# Validation and release gate — 0.1.11

## Update and compatibility coverage — 2026-10-05

The update controller is checked for complete release selection, draft/foreign archive rejection, numeric version ordering, daily cache persistence, concurrent requests, declined confirmation, duplicate installation gestures, local restore, pinned-source migration, offline/missing market responses, and failed/stale/blocked installation results. It uses the independent market's existing HTTP contract and never launches a package manager itself. Successful updates preserve the loaded version until a restart.

Official host-package checks verify the declared web service and client entry packages, and, when provided by the host, its unmodified compatibility evaluator. Modern host types must retain the public assistant stream contract. The CI matrix uses 0.1.1-rc.2, 0.2.0-rc.2 and the moving alpha channel; this is contract validation, not complete GPU/browser acceptance on every alpha build.

The historical real-model/browser evidence below remains applicable to the unchanged synthesis path. New update behavior does not establish real GPT-SoVITS model availability, physical listening quality or compatibility with 0.3.

Local 0.1.11 validation passed 91 Node tests, 57 client-load checks, 6 bilingual checks and 14 Python adapter tests. The contract tool passed against the complete official 0.1.1-rc.2, 0.2.0-rc.2 and 0.2.1-alpha.1 installations. The update settings panel loaded in the real 0.2.0-rc.2 browser; a release check returned the public candidate metadata and kept installation disabled when no higher published version was available.

## Real Harness compatibility acceptance — 2026-10-05

The final 0.1.10 package was installed in the existing official Harness **0.1.1-rc.2** web profile and a separate complete official **0.2.0-rc.2** web profile. The older running installation was checked directly; it was not assumed to be 0.2. No Harness core files or model weights were changed.

- Both real hosts loaded the plugin and displayed its settings, message actions and automatic reading controls. IndexTTS used an existing installation with `examples/voice_01.wav` as the reference.
- A real 0.1 preview returned a valid 114220-byte WAV. On 0.2, two separate sentences produced successful real synthesis requests and sequential browser playback. A captured response was mono 22050 Hz / 16-bit WAV, 138796 bytes, 3.146 seconds, with nonzero samples.
- The complete 0.2 Agent used a local OpenAI-compatible fixture for answer generation. Its real start/chunk/end frames and durable final message triggered one real IndexTTS synthesis for a short reply, without replaying the final message. Only the answer source was a fixture; the host, TTS model and browser were real.
- Manual message reading, pause before the first buffer arrived, resume and stop were exercised in the actual 0.2 browser. The pending-buffer pause defect discovered during this run is fixed in 0.1.10 and has a regression test.
- Final local automated checks: **82 Node tests, 57 client load/render checks, 6 bilingual checks and 14 Python adapter tests** passed on Windows / Node 24.13.1.

This verifies the loading/interface repair and IndexTTS path on those two host versions. It does not establish physical speaker listening quality, uninterrupted playback when synthesis is slower than real time, a real GPT-SoVITS model run, or every original-plugin coexistence scenario. GPT-SoVITS remains covered by adapter fixtures, so this release retains candidate status. Historical preparation notes below describe the evidence available at their dates; they are superseded by this section where applicable.

## Initial 0.1.9 compatibility checks — 2026-10-05 (before real-host acceptance)

The original `dsh-web: ^0.1.0-rc.6` peer is rejected by Harness 0.2's `evaluatePluginCompatibility`: it checks all `@deepseek-ai/dsh-*` peers against the running DSH version with `includePrerelease: true`, and denied bundles/entries are omitted before activation. Enabling a denied entry cannot make it load. The plugin actually consumes `webServer`, not the search/fetch `dsh-web` capability.

The 0.2 package split also removed `dsh-client-runtime`, moved messages to the public Chat selector, and moved transient chunks out of the durable Session log into `agent/assistant-stream`. Version 0.1.9 corrects the runtime declaration and adapts these public interfaces while retaining the 0.1 interfaces.

