feat: /rewind — jump back to a prior user message and re-prompt (#21910) by SaguaroDev · Pull Request #23445 · NousResearch/hermes-agent

SaguaroDev · 2026-05-10T22:48:07Z

Closes #21910

Adds a first-class /rewind affordance — jump back to a prior user message in the current session, re-prompt from there with the original text pre-filled and editable, while keeping the truncated rows on disk for audit. Equivalent to Claude Code's double-Esc behavior, scoped per the issue to CLI + TUI only (gateway is a v2 follow-up).

What ships

Schema (v11 → v12)

messages.active INTEGER NOT NULL DEFAULT 1 — soft-delete flag
sessions.rewind_count INTEGER NOT NULL DEFAULT 0 — audit counter
Deferred idx_messages_session_active index (created after _reconcile_columns so legacy DBs don't trip on the missing column)

Declarative reconcile + a belt-and-suspenders UPDATE messages SET active = 1 WHERE active IS NULL covers existing databases.

SessionDB API

rewind_to_message(session_id, target_message_id) → {rewound_count, target_message, new_head_id} — refuses non-user targets and cross-session targets with ValueError; idempotent
restore_rewound(session_id, since_message_id) → int — reverse op for a future /unrewind
list_recent_user_messages(session_id, limit=20) — picker source
get_messages, get_messages_as_conversation, search_messages gain include_inactive: bool = False. By default rewound rows are excluded everywhere — including session_search — so re-prompts and search hits see a clean transcript. Opt in for audit / debug tooling.

CLI (`/rewind`)

Slash command registered in hermes_cli/commands.py; handler in cli.py mirroring _handle_branch_command:

Picker (existing _run_curses_picker) shows the last 10 user messages
Soft-truncate via SessionDB.rewind_to_message
Reload active-only history into conversation_history
Mirror /branch's agent surgery (reset_session_state, _invalidate_system_prompt, _last_flushed_db_idx)
Notify _memory_manager.on_session_switch(..., rewound=True) so per-turn caches invalidate (Hindsight etc.)
Pre-fill the prompt_toolkit buffer with the selected message text — edit and resubmit

TUI (Ink + tui_gateway)

command.dispatch gets a name == "rewind" branch returning the new {"type": "prefill", "message": <text>, "notice": "↶ Rewound N message(s)…"} payload
/rewind added to _PENDING_INPUT_COMMANDS so slash.exec routes it correctly
New prefill variant on CommandDispatchResponse + asCommandDispatch validator; createSlashHandler handles it by calling ctx.composer.setInput(d.message) (no auto-submit, unlike send)
v1 auto-picks the most recent user turn in the TUI — a dedicated Ink picker overlay for multi-step rewind is tracked as a follow-up

Memory hook (#6672 extension)

MemoryProvider.on_session_switch documents rewound: bool = False. MemoryManager.on_session_switch keeps it in **kwargs (not an explicit named kwarg) so existing tests that assert exact extra dict equality stay green. Existing callers (/branch, compression, etc.) are untouched.

Test coverage

Suite	Tests	Status
`tests/test_hermes_state.py`	+10 new (rewind primitives, migration v11→v12, FTS filter on/off, ValueError paths)	✅
`tests/cli/test_rewind_command.py`	+6 (handler, mocked picker, memory notify, guard paths)	✅
`tests/tui_gateway/test_rewind_command.py`	+7 (prefill payload, in-memory truncation, DB soft-delete, memory notify, busy-session refusal, registry placement)	✅
`ui-tui/src/__tests__/asCommandDispatch.test.ts`	+3 (prefill variant + malformed rejection)	✅

All existing suites in the blast radius (memory, cli, branch, hermes_state, tui_gateway) — 253 / 253 passing post-rebase.

Pre-existing failures on origin/main that are unrelated and unaffected by this PR:

test_anthropic_adapter.py::test_returns_token_from_credential_files — mock setup issue
test_cli_save_config_value.py::TestSaveConfigValueAtomic::* — missing ruamel module in CI env

Acceptance checklist (from #21910)

/rewind command in CLI and TUI
Truncated transcript no longer shown to the model on the next turn (active-only filter in get_messages_as_conversation)
Rewound rows preserved in SessionDB (active=0), not deleted
Memory providers notified on rewind (on_session_switch(rewound=True))
Tests covering: rewind to first user message, rewind mid-session, rewind then continue normally, rewind across tool-call boundaries
Double-Esc keybind — deferred per design discussion (/rewind slash command is the canonical entry; keybind can land in a follow-up after we audit Esc Esc for conflicts with prompt_toolkit's existing keymap)

