dsh-plugin-bridge
English | 中文
Move a produced DeepSeek Harness session to another tool preset through a previewable, fixed-schema handoff. The original session stays untouched.
Quick start · Safety and cost · Accuracy · How it works · Compatibility
Quick start
dsh plugin --profile web add github:Totoro-qaq/dsh-plugin-bridge#v0.2.5
# restart dsh web onceThen type in the official WebUI:
/bridge list available target presets
/bridge code preview the handoff; changes nothing
/bridge code --go migrate, restate, then wait for confirmation
/bridge code --go --continue restate and start work in the same target turnThe preview is editable. If a number or path is wrong, edit the printed summary file and run:
/bridge code --go --file <path>Uninstall with dsh plugin --profile web remove dsh-plugin-bridge, then restart dsh web.
Verified on the official rc.8 WebUI (2026-08-20): clean install, plugin load,
standard → minimalpreview/migration, paused-goal persistence, and host restart. The package ships prebuiltlib/, so git installation needs no pnpm build-script allowlist.
Safety and cost
Bridge optimizes for high-fidelity, bounded-cost migration:
- Default — accuracy first: the target restates the handoff in one short turn and waits. You verify it before any work starts.
--continue— lower latency/cost: the target restates and begins the next step in the same model request.- In both confirmation modes, any stored goal is paused before kickoff. The goal driver cannot silently queue another model round.
- If pausing fails, Bridge fails closed: it cancels automatic startup and sends no kickoff.
- Installing the plugin adds zero prompt tokens to normal sessions.
/bridgeis a host slash command, not a model tool or skill. - Existing image analysis is copied verbatim outside the compressed summary. Only unresolved images may use vision tokens, and only when the target accepts images.
The release-acceptance run used one summary request per fixture. Confirm always reached first useful work in two target requests; --continue always did so in one. Token totals varied sharply with preset, output length, and cache state, so Bridge does not claim a universal percentage saving.
Observed across the fixed 12-cell release acceptance:
| Token measure | Nominal | Processed |
|---|---|---|
Confirm extra vs --continue, pooled summary + first useful work |
+12.82% | +52.56% |
Confirm extra vs --continue, paired median |
+8.1% | +65.6% |
| Summary worker share of the clean acceptance suite | 20.74% | 16.71% |
Nominal is the primary comparison. Processed weights cache-read tokens equally and is not a bill; the worker row is a composition share, not causal overhead versus a no-Bridge baseline.
Token definitions, ranges, and raw totals
`Nominal = uncached input + output`; `processed` also includes cache-read/cache-write tokens and is a sensitivity measure, not a bill. “Summary + first useful work” excludes the pre-existing source conversation and official compaction cost. Across the six paired fixtures, confirm's nominal extra had a median of +8.1% and a range of -47.9% to +206.7%; target-only nominal extra had a median of +11.0% and a range of -74.7% to +1059.1%. The wide spread is why request count is the stable product claim—not a fixed token-saving percentage. The six fixed summary workers used 19,551 nominal / 26,463 processed tokens. The 12 target sessions used 74,716 / 131,932. Counting each shared summary once, the clean acceptance components totalled 94,267 / 158,395 tokens. See the [full report](https://github.com/Totoro-qaq/dsh-plugin-bridge/blob/main/reports/v0.2.3-e2e-report.md) and [raw JSON](https://github.com/Totoro-qaq/dsh-plugin-bridge/blob/main/reports/v0.2.3-e2e-2026-08-20T13-19-13-924Z.raw.json).In short: the safe default deliberately spends a separate confirmation turn; --continue combines confirmation and useful work into one target turn without enabling a background goal round.
Accuracy
The default compression tier is pro. The latest rc.8 release acceptance covered 6 frozen fixtures and 12 target sessions:
| Measure | Result |
|---|---|
| Summary facts | 30/30 |
| Target restatement facts | 60/60 |
| First useful work facts | 60/60 |
| Critical facts | 90/90 |
| Obsolete-value resurrection | 0 |
| Exact confirm / continue request count | 6/6 · 6/6 |
This is a small, repair-driven release gate—not a statistical accuracy guarantee. It includes three 21-message sources with real compaction reuse plus three short sources, across minimal, standard, and code targets. Full methodology, raw token deltas, variability, and archive evidence are in the v0.2.3 baseline + fix report.
