **Test.** The unwritable-npm test skips when run as root.

Two other CodeRabbit findings on #4096 were answered there. The
protected-root boundary was already fixed by #4105. Windows bun
shim-target ownership was declined with evidence.

## Verification

- `tests/test_install_agents.py` passes, including the PowerShell cases
under `pwsh`.
- The new tests cover each case, and the bun-relocation and mise-shim
tests fail on the previous code.
- shellcheck, shfmt, PSScriptAnalyzer and the prose gate are clean.
- Three recorded local passes ran. The last found nothing open.

## Pre-existing Findings, Not Filed

- `host-setup/linux/install-agents.sh` `manager_plan` (`pre-existing`,
from #4095): the Node-runtime exception covers only node. A tool
pip-installed into a mise-managed Python would get `mise upgrade
python`. Among the managed tools only cswap could reach it, and only if
installed with pip instead of the documented `uv tool install`.
- `host-setup/linux/install-agents.sh` (`pre-existing`, consolidation):
`mise_key` runs from both classification and the plan, so `mise ls
--json` runs twice per mise tool. It is harmless, and each run is
bounded by `timeout 30`.

## Not Verified

- Real package-manager installs, a `verification` gap. The tests use
stub managers and modeled layouts. Neither real bun, mise and scoop
installs, nor PowerShell's handling of their shims, has been exercised.
The maintainer's Windows 11 retest after this release covers the
npm-installed opencode case.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>

… Dead Bound (#4112)

## Summary

Answers the attested local review of promotion PR #4096 at 6ba6179.
Copilot reached its round cap there, so the policy hands that head to an
attested local pass.

- **herdr report vs upgrade.** The report now shows herdr's
outdated-integration and restart notes for a manager-owned herdr too, on
both platforms, but only when the manager plan will run. That matches
the post step `--upgrade` runs since #4111.
- **`Hide-Home` in `install-tools.ps1`.** It had the unbounded profile
match that #4105 fixed only in `install-agents.ps1`, so it now gets the
same component boundary. The two copies are textually identical, as
`host-setup/README.md` requires, and a test holds them equal.
- **One path-containment helper.** The new `Test-PathUnder -Path -Root
[-Below]` in `install-agents.ps1` replaces seven hand-written "at or
under this root" checks, each of which normalized separators in its own
way. Behavior is unchanged, including a scoop root written with a
trailing separator or on an absent drive. `Hide-Home` stays
self-contained so its two copies remain identical.
- **`Resolve-FinalPath`.** Its depth bound and comment were dead code: a
link loop throws from `ResolveLinkTarget` first, and the threat model
rules out guarding against it. They are removed.

## Verification

- `tests/test_install_agents.py` passes, including the PowerShell cases
under `pwsh`. `tests/test_install_tools.py` and
`tests/test_cli_usage.py` pass too.
- New tests cover the manager-owned herdr notes, the identical
`Hide-Home` copies, `Test-PathUnder`, and a scoop root with a trailing
separator. The scoop-root test fails on the intermediate commit.
- shellcheck, shfmt, PSScriptAnalyzer and the prose gate are clean.
- Three recorded local passes ran. The last found nothing open.

## Not Verified

- Real package-manager installs, a `verification` gap. The pure path
helpers, `Test-PathUnder` included, run on the native Windows CI runner,
which passed on this head. Real scoop, bun and npm installs have not
been exercised. The maintainer's Windows 11 retest after this release
covers the npm-installed opencode case.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>