# Omni > [δΈ­ζ–‡](./README.md) > **Omni adds a task-level reliability layer to DeepSeek Harness:** > decide when to intervene, prepare the right context, constrain completion criteria, verify real evidence, and change strategy on failure. > It reuses existing DSH Skills, Tools, and Plugins instead of re-implementing them. **DSH gives agents capabilities. Omni makes their outcomes dependable.** **Fast by default. Smart when needed. Proven when done.** ```bash dsh plugin --profile web add dsh-omni-router ``` --- ## Why Omni Coding agents commonly suffer from: - Wrong code or wrong context - Losing track in complex tasks - Forgetting acceptance criteria - Tests not actually running - Repeating the same failed strategy - Saying "done" without actually finishing Omni exists to solve these engineering problems, not to be another router. --- ## What Omni does | Core value | What Omni actually does | |---|---| | 🧠 **Intervention Intelligence** | Simple tasks are true NOOP; complex tasks get Assist / Guard | | 🎯 **Context & Task Intelligence** | Task Contract + focused context + capability sufficiency | | βœ… **Proof & Recovery** | Evidence / freshness / acceptance coverage / recovery | | ♻️ **Reuse first** | Design principle: prefer installed Tools / Skills / Plugins | | 🧩 **Capability Gap Management** | Fill gaps only when they block acceptance | | πŸ”„ **Failure discipline** | retry β†’ repair β†’ expand context β†’ change hypothesis | --- ## 30-second demo ### Without Omni ```text User: Fix refreshSession firing three concurrent refreshes Agent: modified session.ts ran some checks Done βœ… ``` Problems: - The real concurrency test never ran - A caller was missed - No regression test was added - No evidence that "done" is true ### With Omni ```text Task Contract C1 one refresh request per session C2 concurrent callers behave correctly C3 regression test passes Relevant Context session.ts middleware/auth.ts websocket/reconnect.ts session-concurrency.test.ts DSH executes... Evidence βœ“ concurrency test: 28/28 βœ“ full auth suite: 143/143 βœ“ relevant files verified Proof: 3/3 ``` --- ## One plugin, reuse the whole DSH ecosystem > This is not the core of Omni; it is the ecosystem strategy: **Reuse first. Don't reimplement host capabilities.** ``` Omni Capability + Reliability β”‚ β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”Όβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β”‚ β”‚ β”‚ β–Ό β–Ό β–Ό Skills Plugins Tools β”‚ β”‚ β”‚ β”œβ”€β”€ TDD β”œβ”€β”€ Browser β”œβ”€β”€ Shell β”œβ”€β”€ Git β”œβ”€β”€ Router β”œβ”€β”€ Editor └── Verify └── Others └── Tests β”‚ β”‚ β”‚ β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”Όβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ β–Ό DeepSeek Harness β”‚ β–Ό Execute β”‚ β–Ό Evidence β†’ Verify ``` > **Omni does not try to own the most capabilities. It makes the good capabilities you already have work together.** ### Example: Frontend UI fix ```text User: Inspect this webpage and fix the UI. Omni: 1. Detect frontend task 2. Check current DSH capabilities 3. Browser already installed β†’ reuse 4. UI/Verification skill exists β†’ let DSH use it 5. Do not install duplicates 6. Generate UI acceptance criteria 7. DSH executes 8. Collect screenshot / tool / test evidence 9. Verify ``` ### Example: Complex refactor ```text Current environment: βœ“ Git βœ“ Testing βœ“ Skills βœ— Some task-specific capability Omni: Existing capabilities β†’ all reused Missing capability β†’ only fill that gap ``` --- ## Under the hood Omni's surface is simple, but it is not a prompt template. ```text Task β”‚ β–Ό Task Classification β”‚ β–Ό Intervention Gate ───────────→ NOOP β”‚ β–Ό Task Contract β”œβ”€β”€ Acceptance Criteria β”œβ”€β”€ Risk β”œβ”€β”€ Context Budget └── Capability Requirements β”‚ β–Ό Context Engine β”œβ”€β”€ multilingual tokenizer β”œβ”€β”€ lexical retrieval β”œβ”€β”€ symbol retrieval β”œβ”€β”€ dependency graph └── freshness tracking β”‚ β–Ό Capability Audit β”œβ”€β”€ existing tools β”œβ”€β”€ native skills β”œβ”€β”€ installed plugins └── missing capabilities β”‚ β–Ό DeepSeek Harness β”‚ β–Ό DSH Events β”‚ β–Ό Evidence Trust β”œβ”€β”€ T0 model claim β”œβ”€β”€ T1 observation β”œβ”€β”€ T2 host/tool β”œβ”€β”€ T3 deterministic execution └── T4 independent verifier β”‚ β–Ό Criterion Verification β”œβ”€β”€ PASS β†’ Verified └── FAIL β†’ Recovery β”‚ └──→ DSH ``` --- ## Reliability Engineering Omni handles these real agent failure modes: ```text βœ“ Model claims tests passed β†’ model evidence is only T0/T1 βœ“ Tool output confused with model text β†’ provenance-aware Evidence βœ“ Old test results used after files changed β†’ Evidence freshness / invalidation βœ“ permission / dependency / timeout treated as success β†’ success whitelist + failure taxonomy βœ“ Chinese task cannot find English symbols β†’ multilingual query tokenizer + semantic expansion βœ“ Same failure retried forever β†’ failure-aware recovery βœ“ Simple tasks slowed by complex frameworks β†’ true L0 NOOP βœ“ Rebuilding capabilities that already exist β†’ capability reuse first βœ“ Users must manually install missing capabilities β†’ approval-gated provisioning (recommend by default, auto-trusted only on explicit opt-in) βœ“ Two routers controlling the same model β†’ compatibility delegation ``` --- ## Omni is not just a router ```text Normal Router Task β†’ choose mode/model/agent Omni Task β†’ decide whether to intervene β†’ reuse existing capabilities β†’ fill missing capabilities β†’ prepare correct context β†’ define acceptance β†’ host executes β†’ collect real evidence β†’ verify completion β†’ recover from failure ``` > **Routing is one decision. Reliability is the whole lifecycle.