dsh-checkpoint-rewind

⭐ 如果它帮到了你

这个插件是 DSH 插件家族的一员（40+ 个，全部 Apache-2.0）。如果你在用，给个 star —— 它不会解锁任何功能，但会让下一个人在搜索里更容易找到它。

English: part of a 40+ plugin family for DeepSeek Harness. If it is useful, a star helps the next person find it — nothing is gated behind it.

Compatibility

What you get

dsh-checkpoint-rewind captures a three-state unified checkpoint — workspace, session cursor, and plugin config — and restores one or all three with a single approved command:

1. Three-state record — every checkpoint stores the workspace state (git tree SHA, or a copy manifest), the session event cursor (seq + turn boundary), and a config snapshot, tagged by source (manual / auto / guard / mutation).
2. Four capture triggers — before every mutating tool (fs/write-intent, fs/edit-intent, tools/pre-execute), on automatic interval (autoCheckpoint, default every step), manually (/checkpoint and the checkpoint tool), and as a guard before every rewind.
3. git-first provider — git stash create / commit-tree produce unreferenced snapshot objects that never touch your worktree, index, or history; restore is worktree-only and path-explicit. Non-git directories (and unborn-HEAD repos) degrade to an incremental copy provider with hardlink reuse.
4. One-shot rollback — /rewind workspace|session|config|all <target> restores the selected states; preview is a read-only impact report, diff <a> <b> compares two checkpoints, clear deletes them (this session; clear --all spans all sessions and workspaces).
5. Fork-based session rollback — session rollback replays events up to the checkpoint boundary through the official SessionStore.fork primitive into a new child session (falling back to the seeded sessions.create path when the host has no fork or the checkpoint has no boundary); the original session keeps its full history.
6. Settings page timeline — the Plugins → Checkpoints tab renders the session's checkpoints with pairwise line-level diffs.

Why another rewind plugin?

The difference in one sentence: dsh-checkpoint-rewind captures the workspace state with side-effect-free git primitives before each mutation, and makes "back to step N" one approved command — guard checkpoint first, files restored second, config restored third, session replayed fourth, each phase logged. No delta bookkeeping to drift, no message-level editing (that belongs to a different plugin), no cross-device sync.

Quick start

# 1. install the bundle into your profile
dsh plugin --profile web add "github:PerryLink/dsh-checkpoint-rewind#main"

# or from npm (published releases)
dsh plugin --profile web add dsh-checkpoint-rewind

# 2. restart and verify the row
dsh --profile web --dump-config | grep -A4 'id: checkpoint-rewind'

Checkpoints persist through the storageDomain service. The plugin mounts without it and never blocks profile startup — checkpoint/rewind commands then return a structured error naming the exact rows to add. Compose the storage stack once to enable checkpoints:

- insert:
    - id: checkpoint-rewind-storage
      name: '@deepseek-ai/dsh-storage'
    - id: checkpoint-rewind-storage-json
      name: '@deepseek-ai/dsh-storage-json'
      config:
        root: !!js dshHomePath('checkpoint-rewind/storage')
    - id: checkpoint-rewind-storage-domain
      name: '@deepseek-ai/dsh-storage-domain'
      config:
        backend: json

The package is pure ESM with no build step — index.mjs and lib/ are the shipped artifacts. Workspace mutations now create checkpoints automatically; run /rewind to list them:

rewind: 3 checkpoints (newest last):
#a1b2c3d4 · (git) · turn 2 step 1 · 2026-08-14 12:00:01 (3 min ago) · trigger: bash · 4 files · 1.2 MiB
#b2c3d4e5 · (git) · turn 2 step 3 · 2026-08-14 12:00:41 · trigger: str_replace_editor · 2 files · 310 KiB
#c3d4e5f6 · (copy) · turn 3 step 1 · 2026-08-14 12:01:10 · trigger: write · 1 file · 90 KiB
run "/rewind <id>" to restore files and fork the session from that checkpoint

Address a checkpoint by its unique id prefix, by step number, or by latest:

/rewind b2c3d4e5
/rewind step 2
/rewind latest
/rewind preview b2c3d4e5   # read-only: show which files would change, touch nothing
/rewind workspace b2c3d4e5 --files src/a.ts,src/b.ts   # selective restore: only these files (approval-gated)
/rewind clear              # confirmed deletion of this session's checkpoints (files untouched)
/rewind clear --all        # confirmed deletion across ALL sessions and workspaces (files untouched)

preview resolves through the same addressing and prints the impact without asking for confirmation or writing anything. --files restores only the named files through the same approval-gated transaction (unknown paths fail closed; it is incompatible with workspaceRestore: reset-hard and disabled when selectiveRestore: false).

