DSH ClinePass

ClinePass plugin for DeepSeek Harness, maintained in this repository: xiangnan0811/dsh-clinepass.

中文说明 · Русский

The settings namespace is dsh-clinebot, the provider id is clinebot, and the HTTP prefix is /dsh-clinebot. Do not install another plugin that also registers provider id clinebot in the same profile. Both would write llm-pi-ai.providers.clinebot.

The package name is dsh-clinepass. private: true blocks npm publish.

Install from this repository

Version 0.8.0 requires DeepSeek Harness ^0.2.1-rc.2.

DeepSeek Harness installs plugins with pnpm. A GitHub spec is enough:

dsh plugin --profile web add github:xiangnan0811/dsh-clinepass

Restart dsh. ClineBot is in the settings sidebar. The sidebar Plugins page also opens it from the dsh-clinebot row inside dsh-clinepass. Settings → Built-in plugins is a read-only inventory and no longer hosts this card.

npm publish is not required for the DSH plugin market in the sidebar (dsh-market). That market installs with dsh plugin --profile web add github:owner/repo. Publishing to npm is only needed for dsh plugin --profile web add <npm-name>.

This repository is not a GitHub fork. The market lists it after the topic dsh-plugin is on the About page and cordis.patch.yml is in the root. The index runs around 06:00 and still checks that the repo is a DSH plugin. A submission issue on 2BingLing/dsh-market is the other way in.

Do not export CLINEBOT_API_KEY in the shell that starts dsh. DSH treats that variable as read-only and rejects a key saved from the settings page. Save the key in the plugin card. The browser calls http://127.0.0.1:<port>/dsh-clinebot/... on the local DSH server. The plugin then calls baseUrl, which should be https://api.cline.bot/api/v1.

What the settings card does

• Saves the API address and the ClinePass key.
• Shows 5-hour, weekly, and monthly usage from GET /users/me/plan/usage-limits, with the percent, a meter, and a reset date. Account, plan, and last fetch time are separate rows. The account email is masked until you choose Show email. A cancellation timestamp from the plan is an extra row. When more than one key is saved, the other accounts are listed with their three percents. That list does not change which account is chosen after a rate limit.
• Registers the selected models into DSH. The output cap is not sent as max_tokens. The request budget is. The default budget is 65536.
• Adds selection shortcuts, search, optional verified-plan filtering, batch defaults and custom model forms. An explicit empty selection registers none. Individual overrides take precedence over batch defaults. Restoring model capabilities leaves the request budget unchanged; it has its own reset.
• Validates an entered key without saving, or scans known Cline CLI, VS Code, Cursor files and Cline environment variables when requested. Discovery shows masked values; five-minute server handles import selected keys into the account pool. Encrypted editor secrets may not be discoverable.
• Lets you pick the smoke-test model. Muse Spark uses reasoning_effort: low and a 1024-token cap so the short reply is not swallowed by reasoning.
• Saves request statistics per DSH profile: results, attempts, latency and reported token counters. Filter by UTC date, model or account, export CSV, or reset statistics from the settings card. /cline stats and /cline telemetry show the same profile totals. Prompts and credential values are never stored.
• A sidebar quota button opens the host's detail modal with quotas, statistics and full configuration, including a narrow-screen entry. Reset countdowns keep missing times and usage unknown. Sidebar and composer displays default on and have separate switches.
• Composer usage comes from the current session's ClinePass reply/attempt projection, excluding other providers. Cost estimates price uncached input, output, cache-read and cache-write tokens per million, using a manual rate, a bound reference or an automatically matched models.dev USD reference, in that order. USD is the default unit; changing currency performs no conversion, and other currencies need manual rates. A missing cache price uses the input price; unreported counters count as zero and models without a price are left out. The ClinePass subscription does not bill per request, so this is only a rough reference. Reasoning is already within output. Missing counters remain unknown in the token totals, including in native session replay. Estimates are references at current rates, not historical bills.
• Streaming probes allow cancellation, request at most 1024 output tokens and expire after 30 seconds. They measure first text/reasoning/tool output and total duration, and count in profile statistics rather than session reply usage. /cline ping <model> or /cline ping default runs a probe; bare /cline ping retains API health checks. /cline account lists accounts, /cline use <name> pins one, and /cline use auto restores automatic routing.
• Saves extra CLINEBOT_API_KEY_* accounts on the same card. The pool name is saved before the secret. If the secret is not stored, that new name is removed. A direct DSH turn with HTTP 429 or confirmed account quota exhaustion selects another account for later requests. HTTP 402 reports insufficient credits for this model and leaves the global account unchanged. Switching, that rotation, a new API address, a new primary credential name, or replacing the active key drops the previous plan model list until the next successful plan read. Replacing the active key also removes the saved plan cache, so the old list is not restored from disk. A plan read that has already started cannot put that list back. /cline prints plain text.