- Node regression suite: 81 tests; client load/render: 57 checks; bilingual keys: 6 checks; Python adapters: 14 tests passed on Windows / Node 24.13.1. No real model was started for this repair.
- The unmodified compatibility evaluator extracted from the official `dsh-app-boot@0.2.0-rc.2` package rejects the old manifest and accepts the new manifest for `0.1.0-rc.6`, `0.1.1-rc.2`, `0.2.0-rc.2`, and `0.2.0`. It continues to reject below-minimum and future `0.3.0-rc.1` versions. This is a gate check, not execution on every listed host.
- The official 0.2 SlotRegistry and SlotCore with Cordis 4.0.4 registered the plugin's overlay, input toggle, assistant actions and settings tab, then removed them on plugin disposal. React was supplied by the existing installation and the DOM was a lightweight fixture; actual visual/audio acceptance is still pending.
- New tests cover legacy/new message selectors, start/chunk/end framing, independent transient revisions and durable seq, duplicate/end-frame suppression, final-message deduplication, retry residual cleanup, Agent replacement, session separation and manual read priority.

These initial fixture checks preceded the real-host acceptance recorded above and do not alone prove browser playback or compatibility with arbitrary future versions.

## Automated coverage

Run `npm test` and `npm run test:adapter`. Tests require no GPU, downloaded model, cloud API, login credential or newly installed production dependency.

- Configuration validation, localhost defaults, custom ports/base paths, private LAN consent, public URL/credentials/query rejection.
- IndexTTS and GPT-SoVITS process worker contracts, JSONL health/voices/synthesis, worker startup and cleanup; legacy HTTP adapters remain covered for compatibility.
- Offline service, full-response timeout, HTTP cancellation, redirects, malformed payloads, error redaction, explicit debug logging.
- Three Chinese/English sentences → exactly three Runtime requests; short sentences never merge; 55/70-character limits; streamed decimal/tail handling; visual dash separators are removed before synthesis.
- Nonblocking event ingestion, duplicate `seq`, final-message deduplication, repeated React session snapshots, session separation.
- Failure recovery, cancellation and turn suppression, switching engine/voice, bounded prefetch, pause backpressure, client expiry.
- Browser controller schedules three consecutive Web Audio buffers and acknowledges them after completion (mock AudioContext).
- Existing UI bundle load/render, bilingual keys and settings persistence; simplified engine/project/reference controls, per-engine path restoration, output-language selector, model speed control and collapsed advanced paths.
- Bounded read-only IndexTTS/GPT-SoVITS discovery, wrapper/Unicode layouts, missing or ambiguous candidates, incompatible Index config version, directory-link escape and audio count/depth limits; same-origin route validation with no client/worker creation.
- Discovery request deduplication, preservation of manual overrides, late-response rejection, engine/provider synchronization, GPT reference synchronization and disabling stale Host Auto Read during incomplete configuration.
- Unmodified upstream 0.3.0 Host/client fixtures loaded in both orders: no route/slot/style collisions, no original settings/global changes, and independent disposal.
- Public inventory detection, missing inventory fallback, enable/disable, forced companion mode, one-time migration/reload/reset, Auto Read state guard and manual playback preservation.
- Optional stdlib Python HTTP facade exercised with mock existing objects for both engines. No actual model object is created.
- Official backend launch guards: source-only entry inspection, IndexTTS/GPT API-shape selection, Gradio 4/5 metadata and loopback checks, explicit saved GPT weight selection, occupied-port refusal in auto mode, and owned-process EOF cleanup.
- Generated client consistency, package identity, original LICENSE preservation and release artifact scan.

Preparation results: 75 Node tests (including discovery, settings, process-worker, coexistence and runtime-pool tests), 57 client-load checks, 6 i18n checks, and 14 Python adapter tests passed on Windows, Node 24.13.1 / Python 3.11.9. The Gradio 4/5 official-shape fixture smoke also passed. This is not a claim that GitHub Actions has already run remotely.

