Skip to content

Releases: HaowenCang/dsh-turn-performance-meter

v0.1.7 — Session recovery & DSH 0.2.1 compatibility

Choose a tag to compare

@HaowenCang HaowenCang released this 24 Oct 12:32

Session navigation recovery

Fixed a case where the live dashboard could stay stuck on a previous tool call or a "waiting for model" state after switching from the main agent to a subagent and back (or when switching sessions while a turn was live). The plugin now re-binds to the visible session's current event-source generation using event-source identity-aware rebind and performs an authoritative resync, so the live meter reflects the session that is actually on screen.

Resource lifecycle

When a session scope is retired (navigating away), its retention watcher and old event-source subscription are released automatically instead of accumulating one per visited session. Re-visiting the session restores exactly one observer. Historical completed cards and accumulated statistics are retained.

DSH compatibility

This release declares engines.dsh: ">=0.2.1-rc.2 <0.2.2-0" instead of an exact DSH peer dependency. It was verified on DSH 0.2.1-rc.2 and DSH 0.2.1-rc.3. Later versions inside the declared range are allowed to install by the range decision but have not been device-verified, and no runtime compatibility claim is made for them. The DSH 0.2.0 series and the DSH 0.2.2 series are not supported.

Metric stability

No change to TPS, TTFT, token accounting, tool timing, rate publication, peak statistics, completed-curve geometry, or the 100 ms presentation cadence. The shipped bundles are byte-identical to the previously accepted artifact (client.js / lib/client.js, 651,961 bytes, SHA-256 07F717FA4BA11425A034B3DD81E122C208D99D1783E4C0F919C409118137ED69).

Historical releases

v0.1.6 remains the historical release that uses the exact 0.2.1-rc.2 DSH peer; its npm package, Git tag, GitHub Release, and canonical tarball are unchanged. Version 0.1.7 does not support the 0.2.0 series or the 0.2.2 series.

Artifact

  • dsh-turn-performance-meter-0.1.7.tgz — 408,179 bytes
  • SHA-256: C401A29F08B5E6BD091B0603798F989CC6614C34A5FD5DF57C3EB394B9763DDC
  • npm: npm install dsh-turn-performance-meter@0.1.7

v0.1.6 — DSH 0.2.1-rc.2 compatibility

Choose a tag to compare

@HaowenCang HaowenCang released this 18 Oct 05:42

Compatibility and runtime acceptance release for DeepSeek Harness 0.2.1-rc.2.

This release establishes verified compatibility with DeepSeek Harness 0.2.1-rc.2 and updates the exact runtime peer dependency declaration.

Highlights

  • DSH 0.2.1-rc.2 compatibility: Adds verified compatibility with DSH 0.2.1-rc.2 following full runtime audit and acceptance testing.
  • Exact peer dependency: Updates the exact DSH peer dependency in package.json to "@deepseek-ai/dsh": "0.2.1-rc.2".
  • Preserved core telemetry: Preserves existing TPS, TTFT, tool timing, and completed-curve behavior without modification.
  • Preserved presentation & lifecycle stability: Preserves multi-attempt continuity connectors and background/foreground-recovery behavior.
  • No substantive algorithm changes: No substantive telemetry calculation or rendering algorithm changes are introduced; metric semantics and presentation cadence remain unchanged.
  • Regression test coverage: Includes updated compatibility regression suites verifying runtime contracts against DSH 0.2.1-rc.2.
  • Historical version availability: v0.1.5 remains published and available for DSH 0.2.0-rc.2. Compatibility is not claimed for untested DSH releases.

Compatibility

  • Supported: DeepSeek Harness 0.2.1-rc.2 only (exact peer dependency: @deepseek-ai/dsh@0.2.1-rc.2).
  • DSH 0.2.0-rc.2 users should continue using v0.1.5.
  • Do not install on untested DSH releases.

