Skip to content

Releases

dsh-plugin-v1.8.0

Choose a tag to compare

@github-actions github-actions released this 21 Nov 14:21

Added — external memory supervisor

  • lib/supervisor.js: pure decision logic — thresholds derived from the host
    heap cap (warn 73% / kill 93%), one-shot request-exit, graceMs before
    kill, and seven-way post-mortem attribution.
  • scripts/aiprison-supervise.mjs: standalone process that samples the host
    RSS and CPU seconds from the OS, writes a JSONL curve plus
    supervisor-verdict.json. Hidden by default: observe-only unless
    --confirm-kill is passed.
  • Covers the two blind spots of the in-process watchdog: growth while the JS
    thread is synchronously blocked, and growth before the plugin loads.

Fixed — real-machine attribution

  • The CLI now actually calls classifyExit when the host disappears; before
    this it only wrote host-exited, so attribution existed solely in unit tests.
  • sync-blocked no longer requires a sampling gap of three intervals, which
    cannot happen in practice (the next tick after death is one interval away).
    The reachable signature is a peak between the warn and kill lines with the
    CPU pegged (≥ 0.8), which is what a stalled host looks like from outside.
  • Added exit-requested (we asked it to quit and it did — not a crash) and
    supervisor-kill (a forced kill is attributed to the supervisor).

dsh-plugin-v1.7.7

Choose a tag to compare

@github-actions github-actions released this 19 Nov 16:13

Fixed

  • The crash-plan level is recomputed from the crashes inside the 10-minute
    window instead of showing a stale persisted level (a fresh profile with no
    recent crash now shows L0, not yesterday's L2).
  • ai_prison_watchdog_status now returns the crashPlan object (level, window,
    reset time, next level, steps and last plan), matching the panel.

dsh-plugin-v1.7.6

Choose a tag to compare

@github-actions github-actions released this 19 Nov 15:32

Added — crash degradation plan (crashForensics v2)

  • Explicit L1–L3 plan preview in the snapshot and the Enforcement card:
    current level, crashes inside the 10-minute window, the 30-minute stability
    reset, the next level and which steps are active.
    • L1 isolate the attributed (or in-flight) plugin; unknown attribution
      records the plan without killing anything.
    • L2 L1 + close plugins changed since the clean baseline (without a
      baseline only the attributed target is closed).
    • L3 safe mode: keep the clean baseline and close every newer plugin;
      without a baseline, close every non-official third-party plugin.
  • Escalation raises a forced crash-escalation notification; 30 minutes
    without a crash records crash-reset and returns to L0.
  • Plans now record targets, keep, applied, executedAt and a
    fallback (observe-only / changed-set / safe-mode).
  • pending-boot.json never downgrades an existing plan level.
  • Crash enforcement waits for pluginManager and the first scan runs on the
    next microtask, so the initial plan can close plugins instead of seeing an
    empty bundle list.

Added — per-plugin memory tiers in the plugin prison

Added — per-plugin memory tiers in the plugin prison

Occupancy Action
≥ 300 MB warning incident (plugin-memory warn)
≥ 500 MB throttle: add-memory plugin operations are denied (target growth ≤ 1 MB/s)
≥ 1000 MB force close (quarantine + pending-boot)
≥ 2000 MB force uninstall via pluginManager.removeBundle (fallback: force close)
≥ 3000 MB close every third-party plugin (official/installation plugins excluded)
  • The panel Enforcement Top5 table gains a memory tier column with the current
    occupancy; the incident list and remediation text explain each tier.
  • Occupancy uses the measured module-load budget plus child RSS when available,
    falling back to the operation estimate. Actions only run in enforce mode;
    observe mode records tier changes.
  • Tiers are recomputed after each heap sample; escalation is one-way per run
    (a lower reading does not release a closed/uninstalled plugin).

dsh-plugin-v1.7.4

Choose a tag to compare

@github-actions github-actions released this 19 Nov 14:39

Changed

  • Enforcement card: the Top5 action button now reads 立即缉拿 / Arrest now
    (previously 立即关闭 / Close now), and its confirmation prompt uses the same
    wording. The underlying action is unchanged (close the plugin); only the
    AI-Prison-themed label changed.
  • Enabling a quarantined plugin from the UI is immediately closed again and raises
    a tray notification: 请到 AI 监狱释放该插件!再开启。 The reminder is forced
    even when general notifications are off (rate limited per kind).
  • Header version pill shows the plugin version (AI Prison vX.Y.Z) instead of
    the bundled aiprison CLI version; the CLI version stays in the footer. Before,
    the pill always read aiprison 1.0.9 regardless of plugin releases.