As checked on 2026-10-26, Cline exposes public model lists at /ai/cline/recommended-models and /ai/cline/models. The first lists ClinePass IDs; the second supplies API names and reference capabilities. /models now works and serves as the ID-only fallback. Authenticated Sync plan models still reads /users/me/plan separately. The default smoke model remains cline-pass/deepseek-v4.1-flash.

The current default selection includes GLM 5.3 / 5.3 Flash, Kimi K3.1 Pro, MiniMax M3.1 Flash, MiMo V2.6 Flash / Pro, Qwen 3.8 Max, DeepSeek V4.1 Flash (stable and 1017), DeepSeek V4 Pro, and Muse Spark 1.3 Contributor. Older generations remain editable but are no longer implicitly enabled. Explicit selections and a saved legacy default remain usable; IDs, budgets and upstream pins are not rewritten. In all-except-disabled mode, explicitModels also retains legacy opt-ins. Enable all includes old generations; Use current defaults clears opt-ins and disabled IDs, retaining a saved default model.

Under Custom models → Get models from API, search and select a model to fill its exact ID, display name and published context/output/image parameters. Only the configured API receives the active credential. Ordinary API IDs are shown separately because they do not use the ClinePass subscription lane. For an already listed Pass ID, a unique exact vendor-model suffix can supply reference parameters; ambiguous matches stay unknown and never create a Pass ID. Built-in models can be enabled directly. Saving a custom model also selects it, including from an explicit empty selection, and preserves it across account changes. The verified-plan filter still applies and is explained beside the selection. Manual entry needs only an ID; name and advanced parameters are optional. Blank parameters inherit batch/default settings; unknown limits remain unknown. The prefilled request budget is at most 65536 and is capped by any published lower context/output ceiling. Reasoning levels are never inferred from a generic reasoning flag. Discovery makes no inference calls and does not verify plan access.

Version 0.8.0 targets DeepSeek Harness 0.2.1-rc.2 (next on 2026-10-19). Earlier 0.1.7-rc.2 validation applies to previous versions. The settings sidebar entry is settings.section dsh-clinebot. The same page is plugins.row.config at dsh-clinepass#dsh-clinebot.

On hosts providing connection.requestRejection, all /dsh-clinebot/* routes require DSH authentication, including reads and update checks. Open the URL printed by dsh web to sign in. Unauthenticated requests return 401; host or origin rejection returns 403. Settings writes also retain their origin checks. Writes require Origin or Referer. Fetch Metadata alone is not authorization; forwarded hosts are trusted only from a loopback proxy, and DSH authentication is still required.

Version 0.8.0 has also passed isolated DSH 0.2.1-rc.3 Web and direct headless fixture checks. Headless disables the managed credentials service by default; this plugin requires that service. In the headless profile patch, enable {id: credentials, disabled: false} and set agent-default-model to provider clinebot and an enabled model. Launch-environment keys still work. The optional upstream TUI settings integration is deferred because the official host has no verified tuiSettingsSections consumer.

Configuration

Catalog synchronization starts at about 0.5 seconds, with conditional retries at 3 and 15 seconds. An hourly check refreshes results older than six hours. Manual sync uses the same writer and freshness timestamp. Failures retain valid data; disable/unload cancels reads and invalidates late writes. Quota polling does not fetch the full catalog. Cache v4 isolates the API address, account and credential fingerprint; older caches are refetched. A failed cache save shows a warning while settings and model calls remain usable. Custom modelsCachePath writes must stay inside DSH_HOME as a safe .json file (profiles/ entries must start with clinebot-). Invalid or failed image_url blocks in chat payloads are downgraded to plain text blocks before upstream calls in both proxy and direct modes.

FREE badges/filtering use exact IDs from the gateway's explicit recommendation free list, verified on 2026-10-29. Names and zero reference prices are not free evidence; directory metadata does not verify account access or billing. Dated models use exact listed IDs and version-specific metadata, without rewriting vendor IDs to Pass IDs, extending verified plan membership or importing models.

