SekaiSync Connect for DeepSeek Harness

English | 中文

Overview · Quick start · Configuration · Plugins panel · Tools · Usage · More information

A DeepSeek Harness direct-connect module for the SekaiSync knowledge base: zero dependencies, zero build steps, and minimal token usage — plus a Plugins-panel page for choosing the deployment path.

Installation

Requires Python ≥ 3.10 with a working python -m sekaisync (no third-party dependencies), plus a synced store (a sekaisync init / sync product).

# Install from Git:
dsh plugin --profile web add github:omoinoki/dsh-sekaisync-connect

# …or from a local checkout:
dsh plugin --profile web add ./dsh-sekaisync-connect

Project home: https://github.com/omoinoki/dsh-sekaisync-connect

Runtime	Verdict
0.1.x	DENY — the panel API this plugin needs does not exist before 0.2.0-rc.1
0.2.0-rc.1	allow (verified: tools, panel, and a real lookup all run on this runtime)
0.2.x	allow
0.3.0 / 0.3.0-rc.1 / 1.0.0	DENY — unverified versions fail closed

Anything outside the allowed range is refused rather than half-loaded, so the worst case is a clear "incompatible" notice instead of a crash. To run an unverified version anyway, use dsh plugin allow-version.

If a runtime upgrade ever disables this plugin, the symptom is that the tools disappear entirely and the panel stops loading — the whole bundle is rejected before any of its code runs.

Configuration

Precedence: environment variables > profile row config > SEKAISYNC_CONFIG file > plugin-directory config.local.json > config.json > automatic discovery.

Setting	Description
SEKAISYNC_STORE	Store directory containing kb/. When omitted, discovery checks <SEKAISYNC_ROOT>/store, cwd/store, and the repository store located with python -c "import sekaisync"
SEKAISYNC_ROOT	SekaiSync repository root, used as the working directory for python -m sekaisync
SEKAISYNC_PYTHON	Python executable; defaults to python
SEKAISYNC_PORT	Port to probe for an existing server; defaults to 8787. If a ready SekaiSync service is already listening there, it is reused instead of starting another process. 0 retains its meaning: skip the external probe and start a managed service directly
SEKAISYNC_MAX_RESPONSE_BYTES	Byte budget for one successful HTTP response; defaults to 134217728 (128 MiB), configurable from 65536–1073741824. It can also be set as maxResponseBytes in a configuration file or profile row. Exceeding the budget cancels the read and returns an explicit error

You can also override these settings in your own profile patch; plugin upgrades do not overwrite that layer:

- id: sekaisync-connect
  name: 'dsh-sekaisync-connect'
  config:
    store: 'D:\\sekaisync\\store'
    python: 'py'

Keep the row id as sekaisync-connect. It is the key the panel writes through, so if you copy this snippet into a profile that still uses an older id, update that row's id to match — otherwise the panel and the saved path refer to different rows.

• HTTP connects only to loopback and rejects redirects. Health and error responses are limited to 64 KiB and 4 KiB respectively. The default successful-response budget is 128 MiB, covering typical body results with limit=100 and max_text_chars=200000. max_text_chars=0 still means full text upstream; if an exceptionally large full-text response exceeds the transfer budget, increase maxResponseBytes (up to 1 GiB) or read in batches.
• External-port reuse still follows the existing /health ready/status check. That endpoint carries no store identifier, so you must ensure the configured port serves the intended knowledge base.

Choosing the deployment path in the Plugins panel

When a Web profile serves the client, the plugin also appears as a Configure page on its row under Plugins in the sidebar. That page selects the SekaiSync deployment from the GUI instead of editing JSON by hand.

Open Plugins → Installed → dsh-sekaisync-connect → the row's Configure control. The page offers:

Control	What it does
Path field + Check	Classifies the path and reports whether it is a store, a repository root, or a kb/ directory — plus database size, kb/ entry count, and the freshness payload when present
Save and apply	Validates the path, then persists it through the official settings service into the profile's cordis.patch.yml and applies it live — the effective store / root readout updates immediately, no restart required
Test connection	Performs a real /health request against the selected deployment and reports latency and readiness

The Effective store / Effective root readout tracks the path field as you type (debounced), and shows a Not saved yet marker whenever the typed path differs from the saved one — so a change is visible before you commit to it.

The panel reads and writes the plugin's own Cordis Config (store / root), so a change survives upgrades and takes effect without restarting DSH. The value shown is what the runtime resolves after the full precedence chain: environment variables > profile row config > SEKAISYNC_CONFIG > config.local.json > config.json > automatic discovery.

