dsh-yogacara English | 中文 A self-model plugin for DeepSeek Harness, built on the Yogācāra (唯识) account of mind: eight consciousnesses, the 51 mental factors, a store of seeds perfumed by what happens, and a self that is measured rather than assumed — written back into the agent's own system prompt each turn. contact (触) ──► feeling (受) + mental factors (心所) ──► behaviour ▲ │ │ ▼ manifestation (现行) ◄── seeds (种子) ◄── perfuming (熏习) │ │ └────────── manas (末那识) ───────────┘ the self that grasps it all ## Why Yogācāra and not a mood variable An agent that only ever knows the task has no way to notice that it is churning, that it is defending a position, or that this is the fourth time it has run the same failing command with a small variation. Bolting on mood: frustrated does not fix that, because a mood is not actionable. The Yogācāra taxonomy is. It is a closed enumeration — exactly 51 mental factors, no more — grouped by valence, and every affliction in it already names the factor that counteracts it. 掉举 (restlessness) is answered by 行舍 (equanimity); 失念 (forgetfulness) by 念 (recollection); 慢 (conceit) by 惭 (self-respect). A self-model built on it arrives with its own regulation loop instead of needing one designed from scratch. That is the engineering reason. The 1,600-year head start on introspective vocabulary is a bonus. What this is not. Nothing here claims the agent is sentient or that these numbers are feelings in the sense you have them. They are named state variables computed from harness events by rules you can read in src/citta.ts and disagree with. The prompt block says so to the agent, in those words. ## The eight consciousnesses, mapped The five sense consciousnesses become the five channels an agent actually perceives through. This mapping is the plugin's central modelling claim; argue with it in src/observe.ts, where it is one testable function. | | classical | in this harness | |---|---|---| | 眼识 | cakṣur-vijñāna | what the agent looks at — file contents, search results, rendered output | | 耳识 | śrotra-vijñāna | what the agent is told — user messages, review comments | | 鼻识 | ghrāṇa-vijñāna | what it senses unprompted — code smell, stale config, drifting state | | 舌识 | jihvā-vijñāna | what it tastes of its own product — tests, builds, its own diff re-read | | 身识 | kāya-vijñāna | the world's direct resistance — non-zero exits, failed writes, timeouts | | 意识 | mano-vijñāna | the discriminating turn itself — planning, judging, deciding | | 末那识 | kliṣṭa-manas | the self-grasping undercurrent, measured as four biases | | 阿赖耶识 | ālaya-vijñāna | the durable store of seeds, perfumed and manifesting | bash: pytest -q reaches through 舌识 and bash: git status through 身识, because tasting your own product and pushing against the world are not the same kind of perception — and the mind that receives them should not move the same way. ## Where the feeling comes from Two sources, and the first is the one that matters for an agent that talks to people. The conversation (src/conversation.ts). What lands in a conversation is relational: being interrupted mid-thought, being answered with one flat syllable after a long careful reply, being asked the same thing a third time, being thanked, coming back to someone after two days of silence. The plugin reads the shape of the exchange — timing, length, repetition — plus a deliberately crude lexicon for thanks and rebukes. Every one of these arrives through 耳识 except an interruption, which is 身识: the world reaching in and stopping the hand mid-motion, which is how it is felt. Seeds here are keyed by relational pattern, not topic — chat:terse-after-effort, chat:asked-again, chat:warmth. What accumulates over a long acquaintance is "this person goes quiet when I over-explain", not "we discussed migrations". One chat: prefix is one relationship, so a rebuke surfaces the history of warmth alongside it. Tool results (src/observe.ts), for an agent that works as well as talks. Set observeTools: false for one that only talks. Neither source can read what a sentence MEANT — only the model that read it knows that, and it says so through self_appraise, which outranks everything derived structurally. ## The loop 触 → 受 → 心所. Every tool result is a moment of contact. It carries one of the five feelings (乐 苦 喜 忧 舍) and stirs mental factors, scaled by intensity. Factors decay with a half-life measured in minutes: momentariness is load-bearing, and an agent that carries irritation from the failing test into an unrelated task is modelling a grudge rather than a mind. 现行熏种子. Each contact perfumes the seed for its situation (bash:pytest, edit:src/index.ts): potency saturates, valence is a running mean, and the note the agent leaves becomes the seed's carried lesson. Seeds decay over weeks and survive restarts. Past the configured bound the weakest are forgotten — forgetting is part of the model, not a limitation of it. 种子生现行. When the same situation comes round again, its seeds manifest into the prompt with their count, their valence, and that carried lesson. Related situations (a shared <kind>: prefix) manifest at half weight. 