The static UI fixture was opened in a real browser and inspected after this change. Both engines show project/reference fields and correctly selected providers; advanced paths are hidden by default and the native details control expands. The fixture uses the upstream lightweight DOM renderer, not a running Harness/React application, so it is visual QA only; discovery/editing behavior is covered separately by controller tests. Run `node tools/gen-ui-preview.mjs` then `node tools/preview-server.mjs` to reproduce; generated HTML/screenshots are excluded from Git.

Read-only discovery was also run against the existing local IndexTTS wrapper on 2026-08-28. It resolved `app`, its `.venv` Python, `checkpoints` and 14 audio candidates, including example voices and a nested preset reference. It deliberately did not auto-select a voice among multiple candidates. This check did not launch Python, read audio/weights or establish GPT-SoVITS availability. GPT integrated-package discovery was tested with filesystem fixtures only.

## Historical model preparation — 2026-08-28, PARTIALLY COMPLETE

During preparation, installed Harness event types and both existing Python Runtime method signatures were inspected read-only. The existing GPT-SoVITS HTTP bridge contract was inspected read-only. No Runtime implementation or inference parameter was changed.

The IndexTTS 2.5 installation and a usable local worker were confirmed on 2026-08-28. The direct worker smoke passed on CPU, and the real GPU regression passed (`1 passed, 16 deselected`) with the existing `雷米埃尔` cache. No model code, weights or inference settings were changed. A local headless WebUI process also exposed its real Gradio 5 schema and generated a valid WAV from `examples/voice_01.wav`; this was a test of the current local `app/webui.py`, not proof of an unmodified upstream checkout. A running Harness web endpoint, browser playback session, and a real GPT-SoVITS installation with weights are still not confirmed. The requested full end-to-end acceptance below has therefore **not passed**, and the project remains a release candidate. Fixture audio must not be reported as real generated speech.

### Local role and WebUI checks outside DSH — PARTIAL

The existing IndexTTS preset registry recognized `甘雨` and `雷米埃尔`; both caches contained the required tensors. Loading each cache with `IndexTTSRuntime` and synthesizing `你好，这是本地角色音色测试。` produced nonempty mono 22050 Hz WAV files with nonzero RMS: `outputs/verification/甘雨-verification.wav` (3.506 s, 142380 bytes) and `outputs/verification/雷米埃尔-verification.wav` (3.228 s, 154668 bytes).

The local WebUI started headlessly on loopback, returned Gradio `5.45.0 / sse_v3` metadata, and `/gen_single` produced a valid 22050 Hz WAV (177708 bytes, 4.03 s) when given the real RIFF `examples/voice_01.wav` reference. Both saved preset files named `prompt.wav` are actually MP4/ISO files (`ftypisom` and `ftypiso4` headers). The unmodified path fails on those files, while the plugin's new launch hook matched their SHA-256 fingerprints, loaded the existing voice caches, and generated a valid WAV through the official inference path. The original files were not altered. The GPT-SoVITS tree currently present under restore staging contains source files but no model weights or usable environment, so it was not a real synthesis test.

Earlier direct health probes to the documented example endpoints (IndexTTS port 8765 and GPT-SoVITS port 9881) both returned `OFFLINE / ECONNREFUSED`; those probe ports are not assumed to be the user's actual deployment ports. This is separate from the later direct IndexTTS process-worker smoke below.

### Direct IndexTTS 2.5 process-worker smoke — PASSED

Using the existing installation's Python environment, `app/checkpoints`, and the repository's example `voice_01.wav`, the plugin's `ProcessTTSProvider` started `adapters/index-tts-worker.py`, imported `indextts.infer_v2_5.IndexTTS2`, passed `health` and `voices`, then synthesized these three sentences independently:

| Request | Result |
| --- | --- |
| 1 | WAV, 121388 bytes, 22620 ms |
| 2 | WAV, 124972 bytes, 16750 ms |
| 3 | WAV, 119340 bytes, 16792 ms |

All responses had a valid RIFF/WAVE header. Total synthesis time was about 56 seconds on CPU. The worker's temporary output files were removed and the child process was disposed after the run. This proves the local process/adapter/model boundary only; it does not prove Harness event ingestion, sentence streaming timing, browser Web Audio playback, pause/resume, or audible quality.

