A state-of-the-art, semi-autonomous multi-agent system built in Rust. It is designed for complex technical tasks, autonomous problem-solving, and seamless human-AI interaction via text and voice.
Rust
3
209 commits
updated Jul 1, 2026
This repository is the Rust runtime substrate for the Nexus Control Plane. It is being collapsed from the older Agency harness into a smaller NanoClaw-shaped runtime while preserving useful Agency ideas through clean-room adoption.
The current goal is not to make Agency disappear. Agency is the native Rust harness lineage. NanoClaw is the claw-behavior reference model. The active runtime should keep the useful harness capabilities, but only after they are distilled into clear Rust NanoClaw contracts.
src/foundation/ contains reusable domain primitives.src/nanoclaw/ contains the active runtime implementation.src/bin/nanoclaw.rs is the only active binary target.Cargo.toml disables Cargo auto-discovery for legacy bins, tests, examples,
and benches.make verify runs scripts/check-legacy-source-gates.sh, which fails if old
Agency modules, tests, or parked root binaries leak back into the active Cargo
surface.graveyard/agency-harness/ preserves old Agency binaries and integration
tests as reference material outside the active compile path.start_agency.sh and scripts/bundle_onnx.sh
are compatibility guards only; they do not launch old Agency services or
inactive binaries.The old distributed Agency microservice suite is not the current operator entrypoint. It remains source material for adoption decisions, not proof of the runtime.
The active Rust runtime is responsible for:
Provider/model access must pass through the existing adapter, gateway, and runtime contracts. Do not bypass those paths for Azure, ZAI, Codex, OpenClaw, or any other provider except for an explicit one-off diagnostic.
cargo check --all-targets
cargo test --all-targets
cargo run --quiet --bin nanoclaw -- show-config
Expected active Cargo targets:
rust_agencynanoclawbuild-script-buildShow the resolved runtime configuration:
cargo run --quiet --bin nanoclaw -- show-config
Inspect the active runtime channel set:
cargo run --quiet --bin nanoclaw -- runtime status
The status report includes a typed runtimeChannels registry covering local,
scheduler, Slack, webhook, OpenClaw gateway, and discontinued legacy PM
automation ownership. It also includes runtime PID files under data/runtime/,
so local production profiles can be inspected and controlled without the old
Agency service scripts. Startup and preflight events are appended to
data/runtime/startup-events.jsonl for durable operator evidence.
Inspect runtime state, recent tasks, and recent execution provenance:
cargo run --quiet --bin nanoclaw -- runtime inspect --limit 5
Inventory the active runtime state roots without deleting or migrating anything:
cargo run --quiet --bin nanoclaw -- runtime state --limit 5
That report includes stateResidue, which separates active runtime roots from
legacy memory/vector/cache candidates and an operatorActions plan for each
legacy candidate. To include the same report beside stale-PID cleanup output
without deleting state:
cargo run --quiet --bin nanoclaw -- runtime cleanup --state-residue
Plan or apply guarded state-residue actions:
cargo run --quiet --bin nanoclaw -- runtime state-action --plan --limit 10
cargo run --quiet --bin nanoclaw -- runtime state-action --apply \
--confirm state-residue:agency_history_jsonl:archive_legacy_state
runtime state-action --apply mutates exactly one confirmed action id from a
prior plan. The current apply registry is non-destructive: it can archive
legacy memory/vector/history residue into data/runtime/state-archive/ and
writes a receipt under data/runtime/state-actions/. It does not delete source
references, purge caches destructively, or migrate legacy state without a
separate runtime contract.
Classify remaining legacy Agency source material against its active Rust NanoClaw descendant surface:
cargo run --quiet --bin nanoclaw -- runtime source-disposition --limit 10
This report is deterministic and report-only. It fingerprints parked source
directories such as src/fpf, src/orchestrator, src/agent, src/memory,
and src/tools, then records whether each path should be distilled further,
left parked, or held for a future runtime contract.
Get a deterministic health report over runtime directories, PID files, startup/preflight failures, runtime-channel ownership, gateway/webhook auth posture, task backlog, recent execution evidence, and known recovery suggestions for missing runtime config:
cargo run --quiet --bin nanoclaw -- runtime health --limit 5
cargo run --quiet --bin nanoclaw -- runtime health --strict
cargo run --quiet --bin nanoclaw -- runtime health --notify-local ops
Convert health recovery evidence into a guarded repair plan, or attempt only safe deterministic repairs:
cargo run --quiet --bin nanoclaw -- runtime repair --plan --limit 5
cargo run --quiet --bin nanoclaw -- runtime repair --apply --limit 5
runtime repair --apply will not write secrets, choose ports, enable legacy
lanes, or restart long-running profiles automatically. Those remain explicit
operator actions.