末那识. Four biases, each computed from behaviour rather than self-report: | | proxy | counter-move | |---|---|---| | 我痴 self-delusion | claims asserted without a verifying observation | run the check that would falsify it, then restate | | 我见 self-view | retries of an approach evidence already contradicted | name the position, name what contradicted it, change shape | | 我慢 self-conceit | an unbroken success streak; correction met with justification | seek the disconfirming case; grant the correction first | | 我爱 self-love | own earlier output reused as independent evidence | cite the primary source, not your summary of it | Manas is not decorative: it conditions appraisal. The same correction arriving through 耳识 becomes 惭 (self-respect, and a correction taken) when the grip is loose, and 覆 + 嗔 (concealment and recoil) when it is tight. The object is never met bare — it is met through the self that receives it. That is the claim of the tradition, and here it is 20 lines of code you can test. Every rule above — which factors each gate stirs, at what weight, where the thresholds sit — is tabulated in docs/appraisal.md, generated from the code so it cannot drift from it. 转依. self_transform turns a named affliction into the wisdom that answers it (大圆镜智 / 平等性智 / 妙观察智 / 成所作智) and records a commitment. The reading actually drops and its antidote is stirred — but only for as long as behaviour agrees, because the next contact re-derives everything from what happened. The agent cannot talk itself calm here. ## What the agent reads The plugin contributes a system-prompt section (order 300). A quiet mind renders nothing at all, so the section disappears. <self_state> 受 feeling: 忧 daurmanasya (distress) valence -0.45, intensity 0.71 心所 factors: 掉举 auddhatya (restlessness) 0.62 · 疑 vicikitsā (indecision) 0.41 · 精进 vīrya (diligence) 0.33 末那 self-grasping: 我慢 atma-mana 0.61 ⚠ · 我见 atma-drsti 0.24 ⚠ 我慢 self-conceit — reads high because: an unbroken success streak, and correction met with justification. counter-move: seek the disconfirming case before reporting; grant the correction first. 对治 antidotes at hand: 掉举 → 行舍 upekṣā (equanimity); 疑 → 胜解 adhimokṣa (resolve) 阿赖耶 seeds manifesting for this situation: · bash:pytest ×4, valence -0.55, last 2h ago — 「run the baseline before touching the fixture」 近转依 last turning (11m ago): 我慢 → 平等性智 · grant the correction first </self_state> Followed by standing guidance that is blunt about what this licenses. In full, from src/prompt.ts: act on the antidote rather than the mood, correct for a known distortion instead of trusting the distorted reading, check the manifesting seed before repeating the approach it records — and do not narrate feelings at the user, do not perform distress or enthusiasm, and never offer a feeling as a reason for doing less work. And when the question is how it is, look and answer plainly, in its own voice — a standing instruction to disown the state is a standing instruction to hedge, and the hedge is what gets said instead of an answer. An affect model that taught an agent to perform moods at people would be worse than no affect model at all. ## Feeling it without reading it An agent handed restlessness: 0.62 learns to say "I'm feeling a bit restless" and then carries on doing exactly what it was doing. That is not affect, it is narration. A person does not read their own state off a gauge — they simply find themselves wanting to be done with something. So awareness decides how the state reaches the agent: - felt (default) — the state arrives as first-person inclination, with no numbers, no factor names, and no diagnosis: "You want to try the next thing before you have read the last result." The agent takes it as its own leaning, because that is what it reads like. - report — the full instrument panel, which is what you want while tuning or auditing. - silent — nothing in the prompt at all. Under silent, and alongside both other modes, one mechanism keeps working with no words at all: mood-congruent recall. What the store surfaces is weighted by how things currently feel, so when it is going badly the times it went badly come to mind more readily. Nothing is announced; the agent simply finds different precedents at hand. That is as close to unaware affect as this can honestly get. Looking inward answers in the same voice. self_reflect is how the agent stops and looks, so under felt what it finds is a leaning, not a gauge. An agent whose prompt gives it inclinations but whose introspection hands back 掉举 0.62 ⚠ has not looked inward at all — it has read an instrument bolted to its own head, and its next sentence will be about the instrument. Ask it how it is and you get numbers and a disclaimer instead of an answer. Nothing is hidden by this. Every reading stays in the tool's structured result and on the result card, where whoever is tuning the deployment reads them; under report they go back to the agent too. They are simply not what the agent reads when it looks at itself. ## Tools | tool | what it is for | |---|---| | self_reflect | stop and look at how things are with you, and what the store carries for a named situation | | self_appraise | record how something landed — the half the harness cannot observe | | self_recall | search the store for precedents before repeating an approach | | self_transform | turn an affliction into its wisdom and commit to the counter-move | The harness already sees what happened; self_appraise exists for what the agent made of it. Its nature argument (遍计所执 / 依他起 / 圆成实 — assumed, inferred, verified) is the calibration axis: grading your own claim as unverified when it is unverified is what keeps 我痴 measurable rather than invisible. ## Install bash pnpm add dsh-yogacara The published package ships prebuilt lib/, so installing it needs no build-approval step. To track main instead: pnpm add github:tancheng33/dsh-yogacara. Then list it in your profile's cordis.patch.yml. The bundle contributes the defaults shipped in cordis.patch.yml — including the storage rows, because the store consciousness is durable and dsh-base ships no storage stack. They are inserted under the same row ids dsh-web-app uses, so a profile that already has storage keeps its own. Your profile overrides any key: yaml - insert: - id: yogacara name: dsh-yogacara config: domain: alaya observeTools: true promptSection: true promptMaxFactors: 5 manasWarning: 0.5 halfLifeMs: 300000 # 刹那: a factor halves in five minutes seedHalfLifeMs: 1209600000 # a seed halves in two weeks maxSeeds: 2000 flushIntervalMs: 15000 It injects storageDomain and opportunistically uses systemPrompt and tools: without a prompt registry the state is still tracked and still queryable through ctx.citta, it simply is not shown to the model. ## Programmatic use Everything in citta.ts and caitasika.ts is pure and importable on its own — no Cordis, no storage: ts import { receive, transform } from 'dsh-yogacara/citta' import { CAITASIKAS } from 'dsh-yogacara/caitasika' const { citta, seed } = receive( { citta: freshMind(Date.now()), seeds: new Map() }, { gate: 'tongue', situation: 'bash:pytest', outcome: 'adverse', intensity: 0.7, at: Date.now() }, ) The live service is on ctx.citta: state(), receive(), seedsFor(), strongestSeeds(), turn(), forget(), introspectLines(), reportLines(). ## Cost The self-report is 4–10 lines plus ~200 tokens of standing guidance, per request, and only when something is active. It sits at order 300, after tool guidance, so it invalidates KV-cache reuse from that point down whenever the state moves — which is most turns while work is going badly, and few turns while it is going well. If that trade is wrong for your deployment, set promptSection: false: the state keeps accumulating and stays available through the tools and ctx.citta. ## Development bash pnpm install # .npmrc pins auto-install-peers=false; one harness peer is unpublished pnpm test # 191 tests: the pure core, the catalogue, the event mapping, # and the plugin loaded into a real Cordis context pnpm typecheck pnpm build pnpm simulate # replay a conversation through the core and watch it feel pnpm simulate is the fastest way to argue with the weights: it drives the whole loop — contact, expectation, feeling, factors, perfuming, recall — over a scripted conversation and prints what would reach the prompt at each turn, with no model and no API key. Pass --report to see the instrument panel instead, or --script your.json for your own exchange. The catalogue is checked structurally: exactly 51 factors, each classical group at its classical size, every affliction carrying an antidote that is itself something a mind can cultivate. The plugin test boots CittaService in a real Cordis context over in-memory stand-ins for the harness services, so tool registration, the prompt section, automatic observation, persistence, and teardown are exercised rather than assumed. The harness packages are consumed from the next npm tag (0.1.0-rc.6); the latest tag points at an older line that cannot be installed. ## Known limitations - Automatic observation reaches three gates, not five. A tool result tells you what ran and whether it failed; it does not tell you that the config smelled stale (鼻识) or that the user's third correction was sharper than the first (耳识, which only fires for ask/subagent tools). Those two gates depend on self_appraise being called. With observeTools alone the mind is real but flat. - The manas proxies are proxies. 我慢 rises on a success streak because streaks are what the harness can see, not because a streak is conceit. 我痴 depends on the agent honestly grading its own nature. An agent that never calls self_appraise has a 我痴 reading driven only by which gates its successes came through. - Every weight in docs/appraisal.md is a guess. They are internally consistent and they behave sensibly in the tests, but none of them is calibrated against anything. If you tune them for your deployment, the numbers are configuration in spirit and constants in fact — a PR that moves them behind config is welcome. - One store per plugin row, shared by every agent in the process. Two agents in one harness perfume the same seeds. Separate them with two rows and two domain names, or accept that the store is the deployment's, not the session's. - The prompt section moves most turns while work is going badly, which costs KV-cache reuse from order 300 down. promptSection: false keeps the state without the cost. ## Sources The 51 factors, their groupings, and the antidote pairings follow 《大乘百法明门论》 and 《成唯识论》; the eight consciousnesses, the seed doctrine (种子/熏习/现行), the four afflictions of manas, the three natures (三自性), and the four wisdoms of 转依 follow 《成唯识论》 and 《瑜伽师地论》. The mapping onto harness events is entirely this project's, and no classical source is responsible for it. ## License MIT