Installation

Pinned install (recommended during release propagation):

dsh plugin --profile web add dsh-turn-performance-meter@0.1.6

Standard install:

dsh plugin --profile web add dsh-turn-performance-meter

Note on release propagation: Newly published npm versions may temporarily be excluded from unpinned resolution due to pnpm minimumReleaseAge configuration. Using the pinned form (@0.1.6) ensures immediate deterministic resolution.

Artifacts

  • dsh-turn-performance-meter-0.1.6.tgz (canonical npm tarball, 386,334 bytes)
  • dsh-turn-performance-meter-0.1.6.tgz.sha256 (SHA-256 sidecar checksum)
  • SHA-256: 6d1196d226ed93605263407e37ce7c44120576fb57ab476cd4d3f88f2ccb7884
  • SHA-1: 6aedd1b4362df6faaa2b7aa3c6f22ff3aa46f53b
  • SHA-512 integrity: sha512-k8LQL38hccmRal+Wu1RwCsa49RTrWIYK+KppGdgpRPhVBR+D03BFmrHS0Wnb3ocTAatpUv9GdgSboGw1CEfuYQ==
  • Built from tag v0.1.6 (568ebaaeab2c9362a2c265f9d5173bac73b264e5).

v0.1.5 — Multi-attempt curve continuity and foreground recovery

Choose a tag to compare

@HaowenCang HaowenCang released this 07 Oct 08:49

Multi-attempt curve continuity and foreground recovery for DeepSeek Harness 0.2.0-rc.2.

This release introduces visual throughput curve continuity across multi-attempt generation boundaries and adds robust foreground recovery with timer suspension when the DSH page is backgrounded.

Highlights

  • Visual curve continuity: Completed throughput curves now visually stitch eligible same-phase generation segments across multi-attempt/tool boundaries.
  • Presentation-only connectors: Continuity connectors are presentation-only and do not modify TPS, peaks, token totals, durations, or attempt-local estimator resets.
  • Stale live-meter state fix: Fixes stale live-meter state after the DSH page is backgrounded.
  • Timer churn prevention: Background presentation timers are suspended instead of continuing timer churn while hidden.
  • Authoritative foreground resync: Returning to the foreground performs an authoritative state resynchronization before the recovered view is shown.
  • Full-window recovery on missed revisions: Missed session-event revisions trigger full-window recovery, preventing transitions such as tool/result, new attempts, or turn/end from being lost.
  • Generation-safe lifecycle handling: Foreground lifecycle handling is generation-safe, ensuring rapid tab/window switching cannot suppress a genuine hidden -> visible recovery.
  • Exact compatibility: Compatibility remains exactly DeepSeek Harness 0.2.0-rc.2.

Compatibility

DeepSeek Harness 0.2.0-rc.2 only (exact peer dependency: @deepseek-ai/dsh@0.2.0-rc.2).

Install

dsh plugin --profile web add dsh-turn-performance-meter

Pinned:

dsh plugin --profile web add dsh-turn-performance-meter@0.1.5

Release-Window Note

Immediately after publication, an unpinned install may temporarily resolve the previous version because DSH/pnpm uses minimumReleaseAge. Pinned @0.1.5 is the reliable immediate-install form.

Artifact

dsh-turn-performance-meter-0.1.5.tgz is the canonical tarball published to the npm registry, built from the v0.1.5 annotated tag (06ffa9855beb1fb64155203281e873a7ba63efdd).
SHA-256: 9fe55298da4972399ce6bd6046841c0a228a5e4b9577b01788c5c7d9179a1b6c; the dsh-turn-performance-meter-0.1.5.tgz.sha256 sidecar carries the same digest.

v0.1.4 — Live panel alignment and 10 Hz display precision

Choose a tag to compare

@HaowenCang HaowenCang released this 05 Oct 14:08

Live panel alignment and 10 Hz display precision for DeepSeek Harness �.2.0-rc.2.