Earlier compression-tier benchmark
The earlier August 2026 benchmark remains useful for comparing compression tiers: | Measure | Result | |---|---:| | Summary fidelity | 97.5% test / 96.7% validation | | Migrated fact usability | 87.5% test / 83.3% validation | | `pro` fact usability (8 runs) | **95%** | | `flash` fact usability (8 runs) | 80%, including one total-loss run | | Five-section schema compliance | 100% | `pro` and `flash` cost nearly the same here; the reason for defaulting to `pro` is lower failure variance, not a statistically proven mean advantage. Numbers and ports remain the main drift risk, which is why preview and restatement are part of the product path. See [the earlier benchmark, A/B control, and known weaknesses](https://github.com/Totoro-qaq/dsh-plugin-bridge/blob/main/docs/benchmark.md).How it works
fold history → generate five-part summary → preview/edit → create target session
→ pause stored goal → inject summary → restate (and optionally continue)
image history → copy associated assistant text verbatim; unresolved raw images use rc.8 capability fallbackThe five sections are Goal, Current state, Key decisions & conventions, Key files, and Next step. Tool traces from the old preset are intentionally dropped: Bridge moves state, not incompatible tool history.
The original session is never rewritten. If the migration is unsatisfactory, return to it and archive the new session.
Images in rc.8: raw transfer vs verbatim evidence
Bridge uses an automatic, accuracy-first image policy: - An image-only user message never disappears from source material. - If the source turn has associated assistant text, Bridge appends that text **verbatim** under `Visual evidence`; the summary worker cannot rewrite it. The preview deliberately calls it an associated response—not proof that every pixel was understood. - If no assistant text follows the image, Bridge marks it `Unresolved`. On rc.8 it reads the durable source attachment and tries to include the raw image in the target kickoff. - An image-capable target receives the raw image plus the handoff. A text-only DeepSeek target rejects the image during host admission, before a message or model request is created; Bridge then sends the text handoff with an explicit unresolved warning. - Raw images are not resent when verbatim evidence already exists, avoiding unnecessary visual-token and cache cost. Reattach the source image when the preserved response is insufficient for the next step. The normal five-part summary remains bounded by `summaryCharBudget` (2,400 characters by default). Verbatim visual evidence has a separate 60,000-character budget and is admitted as whole blocks: Bridge may omit an older block with a visible warning, but never cuts an image-derived response in the middle. Copying existing text adds no model round; a successfully attached raw image is priced by the selected vision provider. rc.6/rc.7 remain compatible through the text path. Raw attachment recovery is an optional rc.8 gateway capability and does not become a required `/bridge --doctor` method.Why not switch the preset in place?
A preset is a system prompt, tool set, and plugin composition—not a tone setting. Tool calls already stored in the session are only valid under the composition that produced them. DeepSeek Harness therefore allows `agentPreset.select` only while a session is blank; changing it later would leave ghost tool calls that the new preset cannot execute. Bridge respects that boundary by opening a clean target session and carrying only a bounded handoff. Model and thinking-effort changes remain separate and can still happen inside a session.Official preset factory, @ session references, and TotoroPilot
- The official Agent Preset factory creates or configures presets for blank/future sessions. Bridge moves existing work into one of them.
- rc.8's `@` reference attaches a bounded read-only snapshot of another session to the current session. Bridge creates a new session and removes old-preset tool traces.
- **TotoroPilot** exposes the same migration pipeline through a GUI modal. The plugin itself remains usable directly in the official WebUI.
Compatibility and limits
Verified against DeepSeek Harness 0.1.0-rc.6, rc.7, and rc.8, Node.js 22 and 24. Run /bridge --doctor after a Harness upgrade; it names missing gateway methods instead of failing vaguely.
- The official WebUI currently needs one restart after install and cannot let a plugin navigate to the session it creates. Bridge prints the exact title and session ID.
- Default DeepSeek text models still cannot inspect images. Bridge preserves prior visual evidence and fails visibly when no reusable rc.8 attachment is available; it does not run a hidden local vision model.
- Preview normally takes 20–60 seconds and is bounded by
previewTimeoutMs. - Release acceptance now covers real compaction reuse, but only one run per cell; it should not be read as a population guarantee.
- The older tier-comparison benchmark predates prompt+goal dual injection and remains labeled as historical evidence.
For the full walkthrough, rollback checklist, configuration table, CLI path, and FAQ, read the Chinese usage guide.
Advanced CLI and configuration
```bash dsh-bridge doctor dsh-bridge preview --to code --sessionDevelopment
npm test # build + typecheck + 120 tests
npm run pack:checkCI verifies Node 22/24, generated lib/, datasets, package contents, and the narrow upstream RPC/command contract. Tests use a fake host and spend no model tokens. Evaluation is separate and uses your own local model credentials; see docs/benchmark.md.
License
MIT