Out of scope for v1 (tracked as follow-ups to #21910)

Dedicated Ink picker overlay for multi-step rewind. The TUI's v1 auto-picks the most recent user turn, which covers the common case. The CLI ships the full picker.
Gateway platforms (Telegram, Discord, etc.) — issue body explicitly scopes v1 to CLI/TUI. xavierchoi's comment on the issue outlines the v2 work (best-effort message deletion, permission probe, thread-scoped boundaries).
Compression-fork boundary handling — rewinding across a context-compression session split would need the parent-chain logic in resolve_resume_session_id extended; not addressed here.

SaguaroDev · 2026-05-10T22:49:55Z

@xavierchoi — you volunteered to help test on the original issue (#21910). v1 is up here when you have a moment. A few notes specifically on the gateway concerns you raised:

This PR is intentionally scoped to CLI + TUI only, matching the issue body. Your v2 ideas (best-effort message deletion, per-platform message ID tracking, permission probes, thread-scoped boundaries) are not addressed here — they're the natural follow-up.
The memory-provider notification path is in place (on_session_switch(..., rewound=True)), so when v2 lands on gateway it can hook the same seam without rework.
In the TUI, v1 auto-picks the most recent user turn rather than rendering a dedicated Ink picker overlay. The CLI ships the full curses picker (last 10 user messages). I'd value a second opinion on whether the auto-pick TUI behavior is acceptable or whether the picker overlay should block v1.

What would help most:

Smoke-test /rewind in the CLI — pick a mid-session message, verify the prompt prefill is correct and that the next turn sees the truncated transcript.
Same in the TUI (hermes --tui) — confirm the prefill lands in the composer and that you can edit before submitting.
Sanity-check that session_search no longer surfaces rewound content (rewound rows are excluded by default; opt-in via include_inactive=True for audit).

Happy to iterate on anything that feels off. Thanks for the detailed v2 scope on the issue — it shaped how the memory hook signature is laid out so this PR doesn't paint v2 into a corner.

xavierchoi · 2026-05-10T22:54:03Z

Nice, glad to see this moving. The active-flag soft-delete approach is cleaner than I expected — keeping rewound rows queryable for audit while excluding them from search by default is the right call. Will keep an eye on the v2 gateway scope.

SaguaroDev · 2026-05-11T09:42:30Z

v2 plan is mapped out and committed to a separate branch (feat/rewind-v2-gateway-21910) — will open the follow-up PR once this one merges. Plan doc: docs/plans/2026-05-11-rewind-v2-gateway.md on that branch.

Quick scouting summary in case it's useful for this review:

SendResult.message_id already returns from every send → no new contract needed
base.delete_message(chat_id, message_id) -> bool already exists (Telegram implements; ephemeral-TTL path uses it) → Discord/Slack/Matrix just need adapter overrides
/rewind is already in GATEWAY_KNOWN_COMMANDS via the central registry → handler is the only gateway-side gap
v1's on_session_switch(rewound=True) hook is what v2 reuses for memory-provider notification on gateway rewinds

v2 scope (9 TDD-shaped tasks):

outbound_messages table in hermes_state.py linking (session, seq) → (platform, chat, thread, msg_id)
Wire record_outbound_message into the gateway send-finalisation path
3-5. delete_message overrides for Discord / Slack / Matrix
Gateway /rewind handler — calls rewind_to then iterates outbound rows in the rewound range and best-effort deletes
Ephemeral summary reply (Removed N, kept M (missing Manage Messages on Discord))
Docs + hindsight regression in gateway mode
Changelog + PR

Thread-scoped by design ((chat_id, thread_id) filter on the outbound query), so the rewind blast radius is the same as the session boundary — matches @xavierchoi's v2 suggestion.

Holding the v2 branch until this one lands so I can rebase cleanly if anything shifts in review.

