DeepSeek Harness Plugin Hub

Publish and manage complete Harness Profiles. Discover Plugins for your next setup.

Explore

PluginsPresetsDocsNews

Community

Publish a pluginContactReport an issue

Resources

Plugin Hub on GitHubDeepSeek HarnessSystem statusPrivacy notice
© 2026 DeepSeek Harness Plugin HubPowered byPaxTech

Independent and unofficial. Not affiliated with, authorized by, or endorsed by DeepSeek.

Session S3 — DSH Plugin for DeepSeek Harness
DeepSeek Harness Plugin Hub
ProfilesPluginsCategoriesNewsDocsSign inManage Profiles
ProfilesPluginsCategoriesNewsDocsSign in
← Plugins
S

dsh-session-s3

Session S3

Community DSH plugin: S3-backed SessionPersistence (wal3-Lite). Immutable fragments + CAS manifest.

The plugin will be installed here. Keep web if you are unsure.

npx -y @deepseek-ai/dsh plugin --profile web add github:gengmao/dsh-session-s3#466331cb4519496536ce0889749729986b933a6e
READMECompatibilityVersions

Compatibility and provenance

Session S3 is published as dsh-session-s3 and currently resolves to version 0.1.0. The Hub verifies its manifest and preserves the exact installation source for reproducible installs.

DSH compatibility
*
Runtime surfaces
any
Release source
github
Registry updated
8/20/2026

Versions

0.1.0stable
8/20/2026

Related plugins

Loading related plugins…

Latest
0.1.0
DSH
*
HMR
Process restart
Tree shaking
Safe tree shaking not declared
Unpacked size
Unavailable
Files
Unavailable
Surface
any
License
MIT
Source
github
GitHub
★ 1
Weekly downloads
0
Last push
9/3/2026
View source ↗
README badge

Click the badge to copy Markdown for your README.

Do you maintain this Plugin?Claim benefit · Priority security scan

Verify the GitHub repository declared in package.json to manage this listing. After you claim it, Hub will prioritize a security scan of the current version and publish the result when it passes.

Claim this Plugin →
Report an issue

Related plugins

More verified plugins in memory-context.

Contextdsh-contextA DeepSeek Harness plugin for context insight and management, with context dashboard and context command, for understanding how the context is made of, and how it evolves.Weknora@wxg-prc-cpg/dsh-weknoraWeKnora knowledge retrieval tools for DeepSeek Harness (dsh): semantic search, document reading and RAG/agent answers over your own knowledge bases.Memsearch Dsh@zilliz/memsearch-dshMemSearch plugin for DeepSeek Harness: shared markdown memory across agents, with capture, pre-step context injection, memory-recall skill, and a skill-candidate review panel.Memory@furongjun1999/dsh-memoryLingshu (Lingshu·líng shū) DeepSeek Harness plugin: a complete brain—long-term memory/knowledge flywheel/self-awareness/recursive reflection integrated with DSH, with conversations automatically distilled into the md_cg cognitive graph (md documents)

README

dsh-session-s3

Unofficial community DeepSeek Harness plugin: an S3-backed SessionPersistence provider.

wal3-Lite: immutable JSONL fragments + a CAS manifest (If-None-Match / If-Match). Per-fragment SHA-256. No setsum in Phase 1.

Fixes the class of JSONL durability bugs that come from torn writes, missing fsync, and concurrent writers — by never mutating a fragment and coordinating writers on a single compare-and-swap object.

The 7 field-confirmed DSH corruption discussions (#1333, #1452, #1497, #1473, #1586, #2167, #2342) are encoded as tests in test/corruption-scenarios.test.ts. Results: docs/corruption-scenarios.md.

For the correctness boundary, commit point, crash states, sequence domains, and rejected alternatives, see Design rationale.

Install

Plugins install into a profile ($DSH_HOME/profiles/<name>), not globally. dsh plugin is a pnpm forwarder: it adds the package as a dependency, and because this repo declares dsh.bundle.patch, it also appends dsh-session-s3 to dsh.profile.bundles.

From GitHub (typical):

dsh plugin --profile web add github:gengmao/dsh-session-s3

or without a global dsh:

npx -y @deepseek-ai/dsh plugin --profile web add github:gengmao/dsh-session-s3

pnpm ≥10 blocks git-hosted prepare scripts (this package runs tsc on install). If add fails, allow the build and re-run:

# $DSH_HOME/profiles/web/pnpm-workspace.yaml
allowBuilds:
  dsh-session-s3: true

From a local checkout:

dsh plugin --profile web add ./dsh-session-s3
# or
dsh plugin --profile web add "link:$PWD"

From npm (once published):

dsh plugin --profile web add dsh-session-s3

Same action in the Web UI: Settings → Plugins. Restart dsh web (or dsh --profile web) after adding.

