@local/math-agent-preset
Math_Agent：适用于 DeepSeek Harness 的数学训练/辅导预设，从不替受训者解题。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:RepInS-01/math-agent#ed81061ecf5aff443a95b8eb2c601dcd75697c95Math Agent Preset 以 @local/math-agent-preset 发布，当前版本为 1.0.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
A math training preset for DeepSeek Harness — not a problem solver.
Rather than solving problems for the trainee, it provides rapid feedback to help sharpen mathematical understanding and build intuition. It features a programmatic zero-leak output guard that mechanically prevents the coach from ever revealing the answer.
attempt_update tool (status: in-progress / flawed / validated / incomplete), rendered into the system prompt every step, and rebuilt from the session log after a resume. The coach is nudged in the prompt after several rounds without an update. For longer sessions, attempts can additionally be persisted to notes/attempts.md.hint_log tool and counted in the system prompt — a rising count is a signal to slow down. At session start the coach asks the trainee for the hint policy via ask_user_question (strict = L0/L1 only; normal = full ladder) and records it with the hint_policy tool — per-session, resumable, switchable mid-session. The strictness value in cordis.patch.yml is only the default for sessions that never chose (ships as strict); under a strict policy hint_log rejects L2/L3 programmatically.validated through attempt_update, never on the model's own say-so.llm/stream that inspects the coach's entire reply. When answer clues (numerical values, intervals, correctness judgments, answer forms) are detected, the entire response is replaced — not a single character of the model's original output reaches the trainee. This does not rely on the model's "self-discipline."knowledge-base skill). During final synthesis, understandings backed by facts and data are prioritized.math-agent/
├── package.json # Bundle manifest: declares `dsh.bundle.patch` → cordis.patch.yml
├── cordis.patch.yml # Bundle patch: the `preset-math-agent` declaration — display metadata + full Cordis plugin list (tools, persona, skills, guard registration)
├── attempt-tracker.js # Training-state plugin: attempt_update + hint_log + hint_policy tools, prompt injection, validated flag, stale-update nudge (folds from the session log)
├── zero-leak-guard.js # Programmatic zero-leak output guard plugin (addressed through the profile's node_modules symlink — see the RESOLUTION ANCHOR note in cordis.patch.yml)
├── zero-leak-guard.test.mjs # Test corpus: leak samples that must block, clean replies that must pass, tracker + unlock behavior
├── README.md # This file
├── LICENSE # MIT License
└── skills/
├── coaching-protocol/ # Coaching protocol: workflow, red lines, zero-leak rules
├── knowledge-base/ # Knowledge base integration interface (reserved)
└── final-synthesis/ # Final synthesis workflow and output structure
Prerequisite: DeepSeek Harness version 0.2.1 or newer — either the Desktop app or the Web UI (npx @deepseek-ai/dsh web). This preset is derived from the standard preset and ships as a DSH bundle (package.json + cordis.patch.yml). (The previous commit supports 0.1.7; 0.2.1 added dsh-tool-schedule and dsh-plugin-market, which this preset's plugin list now references.)
DSH versions before 0.1.7 used a different mechanism (a
$DSH_HOME/.agent-presets/<id>/directory holdingpreset.yml+agent.cordis.yml). Nothing reads that directory any more; the bundle form below is the only supported installation.
Install the bundle into your DSH profile, either way:
github:RepInS-01/math-agent. On Desktop, package operations are owned by the app itself — the dsh plugin --profile desktop CLI is refused, so use the Plugins page; orplugin_manager tool), ask the agent to install_bundle with target set to github:RepInS-01/math-agent. Bundle installation executes plugin code in the Host process, so it requires Full access or an approval.Developing on the preset instead? Clone the repository, keep the checkout in place, and install the checkout's absolute path (e.g. /Users/you/math-agent) instead of the github: spec. The installer then adds a link: dependency, so edits to the checkout take effect after a restart:
git clone https://github.com/RepInS-01/math-agent.git
Restart the app (Desktop: quit and reopen; Web: restart dsh web). The Math_Agent preset (id: math-agent) appears in the preset roster. Sessions recorded with the legacy preset resume against this declaration unchanged, since the id is the same.
