feat(tui): add /undo and /revert commands

[hshinosa]

What

Two new TUI slash commands that reuse the dashboard's edit-message primitives:

• /undo — rollback the most recent user turn (files + messages).
• /revert — picker over the session's user messages; rollback to any earlier point. Also accepts /revert <message-id> for scripts and copy/paste from the dashboard.

Why

The dashboard's "edit message" flow already lets an operator drop a turn and revert every file the agent touched in it. The TUI had no way to reach that machinery — operators had to leave the terminal, open the dashboard, find the message, and click a button, or hand-edit ~/.antares/checkpoints/ and prune the message log by hand.

UX

Inline confirm block, no modal. y on an empty composer commits, Esc discards, everything else discards and falls through so the keystroke lands normally.

Revert to the last turn?
  3 file(s) to restore (1 created files will be removed).
  1 file(s) edited outside this session will be skipped.
  [revert] internal/agent/agent.go
  [delete] internal/agent/foo_new.go
  [skip  ] README.md
Press y to confirm, any other key to cancel.

/revert uses the same modal picker /model and /theme already use — one keyboard idiom for every selection. Preview label carries the message text (single-line elided at 60 chars); right column shows latest / N turns ago.

Commit path

Files first, then messages, then drop the persisted context summary:

1. agent.RollbackSince(sessionID, marker, skipExternal=true) — restore snapshots, remove agent-created files, honour "externally changed" (leave user-edited files alone).
2. store.DeleteMessagesFrom(sessionID, marker) — drop the message and everything after it.
3. Drop meta.context_compact — its through_seq now points past deleted rows; leaving it would either replay a stale summary or hide live messages.
4. Reload the transcript from storage — the message log is authoritative, so the composer no longer shows the reply we just deleted server-side.

If file restore fails, the message log stays intact — a half-done state where messages are gone but files are still at their post-turn contents is worse than "nothing happened, try again".

Every primitive is the same one server/handlers_chat.go's handleEditMessage uses, so the dashboard and TUI stay in sync — a fix in the checkpoint layer lands in both surfaces at once.

/redo — intentionally not included

OpenCode and omp implement /redo by recording reverse-ops per tool call into a session journal. Antares has append-only messages plus per-turn checkpoint tarballs; no branch semantics. Adding /redo would need a schema change (session_undo_log table + reverse-op recording per tool call). Left as future work — same conclusion as the "known bugs" gap OpenCode issue #15391 documents.

Files

File	LOC	Purpose
internal/tui/undo.go	327 (new)	Both commands, pending state, commit, transcript reload, previewText
internal/tui/undo_test.go	128 (new)	previewText, confirm message shape, cancel state, commands registered
internal/tui/tui.go	+5	pending *pendingRevert field
internal/tui/tui.go	+22	Key handler intercept for y/Esc after a stage
internal/tui/commands.go	+2	Register /undo + /revert in palette

Not touched

• Dashboard, HTTP handlers, agent, store, checkpoint — all reused as-is.
• No schema change.
• No config field.
• No new dependency.

Verify

• GOTOOLCHAIN=go1.26.3 go build ./... clean
• GOTOOLCHAIN=go1.26.3 go vet ./... clean
• go test ./internal/tui/... ./internal/agent/... ./internal/server/... ./internal/checkpoint/... ./internal/store/... -count=1 — 5 packages green
• 4 new subtests
• Live smoke: go run ./cmd/antares tui → /undo, /revert, /revert <id> all work; empty-session guards fire the expected notice; cancel path leaves state untouched.

[enowdev]

Merged into main via 4dd05ab, with two fixes applied on top for defects found during review. Both were verified against the real checkpoint store before and after.

1. /revert <message-id> with an unknown id could destroy the session log. The id went straight to PreviewChangesSince with no validation. touchedSince matches nothing for an unrecognised marker, so PreviewSince returns (empty, nil) — indistinguishable from "this turn changed no files". The operator saw "No file changes to revert — only the chat history will be trimmed", pressed y, RollbackSince silently no-opped, and DeleteMessagesFrom returned ErrNotFound, surfacing "files restored, but trimming messages failed" when neither had happened. (DeleteMessagesFrom is correctly session-scoped, so no rows were actually touched — the damage was the false report, not data loss.) Now the id is resolved against the message log first and rejected before anything is staged; assistant ids are rejected too, since truncating from one would cut the log at the wrong point.

2. The "N skipped" count was always zero. commitPending read len(res.Failed), but RestoreSince drops externally-changed files with a bare continue and never records them in Failed. Confirmed by probe: agent writes a file, user edits it outside the session, revert → Restored=[] Deleted=[] Failed=map[]. The confirm block correctly warned "1 file(s) edited outside this session will be skipped", then the summary silently dropped it. The count now comes from the staged preview, and genuine restore failures are reported separately rather than conflated with skips.

Three regression tests added in internal/tui/undo_guard_test.go, backed by a real in-memory store rather than a fake.

The design itself is good — reusing handleEditMessage's exact primitives, files-before-messages ordering, and dropping context_compact all match the dashboard, so the two surfaces stay in sync. Agreed on leaving /redo out.

Verified: go build, go vet, full go test ./..., and go test -race over the affected packages.