返回目录
其他 待识别

dsh-session-pruner

mrzhangkris/dsh-session-pruner

DSH 会话生命周期管理插件:one-shot 子代理自动清理 + 容量保底 + 连带清理 projcache,从源头杜绝缓存膨胀卡顿

Stars
3
Forks
0
Issues
1
更新
19 天前

PROJECT TOPICS

项目标签

PROJECT README

README

dsh-session-pruner

DSH session lifecycle management plugin — full-type session lifecycle management: one-shot subagents archived on completion, continuable subagents and main sessions archived when idle, a capacity cap, and projection-cache cleanup. Prevents session-library accumulation stalls at the source.

Every session type has a defined destination: finished one-shot subagents are archived automatically, idle continuable subagents / main sessions are archived, and overflow is recycled by priority. Archive first (recoverable), delete after expiry — the GUI syncs within 30s, fully panel-configured with hot reload.

简体中文 · Apache-2.0 · npm · npm version · Changelog

Why

DSH (DeepSeek Harness) caches a full projection of every session in session_projcache.json (token stats, context pressure, ...), and the storage backend rewrites the whole file atomically on every write. When the session library accumulates thousands of subagent sessions:

  • The cache balloons past 100MB and each checkpoint fully re-serializes → main process CPU 250%+
  • The single-threaded event loop is saturated → every session load stalls, even GET / times out

Managing session lifecycle (this plugin) is the root fix: no session accumulation → no cache rows → no stalls.

Features: full-type lifecycle

Session type Trigger Action Default
one-shot subagent subagent/end / agent/disposed event (+ grace) archive within seconds (event-driven) event + 3min grace
continuable subagent idle over N days archive (recoverable) off (0 days)
main session idle over N days archive (recoverable) off (0 days)
any type total exceeds capacity cap recycle by one-shot → continuable → main + oldest 400
archive directory kept over N hours physically deleted 24 hours

Behavior note (v0.2.3+): one-shot subagents are archived uniformly by the oneShotMinAgeMinutes idle threshold (default 3 minutes), with or without end-seed. The earlier fallback — "one-shot sessions without end-seed must stay idle a full hour before archiving" — has been removed.

Archive mechanism (recoverable)

Cleaned sessions are moved to ~/.dsh/sessions-archive/ first (workspace/session-id structure preserved) — they disappear from the GUI immediately (the list only reads the sessions directory), but the files remain and can be restored manually:

# Restore: mv back into the sessions directory
mv ~/.dsh/sessions-archive/<workspace>/<session-id> ~/.dsh/sessions/<workspace>/

# ⚠️ After restoring, pin it (or open it) immediately — until the session is
# opened it is not live-protected, and an idle hit (e.g. one-shot over the
# threshold, main over mainIdleDays) within one scan cycle may archive it again.
# Add its session ID to the "Pin whitelist" field in the settings card.

A "delete directly" mode (no archive, irreversible) is also available.

Safety (double protection)

  • Running sessions are never touched: live sessions (still held in the in-memory session store, open/loading) are skipped — and the live check is fail-closed: a store query error treats the session as live, never deleting on uncertainty
  • Idle = last log write: idle is judged by the session log file mtime (last write time), not the directory mtime — DSH appends to session.jsonl.zstd, so active sessions keep refreshing their mtime and are never misjudged idle
  • one-shot: finished one-shot subagents are archived uniformly by the oneShotMinAgeMinutes idle threshold (with or without end-seed, same threshold); the capacity cap additionally skips sessions lacking session/end-seed
  • Main sessions do not participate in capacity recycling by default (configurable)
  • Per-action failure isolation: every action is try/catch wrapped

How it works

Dual-track triggers (events = hot path, disk = source of truth)
  ┌─ Event-driven (seconds): subagent/end + agent/disposed
  │     ├─ 500ms batch window merges storms → oneShotMinAge grace re-check
  │     └─ single-session check (memory-first, at most one zstd decompress) → archive
  └─ Scheduled reconcile (fallback, default 60min)
        ├─ pruneArchive: physically delete expired archive sessions
        ├─ iterate ~/.dsh/sessions/*/ decompress log (system zstd, multi-frame)
        │     ├─ origin: main | subagent       (session header)
        │     ├─ mode: one-shot | continuable  (subagent/descriptor event)
        │     └─ ended: contains session/end-seed
        ├─ one-shot idle over threshold ──→ archive (archiveMode)
        ├─ continuable/main idle N days ──→ archive
        ├─ total > cap ──→ recycle by priority + oldest (skip running/live)
        └─ each archive also: purge projcache row + workspace accounting

GUI sync — change-driven primary, full-refresh fallback:

  • dirty-flag (primary): host keeps an in-memory monotonic archive log; the client polls /plugins/dsh-session-pruner/archived every 3s and only issues refreshList() + refreshSubagents() when a change is reported — sidebar and task panel stay consistent within seconds, zero RPC when nothing changed.
  • Full fallback: every uiRefreshSeconds seconds the client refreshes both data sources anyway (main list via refreshList(), each known parent's subagent catalog via refreshSubagents()), covering dirty-flag failures (older host / route unavailable). No page reload needed.