dsh-plugin-v1.7.3

Choose a tag to compare

@github-actions github-actions released this 19 Nov 11:08

Added

  • Official tool API: tools register through ctx.tools.register(plainDefinition)
    first (Host ToolDefinition contract); defineTool from
    @deepseek-ai/dsh-tools is now only a legacy fallback, so registration no
    longer depends on installation-path guessing.
  • Artifact attestation: the release workflow attests the npm tarball, the
    six platform-lite tarballs and SHA256SUMS (actions/attest-build-provenance).
  • SignPath OSS signing wired: conditional workflow steps upload the Windows
    CLIs, request Authenticode signing (wait for completion), replace the binaries
    and continue; without SignPath variables the build stays unsigned with an
    explicit warning. Setup, variables and verification are documented in
    docs/SUPPLY-CHAIN.md.
  • Full-platform distribution: the release also builds six platform-lite
    tarballs (1.37–1.56 MB each) next to the all-platform tarball, with checksums
    and attestation; the npm package layout is unchanged.

Fixed

  • scripts/pack-platform.mjs gained --all, exported argument parsing, target
    suffixes and three tests; release assets no longer overwrite each other.

dsh-plugin-v1.7.2

Choose a tag to compare

@github-actions github-actions released this 19 Nov 10:49

Fixed

  • Entry-based installation anchors accept a non-absolute argv[1]; the 1.7.1
    tag accidentally shipped without this portability change, so the harness
    tests failed on ubuntu CI.
  • manifest.test.mjs normalizes CRLF before comparing the embedded market icon
    with icon.svg; patch-docs.test.mjs accepts \r?\n row separators, so a
    Windows checkout with autocrlf no longer fails those assertions.

Note

  • 1.7.0 and 1.7.1 reached npm but their tag-triggered GitHub Releases failed on
    the two test portability issues above; 1.7.2 is the complete three-channel
    release.

dsh-plugin-v1.6.1

Choose a tag to compare

@github-actions github-actions released this 18 Nov 16:14

Fixed

  • 1.6.0 shipped watchdog-runtime.js calling createBootWatch without importing
    ./boot-watch.js, so the watchdog degraded with "createBootWatch is not defined"
    and the startup watch never started. The import is restored and a regression
    test (test/watchdog-imports.test.mjs) now asserts every helper module the
    runtime calls is imported.

dsh-plugin-v1.6.0

Choose a tag to compare

@github-actions github-actions released this 18 Nov 15:51

Added — startup watchdog, safe boot and panel 6 GB control

  • bootWatch: 500 ms heap sampling from watchdog activation to the clean
    baseline, written to .ai-prison-watchdog/run-<start>.jsonl (last 5 kept) so a
    post-crash boot or the guard can inspect the curve even when the host died.
  • Startup thresholds: growth > 150 MB/s or total > 2.0 GB (4 GB host) /

    3.0 GB (6 GB host) writes a safe-boot plan (base + web-app + ai-prison) and
    records a startup-memory incident; the guard enforces it on next launch.

  • Guard v2: startup-phase OOM in the newest crash log forces that safe boot,
    moves the ai-prison bundle to position 2, applies pending-boot plans, updates
    the 6 GB heap patch from a flag file, and logs the last startup curve.
  • Panel 6 GB control: the Enforcement card can request/cancel "6 GB heap on
    next launch"; the request is a flag file the guard consumes.
  • Safer defaults: RSS child scan and automatic budget probes are now off
    unless explicitly enabled (rssScan: true, probeAuto: true); notifications
    stay off. This also removes the 1.5.x periodic subprocess spawns from startup.
  • ai_prison_watchdog_status now reports boot (run file, peak, trigger) and
    heap6g.

Note

  • The panel heap toggle only takes effect on the next launch through the guard,
    because the archive is locked while DSH runs.

dsh-plugin-v1.5.4

Choose a tag to compare

@github-actions github-actions released this 18 Nov 13:30

Fixed — RSS diagnostics were invisible

  • state.json and ai_prison_watchdog_status now carry rss, rssAt,
    rssError, rssStats, rssServiceError and rssAttempts, and the values are
    restored on reload. Before this, a working (or failing) scan could not be
    observed outside the panel route.

dsh-plugin-v1.5.3

Choose a tag to compare

@github-actions github-actions released this 18 Nov 13:10

Fixed — CI

  • The RSS scanner tests exercise a Windows-only code path; they now skip on the
    ubuntu release runner instead of failing there. The parser test still runs on
    every platform.