| Required acceptance | Result |
| --- | --- |
| Direct IndexTTS worker: three requests and valid audio | **Passed 2026-08-28 on CPU**; see the smoke record above |
| Real Harness answer → IndexTTS | Passed 2026-10-05 on 0.2 with fixture answer generation and real synthesis; two-sentence manual playback also passed. Three-sentence continuous listening quality remains unverified. |
| Real Harness answer → GPT-SoVITS: three requests and continuous browser speech | Pending: running Harness + configured GPT-SoVITS worker required |
| Real browser pause / resume / stop | Passed 2026-10-05 on official 0.2; physical speaker listening quality remains unverified. |
| Both plugins enabled in actual Harness; original Edge/RVC preserved and local Auto Read ownership verified | Pending; both load orders tested only with captured original bundles and mocked Harness/browser APIs |
| Node 22/24 GitHub Actions matrix | Workflow prepared; remote CI has not been run here |

The tool `npm run smoke:runtime -- --engine … --endpoint … --voice …` performs three real HTTP syntheses without saving audio. Even a successful command does not replace browser listening acceptance.

## Exact real acceptance procedure

1. Prepare the two **existing** model installations. Supply each project folder, run automatic discovery and select its reference audio; use Advanced only for missing/ambiguous paths or GPT reference transcript/languages. Do not change model code/weights/inference parameters for this test.
2. Install this checkout into the existing web profile, restart Harness and refresh the browser. The original TTS plugin may remain installed/enabled. With both enabled, confirm distinct original and **Local voice** tabs; turn OFF original Auto Read before enabling Local Auto Read. Also test the reverse installation order and independent disable/re-enable as described in [COEXISTENCE](COEXISTENCE.md).
3. Select IndexTTS 2.5 (local), discover the project/reference and Test Connection. Confirm the worker responds, then try a short preview. Click the page once to unlock audio.
4. Enable Auto Read and ask Harness: “请只回答三句：第一句，这是本地测试。第二句，正在逐句播放。第三句，测试结束。”
5. Observe exactly three successful local synthesis log records (request ID, chars, duration, status; Debug OFF). Listen for order/continuity, and verify sentence 1 can start before the full answer is complete. Record whether the worker is slower than real time.
6. Repeat with a new reply, pause/resume, stop mid-reply, then manually read a message. Check stop does not play stale audio. Refresh and verify config persisted but old replies do not replay.
7. Switch to GPT-SoVITS (local), discover its project/reference, set the required transcript/languages in Advanced and repeat steps 3–6. Check no old engine/voice audio remains.
8. Temporarily break the worker command, verify a clear spawn/import/timeout failure, then fix it and manually retry. Do not let tests silently use Edge.

Record date, Harness version, engine/bridge version, browser, protocol, request count, playback ordering/continuity and remaining issues. Keep personal endpoints, voice IDs, logs, audio and screenshots in an ignored local artifact directory; publish only anonymized results.

## Before publishing

- [ ] Complete GPT-SoVITS real-model acceptance and continuous listening checks before claiming full engine acceptance. IndexTTS and the actual 0.2 UI were checked for this candidate.
- [x] Confirm repository owner/name: `abbccdd/dsh-localtts`; installation is pinned to `v0.1.10` or its prebuilt archive.
- [ ] Confirm the existing Runtime HTTP contracts and its operator's model/voice usage rights.
- [ ] Inspect `git diff`, `git status`, and `npm pack --dry-run`; verify no private files were force-added.
- [ ] Review upstream inherited Edge/RVC security assumptions. Do not expose their trusted-local management routes to untrusted users.
- [ ] Run CI on Node 22 and 24, then tag/release. This upload contains source only; models, audio and caches remain excluded.

Upstream RVC developer guides/tests remain for provenance; their historical real-machine results are **not validation of this fork**. The new README and this file govern Local Runtime release readiness.