Install

From npm (recommended)

dsh plugin --profile web add dsh-session-pruner

From source (development)

dsh plugin --profile web add /path/to/dsh-session-pruner

After installing (or upgrading), restart the dsh web daemon to load the new version (launchctl kickstart -k gui/$(id -u)/com.deepseek.dsh-web) — config changes alone hot-reload without restart.

Configuration (settings panel, hot reload)

settings panel

After install, open Settings → Plugins → 会话生命周期管理 card. All 10 options save with hot reload (no restart). Six everyday options are visible by default; the four low-frequency fallbacks are tucked into an "Advanced" collapsible section (its title shows an unsaved-changes badge when applicable):

Field Default Description
Scan interval (min) 60 reconcile fallback (events are primary)
Capacity cap (sessions) 400 recycle by priority + oldest when exceeded
UI fallback refresh interval (s) 30 dirty-flag primary (3s change poll); full refresh fallback
Archive retention (hours) 24 physical delete after retention
Archive mode archive archive (recoverable) / delete directly (irreversible)
Continuable idle archive (days) 0 archive after N idle days, 0 = off
Main idle archive (days) 0 archive after N idle days, 0 = off
Clean main on overflow off main participates in capacity recycling
One-shot min survival (min) 3 newly finished subagents are not cleaned within N minutes (protects finishing/references)
Pin whitelist (session IDs, one per line) empty pinned sessions are never auto-cleaned (pin restored sessions immediately)

The card also shows a live status line (30s poll): archive count + earliest expiry, session total (+ overflow), last cleanup (count + time), pinned count.

Env vars (fallback, panel wins): DSH_SESSION_PRUNER_INTERVAL_MS / _MAX / _CLEAN_MAIN / _ARCHIVE_HOURS / _ARCHIVE_MODE / _CONTINUABLE_IDLE_DAYS / _MAIN_IDLE_DAYS / _ONE_SHOT_MIN_AGE_MINUTES / _PINNED_IDS (comma-separated).

Logs

Output in guard server-*.out.log:

[dsh-session-pruner] armed: interval=60min cap=400 cleanMain=false
[dsh-session-pruner] hot-reloaded: interval=60min cap=400 ... contIdle=0d mainIdle=0d pinned=0
[dsh-session-pruner] archived a1b2c3d4 (subagent/one-shot) one-shot idle cache=true
[dsh-session-pruner] archive pruned: 2 expired

cache=true/false tells whether the projection cache row was purged along with the session.

Tests

npm test               # regression suite: audit PoC checks + full e2e (isolated tmp DSH_HOME)
node test/dry-run.js   # read-only full-library scan, verify classification (no deletion)
node test/e2e.js       # create a fake one-shot session, verify the real cleanup path
node test/poc-audit.js # audit regression: ended misjudgment / dual-source drift / archive orphans / pin

Implementation notes

  • Multi-frame zstd: DSH session logs are concatenated zstd frames (append writes); Node zlib decodes a single frame only, so the plugin shells out to the system zstd CLI (brew install zstd on macOS)
  • Cache row purge: storageDomain.get('session_projcache').table('sessions').delete(id) — the official write chain (atomic persistence + in-memory sync)
  • Workspace accounting: the session id is removed from the workspace domain on archive, keeping the data source consistent with disk
  • Zero bundled deps: runtime modules (@deepseek-ai/dsh-settings, schemastery) are provided by the DSH host; the plugin ships no dependencies of its own
  • Panel + hot reload: installSettingsSection + hand-written client card (__ModuleLoader__ bundle), onChange re-schedules the timer instantly

Developer guide

  • docs/DEVELOPMENT-GUIDE.md — DSH plugin development practice guide (architecture, Host/Client, settings panel, deployment ops, pitfalls with fixes), the foundation for future plugin work
  • docs/DESIGN.md — design decisions and rationale (three-tier strategy, dual-track triggers, fail-closed safety, invariants)
  • docs/TESTING.md — test matrix, verification pyramid (V0/V2/V3), release checklist
  • docs/PROJECT-STATUS.md — current status snapshot and backlog, for new contributors/sessions

Known limits

  • If completion events are lost (host restart mid-run), a finished one-shot subagent is only picked up by the next reconcile scan — worst case one intervalMinutes (default 60min)
  • Restored (mv-back) sessions are not live-protected until opened — pin them to survive the window (see Archive mechanism)
  • Requires the system zstd CLI
  • The root fix lives upstream: projcache stale-session eviction / incremental storage writes, see deepseek-harness Discussion #1550

License

Apache-2.0

CLASSIFICATION EVIDENCE

分类依据

项目类型待识别
功能分类其他
规则置信度

系统优先读取 GitHub Topics,再与站内分类词典和词根规则比对。当前命中: 无有效分类标签。