The proxy cools HTTP 402 for five minutes per account and exact model, and can try another account before the first output. Other models on that account remain available. Recovery can target one model; account-only API requests still clear all its cooldowns. Direct calls, probes and smoke tests do not rotate the global account on 402. HTTP 413 and recognized request-size errors fail without retry or account rotation, with advice to reduce messages/images or start a new session. Context token overflow remains separate. New statistics distinguish insufficient credits, rate limits and oversized requests; historical aggregates are not reclassified.

dsh-clinebot:
  enabled: true
  baseUrl: https://api.cline.bot/api/v1
  apiKeyEnv: CLINEBOT_API_KEY
  defaultModel: cline-pass/deepseek-v4.1-flash
  selectionKind: all-except-disabled
  timeoutMs: 15000
  smokeTimeoutMs: 25000

Field	Default	Role
baseUrl	https://api.cline.bot/api/v1	API the plugin calls
apiKeyEnv	CLINEBOT_API_KEY	Credential name
defaultModel	cline-pass/deepseek-v4.1-flash	Smoke test and provider default
selectionKind	all-except-disabled	Enables current models plus explicit legacy opt-ins, excluding disabledModels; explicit mode registers only explicitModels
planFilterEnabled	false	Intersects the saved manual selection with a successfully read plan; missing plan data does not clear the selection
batchModelDefaults	inherit	Context, output capability and request budget defaults, below individual model overrides
proxyEnabled	false	Authenticated loopback proxy with account retries before the first model output
accountMode	least-used	Lowest measured quota first, or round-robin; unknown quota stays unknown
modelRoutes	[]	Explicit model/account routes, ahead of manual pin, session binding and strategy
upstreamPins	[]	Strict per-model ClinePass backend provider pins; separate from account selection
historyImageOffload	false	Ask the native host to log and replay historical image offload decisions
historyImagesToKeep	2	Historical retained images; the latest user/tool input is retained
timeoutMs	15000	Usage and plan requests
smokeTimeoutMs	25000	Smoke test
webSearchEnabled	false	Register native search provider clinebot without changing global selection
searchTimeoutMs	15000	Search deadline, including the response body
statsRetentionDays	95	Minute aggregates retained for 0–3650 days; 0 is unlimited; daily totals remain
quotaForecastEnabled	false	Estimate quota pace from comparable samples at least a minute apart
priceAutoSync	false	Refresh public reference prices on model synchronization, at most once per six hours
priceBindings	[]	Explicit {model, provider, sourceModel} reference-price choices; manual costRates take priority
dailyBudget, monthlyBudget	0	Optional UTC day/month estimated-cost limits in costCurrency; 0 disables each limit

Checked models are what Settings → Models lists. Per-model overrides for context, output cap, request budget, input, and reasoning are edited on the card. Empty override fields inherit batch defaults, then the catalog. Native Models edits are captured in hostModelOverrides and survive hiding models and restarts. An explicit edit or restore on this card clears the corresponding captured field. Per-model default effort is applied before the native request freezes; explicit caller choices, including Off, take precedence. On the next LLM service, only clinebot transfers to the ClinePass native adapter so statistics can observe actual provider usage. Its model parameters remain editable in Settings → Models. The adapter stays registered when proxy mode returns to direct mode.

Contradictory token counts are excluded from measurements and cost estimates; the model reply and request outcome remain intact. Session projection caches are rebuilt from the existing log under these rules; missing historical wire data is not reconstructed.

To keep a model on one backend, use the dedicated Pin upstream provider section in settings (also available in the model editor). Select a model and set the provider to an ID such as deepseek, then save the upstream pin. Leave it blank or choose Use automatic routing to clear it. If the selected model is removed, choose a model again before saving; the form does not silently select another model. For configuration files, use:

upstreamPins:
  - model: cline-pass/deepseek-v4.1-flash-1017
    provider: deepseek

Version 0.7.2 adds DeepSeek V4.1 Flash 1017 as a separate model. Existing model IDs, defaults, session selections and pins are not migrated. The dated ID appears in Cline's recommended catalog (2026-10-25). Vision, reasoning, 1M context and 384K output capabilities follow DeepSeek's vendor documentation, not measured ClinePass limits. The default request budget remains 65536.

Stable and dated IDs can use different gateway routes. Enable the dated model, save a separate deepseek pin for its full ID, then explicitly select it in the streaming probe. Changing the smoke-test default does not switch existing sessions; select the snapshot in each session when ready. Verified-plan filtering can hide it if the account's discovered plan list omits it. Gateway enforcement still needs account-specific verification; no automatic model rewrite or probe is performed.