** --- ## Works better with good plugins Omni does not ask the ecosystem to rewrite itself around it. If `dsh-router-standard` is installed, Omni automatically creates the combined preset: ```text Omni Router + Router Standard ``` Router Standard keeps handling DeepSeek persona / attention / workflow routing; Omni handles capability, context, evidence, and reliability. --- ## Built like infrastructure, not a prompt pack Omni has independent test coverage for routing, Context, Evidence, Capability, Recovery, Host Adapter, Installer, and OmniBench (~60 test files). Every RC requires a full `npm test` regression; benchmark supports multi-arm comparisons (raw / omni / mid / frontier / stack / stack_omni…). ## Measured Reliability ```text Evidence Federation third-party results β†’ EvidenceRecord v1 (no self-reported trust) Context Metrics Recall@5/10 Β· Precision Β· token overhead Β· irrelevant ratio Intervention KPIs NOOP precision Β· missed/false intervention rate Recovery Telemetry failure fingerprint Β· retry/repair/hypothesis funnel Β· success rate Multi-arm Benchmark comparePair(baseline, candidate) + release gates ``` --- ## Proof of Completion Every acceptance criterion has an id (e.g. `C1`, `C2`). Every criterion needs at least one fresh evidence record at the required trust level: | Trust | Source | |---|---| | T0 | model claim | | T1 | agent observation / model-written JSON | | T2 | host/tool output | | T3 | deterministic command/test execution | | T4 | independent / hidden verifier | ```text Completed. βœ“ C1 Root cause fixed βœ“ C2 Regression test added βœ“ C3 143 tests passed Proof: 3/3 ``` If DSH has no reliable event hook yet, Omni honestly reports: ```text Verified Partially Verified Unverified ``` --- ## Compatibility | Project / Capability | Omni strategy | |---|---| | `dsh-router-standard` | **Delegate** persona/reasoning routing | | DSH native Skills | **Reuse** automatic activation | | Browser plugin | **Reuse** browser capability | | Testing/verification Skills | **Reuse evidence/results** | | Git plugins | **Reuse** native workflow | | MCP servers | **Treat as host capabilities** | | Missing capability | **Provision only when necessary** | --- ## Evolution ```text 1.x Router ↓ 2.x Planner / Context / Capability / Evidence exploration ↓ 2.4 Reliability Kernel ↓ 3.0 Convergence: Decide Β· Prepare Β· Verify ``` > 3.0 intentionally removed or hid overlapping runtime/router features instead of continuing feature growth. --- ## OmniBench ```text 3.0 RC benchmark validation is in progress. Target: β‰₯50 repos β‰₯100 tasks β‰₯3 paired runs/task hidden verifier 100% Medium/Hard uplift β‰₯10pp False completion <3% Cost ≀2.5Γ— ``` No frontier parity claim will be made until these gates pass. ```bash npm run omnibench:v2:generate -- benchmark/omnibench-v2/manifest.local.example.json node benchmark/omnibench-v2/run.mjs benchmark/omnibench-v2/manifest.local.example.json --exec node benchmark/omnibench-v2/matrix.mjs node benchmark/omnibench-v2/gates.mjs ``` --- ## Architecture ```text kernel/ task-contract Β· intervention-gate Β· omni-event Β· evidence-trust context/ query-tokenizer Β· capsule Β· freshness recovery/ failure-taxonomy Β· strategy-shift Β· recovery-policy capability/ auditor Β· quality Β· solver Β· provisioner Β· performance mission/ mission-ir Β· planner-dag Β· mission-dag (compat) host/ interface Β· dsh-adapter benchmark/ omnibench-v2 runner Β· matrix Β· gates ``` --- ## FAQ **Q: Does it slow down simple tasks?** A: No. L0 is a true NOOP; Omni injects no prompt. **Q: Does it replace DeepSeek Harness?** A: No. DSH keeps execution, tools, skills, plugins, and runtime. **Q: Does it take over my plugins?** A: No. Omni defines what "success" means; it does not micromanage plugin calls. **Q: Can it auto-install skills/plugins?** A: Yes, but default is `recommend`: detect gaps β†’ search β†’ recommend β†’ user/host approval before install. Auto-install only with explicit `auto-trusted`, and every change is recorded in `.omni/capability-audit.json`. **Q: Does it conflict with Router Standard?** A: No. When Router Standard is detected, Omni auto-creates a combined preset and delegates reasoning/persona routing. --- ## Credits / Inspirations Omni is inspired by several DeepSeek Harness community projects, especially: - `dsh-router-standard / dsh-routing-suite` - persona routing - attention engineering - progressive disclosure - real Harness assembly-chain research Omni does not replace these capabilities; it makes them work together in one reliability loop. --- ## License [MIT](./LICENSE)