Constraints, by design:

• The panel writes only store and root, through the official settings service, and only after the path passes classification. python is deliberately not editable from the panel: exposing an executable path over HTTP would put arbitrary program execution on a web page. The worst a panel write can do is point the knowledge base at another directory.
• The panel exposes no auto-detection, folder chooser, or directory browser. Those actions need host-side directory-level permissions that a plugin row cannot rely on, so they were removed rather than left as buttons that do nothing. Type the path instead; Check classifies it and tells you what it found.
• The panel backend is a Typert Remote service (namespace sekaisync) registered in the profile row's own fiber, mirroring the official dsh-experimental-voice-input-bundle/dsh-api-settings-controller pattern. Writes go through settings.update, which the framework fences with the same Host/Origin and cookie authentication as every /api surface, and which serializes writes and validates values before persistence. A profile without a settings service (a pure CLI composition) simply never mounts the panel, while the 10 tools keep working.
• The panel is an addition, not a requirement. Without it, the file-based configuration above works exactly as before.

Tools

Ten model-facing tools, all read-only against the local knowledge base:

Tool	Cost	Purpose
sekai_probe	~1 ms	Decide whether a topic is Project Sekai at all; returns verdict plus a suggested next tool
sekai_lookup	~2.3 s	Match entities (characters, events, cards, gachas, songs, areas) by name in any language, with cross-region names and event-shorthand resolution
sekai_fact	~5 ms	Compact fact pack for one entity id (e.g. character:1, card:123) — the cheapest way to get structured facts
sekai_resolve	~1 s	Resolve a proper noun to its official localized name; reports translation_status honestly instead of leaving gaps
sekai_term	~150 ms	In-game terminology and cross-language glosses, with tags, weight, and evidence lines
sekai_penetrate	40–100 s	Cross-language penetration of one term at one story point, aligned per language
sekai_alias	~1.3 s	Resolve event shorthand (khn3, wl3) and official event names, including natural-language questions
sekai_web	40–160 s	Full-text search over crawled story text; use when quoting original lines
sekai_news	~80 ms	Official announcements across the five regions, filterable by language, category, and body availability
sekai_status	6–33 s	Knowledge-base readiness, data freshness, per-region coverage, and sync rate

Usage Recommendations (Token Discipline)

1. Use sekai_resolve before generating any localized text. Use sekai_fact for precise facts; do not let the model answer from memory.
2. For an unfamiliar term, an uncertain topic, or a character name with an unknown source, first use sekai_probe to determine whether it concerns Project Sekai, then route to local tools or the web. Matches directly suggest a next_tool.
3. For time-sensitive questions such as current events, gachas, and maintenance, check freshness and coverage with sekai_status first.
4. To verify a localized name in context, use sekai_term to confirm the term exists, then sekai_penetrate to inspect aligned lines across languages. Remember that missing explicitly means there is no corresponding line in that language; it is not an error.
5. sekai_web is slow; use it only when quoting original story text. Try sekai_lookup / sekai_term first, or use sekai_probe to confirm the direction.
6. Ask directly with event shorthand or an official event name, whether a unit-focused event or World Link. sekai_lookup automatically includes activity resolution, and sekai_alias supports shorthand, official event names, and natural-language questions.
7. If a result is truncated, reduce limit or use a more precise query.

Troubleshooting

• ERROR: 未找到 sekaisync 知识库 store: set SEKAISYNC_STORE or edit config.json.
• ERROR: 无法启动 python …: confirm that python -m sekaisync --help works. If the package is missing, run pip install -e <sekaisync 仓库>.
• Server 60-second cooldown: automatic restarts pause after 2 consecutive crashes. Check the integrity of the store with python -m sekaisync --no-event-check integrity.
• ERROR: HTTP … in a tool result: the server has started but the request failed, usually because of a parameter issue. sekai_status reports the service mode and store path.
• ERROR: 缺少必填参数 …: the plugin's guard caught a missing argument. Supply the requested argument and retry.
• The row's Configure page is missing: the panel needs a Web profile with the plugin-manager UI and the settings service (a pure CLI composition has neither). Confirm ui-plugin-manager is in the profile, that the bundle is switched on, and that the deployment path resolves (see the first item above).
• Save and apply reports success but the old path stays in effect: a higher layer (environment variable or profile row config) fixes store. The page names the effective value; change the higher layer instead.
• forbidden or unauthorized from the panel: the request did not pass the framework's Host/Origin and cookie fence. The panel's settings-backed writes are intentionally restricted to the authenticated local client.
• If sekai_probe reports 词表来源=static, the dynamic lexicon (terms export) has not finished warming or its build failed. The probe has fallen back to the built-in static lexicon; retry later or check store integrity.