Probes and request history expose response routing details when reported: response ID, returned model, canonical gateway model, resolved provider and the DeepSeek metadata fingerprint. These bounded identifiers aid diagnosis but never replace the final-provider check or make an anonymous private provider match deepseek. Full planning text, headers and content are excluded; old records are not inferred. Smoke statistics retain the requested model; the returned model is a separate field.

Pins apply while this plugin is enabled to native direct/proxy chat, auxiliary model calls, streaming probes and smoke tests. Direct calls freeze the pin when prepared; proxy calls use the configuration captured at proxy ingress. Account retries retain the same provider restriction. No provider fallback is requested. The request carries both provider.only (OpenRouter) and providerOptions.gateway.only (gateway planner); zai/z-ai use each router's spelling. Unknown models/providers are not automatically tested or discovered.

For cline-pass/deepseek-v4.1-flash-1017 pinned to deepseek, the plugin also adds vision and implicit-caching to providerOptions.gateway.has, retaining additional caller capability filters without duplicates. Missing, null or empty filters receive both tags. This keeps the ClinePass model ID and credential; no model selection or saved configuration is rewritten. On 2026-10-26, only alone reached an anonymous private candidate, while this combination reached DeepSeek. The stable ID still reached the private route with both tags, so the workaround applies only to the dated ID and this pin. Other pins and automatic routing are unchanged. Extra capability conditions can leave no matching candidate.

Use Run streaming probe to compare the requested and reported provider; it makes a real request. Only structured finalProvider or direct provider metadata counts as confirmation. A different provider fails with UPSTREAM_PIN_MISMATCH; missing metadata fails with UPSTREAM_PIN_UNVERIFIED. Some accounts ignore routing restrictions (including private routes); HTTP 200 alone does not prove a pin. Metadata can arrive after output has begun: a late failure ends the call without replay and cannot undo output or charges already incurred. Pinning cannot ensure cache hits or override the service's routing policy. Small text requests on one live ClinePass account verified the dated-model combination on 2026-10-26; this does not verify images, long context, quota accounting or future routing. Local fixtures exercise the installed DSH adapter and failure paths. The wire protocol follows the public findings in cline-pass-switcher and dsh-cline-pass.

Ordinary chat shows the last response's reported upstream in the composer. Expand it for this session's retained requests; profile statistics show the latest 10 requests (expandable to 50) of the 100 retained, including failures, automatic routing and individual account attempts. Each record persists the requested provider, whether a restriction was attached, the reported provider and its observation: matched this time, mismatch, unverified, not attached, automatic, or unrecorded. Attaching a restriction does not prove delivery; a matching response does not prove future enforcement. The model badge only describes the requested provider. Keys and chat content are not recorded. Session records refresh on settlement, opening details or manual refresh, and are filtered strictly to this session within those 100 profile requests. Missing evidence never falls back to another session or current settings. CSV remains an aggregate export; per-request routing is available in the stats API's recent.

Proxy credentials are ephemeral references in DSH credentials. Disabling proxy mode restores the upstream address, account reference and original retry/timeout controls. Proxy requests use their own retry policy; direct requests retain the existing next-turn rotation. Connect, first-output and JSON-body deadlines default to 15 seconds, stream idle to 30 seconds, total to 300 seconds, and the complete request body to 32 MiB. Edit these under Proxy and account routing. Native image size limits are separate; historical offload requires the host image-offload plugin and keeps original session attachments. The proxy accepts exact IDs from the plugin model list; add a new model before calling it. See the validation record for observed coverage.

Statistics use <current profile directory>/clinepass/stats.json, with atomic replacement and the host's file lock for concurrent writers. Daily/model/account aggregates persist; only the latest 100 individual records are retained. A request retried by the proxy counts once, with each upstream attempt recorded separately. Request/token aggregates belong to the final account; failed attempts retain their own account aggregates. CSV scope=request|attempt distinguishes those rows: do not add both scopes together as logical requests. Missing usage is unknown, cache input is disjoint from uncached input, and reasoning is already included in output. Counts show measured/all request samples. Latencies are mean completion and first-output times; non-streaming smoke tests have no TTFT. These are new measurements, not reconstructed history or billing estimates. Storage errors remain visible while model calls continue; unsaved measurements remain in memory. Reset clears the entire profile's statistics, including filtered dates, and archives a corrupt file before replacing it. Accounts and models remain.

