hooks/builtins: name the resolved shell in the env block

[trungutt]

The problem

On Windows, docker-agent sessions repeatedly emit POSIX syntax — &&, grep, head, 2>/dev/null, ls -la, $(pwd) — under Windows PowerShell 5.1 or cmd.exe. Each attempt fails with a distinctive error, the model re-derives the dialect from scratch, and often emits the same wrong syntax on the next turn. Long sessions burn many tool calls this way.

Why

The env block at the top of every session says:

<env>
  Operating System: Windows
  ...
</env>

Windows alone doesn't disambiguate the shell (PowerShell 5.1? 7? cmd.exe? — different dialects). The model defaults to POSIX by habit since that's most of its training data. The shell tool description does carry a shell-specific hint (shellSyntaxHint), but it sits far from where the model is paying attention.

The change

Add a Shell: line to the env block, resolved via a new shellpath.ShellBaseName helper:

<env>
  Operating System: Windows
  ...
  Shell: powershell (C:\Windows\System32\WindowsPowerShell\v1.0\powershell.exe)
</env>

Same treatment for pwsh, cmd, zsh, bash, etc. Prompt-cached, so no per-turn cost.

[Sayt-0]

The Shell: line in the env block addresses a real gap: sessions on Windows repeatedly emit POSIX syntax under PowerShell 5.1 / cmd.exe, and the existing shellSyntaxHint sits too far from the model's attention. The approach (reuse shellpath.DetectShell, prompt-cached line) is sound. Build and tests pass on the branch.

Two items to resolve before merge:

#	Issue	Files	Severity
1	ShellBaseName duplicates an existing private helper	pkg/shellpath/shellpath.go, pkg/tools/builtin/shell/shell.go	blocking
2	Doc comment references shellDialectHint, which does not exist in the repo	pkg/hooks/builtins/add_environment_info.go	blocking

Details in the inline comments. Two minor nits included as well, take or leave.

[aheritier]

CI is green on 3609910 (build-and-test, lint, license-check, windows-tests, build-image linux/amd64+arm64, CodeQL go/js/actions), and all three commits are signed and verified. Verified independently on the head commit: go build ./..., GOOS=windows GOARCH=amd64 go build ./pkg/..., and go test ./pkg/shellpath/... ./pkg/hooks/builtins/... ./pkg/tools/builtin/shell/... all pass.

The change does what the description says, and the previous round is fully addressed:

• shellpath.ShellBaseName is now the single implementation: grep -rn '\bshellBaseName\b' returns nothing, both call sites (shell.go:371, shell.go:412) delegate, and the moved table test in pkg/shellpath/shellpath_test.go:10 is a superset of the removed one (adds /bin/sh and "").
• grep -rn shellDialectHint returns nothing — the dead pointer is gone.
• The "No-op when Cwd is empty." note and the argsPrefix clarification are back.
• The core invariant holds: the hook resolves via shellpath.DetectShell(), the same source as shellHandler.shell (shell.go:294), so the env block cannot name a shell different from the one the tool executes.
• The prompt-cache claim checks out — AddEnvironmentInfo output sits behind the cache_control marker (pkg/session/session.go:2227).

Findings

[should-fix] Docs enumerate the env block without the new shell field.
Two reference tables list the injected fields exhaustively and are now incomplete:

• docs/configuration/hooks/index.md:209 — "Adds the working directory, git-repo status, OS, and CPU architecture."
• docs/configuration/agents/index.md:96 — "injects working directory, OS, CPU architecture, and git info into context."

Adding the resolved shell to both keeps them accurate in the same change.

[should-fix] 465e8955's commit body will land in main with stale references.
Squash merge is disabled on this repo (allow_squash_merge: false), so all three commit messages are permanent history. That body cites shellDialectHint — which does not exist in the repo — and "follow-up commit", which stops resolving once the message is in history; it also states "the pre-existing copy in the shell toolset moves over here too", which only became true in 715df7ba. The code comment was corrected in 3609910, but the commit body still carries the dead pointer. Worth a reword on the next rebase.

[optional] The env-block / shell-tool agreement isn't tested.
add_environment_info_test.go:54,61 builds the expected string from the same shellpath.DetectShell() and shellpath.ShellBaseName() calls production uses, so it pins the line's placement and format but cannot fail if the hook's resolution later diverges (e.g. someone swaps in DetectUnixShell) — which is the invariant this change exists to protect. It matches the existing displayOS()/displayArch() style in that test, so take or leave.

[optional] pkg/shellpath's package comment (shellpath.go:1) still scopes the package to "safe shell binary resolution to prevent PATH hijacking attacks (CWE-426)", while ShellBaseName is a pure formatting helper. A one-line broadening keeps the doc honest now that the package hosts both.

Leaving this as a comment rather than an approval so the two [should-fix] items — the docs enumerations and the 465e8955 commit body — can be picked up first; neither is a correctness problem, and the code itself looks good to me. Note that merge is separately gated by the earlier CHANGES_REQUESTED review, which resolving threads does not dismiss; the original reviewer needs to re-review or that review needs dismissing.