SaguaroDev · 2026-05-12T01:48:02Z

Review in 5 minutes

Quick verification recipe if you want to smoke-test before reading the diff:

gh pr checkout 23445
python -m pytest tests/test_hermes_state.py tests/cli/test_rewind_command.py tests/tui_gateway/test_rewind_command.py
# expect: 241 passed

Then in two terminals, one CLI and one TUI:

CLI (hermes):

Send 3-4 user messages, get responses
/rewind → curses picker shows last N user turns
Pick one → transcript truncates, picked message lands prefilled in the prompt, edit and submit
session_search "..." should NOT surface rewound content (opt-in via include_inactive=True for audit)

TUI (hermes --tui):

Same setup, then /rewind
v1 auto-picks the most recent user turn (single-step undo, Claude-Code-style)
The picked message lands in the composer for editing — does NOT auto-submit
Dedicated picker overlay = v1 follow-up

The one open design call I'd value a second opinion on: the TUI auto-picks the most recent user turn rather than rendering a dedicated Ink picker overlay. My read is this matches the most common use case (single-step undo) and the picker overlay is a clean v1.1 follow-up; happy to block on the overlay if you'd rather ship them together.

Scope: CLI + TUI only, as the issue body specifies. Gateway (Discord/Telegram/Slack message deletion) is v2, mapped out separately and reuses the on_session_switch(rewound=True) seam this PR adds.

Size note: +1026 / −10 across 13 files but the diff splits cleanly along the 6 commits — the schema/primitives commit is independently mergeable as a no-op-from-user-perspective base if you'd prefer to land it first and layer CLI + TUI as follow-ups. Happy to split.

@teknium1 — could I get a look at this when you have a window? Most of the surface area is hermes_state.py + cli.py + tui_gateway/server.py; CI is green locally on 241 rewind-specific tests and the branch is rebased clean on main.

xavierchoi · 2026-05-12T02:34:54Z

Appreciate the ping, and honestly impressed at how fast this came together.

Have to be upfront: I can't smoke-test on my end right now — I'm running Hermes gateway-only on Discord, no local CLI/TUI session set up. So no signal from me on (1) and (2). For (3), I'll trust the test coverage — the include_inactive default-off behavior on search_messages reads correctly in the diff.

On the TUI auto-pick vs dedicated picker overlay design call: my vote is ship v1 with auto-pick, track the overlay as a follow-up. Single-step undo covers the muscle-memory case (it's what double-Esc trains you to expect), and the CLI already has the full picker for the multi-step case. Blocking v1 on the overlay feels like scope creep for a marginal UX gain.

Also — thanks for mapping the v2 gateway branch already. outbound_messages keyed by (session, seq) → (platform, chat, thread, msg_id) and the thread-scoped boundary are exactly the shape I had in mind, and reusing the on_session_switch(rewound=True) seam means v2 won't need to retrofit anything. Looking forward to that PR.

SaguaroDev · 2026-05-12T21:38:03Z

@teknium1 @OutThisLife — could I get a review when you have a window? Diff splits cleanly:

CLI + SessionDB half (cli.py, hermes_state.py, hermes_cli/commands.py, agent/memory hook) — @teknium1 if you're the right set of eyes
TUI half (tui_gateway/, ui-tui/src/, asCommandDispatch validator) — @OutThisLife since you own most of ui-tui/

253/253 tests green in the blast radius. Acceptance checklist 5/6, with the deferred Esc-Esc keybind explained in the PR body. Happy to split into two PRs if that's easier to review.

…ch#21910) Schema v12 adds: - messages.active (default 1) — soft-delete flag for /rewind - sessions.rewind_count (default 0) — audit counter - idx_messages_session_active deferred index New SessionDB methods: - rewind_to_message(session_id, target_message_id) — soft-deletes rows >= target_id, refuses non-user targets, increments rewind_count - restore_rewound(session_id, since_message_id) — undo for stretch goal - list_recent_user_messages — picker source Existing methods get include_inactive kwarg (default False): - get_messages, get_messages_as_conversation, search_messages. Rewound rows excluded from session_search by default — opt-in for audit. The deferred index pattern (DEFERRED_INDEX_SQL run after _reconcile_columns) avoids 'no such column: active' on legacy pre-v12 databases, since executescript(SCHEMA_SQL) runs before column reconciliation.

Adding rewound as an explicit named kwarg on MemoryManager.on_session_switch broke existing tests that assert exact equality on the extra kwargs dict captured by providers (test_manager_fans_out_to_all_providers, test_manager_reset_flag_preserved). The MemoryProvider base class still documents rewound as a known kwarg, but the manager now passes it through **kwargs like any other forwarded flag. The CLI rewind handler already passes rewound=True keyword-style, so behavior is identical.