Remove with dsh plugin --profile web remove dsh-session-s3.

Requires Node >= 22 (PersistenceCoordinator uses Promise.withResolvers; a polyfill is loaded at import so 18–21 do not crash, but DSH itself wants 22).

Peer packages (@deepseek-ai/dsh-session, dsh-session-persistence) come from the DSH profile. A standalone npm install of this repo uses .npmrc legacy-peer-deps=true because @deepseek-ai/dsh-type-meta is unpublished on npmjs.

Config

KeyDefaultNotes
bucket(required)S3 bucket
prefixdsh/Keys live at {prefix}sessions/{sessionId}/
regionunset (SDK / AWS_REGION); auto if endpoint is setSet explicitly for AWS; auto for R2
endpoint—R2 / Tigris / MinIO / SeaweedFS / GCS interop (must be http(s))
forcePathStyletrue when endpoint is setPath-style URLs
accessKeyId / secretAccessKeyunsetOnly for static keys. Prefer the SDK default chain (see below)
flushThresholdEvents50Library createProvider() only: flush its in-memory buffer after N events
flushThresholdBytes262144Library createProvider() only: flush after 256 KiB
preparedSessionCacheSizecoordinator default (5)DSH seam: LRU of unpublished preparations
writeBatchMaxDelayMscoordinator default (200)DSH seam: coordinator write-behind delay

Invalid config fails loud at load, listing every problem (not just the first).

The bundle patch replaces the profile sessions row and the default export is a Cordis Service registered as ctx.sessionPersistence, so Harness resume/list/the agent loop talk to S3 instead of JSONL.

After install, set the bucket (and optional endpoint) in the profile overlay $DSH_HOME/profiles/web/cordis.patch.yml:

- id: sessions
  name: dsh-session-s3
  config:
    bucket: my-sessions
    region: us-west-2
    prefix: dsh/
    # endpoint: https://<account>.r2.cloudflarestorage.com   # R2 / MinIO / Tigris

Do not put access keys in the patch. See AWS credentials.

Library helper (no DSH, not the seam):

import { createProvider } from "dsh-session-s3";

const persistence = createProvider({
  bucket: "my-sessions",
  region: "us-west-2",
});

await persistence.append("sess-1", { type: "user/message", text: "hi" });
const events = await persistence.read("sess-1");
await persistence.compact("sess-1", 10); // keep the last 10 fragments
await persistence.close("sess-1");

AWS credentials

Resolution order (parseConfig → createS3Client):

  1. Plugin config accessKeyId and secretAccessKey (both required if either is set) — static keys, no session token
  2. Else omit keys and use the AWS SDK default chain (AWS_ACCESS_KEY_ID + AWS_SECRET_ACCESS_KEY + AWS_SESSION_TOKEN, ~/.aws/credentials, AWS_PROFILE, SSO, IAM instance/task role)

Prefer (2) on real AWS. parseConfig does not copy env keys into static credentials — that would drop AWS_SESSION_TOKEN and break SSO / assumed roles / GitHub OIDC. Leave accessKeyId / secretAccessKey unset in the Cordis patch.

Local / laptop

aws configure
# or
aws sso login --profile myprofile
export AWS_PROFILE=myprofile
export AWS_REGION=us-west-2
# then set bucket in the profile overlay, not via a DSH_S3_* env

IAM user / access keys (CI, MinIO, R2)

export AWS_ACCESS_KEY_ID=AKIA...
export AWS_SECRET_ACCESS_KEY=...
# bucket still comes from the profile overlay `config.bucket`

Only set accessKeyId / secretAccessKey in plugin config when you cannot use env or the chain. Do not commit them.

EC2 / ECS / Lambda

Attach an instance or task role. Do not set keys. The SDK picks up the role.

IAM policy (minimum)

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "s3:GetObject",
        "s3:PutObject",
        "s3:DeleteObject"
      ],
      "Resource": "arn:aws:s3:::my-sessions/dsh/*"
    },
    {
      "Effect": "Allow",
      "Action": ["s3:ListBucket"],
      "Resource": "arn:aws:s3:::my-sessions",
      "Condition": {
        "StringLike": { "s3:prefix": ["dsh/*"] }
      }
    }
  ]
}

If-Match / If-None-Match are request headers, not extra IAM actions. list() / listSnapshots() call ListObjectsV2 and need s3:ListBucket.

R2 / MinIO / Tigris

Same keys, plus endpoint. Path-style is on automatically when endpoint is set:

export AWS_ACCESS_KEY_ID=...
export AWS_SECRET_ACCESS_KEY=...
# set endpoint + bucket on the plugin config in the profile overlay

Check CAS

aws s3api put-object --bucket my-sessions --key dsh/_probe --body /dev/null \
  --if-none-match '*'

A retry should 412. If it always 200s, the store is ignoring preconditions and this plugin is unsafe there.