This is a presentation and precision release for the live turn meter. It aligns live text elements across mixed font sizes and matches live stopwatch precision to presentation cadence:

Highlights

  • Shared typographic baseline: Live panel text now shares a typographic baseline across mixed font sizes, eliminating visible vertical drift in waiting, warming, streaming, and tool execution states.
  • Tenths stopwatch precision: Running live stopwatches now show tenths rather than hundredths, matching display precision to the 10 Hz / 100 ms presentation cadence.
  • Timing calculations unchanged: Internal millisecond timing and all TPS and TTFT calculations are unchanged.
  • Completed card unchanged: Completed-card precision and turn-level summary metrics are unchanged.
  • Compatibility unchanged: DSH compatibility remains exactly �.2.0-rc.2.

Compatibility

DeepSeek Harness 0.2.0-rc.2 only

Install

powershell dsh plugin --profile web add dsh-turn-performance-meter

Pinned:

powershell dsh plugin --profile web add dsh-turn-performance-meter@0.1.4

Release-window note

Because pnpm enforces a minimumReleaseAge policy for unpinned package resolution, dsh plugin --profile web add dsh-turn-performance-meter may continue resolving the previous release until the release-age window passes. Use the pinned form (@0.1.4) to install �.1.4 immediately.

Artifact

dsh-turn-performance-meter-0.1.4.tgz is the same canonical tarball published to the npm registry, built from the �0.1.4 annotated tag (1f40933).
SHA-256: cf8ea405c3783e7444cb25808b854d3320ff6b75ae5115d888596acf085ec261; the dsh-turn-performance-meter-0.1.4.tgz.sha256 sidecar carries the same digest.

v0.1.3 — Runtime telemetry correctness and reload parity

Choose a tag to compare

@HaowenCang HaowenCang released this 03 Oct 14:21

Runtime telemetry correctness and reload parity for DeepSeek Harness 0.2.0-rc.2.

This is a correctness and recovery release, not a redesign: it repairs how first-token evidence, phase boundaries and
reload-time reconciliation produce live and completed numbers. No metric definition, UI surface or DSH adapter contract
is redefined.

Highlights

  • fixes first-token/tool-boundary handling
  • removes boundary-origin and short-denominator TPS spikes
  • preserves phase-cut gaps in completed curves/summaries
  • fixes mid-turn reload reconciliation against complete durable settlement streams
  • adds explicit durable temporal-evidence authority
  • prevents incomplete durable streams from claiming reconstructed shape quality
  • preserves null/unavailable rendering semantics

Compatibility

DeepSeek Harness 0.2.0-rc.2 only

Install

dsh plugin --profile web add dsh-turn-performance-meter

Pinned:

dsh plugin --profile web add dsh-turn-performance-meter@0.1.3

Release-window note

Measured in fresh, disposable profiles immediately after publication:

  • dsh plugin --profile web add dsh-turn-performance-meter@0.1.3 installed 0.1.3 straight away. The DSH plugin
    manager recorded dsh-turn-performance-meter@0.1.3 in that profile's minimumReleaseAgeExclude.
  • dsh plugin --profile web add dsh-turn-performance-meter (unpinned) resolved to the previous release, 0.1.2,
    during the release window. The same outcome was reproduced outside DSH by a plain pnpm range resolution against a
    freshly fetched registry metadata cache, so it is a release-window resolution policy in this environment rather than
    a stale-cache artefact.

Until that window passes, use the pinned form to install 0.1.3 immediately.

Artifact

dsh-turn-performance-meter-0.1.3.tgz is the same tarball the npm registry serves, published from the v0.1.3
annotated tag. SHA-256 d29eab7503bd2c0f7d3f535ba4491fd392f1e66c9bb61688e223a19f574acdac; the
dsh-turn-performance-meter-0.1.3.tgz.sha256 sidecar carries the same digest.

v0.1.2 — DSH 0.2.0 and cumulative TPS

Choose a tag to compare

@HaowenCang HaowenCang released this 29 Sep 15:20

dsh-turn-performance-meter v0.1.2 adds the Phase 9 presentation work, MiMo-style phase-cumulative TPS semantics, and verified compatibility with DeepSeek Harness 0.2.0-rc.2.