Report stale or invalid runtime PID files, then remove them only when explicitly applied:
cargo run --quiet --bin nanoclaw -- runtime cleanup
cargo run --quiet --bin nanoclaw -- runtime cleanup --apply
Execution lanes now return and persist a structured ExecutionEvidence
envelope alongside their operator-facing text. The envelope records adapter
type, execution mode, workspace, git state, artifacts, verification, blockers,
and provenance IDs so run evaluation can use typed evidence instead of summary
prose.
Poll the local control-plane runtime once:
cargo run --quiet --bin nanoclaw -- runtime poll
Serve a selected production profile through the NanoClaw entrypoint:
cargo run --quiet --bin nanoclaw -- runtime serve --profile full
cargo run --quiet --bin nanoclaw -- runtime serve --profile gateway
Startup uses the same runtimeChannels registry reported by status and health.
Misconfigured profile channels fail before PID-file creation with
runtime_channel_misconfigured and the missing config/auth fields. Those
preflight failures, startup attempts, running transitions, and startup failures
are recorded in runtime state under runtime.startupEvents. runtime health
derives deterministic recovery suggestions from the same ledger, keeping
secret-bearing fixes operator-gated instead of auto-applying them.
Stop or signal a running profile through the NanoClaw entrypoint:
cargo run --quiet --bin nanoclaw -- runtime stop --profile full
cargo run --quiet --bin nanoclaw -- runtime reload --profile full
List available commands:
cargo run --quiet --bin nanoclaw -- --help
Run a local control-plane command only after confirming the required operator configuration and authentication boundaries for the target environment.
The collapse is source-preserving and evidence-driven:
src/foundation/ or src/nanoclaw/.Current tracking documents:
docs/nanoclaw-rs-migration.mddocs/nanoclaw-rs-collapse-audit.mddocs/agency-capability-adoption-ledger.mdThe legacy README and old service/test targets are preserved under:
graveyard/agency-harness/
Treat that directory as clean-room reference material. It is not part of the active runtime proof, CI target set, or operator startup path.
Prefer small changes with clear validation evidence. When adopting old Agency logic, extract the invariant or runtime contract first; do not revive broad old module graphs just to make legacy surfaces compile.
204 commits
5 commits
Rust
97.9%
JavaScript
1.1%
A state-of-the-art, semi-autonomous multi-agent system built in Rust. It is designed for complex technical tasks, autonomous problem-solving, and seamless human-AI interaction via text and voice.
Rust
3
209 commits
updated Jul 1, 2026
This repository is the Rust runtime substrate for the Nexus Control Plane. It is being collapsed from the older Agency harness into a smaller NanoClaw-shaped runtime while preserving useful Agency ideas through clean-room adoption.
The current goal is not to make Agency disappear. Agency is the native Rust harness lineage. NanoClaw is the claw-behavior reference model. The active runtime should keep the useful harness capabilities, but only after they are distilled into clear Rust NanoClaw contracts.
src/foundation/ contains reusable domain primitives.src/nanoclaw/ contains the active runtime implementation.src/bin/nanoclaw.rs is the only active binary target.Cargo.toml disables Cargo auto-discovery for legacy bins, tests, examples,
and benches.make verify runs scripts/check-legacy-source-gates.sh, which fails if old
Agency modules, tests, or parked root binaries leak back into the active Cargo
surface.graveyard/agency-harness/ preserves old Agency binaries and integration
tests as reference material outside the active compile path.start_agency.sh and scripts/bundle_onnx.sh
are compatibility guards only; they do not launch old Agency services or
inactive binaries.The old distributed Agency microservice suite is not the current operator entrypoint. It remains source material for adoption decisions, not proof of the runtime.
The active Rust runtime is responsible for:
Provider/model access must pass through the existing adapter, gateway, and runtime contracts. Do not bypass those paths for Azure, ZAI, Codex, OpenClaw, or any other provider except for an explicit one-off diagnostic.
cargo check --all-targets
cargo test --all-targets
cargo run --quiet --bin nanoclaw -- show-config
Expected active Cargo targets:
rust_agencynanoclawbuild-script-buildShow the resolved runtime configuration:
cargo run --quiet --bin nanoclaw -- show-config
Inspect the active runtime channel set:
cargo run --quiet --bin nanoclaw -- runtime status
The status report includes a typed runtimeChannels registry covering local,
scheduler, Slack, webhook, OpenClaw gateway, and discontinued legacy PM
automation ownership. It also includes runtime PID files under data/runtime/,
so local production profiles can be inspected and controlled without the old
Agency service scripts. Startup and preflight events are appended to
data/runtime/startup-events.jsonl for durable operator evidence.
