Snapshot & Replay

Freeze an agent run, restore it later, and detect when the underlying world has drifted.

What It Does

A snapshot captures a workspace as a single tar file: mount configs, sessions, history, finished jobs, cache bytes, and one fingerprint per recorded remote read. Loading a snapshot reconstructs the workspace and verifies that the underlying sources have not drifted since capture.

When the backend exposes a stable per-object revision marker (S3 VersionId, Drive revisionId, Git commit SHA), the snapshot also records that revision. At load time, reads pin to the recorded revision and serve the exact bytes the original agent saw, even if the live object has since been overwritten.

The API

Python

# capture — async, stats every touched path on a SUPPORTS_SNAPSHOT mount
await ws.snapshot("run.tar")
await ws.snapshot("run.tar.gz", compress="gz")

# replay — sync construction; drift check fires on first dispatch/execute
restored = Workspace.load("run.tar")                          # STRICT default
restored = Workspace.load("run.tar", drift_policy=DriftPolicy.OFF)

# in-process duplicate (shares remote resources, restores local content fresh)
cp = await ws.copy()

Both snapshot and copy are async because fingerprint capture stats each touched path on a SUPPORTS_SNAPSHOT mount.

What Is And Isn’t Captured

Captured

• Mount configs (creds redacted; restore via resources= override)
• Sessions, history, finished jobs
• Cache bytes for touched paths
• One fingerprint per remote read (ETag-equivalent)
• Optional per-path revision when the backend exposes one

Not captured

• Live state of mounts with SUPPORTS_SNAPSHOT=False (Gmail, Slack, Linear, Notion, …)
• Files the agent never touched
• Raw bytes of remote objects (recoverable only via revision pin)

Drift Detection

On the first dispatch or execute after load, Mirage stats every fingerprinted path against the live source in parallel. If any path’s live fingerprint differs from the recorded one, the workspace raises ContentDriftError:

try:
    await ws.execute("cat /s3/data.csv")
except ContentDriftError as exc:
    print(exc.path, exc.snapshot_fingerprint, exc.live_fingerprint)

Paths that carry a revision pin are skipped — the pinned read serves the exact original bytes, so a fingerprint mismatch is expected and harmless.

Drift Policies

Policy	Behavior on mismatch	Use when
STRICT (default)	Raise ContentDriftError on first mismatch.	Reproducing an agent run; you want to know the world moved.
OFF	Skip drift checks entirely. Evict snapshot cache for fingerprinted paths so reads serve current.	You only wanted the workspace skeleton, not the bytes.

Pass via Workspace.load(..., drift_policy=DriftPolicy.OFF).

How It Composes With Caching

Snapshots interact with two existing caches:

Cache	Holds	Snapshotted?
File cache (Workspace._cache)	raw bytes per virtual path	✅ bytes for touched paths are serialized into the tar
Index cache (Resource._index)	FileStat entries for listings	❌ rebuilt lazily after load

Only files the agent actually read are fingerprinted. Capture walks ws._ops.records for op == "read" and dedups by path. Files that were listed but never opened, or touched only by stat / readdir, do not carry a fingerprint or a revision.

At load time, three pieces of restored state cooperate per read:

1. Cache is consulted first. Snapshot bytes go back into Workspace._cache; a warm read returns them with no network round-trip.
2. Fingerprint verifies the cache. Under STRICT, the eager drift check stats every recorded path against live and raises before any read fires if anything moved. The cache is therefore trusted as authoritative until proven stale.
3. Revision pin is the cold-path recovery. When the cache misses and the backend supports pinning, reads fetch the exact recorded revision (S3 GetObject(VersionId=...)) so you still get original bytes, not the live head.

The same three states under each policy:

Scenario	Cache	Pin	Drift check	Net effect
STRICT, fingerprint matches, cache warm	served	n/a	passes	Original bytes from cache (~0 ms)
STRICT, fingerprint matches, cache cold	miss	n/a	passes	Live GET, current bytes (= original)
STRICT, no pin, fingerprint differs	(n/a)	(n/a)	raises ContentDriftError	Caller informed before reading
STRICT + pin (versioned backend)	served (matches recorded)	installed	skipped for pinned paths	Original bytes from cache or pinned GET
OFF, any state	evicted for fingerprinted paths	not installed	skipped	Live GET, current bytes

The cache is the optimization, the fingerprint is the verifier, and the pin is the recovery — three independent guarantees that “what you replay equals what you captured.”