Highlights

  • completed cards default to a compact collapsed state
  • expanded summary/curve remains available on demand
  • live TPS now uses phase-cumulative averaging
  • reasoning/output transitions reset the phase clock
  • stalls naturally decay by elapsed wall time
  • completed rates are recomputed at settlement
  • completed curve uses the same cumulative estimator family
  • 100 ms presentation cadence
  • exact DSH peer compatibility: 0.2.0-rc.2
  • background-settlement card advancement regression coverage

Compatibility

DeepSeek Harness 0.2.0-rc.2 only.

No support claim is made beyond DSH 0.2.0-rc.2. The compatibility baseline moved from 0.1.7-rc.2 (the v0.1.1 release) to 0.2.0-rc.2, and the exact peer is declared as "@deepseek-ai/dsh": "0.2.0-rc.2". Every DSH declaration this plugin reads was re-audited against the new runtime and is byte-identical to its 0.1.7-rc.2 form, so no adapter, metric or presentation behaviour changed with the runtime itself. No compatibility exemption is required or used.

Install

dsh plugin --profile web add dsh-turn-performance-meter

Pinned:

dsh plugin --profile web add dsh-turn-performance-meter@0.1.2

The unversioned form resolves this version through npm's latest dist-tag. The prerequisite runtime is @deepseek-ai/dsh@0.2.0-rc.2; the v0.1.1 release remains bounded to 0.1.7-rc.2, and the two are separate claims rather than one range.

Release-window caveat (first 24 hours)

pnpm 11 enables minimumReleaseAge by default at 1440 minutes (24 hours), and its default minimumReleaseAgeStrict: false makes the resolver fall back to an older version rather than fail. A freshly published 0.1.2 is therefore too new to be resolved by the unversioned spec, which falls back to 0.1.1 — a version the DSH compatibility gate then correctly rejects for declaring the 0.1.7-rc.2 peer.

Until 0.1.2 is 24 hours old, install it that way:

  • use the pinned form above, which resolves immediately, or
  • add an exclusion to the profile's pnpm-workspace.yaml:
minimumReleaseAgeExclude:
  - dsh-turn-performance-meter

This is pnpm's supply-chain policy rather than a property of this package, and the unversioned command needs no workaround once the release is a day old.

Assets

  • dsh-turn-performance-meter-0.1.2.tgz — the exact tarball the npm registry serves for 0.1.2, retained here as an offline, immutable fallback
  • dsh-turn-performance-meter-0.1.2.tgz.sha256 — SHA-256 of that tarball

Verify before installing:

Get-FileHash .\dsh-turn-performance-meter-0.1.2.tgz -Algorithm SHA256
Get-Content .\dsh-turn-performance-meter-0.1.2.tgz.sha256

Notes

Metric semantics are documented in docs/METRICS_SPEC.md (§6 live TPS, §7 completed rates, §8 curve). The per-phase implementation record, including the superseded trailing-window contract that this release replaces, is in docs/IMPLEMENTATION_LOG.md.

v0.1.1 — npm distribution

Choose a tag to compare

@HaowenCang HaowenCang released this 28 Sep 13:07

dsh-turn-performance-meter v0.1.1

This release adds public npm distribution for the existing v0.1.0 runtime implementation.

本版本为既有 v0.1.0 运行时实现增加公开 npm 分发。

Install

Recommended:

dsh plugin --profile web add dsh-turn-performance-meter

Pinned:

dsh plugin --profile web add dsh-turn-performance-meter@0.1.1

The unversioned command resolves npm's latest dist-tag.

Compatibility

Supported and tested:

  • DeepSeek Harness 0.1.7-rc.2

The package declares the exact peer requirement:

@deepseek-ai/dsh = 0.1.7-rc.2

No compatibility claim is made for later DSH releases. DSH evaluates this peer against the running runtime version before
a profile starts the plugin, so an unsupported DSH is rejected rather than silently admitted.