Inspect runtime state, recent tasks, and recent execution provenance:
cargo run --quiet --bin nanoclaw -- runtime inspect --limit 5
Inventory the active runtime state roots without deleting or migrating anything:
cargo run --quiet --bin nanoclaw -- runtime state --limit 5
That report includes stateResidue, which separates active runtime roots from
legacy memory/vector/cache candidates and an operatorActions plan for each
legacy candidate. To include the same report beside stale-PID cleanup output
without deleting state:
cargo run --quiet --bin nanoclaw -- runtime cleanup --state-residue
Plan or apply guarded state-residue actions:
cargo run --quiet --bin nanoclaw -- runtime state-action --plan --limit 10
cargo run --quiet --bin nanoclaw -- runtime state-action --apply \
--confirm state-residue:agency_history_jsonl:archive_legacy_state
runtime state-action --apply mutates exactly one confirmed action id from a
prior plan. The current apply registry is non-destructive: it can archive
legacy memory/vector/history residue into data/runtime/state-archive/ and
writes a receipt under data/runtime/state-actions/. It does not delete source
references, purge caches destructively, or migrate legacy state without a
separate runtime contract.
Classify remaining legacy Agency source material against its active Rust NanoClaw descendant surface:
cargo run --quiet --bin nanoclaw -- runtime source-disposition --limit 10
This report is deterministic and report-only. It fingerprints parked source
directories such as src/fpf, src/orchestrator, src/agent, src/memory,
and src/tools, then records whether each path should be distilled further,
left parked, or held for a future runtime contract.
Get a deterministic health report over runtime directories, PID files, startup/preflight failures, runtime-channel ownership, gateway/webhook auth posture, task backlog, recent execution evidence, and known recovery suggestions for missing runtime config:
cargo run --quiet --bin nanoclaw -- runtime health --limit 5
cargo run --quiet --bin nanoclaw -- runtime health --strict
cargo run --quiet --bin nanoclaw -- runtime health --notify-local ops
Convert health recovery evidence into a guarded repair plan, or attempt only safe deterministic repairs:
cargo run --quiet --bin nanoclaw -- runtime repair --plan --limit 5
cargo run --quiet --bin nanoclaw -- runtime repair --apply --limit 5
runtime repair --apply will not write secrets, choose ports, enable legacy
lanes, or restart long-running profiles automatically. Those remain explicit
operator actions.
Report stale or invalid runtime PID files, then remove them only when explicitly applied:
cargo run --quiet --bin nanoclaw -- runtime cleanup
cargo run --quiet --bin nanoclaw -- runtime cleanup --apply
Execution lanes now return and persist a structured ExecutionEvidence
envelope alongside their operator-facing text. The envelope records adapter
type, execution mode, workspace, git state, artifacts, verification, blockers,
and provenance IDs so run evaluation can use typed evidence instead of summary
prose.
Poll the local control-plane runtime once:
cargo run --quiet --bin nanoclaw -- runtime poll
Serve a selected production profile through the NanoClaw entrypoint:
cargo run --quiet --bin nanoclaw -- runtime serve --profile full
cargo run --quiet --bin nanoclaw -- runtime serve --profile gateway
Startup uses the same runtimeChannels registry reported by status and health.
Misconfigured profile channels fail before PID-file creation with
runtime_channel_misconfigured and the missing config/auth fields. Those
preflight failures, startup attempts, running transitions, and startup failures
are recorded in runtime state under runtime.startupEvents. runtime health
derives deterministic recovery suggestions from the same ledger, keeping
secret-bearing fixes operator-gated instead of auto-applying them.
Stop or signal a running profile through the NanoClaw entrypoint:
cargo run --quiet --bin nanoclaw -- runtime stop --profile full
cargo run --quiet --bin nanoclaw -- runtime reload --profile full
List available commands:
cargo run --quiet --bin nanoclaw -- --help
Run a local control-plane command only after confirming the required operator configuration and authentication boundaries for the target environment.
The collapse is source-preserving and evidence-driven:
src/foundation/ or src/nanoclaw/.Current tracking documents:
docs/nanoclaw-rs-migration.mddocs/nanoclaw-rs-collapse-audit.mddocs/agency-capability-adoption-ledger.mdThe legacy README and old service/test targets are preserved under:
graveyard/agency-harness/
Treat that directory as clean-room reference material. It is not part of the active runtime proof, CI target set, or operator startup path.
Prefer small changes with clear validation evidence. When adopting old Agency logic, extract the invariant or runtime contract first; do not revive broad old module graphs just to make legacy surfaces compile.
204 commits
5 commits
Rust
97.9%
JavaScript
1.1%