# Harness architecture clawock is an agent-native decision-workflow plugin kit backed by a verifiable harness. It packages reusable skills, tool contracts and decision workflows for external agents. It is not an agent orchestration framework: it does not implement model routing, a ReAct loop or conversation memory. The external agent/runtime remains complete in itself. The runtime is intentionally external. OpenClaw, Hermes, Claude Code, Codex, LangGraph or another runner may own the conversation, model and tools; clawock calls clawock for workflow steps, certified input, deterministic reconciliation, validation, outcome evaluation and publication receipts. A generation is the correlated audit unit emitted by one workflow run, not the product itself. The product-level loop is evidence → debate/workflow → decision → execution/outcome → bounded, reviewable improvement proposal. ## Ownership | Layer | Owns | Current location | |---|---|---| | Plugin/harness core | portable skills/workflows, generation-pinned artifact contract, validation/reconciliation, context assembly, tool schemas, evaluation contracts | `src/clawock/` + portable skill/workflow packages (in progress) | | Profile | portfolio data, schedules, policy values, selected skills/persona, delivery targets and presentation resources | root data + `config/profiles/` + referenced resources | | Runtime adapters | conversations, scheduling, delivery, run history | OpenClaw today; provider interfaces in `src/clawock/providers/` | | Lifecycle runtime | market refresh, `.tmp` artifacts, publication, heartbeats and watchdogs | `src/clawock/{harness,automation}/` in the root wheel | Reusable investment strategies remain product code under `src/clawock/decision/`, even when the first consumer is the live KCNyu desk. Package lifecycle modules bind those strategies into phases and provider-backed delivery; profiles only select values and resources. The full provider → strategy → lifecycle rule is documented in [`product-profile-operations.md`](../reference/product-profile-operations.md#evidence-strategy-and-profile-rule). The public CLI is the stable driver boundary: ```text clawock init clawock workflow list|show|install clawock run prepare --workspace clawock run publish --workspace --request --artifact clawock context audit|assemble clawock brief preflight|postflight clawock report preflight|postflight clawock intraday preflight|postflight ``` `init` and `run` are package-native and work from an installed wheel outside the repository. The control direction is external agent → clawock, like an agent-native business CLI. `run prepare` emits certified input; the calling agent keeps its own model, conversation, memory, skills, tools and repair loop; `run publish` validates/reconciles its files and atomically emits generation-pinned artifacts plus a receipt. clawock never launches the agent. `clawock workflow install investment-decision --workspace ` exports the package-owned pack to `/.agents/skills/investment-decision` as a standard Agent Skill (`SKILL.md`, progressive references and assets). Current OpenClaw also [discovers project-agent skills](https://docs.openclaw.ai/skills) from `.agents/skills`; other skills-compatible runtimes can consume the same directory without a fork. The installed skill tells the current agent how to call the CLI—it does not delegate to a clawock-owned model. The first shipped workflow pins its ID, semantic version, whole-pack certificate and bounded parameters into the prepared request. Its deterministic validator requires supporting and opposing evidence, linked bull/bear cases, thesis invalidation conditions, a bounded action, confidence provenance and exact order/FX arithmetic. A prompt cannot waive those gates. Version 1.1 adds the bounded feedback loop without turning clawock into an agent. An observed outcome is source-linked and reconciled across price and FX; the evaluator reports a direction-adjusted basis-point result with an explicit not-realized-P&L caveat. That evaluation, or a rejected validation receipt, can anchor a proposal that changes only parameters already bounded by the pack. Proposal review certifies an exact hash, apply records before/after overrides, and rollback refuses to overwrite later drift. No command launches a model, rewrites a skill, or lets a proposal accept itself. The shipped parameters govern evidence/provenance strictness only; they do not add or tune factors, catalysts, signals, entries, exits or portfolio rules. The calling runtime's existing investment strategy remains the decision source. The live workflow phase commands dispatch directly to package-owned lifecycle modules selected by a declarative profile. The wheel contains the complete implementation but no user's portfolio or generated state. ## Context contract OpenClaw 2026.7.1 does not have one universal context allowlist. Normal chat injects the five identity/tool bootstrap files plus `HEARTBEAT.md` and `MEMORY.md`; isolated cron injects only the five-file runtime allowlist; heartbeat-light keeps only `HEARTBEAT.md`; bootstrap-pending and subagent runs have their own rules. `src/clawock/context/manifest.json` records each profile, the lazy skill/memory capability roots, conversation-history ownership and the rule that clawock never narrows OpenClaw's tools implicitly. `clawock context audit` fails visibly when an injected file is missing/empty or when moving `skills/`, `MEMORY.md` or `memory/` would silently remove catalog/search capability. With no `--profile` it audits every profile and fails if any one of them does — a file that only the interactive profile requires is not covered by auditing isolated-cron alone. This is broader than comparing prompt text: the skills catalog contains metadata rather than skill bodies, memory search can remain available where `MEMORY.md` is not injected, and session history belongs to the runtime rather than the bootstrap bundle. `clawock context assemble` gives another runner the equivalent bootstrap. Skill bodies remain lazy: the catalog is only an index, and only an explicitly selected `skills/{name}/SKILL.md` is added. This preserves normal chat context and avoids silently loading every skill into every run. The runtime-required root files and capability directories must not move before every active profile passes its parity gate. The manifest owns the portable definition and audit gate; it cannot make an OpenClaw binary consume a different path. The full adapter and cron cutover contract is in [`openclaw-adapter.md`](openclaw-adapter.md). Standalone runs instead list context files in `clawock.json`. Every document and the assembled bundle receive SHA-256 certification in the runtime request and published manifest. The command protocols are specified in [`runtime-protocol.md`](runtime-protocol.md). ## Artifact contract An `ArtifactSet` has one `generation_id`, and every member must carry the same ID. This is the runtime-neutral seam after model deliberation: a runner may use OpenClaw, a direct API or an interactive coding agent, but validation and publish must reject artifacts mixed across generations. ## Workflow pack contract Package data under `src/clawock/workflows/packs//` owns reusable semantics: ```text investment-decision/ ├── SKILL.md ├── workflow.json ├── agents/openai.yaml ├── references/decision-contract.md ├── references/decision.schema.json ├── references/outcome.schema.json ├── references/improvement-proposal.schema.json ├── assets/decision.example.json └── assets/outcome.example.json ``` `workflow.json` is the machine-readable discovery and parameter contract; `SKILL.md` follows the open [Agent Skills specification](https://agentskills.io/specification) and is the runtime-facing procedure; references and assets are loaded progressively. Python validators remain package code so neither a runtime nor a profile can silently edit financial or provenance invariants by changing prose.