feat(subagents): add TUI slash subagent runtime browser

[JimmyDaddy]

What

Ports the custom slash command + slash-subagent workflow to the Go implementation, with the runtime-management UI currently scoped to the TUI.

Skill and command infrastructure

• Parses disable-model-invocation: true so skills can remain slash-invokable while staying out of the model prompt index.
• Loads agent files from .reasonix/agents/ and .claude/agents/ as runAs=subagent slash skills, including CJK file names.
• Supports settings-defined shell slash commands, with project scope overriding global scope.
• Keeps dispatch priority as:

Custom commands (.md) > Shell commands (settings.json) > Skills / Agents

• Agent-style skills are invoked directly as /<name>.

TUI slash subagents

• Bare /subagents is the public task-center entrypoint for active and retained slash-subagent runs.
• Explicit /subagents cancel <id-or-alias> and /subagents clear [completed|failed|canceled|all] remain available as controller-backed management commands.
• Slash arg completion now emphasizes cancel and clear;
• The task-center list view supports filter mode via /, matching alias, skill, or state.
• Renders live slash-subagent output as a structured embedded block inside the main transcript, in conversational order with the parent session.
• Keeps the main composer interactive while a background slash subagent runs.
• Supports collapsing/expanding the live block with Ctrl-O; terminal states auto-expand so the final answer is visible in full.
• Keeps the detail view as an independent read-only replay screen for any retained run.
• Reuses transcript-style rendering for thinking, tool calls, blocked tools, answers, usage, and terminal state in the detail screen.
• Reuses shared picker helpers for selection clamping and rune-safe backspace handling, reducing duplicated TUI logic.
• Localizes new /subagents picker and controller copy through internal/i18n.
• Supports scrolling inside the detail screen, Esc to return to the main agent, and Ctrl-D to quit.
• Preserves completed subagent results without compressing the final answer before handing them back to the parent agent flow.

Desktop scope

• Desktop keeps existing slash skills such as /explore visible.
• Desktop omits /subagents from command autocomplete and blocks hand-entered /subagents runtime-management commands with a localized TUI-only notice.
• Desktop does not expose ListSubagents, SubagentDetail, or CancelSubagent bindings in this slice.

Merge notes

This branch has been merged with the latest upstream/main-v2. The final conflict resolution keeps both sides of the desktop workspace changes:

• upstream recent-workspace persistence/list behavior
• this branch's workspace readability check and macOS Files & Folders permission error guidance

Tests

• git diff --check
• root module: gofmt check, go vet ./..., go build ./..., go test ./..., go test -race ./...
• golangci-lint run --timeout=5m
• make build
• desktop module: gofmt check, go mod tidy, go vet ./..., go build ./..., go test ./...

Ref

• docs/SLASH_SUBAGENT_DESIGN.md

[esengine]

Took this for a thorough pass — the architecture is sound: the slash-subagent surface is strictly user-invoked, task.go's model-facing Description/Schema is unchanged (no new model-driven spawning), and disable-model-invocation keeping slash skills out of the model index is exactly right. Root Go build and the frontend typecheck are clean.

One blocker before this can land: it fails go test on Windows, in internal/control:

--- FAIL: TestSubmitRunsShellCommand
    submit_test.go:78: Submit shell command notices = [\a b"], want [a b]
--- FAIL: TestSubmitRunsShellCommandWithMultilineArgs
    submit_test.go:91: ... = [\a b"], want [a b]

The tests use printf '%s\n' with /echo "a b", which assumes POSIX sh -c semantics. On Windows shellInvocation falls back to cmd /c, where single-quotes and printf behave differently, so the output comes back mangled. Our CI only runs macOS/ubuntu, so this stays green there but breaks on Windows.

Could you:

1. Make those tests OS-aware (skip on Windows, or use a portable command that behaves the same under both sh -c and cmd /c), and
2. Confirm the shell-command feature itself quotes args sanely under cmd /c on Windows (or document that user shell commands are written per-platform)?

Given the size (+5957/45 files), I'll also want to walk through docs/SLASH_SUBAGENT_DESIGN.md once the Windows failures are resolved. Thanks for the design doc — it made the review much easier.

[JimmyDaddy]

Took this for a thorough pass — the architecture is sound: the slash-subagent surface is strictly user-invoked, task.go's model-facing Description/Schema is unchanged (no new model-driven spawning), and disable-model-invocation keeping slash skills out of the model index is exactly right. Root Go build and the frontend typecheck are clean.

One blocker before this can land: it fails go test on Windows, in internal/control:

--- FAIL: TestSubmitRunsShellCommand
    submit_test.go:78: Submit shell command notices = [\a b"], want [a b]
--- FAIL: TestSubmitRunsShellCommandWithMultilineArgs
    submit_test.go:91: ... = [\a b"], want [a b]

The tests use printf '%s\n' with /echo "a b", which assumes POSIX sh -c semantics. On Windows shellInvocation falls back to cmd /c, where single-quotes and printf behave differently, so the output comes back mangled. Our CI only runs macOS/ubuntu, so this stays green there but breaks on Windows.

Could you:

1. Make those tests OS-aware (skip on Windows, or use a portable command that behaves the same under both sh -c and cmd /c), and
2. Confirm the shell-command feature itself quotes args sanely under cmd /c on Windows (or document that user shell commands are written per-platform)?

Given the size (+5957/45 files), I'll also want to walk through docs/SLASH_SUBAGENT_DESIGN.md once the Windows failures are resolved. Thanks for the design doc — it made the review much easier.

Thanks for the careful pass - I've pushed an update for the Windows blocker.

What changed:

• Made the shell-command controller tests platform-safe. The /echo "a b" and multiline-arg cases now use the test binary as a helper command instead of POSIX-only printf '%s\n', so they exercise the same raw slash-arg path without assuming sh syntax.
• Added lower-level internal/command coverage for quoted user args and a plain arg case.
• Updated docs/SLASH_SUBAGENT_DESIGN.md to document the contract: settings slash commands are platform shell snippets (cmd /c on Windows, sh -c elsewhere); Reasonix preserves quoted user args but does not translate POSIX shell syntax to Windows cmd.exe syntax.
• I also included a Windows execution-path change for settings slash commands and the hook shell spawner: on Windows it sets SysProcAttr.CmdLine for cmd /c ... so Go's argv escaping does not mangle the quoted command string before cmd.exe applies its own parsing. Could you sanity-check whether that is an acceptable direction for this codebase? If you'd rather keep the runtime path unchanged and only make the tests/docs OS-aware, I can trim that part back.

Validation run locally:

• go test ./internal/command ./internal/control ./internal/hook
• go test ./...
• cd desktop && go test ./...
• cd desktop/frontend && pnpm build
• git diff --check

[JimmyDaddy]

@esengine latest conflicts are resolved again and the branch is back to a clean, mergeable state.

Before I keep chasing main-v2 conflicts every few days, could you help confirm whether this feature direction is still something you'd like reviewed / considered for merge?

If the overall direction looks good, I'd really appreciate a review pass and any design feedback you have, especially on the TUI-only /subagents management surface

If this is not a direction you want to take, no worries at all — a short "not aligned" reply is enough and I can close the PR instead of continuing to carry the merge burden.

Thanks.