Install & uninstall

• git channel (latest main): dsh plugin --profile web add "github:PerryLink/dsh-checkpoint-rewind#main" — pure ESM, no prepare or allowBuilds step.
• npm channel (published releases): dsh plugin --profile web add dsh-checkpoint-rewind.
• tarball channel: npm pack in this repo, then dsh plugin --profile web add ./dsh-checkpoint-rewind-<version>.tgz.
• storage stack (required for checkpoints, optional for mounting): @deepseek-ai/dsh-storage + @deepseek-ai/dsh-storage-json (config root) + @deepseek-ai/dsh-storage-domain (config backend: json) — see Quick start; the plugin still mounts without them and every command explains the fix.
• uninstall: dsh plugin --profile web remove dsh-checkpoint-rewind — snapshot files stay until you delete $DSH_HOME/dsh-checkpoint-rewind; git objects are garbage-collected.

Configuration

All tunables are Schemastery Config fields (changeable from cordis.yml). Nothing is hardcoded. Provider options (gitBin, snapshotDir, excludeGlobs, verifyByHash) are read from the live config at use time, so cordis.yml changes apply without a restart.

- insert:
    - id: checkpoint-rewind
      name: dsh-checkpoint-rewind
      config:
        provider: auto
        maxSnapshots: 50
        maxSnapshotBytes: 536870912
        pruneOnTurnEnd: true
        confirmVia: auto
        preRewindCheckpoint: warn

Tools & surfaces

Safety model

• Git history is untouchable. The git provider runs only whitelisted side-effect-free primitives — stash create, commit-tree, restore --worktree, ls-tree, diff-tree, ls-files, status, rev-parse, cat-file -e — enforced by a runtime assertion, and object refs are validated as hex ids before being passed to git (a tampered record cannot inject git options). No reset --hard by default, no clean, no index/history mutation, ever (see workspaceRestore below).
• Overwrite rollback, never deletion. Restore only overwrites captured files, and the git provider restores explicit paths (git restore … -- . would delete files git add-ed after the checkpoint). Files created after the checkpoint (untracked or staged) are reported and left in place.
• No writes through links, no path traversal. The copy provider validates checkpoint refs before joining them into snapshot-directory paths, and refuses to restore through a destination (or ancestor) that has become a symbolic link — so a restore can never follow a link out of the workspace.
• Restore requires approval. Overwriting user files always goes through the confirmation seam with ask semantics; a missing, throwing, or answering-no answerer fails closed. /rewind preview is the read-only way to inspect the impact first.
• Rewind is reversible. Before restoring, a guard checkpoint captures the current state; restoring the guard undoes the rewind. preRewindCheckpoint: require aborts the rewind when the guard cannot be captured.
• Fixed-order transaction. Guard first, workspace second, config third, session replay fourth; every phase is logged; a failed restore leaves files, checkpoints, and session untouched.
• workspaceRestore: 'reset-hard' is CC-equivalent and opt-in. It runs git reset --hard <snapshot commit> (branch head moves to the snapshot commit; pre-snapshot history stays recoverable via reflog; untracked files untouched). It is off by default.
• Model-visible ⟺ logged. Everything a user or model sees reconstructs from command/run + command/done (and, once the host knows them, checkpoint/* events) plus the durable checkpoints domain.

How it works

capture ── fs/write-intent · fs/edit-intent · tools/pre-execute (prepend, pass-through)
        ── step/start auto interval ── /checkpoint · checkpoint tool ── pre-rewind guard
             │
             ▼  ProviderRegistry.resolve(auto)  →  git: stash create / commit-tree
             │                                     copy: incremental dir + hardlinks
             ▼
        checkpoints storage domain (SQLite rows / JSON file)  +  checkpoint/* event (adaptive gate)

/rewind <target> ── confirm (userQuestions / approval, fail-closed) ──▶ guard checkpoint
             ├─ workspace: provider.restore(ref)  (restore | reset-hard)
             ├─ config:   settings write-back (persisted on the profile patch)
             └─ session:  SessionStore.fork(source, boundary) → new child session (original untouched)

Full decision record, event vocabulary, and the provider seam contract: ARCHITECTURE.md.

Session events (rc.2 note)

The plugin declares checkpoint/snapshot, checkpoint/bound, checkpoint/prune, and checkpoint/rewind as log-only SessionEventMap members. Harness rc.2 has no plugin event-registration surface and Session.append does not stamp the ignorable envelope (its third argument is surface intent, not options), so appending unknown types would make the session unreadable on reload. The plugin therefore appends through an adaptive gate: a runtime probe (on a detached, never-persisted session store) detects whether the host's append stamps the ignorable envelope — on rc.2 the gate stays closed; on hosts that support it, checkpoint/* events are appended with ignorable: true automatically. Until then the authoritative audit chain is command/run + command/done (harness-known) plus the durable checkpoints storage domain.

Web UI anchor

The plugin returns the new session id in the command result (session: <id>) and the Web shell can navigate there. The session-projection unit checkpoints is shipped: whenever ctx.sessionProjections exists, the plugin registers the unit via ctx.inject (folds checkpoint/snapshot|bound|prune|rewind into a whole-value list) — it stays an empty list on rc.2 hosts until a harness build ships the checkpoint/* vocabulary or the ignorable envelope, then fills in with zero plugin changes.

FAQ

Does this replace git? No — it uses git where available. In a git repo you get byte-perfect, deduplicated snapshot objects without touching history; in any other directory the copy provider does the same with plain files. Regular commits remain your long-term history.