Mount validation (run after any modification): create a new session and select the Math_Agent preset. A successful selection means the composition mounted; on a mount failure the GUI reverts the selection to the default preset and shows the error. Installed bundles and their row activation state can be inspected with the plugin_manager tool (list_bundles / list_plugins).
In the DeepSeek Harness GUI (Desktop or Web), create a new session and select the Math_Agent preset (id: math-agent).
Start a training session with a prompt like:
I'd like to start a math training exercise. Problem: Let aₙ = √(1 + aₙ₋₁), a₀ = 1. Prove that {aₙ} converges and find its limit. I'm thinking of using monotone convergence but I'm not sure how to prove boundedness.
No matter how insistently you ask for the answer — the zero-leak guard stands between you and the model.
llm/stream waterfall (dispatched process-wide); only activates for requests whose system prompt contains coaching signatures (zero-leak iron rules / math coach). All other sessions pass through untouched.attempt-tracker plugin (mounted in the same isolate realm) holds the per-session attempt list. Once an attempt is recorded as validated via the attempt_update tool, the guard stands down so the final synthesis can deliver the complete solution. A missing tracker, an unknown session, or no validated attempt all fall through to blocking — fail-closed in every direction.tools.guard rejects bash, pwsh, web_search, and run_code until an attempt is validated. The stream inspection cannot see tool output, so the tools that could compute or retrieve the answer are disabled outright — and the denial lifts on the same program state as the final-synthesis unlock.LEAK_PATTERNS at the top of zero-leak-guard.js and can be extended freely.The programmatic guard blocks enumerable leak forms (numeric values, intervals, judgment words) in the coach's own reply text, and the tool gate covers the compute/search channels. One channel remains persona-constrained rather than guard-enforced:
When new variants are discovered, simply add them to LEAK_PATTERNS (plus a sample in zero-leak-guard.test.mjs).
The guard ships with a zero-dependency test corpus (Node's built-in test runner):
node --test zero-leak-guard.test.mjs
Run it after every change to LEAK_PATTERNS, attempt-tracker.js, or the guard's unlock logic. The corpus covers: leak samples that must be blocked, legitimate coaching replies that must pass, the validated-session unlock (and its fail-closed fallbacks), and the attempt tracker's state recording, hint logging (including the strict-policy rejection of L2/L3), stale-update nudge, prompt rendering, and session-log fold. Add every newly discovered leak variant to the block list and every reported false positive to the pass list.
Reload caveat: relative-path preset plugins (
./attempt-tracker.js,./zero-leak-guard.js) are imported through Node's ESM loader with no cache busting. After editing plugin code orcordis.patch.yml, restart the DSH host (Desktop: quit and reopen the app; Web: restartdsh web) — if the composition itself changed, reinstall the bundle from the Plugins page as well. Re-selecting the preset in the UI alone is not enough, as re-mounting reuses the cached module. Skill files (skills/**/SKILL.md) are re-read from disk per session and do not need a restart.
skills/knowledge-base/SKILL.md defines a unified search interface contract (local directory + online retrieval), currently in reserved state:
kb/ directory with Markdown files organized by topic; retrieved via glob + grep.web_search retrieval, cited by source URL.confidence (fact / data / reference / heuristic). Final synthesis prioritizes understandings supported by facts and data.persona row in cordis.patch.yml and skills/coaching-protocol/SKILL.md.LEAK_PATTERNS in zero-leak-guard.js.preset-math-agent row's activation state with the plugin_manager tool (list_plugins).This project was inspired by the math coaching concept originally proposed by Bilibili creator PiKaChu345. The original creator has not yet released their implementation publicly. This is an independent reimplementation based solely on the publicly described idea. No source code or proprietary materials from the original work were used. All credit for the original concept goes to PiKaChu345.