…ousResearch#21910) Adds the TUI half of the /rewind feature so the Ink terminal UI gets the same affordance as the prompt_toolkit CLI. Python side (tui_gateway/server.py): - /rewind added to _PENDING_INPUT_COMMANDS so slash.exec rejects it and the TUI falls through to command.dispatch (the only path with access to live session state + memory hooks). - New command.dispatch branch for name == "rewind": v1 auto-picks the most recent user turn (Claude-Code-style single- step undo), calls SessionDB.rewind_to_message, refreshes the in-memory history, fires _memory_manager.on_session_switch with rewound=True, and returns the new "prefill" payload. - A dedicated picker overlay (multi-step rewind) is tracked as a follow-up to NousResearch#21910. TS side (ui-tui/src/): - New "prefill" variant on CommandDispatchResponse + asCommandDispatch validator. Mirrors "send" but does NOT auto-submit; the client drops the message into the composer for editing. - createSlashHandler renders the optional notice via sys() and calls ctx.composer.setInput(d.message), letting the user edit-and-resubmit the rewound turn — the core UX promised by the issue. Tests: - 7 new tui_gateway tests covering prefill payload shape, in-memory history truncation, DB soft-delete, memory-provider notification (rewound=True), busy-session refusal, missing-session error, and registry placement in _PENDING_INPUT_COMMANDS. - Extended asCommandDispatch vitest covering the new prefill variant (with + without notice, and rejection of malformed payloads). Out of scope for v1 (tracked as NousResearch#21910 follow-up): - Dedicated picker overlay in Ink (the multi-step rewind UI). v1 auto- picks the most recent user turn, matching the most common case. - Gateway platforms (Telegram, Discord, etc.) — issue scopes v1 to CLI + TUI only.

SaguaroDev · 2026-05-19T00:18:46Z

Rebased on main and resolved the conflict in hermes_state.py — entirely against c03acca (the AUTOINCREMENT-id ordering fix). Four hunks:

get_messages() — kept the new include_inactive parameter and soft-delete filter; switched ORDER BY timestamp, id → ORDER BY id to match main's clock-regression-safe ordering.
get_messages_as_conversation() inner query — same shape: preserved active_clause, dropped timestamp from the ORDER BY.
search_messages() signature — merged both new params: sort (main) + include_inactive (this PR).
search_messages() docstring — merged both blocks.

Drive-by: the new recent_user_messages() (rewind picker) was using ORDER BY timestamp DESC, id DESC — switched to ORDER BY id DESC for consistency. Same WSL2 clock-regression risk would have applied to the picker.

241/241 in the rewind blast radius after the rebase, plus 67/67 in adjacent session/search/memory tests.

@alt-glitch — would you be able to route this to a reviewer? I keep having to rebuild as main moves; a second pair of eyes on the SessionDB half (the include_inactive + ordering interaction in particular) would unblock the v2 gateway follow-up branch as well.

SaguaroDev · 2026-05-23T19:29:29Z

@alt-glitch is this change not wanted? I will stop dev work on it, but I keep having to update it while there are no responses. That's fine but please tell me to drop it so I can move on with my life. Please

Resolve conflicts in hermes_state.py and tests/test_hermes_state.py with the platform_message_id/observed work that landed on main (PRs NousResearch#27166, #31a01001, #4a91e364). - SCHEMA_VERSION 12 (HEAD) vs 13 (main) -> 14, accommodating both feature bumps (rewind 'active' flag + platform_message_id/observed). - messages table: keep both new columns ('active', 'platform_message_id', 'observed') side by side. - DEFERRED_INDEX_SQL: fold main's idx_messages_platform_msg_id into the same deferred block as idx_messages_session_active (consistent legacy-DB reconciliation path). - _read_messages_for_sessions SELECT: union both column sets and preserve the active-flag filter from the rewind branch. - test_schema_version assertions: switch the three == 12 literals to == SCHEMA_VERSION so the test tracks the constant. Verified with: pytest tests/test_hermes_state.py (227 passed) and pytest -k 'rewind or platform_message or observed or active' (348 passed).

# Conflicts: # hermes_state.py # tests/test_hermes_state.py

SaguaroDev · 2026-05-31T21:11:33Z

@teknium1 This PR is rebased clean on latest main and conflict-free (mergeable: MERGEABLE). The only thing blocking it is the fork-PR workflow gate — all 5 checks (Tests, Lint, Docker, Nix, Supply Chain Audit) show action_required and need a maintainer to click Approve and run workflows before CI can run.

Once CI is unblocked it should go green: the /rewind primitives and migration logic are covered by tests in tests/test_hermes_state.py (243 passing locally), and the merge resolved two conflicts — in hermes_state.py (kept main's FTS5-availability-aware v11 migration plus the new v12 messages.active block) and tests/test_hermes_state.py (kept both new test classes).