Object layout

s3://{bucket}/{prefix}sessions/{sessionId}/
├── manifest.json                 # CAS coordination point
└── fragments/00000001.jsonl      # immutable, SHA-256 in the manifest

Write path (flush):

  1. GET manifest.json. If the tail already has this batch's SHA-256, treat it as a lost CAS response.
  2. Buffer is serialized as a JSONL fragment (seq = last + 1). PutObject with If-None-Match: *. On 412, LIST occupied keys, take seq = max(manifest+1, maxOccupied+1), retry.
  3. Conditional PUT manifest.json with the ETag from step 1 (If-Match, quoted). Reload only after a 412. Idempotent on seq and tail sha256.
  4. Drop only the snapshotted prefix of the buffer after CAS succeeds (events appended during the flush stay).

Uncontended single-writer flush is three S3 requests. A crash between (2) and (3) leaves an orphan fragment. Harmless: the manifest is source of truth.

The successful conditional PUT of manifest.json is the commit point. Fragment sequence numbers order storage objects; they are not DSH event seq values. CAS preserves every writer's bytes but does not allocate cross-process event sequences. See Design rationale for the precise guarantees and remaining races.

S3 compatibility

BackendIf-Match / If-None-MatchPhase 1
AWS S3yes (quoted ETags)intended; live IT env-gated
Cloudflare R2yesintended
Tigrisyesintended
MinIOyesintended; S3_IT=1 against local MinIO
GCS (S3 interop)weak / eventual on some pathsuse with care
SeaweedFSversion-dependentuse with care

CAS is the correctness mechanism. Do not point this at a store that silently ignores conditional puts.

Env-gated integration test (and the CI MinIO integration job):

S3_IT=1 S3_BUCKET=test S3_ENDPOINT=http://127.0.0.1:9000 \
  S3_REGION=us-east-1 \
  AWS_ACCESS_KEY_ID=minioadmin AWS_SECRET_ACCESS_KEY=minioadmin \
  npm test

Caveats

  1. Same composition as JSONL. S3SessionPersistence extends @deepseek-ai/dsh-session-persistence's SessionPersistence and implements PersistenceBackend, then constructs PersistenceCoordinator(ctx, this). Cold load therefore emits synthetic interrupted-turn closers; instanceof SessionPersistence is true. Peer deps (cordis, dsh-session, dsh-session-persistence) are provided by the DSH profile at install time.
  2. One live writer per session. DSH does not support two processes concurrently writing the same SessionId. Manifest CAS is a defensive check: appendBatch revalidates SessionEvent.seq inside the CAS mutate and throws StaleWriterError instead of committing both batches. The loser's fragment PUT is an unreachable orphan. No leases, heartbeats, or fencing.
  3. No setsum (deliberate, Phase 2). Integrity is per-fragment SHA-256 only.
  4. Response ambiguity is bounded, not exactly-once. A fragment PUT can leave an orphan. A lost manifest response is recognized by matching the tail SHA-256 (library helper and DSH backend). Library fragments carry a nonce header so identical consecutive batches do not collapse.
  5. Library createProvider() is a 5-method helper (load/append/read/compact/close) for non-DSH callers. It does not interpret SessionEvent.seq. Concurrent library flushes reallocate a stale fragment ordinal above the committed tail instead of appending out of order. DSH uses the default class export. read() includes the in-memory buffer (not yet on S3). compact / close flush first. trim refuses a session that already has a DSH header.
  6. Trim vs concurrent readers. trim CAS-updates the manifest, then deletes dropped objects. A reader holding an old manifest that GETs a deleted fragment sees FragmentCorruptError. Phase 1 assumes one writer and no trim-during-read.
  7. Manifest rewrite is O(n) bytes per flush. Each commit rewrites the whole fragment list, so bytes transferred grow O(n²) over an untrimmed session. Fine for Phase 1; compact long-lived logs.
  8. Trim guard needs a parseable DSH header. trim refuses a session whose manifest survived schema parse. A DSH manifest whose header is malformed parses as , so out-of-band compact would not refuse it and could still brick . That intersection is not coded around.

Roadmap

  • Phase 2 — setsum / global log verification, verified GC
  • Optional — binary fragments, DynamoDB Streams notifications

Develop

npm install
npm test
npm run build

CI (.github/workflows/ci.yml) typechecks, runs the unit suite, then repeats the suite against MinIO with S3_IT=1.

SPEC.md is the Phase 1 contract.

header
pre-watermark
header: null
appendBatch