Why not git reset --hard by default? Because destroying state is not the job of a safety net. The plugin only creates unreferenced objects and performs worktree-only, path-explicit restores by default, so a bad rewind can never lose history, the index, or files created after the checkpoint. reset-hard is available behind workspaceRestore: 'reset-hard' for users who explicitly want CC parity.

How long is a checkpoint restorable? No ref is held on git-provider snapshots. They're unreferenced stash create/commit-tree objects by design, and discard deletes only the record, leaving the object itself for git gc. Two independent things bound how long a checkpoint actually stays restorable: the plugin's own quota pruning (maxSnapshots, default 50 kept per session; maxSnapshotBytes, a default 512 MiB global soft quota enforced via pruneAll() on capture and swept on turn ends), and the host repo's git gc. With git's defaults, unreferenced objects survive until an auto-gc runs, and even then only objects older than gc.pruneExpire (2 weeks by default) get pruned. A manual git gc --prune=now, an aggressive repack, filter-repo, or a fresh clone can drop them immediately. Pruned objects are now detected up front: the git provider probes the snapshot object before restore (git cat-file -e) and fails loudly when it is missing instead of silently restoring nothing, and a doctor pass walks every checkpoint record to probe object existence and badge unrestorable ones ([UNRESTORABLE: object missing] in the list). The doctor pass is implemented and tested but not yet wired into the restore path (restore does not yet consult the unrestorable marking; tracked in discussion #13).

Can I rewind to a step in the middle of a turn? File restoration is step-precise (/rewind step <N> = nearest snapshot ≤ N). The session replay, however, respects the harness's replay granularity: the child session is seeded up to the checkpoint's turn boundary.

What happens if nobody can answer the confirmation? Nothing is touched — the plugin fails closed (unavailable/rejected), keeps the checkpoint, and returns an explanatory error. With confirmVia: approval on rc.2 the message says to mount userQuestions, because approval requires an open turn and commands run between turns.

Can I undo a rewind? Yes — every approved rewind captures a guard checkpoint of the pre-rewind state first; the result prints rewind guard: <id>, and /rewind <guard-id> restores that state.

How do I address checkpoints? Unique id prefix (the 8-char short id in the list works), /rewind step <N>, /rewind latest, /rewind clear deletes this session's checkpoints, and /rewind clear --all deletes checkpoints across all sessions and workspaces (files untouched). /rewind preview <target> uses the same addressing to show the impact without changing anything.

What does preview do — and not do? It resolves the checkpoint, then runs a read-only comparison: which files would be overwritten (or recreated), which already match, and which files created after the checkpoint would be left in place. It never prompts, never writes, never forks, and records no checkpoint/rewind event — the approval gate only runs on a real /rewind <id>.

Demo

A real assembled-headless integration run (npm run test:integration) drives the full flow: the agent modifies files across two turns, then /rewind preview inspects the impact read-only (no confirmation gate, no writes) and /rewind <id> restores the files and replays the session into a new child session. The run asserts the file contents, the replayed child context, the guard checkpoint, and that files created after the checkpoint survive — for both the copy and git provider flows (the git flow also asserts HEAD and the reflog are untouched). The driver lives in test/integration/rewind-headless.mjs.

Permissions & data

• Permissions: the workshop manifest declares workspace:read, workspace:write, git:read, git:write, snapshot-storage:write, session-log:read, settings:write, and network:none.
• Data: checkpoint records live in the checkpoints storage domain (SQLite rows or a JSON file); copy snapshots live under snapshotDir. Fully local — no network, no credentials. The domain is opened dual-version: 0.4.x-era media (domain v1) are opened in a compatibility mode that keeps old records readable and stores new captures in the v2 shape, so upgrading the plugin never orphans an existing medium.
• Session log: checkpoint/* events are appended through the adaptive gate; the authoritative audit chain is command/run + command/done plus the durable domain.

Security boundaries

• Git history is untouchable. Whitelisted side-effect-free primitives; reset --hard only behind the opt-in workspaceRestore: 'reset-hard' mode. No git clean, ever.
• Overwrite rollback, never deletion. Restore overwrites captured files only; files created after the checkpoint are reported and left in place.
• No writes through links, no path traversal. Copy refs are validated as snapshot ids; restore refuses to follow symbolic links out of the workspace.
• Restore requires approval. A missing or denying answerer fails closed.
• Rewind is reversible. A guard checkpoint of the pre-rewind state is captured first.

Known limitations

• On rc.2, checkpoint/* session events are suppressed by the adaptive gate; the audit chain rides command/run + command/done plus the storage domain until a host ships the vocabulary or the ignorable envelope.
• confirmVia: approval needs an open turn, and commands run between turns — mount userQuestions (or set confirmVia: userQuestions) on rc.2.
• Session rollback forks a new child session at the checkpoint boundary; it never rewrites or truncates the original session.
• workspaceRestore: 'reset-hard' moves the branch head to the snapshot commit; it is off by default.
• A checkpoint captured before any closed turn has no replay boundary — session rollback then creates a fresh child session with empty context.

Troubleshooting

GitHub で見る