The statistics card has one filter bar: a range (all, 24 hours, 7/30/90 days or a custom local minute range, end exclusive), model, account and a failures-only toggle, all applied immediately. Results show request, success, token, reference cost and first-output KPIs; token composition with the cache hit rate (cache read ÷ all input); P50/P95 histogram bounds and mean latency; a weekday/hour heatmap using the browser's current UTC offset; and a date/model/account breakdown with token shares. Fields reported by fewer than 95% of requests are marked. CSV and JSON export use the active filters. Old daily aggregates remain intact; they acquire neither invented minute detail nor percentile samples. Minute/failure filters include retained detail only. Account history keeps the last 200 routing, cooldown, recovery and selection events independently of statistics reset. Additional accounts can be renamed or have their saved key cleared while retaining their entry and pins; environment keys must be removed at their source.

Under Usage & rates or from the cost panel, public USD prices from models.dev can be synchronized. The profile-local cache is clinepass/prices.json; failed refreshes retain the last successful catalog and show an error. The timestamp stays visible. Statistics and session estimates match a reference automatically by model name (ClinePass's own listing first, then the official vendor, then other providers; a trailing -MMDD date is tried without it). A binding to an explicit provider/model overrides that match, and manual rows override both, including incomplete manual rows. Other currencies use manual rates only.

Daily/monthly budgets compare the same reference estimate with the limit before a new native, proxy, smoke or probe call and block only once it is reached; missing prices, usage or statistics count as zero. In-flight calls can exceed a limit; these estimates are not bills or subscription quotas. Resetting statistics also resets budget usage. Changing currency reinterprets manual rates and limits without conversion. /cline cost and /cline budget show the same estimates. Optional quota pace estimates require actual comparable samples from the same account/window and clear on errors, resets or account changes; they do not route requests.

The official Cline code declares provider-executed search for ClinePass models. This does not establish a standalone /alpha/web-search endpoint for DSH. That adapter remains off by default under ClinePass standalone search · experimental, with live channel access unverified. It uses the active account and baseUrl; enable it only for a gateway confirmed to support that endpoint, then select clinebot in the host's web search settings. Existing search/fetch providers are preserved. Each API request asks for at most 10 results (5 if no bound was supplied). Genuine empty results are distinct from HTTP, credential, malformed-response, timeout and cancellation errors. It never falls back to another website or changes the account pin. Disabling or unloading unregisters the provider and cancels pending searches. Isolated next and local mock acceptance have passed; live channel availability remains unverified.

DeepSeek provides independent Search and Fetch APIs. Configure the official provider and a DeepSeek key in the host web settings. Each costs CNY 5 per 1,000 successful requests with results; failed, empty or timed-out requests are not charged. Server-side tool calls start billing at 00:00 Beijing time on 2026-11-01. These costs are separate from ClinePass token estimates. See the official guide.

On DSH next 0.2.1-rc.2, actual plugin disable cancels pending calls, closes the proxy, removes its temporary credential and restores the built-in model adapter. Saved model/account configuration remains available. Re-enabling and hot updates automatically synchronize the active adapter and credential; settings and persisted statistics are retained. These paths have passed isolated CLI Web acceptance; Desktop remains unverified.

Settings shows only the running plugin version and repository link. Updates run from Plugins → ClinePass updates. GitHub targets always use this repository and a checked commit. Local directories, archives and registry installs require an explicit target; local updates accept an absolute .tgz/.tar.gz produced by npm pack. The host inspects and installs the target, owns the profile lock and cancellation, and reports failure or a required restart. Accounts, model settings and statistics are retained.

Running and installed versions are tracked separately, including same-version updates. A matching archive path/version cannot prove matching contents; an unrecorded Git commit remains unverified. Follow the host's restart/refresh prompt and check again. There is no background installation. Archive checks read metadata without extracting files (16 MiB compressed, 64 MiB total entries, 256 KiB manifest). GET /dsh-clinebot/update checks status; installation uses the native dialog, and the former HTTP installation action returns 405.

Develop

npm run build:client
npm test

npm test checks that lib/client.js matches its source without rewriting it, runs the undefined-name check, then the suite. npm pack builds a local tarball. Install that tarball with dsh plugin --profile web add ./<file>.tgz when you are not installing from GitHub.

License

MIT. The copyright notice is in LICENSE.