Distribution

  • npm: dsh-turn-performance-meter@0.1.1
  • GitHub Release .tgz
  • development checkout via link:

Artifact

  • file: dsh-turn-performance-meter-0.1.1.tgz (8 files, 294380 bytes)
  • SHA-256: 38250f6ae546e63b4320c66a835a29aae693068c305437f7e2f6500398f6399d
  • SHA-1: 55efeb484148e46a8749c7e1fb48acfcec5ee5ec
  • sha512 integrity: sha512-n+v7EuNUoqSQXtVH3ztufTLBOjGDFapxcqJ+p6BydWxBtFJlteS5f/nKtG64oAuYmF7EHjJj/+sGzJN1REd7Vg==

This Release asset is the tarball downloaded back from the npm registry after publication, and it is byte-identical to the
tarball packed from the release commit. Its SHA-1 equals the registry's dist.shasum and its sha512 equals the registry's
dist.integrity.

Changes from v0.1.0

No metric, UI, curve, telemetry, or DSH adapter semantics changed.

Changes are limited to:

  • npm registry publication
  • exact DSH peer compatibility metadata
  • runtime-only npm package allowlist
  • installation and release documentation

Verification

  • full local automated suite: PASS
  • sanitization: PASS
  • prepublish tarball install: PASS
  • prepublish cold start: PASS
  • npm registry artifact integrity: PASS
  • exact-version registry install: PASS
  • unversioned standard DSH install: PASS
  • registry-installed cold start: PASS

This repository does not currently use GitHub Actions CI.

v0.1.0 — Initial release

Choose a tag to compare

@HaowenCang HaowenCang released this 28 Sep 11:54

dsh-turn-performance-meter v0.1.0

Initial public GitHub release of dsh-turn-performance-meter.

首个公开 GitHub Release。

Compatibility / 兼容性

  • Supported and tested: DeepSeek Harness 0.1.7-rc.2
  • Public DSH reference commit: 477b4f420553e8a52c2fbccc464d7561b239c443
  • No compatibility claim is made for 0.1.7+, 0.1.x, or later DSH versions.
  • npm distribution is not provided; the package remains private: true.

Highlights / 主要功能

  • Turn-level live TPS using the active attempt's trailing one-second window.
  • Completed-turn Reasoning TPS, Output TPS, Generated Tokens, TTFT, tool timing and model-call count.
  • Ratio-of-sums aggregation across multi-call agent turns.
  • Calibrated throughput curve with reasoning/output phase colouring.
  • Tool execution and inter-attempt waits compressed to zero curve width.
  • Hover and keyboard-focus completed-card interaction.
  • DSH 0.1.7-rc.2 tool-result identity using message.toolCallId / message.isError.
  • Terminal-tail reconstruction for turn/end when no live record is open.
  • Generation-wide durable-event deduplication with bounded retained evidence.
  • 50 ms presentation cadence and no stale TPS during tool/wait states.

Installation / 安装

Download dsh-turn-performance-meter-0.1.0.tgz from this release, then install it into a DSH profile:

dsh plugin --profile web add "file:C:/path/to/dsh-turn-performance-meter-0.1.0.tgz"

For development from a working checkout:

dsh plugin --profile web add "link:C:/path/to/dsh-turn-performance-meter"

Verification / 验证

Release baseline passes:

  • npm run verify: 714 / 714
  • fixture sanitization: PASS
  • client.js / lib/client.js: mirrored and fresh
  • release tarball installation: verified in a disposable DSH profile
  • release tarball cold start: verified against DSH 0.1.7-rc.2

Automated verification is local; this repository does not currently use GitHub Actions CI.

Known limitations / 已知限制

Live TPS and temporal curve points are approximate because DSH does not provide authoritative token counts for individual stream deltas. Provider usage determines the quality of completed token totals and reasoning/output splits. If turn/start is absent from the retained window, TTFT and total turn elapsed remain unavailable rather than inferred.

See README.md and CHANGELOG.md for the full metric and evidence contract.