Would appreciate the workflow approval and a look when you have a moment. Thanks!

teknium1 · 2026-06-01T04:08:51Z

Superseded by #36229. We folded your /rewind work into the existing /undo command instead of a separate command + picker: /undo [N] backs up N user turns (default 1), prefills the backed-up message for editing, and soft-deletes the truncated rows on disk.

Your substantive infrastructure was cherry-picked onto current main with your authorship preserved in git log — the SessionDB rewind primitives (messages.active, rewind_to_message, restore_rewound, list_recent_user_messages), the on_session_switch(rewound=True) memory hook, and the TUI command.dispatch prefill payload are all your commits. Thanks for the clean, well-tested groundwork — the schema reconcile + deferred-index ordering made the rebase onto the v14 schema painless.

alt-glitch added type/feature New feature or request comp/cli CLI entry point, hermes_cli/, setup wizard comp/tui Terminal UI (ui-tui/ + tui_gateway/) P3 Low — cosmetic, nice to have labels May 10, 2026

SaguaroDev force-pushed the feat/rewind-21910 branch from ec6835c to 3f1f5db Compare May 12, 2026 01:41

SaguaroDev mentioned this pull request May 12, 2026

Feature: rewind/edit-and-resubmit (like Claude Code's double-Esc) #21910

Closed

5 tasks

SaguaroDev mentioned this pull request May 13, 2026

feat: /rewind v2 — gateway thread-scoped rewind + best-effort delete (#21910) #25074

Open

4 tasks

SaguaroDev added 6 commits May 18, 2026 19:41

feat(memory): add rewound kwarg to on_session_switch hook

48c465e

feat(cli): add /rewind slash command and handler

7845d4d

test(rewind): cover DB primitives and CLI handler (NousResearch#21910)

572e042

SaguaroDev force-pushed the feat/rewind-21910 branch from 3f1f5db to 05e33f3 Compare May 19, 2026 00:25

SaguaroDev added 2 commits May 23, 2026 15:34

Merge remote-tracking branch 'origin/main' into feat/rewind-21910

a737a3b

# Conflicts: # hermes_state.py # tests/test_hermes_state.py

SaguaroDev force-pushed the feat/rewind-21910 branch from b73fc67 to a737a3b Compare May 31, 2026 20:44

teknium1 mentioned this pull request Jun 1, 2026

feat(undo): /undo [N] backs up N user turns with prefill + soft-delete (#21910) #36229

Merged

teknium1 closed this Jun 1, 2026

SaguaroDev deleted the feat/rewind-21910 branch June 1, 2026 06:49

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: /rewind — jump back to a prior user message and re-prompt (#21910)#23445

feat: /rewind — jump back to a prior user message and re-prompt (#21910)#23445
SaguaroDev wants to merge 8 commits into
NousResearch:mainfrom
SaguaroDev:feat/rewind-21910

SaguaroDev commented May 10, 2026

Uh oh!

SaguaroDev commented May 10, 2026

Uh oh!

xavierchoi commented May 10, 2026

Uh oh!

SaguaroDev commented May 11, 2026

Uh oh!

SaguaroDev commented May 12, 2026

Uh oh!

xavierchoi commented May 12, 2026

Uh oh!

SaguaroDev commented May 12, 2026

Uh oh!

SaguaroDev commented May 19, 2026

Uh oh!

SaguaroDev commented May 23, 2026

Uh oh!

SaguaroDev commented May 31, 2026

Uh oh!

teknium1 commented Jun 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

SaguaroDev commented May 10, 2026

Closes #21910

What ships

Schema (v11 → v12)

SessionDB API

CLI (/rewind)

TUI (Ink + tui_gateway)

Memory hook (#6672 extension)

Test coverage

Acceptance checklist (from #21910)

Out of scope for v1 (tracked as follow-ups to #21910)

Uh oh!

SaguaroDev commented May 10, 2026

Uh oh!

xavierchoi commented May 10, 2026

Uh oh!

SaguaroDev commented May 11, 2026

Uh oh!

SaguaroDev commented May 12, 2026

Uh oh!

xavierchoi commented May 12, 2026

Uh oh!

SaguaroDev commented May 12, 2026

Uh oh!

SaguaroDev commented May 19, 2026

Uh oh!

SaguaroDev commented May 23, 2026

Uh oh!

SaguaroDev commented May 31, 2026

Uh oh!

teknium1 commented Jun 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

CLI (`/rewind`)