Resource Support Matrix

Snapshot support is opt-in per resource via SUPPORTS_SNAPSHOT = True. Resources without it surface in a load-time warning and serve current state with no drift detection.

Legend: ✅ = supported · 🟡 = adapter needed (no work in flight) · 📝 = planned · ❌ = not applicable.

Object Storage

Resource	Drift detection (Py)	Revision pin (Py)	Drift detection (TS)	Revision pin (TS)	Marker	Notes
S3	✅	✅	📝	📝	ETag + VersionId	Pin requires bucket versioning.
R2	✅	✅	📝	📝	ETag + VersionId	Inherits S3 path; pin requires R2 versioning (GA 2024).
GCS	✅	🟡	📝	🟡	ETag + x-goog-generation	TODO(snapshot-gcs): map generation → ContentVersion(kind="revision") in core/s3/stat.py.
OCI	✅	🟡	📝	🟡	ETag + versionId header	TODO(snapshot-oci): same shape as GCS adapter.
Supabase	✅	❌	📝	❌	ETag only	Inherits S3Resource. Supabase’s S3-compat endpoint does not surface object VersionId, so drift detection works but pinning is not available.

Files & Code

Resource	Drift detection (Py)	Revision pin (Py)	Drift detection (TS)	Revision pin (TS)	Marker	Notes
Disk	✅	❌	📝	❌	content hash	Bytes travel inside the tar; pin not meaningful.
RAM	✅	❌	📝	❌	content hash	Same as Disk.
Redis	✅	❌	📝	❌	content hash	Bytes restored via resources= override.
GitHub	📝	📝	📝	📝	commit SHA	TODO(snapshot-github): stat → revision=sha, read via repos.get_contents(path, ref=sha).
Google Drive	📝	📝	📝	📝	md5Checksum + revisionId	TODO(snapshot-gdrive): stat → revision=revisionId, read via revisions().get_media.

Extending To A New Backend

Three steps in Python:

from mirage.observe.context import record, revision_for
from mirage.resource.base import BaseResource
from mirage.types import FileStat


class MyResource(BaseResource):
    SUPPORTS_SNAPSHOT = True  # 1. opt in

    async def stat(self, ...) -> FileStat:
        return FileStat(
            ...,
            fingerprint=my_etag_equivalent,        # 2. for drift
            revision=my_revision_marker_or_None,   # 3. optional, for pin
        )

And in your read function, look up the active pin and pass both the fingerprint and the revision through to record so the snapshot captures whatever the backend served:

async def read_bytes(accessor, virtual_path, ...):
    pinned = revision_for(virtual_path)  # 4. None if no pin installed
    response = backend_get(virtual_path, revision=pinned)
    record(
        "read", virtual_path, "my-backend", len(response.body), start_ms,
        fingerprint=response.etag,        # 5. for drift detection
        revision=response.revision,       # 6. for pinning on replay
    )
    return response.body

At load time the workspace writes each manifest entry straight into mount.revisions — no per-resource hook required. If your backend has no stable revision, skip steps 3 and 6; drift detection still works on the fingerprint alone.

Caveats

• Revision longevity. Pinned reads only work as long as the source still retains the recorded revision. S3 bucket lifecycle rules can age out old versions; Drive keeps revisions for 30 days on non-Workspace files. Treat pins as best-effort.
• First read after load is slower than the rest. Workspace.load() returns immediately, but the first execute() or dispatch() afterwards pauses while Mirage verifies that nothing has drifted upstream. Concretely, it asks the live source “is this path still the bytes I remember?” once per recorded read, in parallel. Tiny for short sessions (tens of milliseconds); a few hundred milliseconds to a couple of seconds for sessions with hundreds of recorded reads. The pause happens once per loaded workspace; subsequent calls are normal speed. Pass drift_policy=DriftPolicy.OFF if you want to skip the check entirely.