rwkv-agent

KO | EN

gitlite — search

#computer-vision #cpp #deep-learning #web #markdown #ai-agents #self-hosted #artificial-intelligence #monitoring #codex

(★ 9)

Local-first realtime web retrieval, reproducible search benchmarks, and evidence-grounded answers for RWKV.

File Explorer

ZIP Download

- - bug.yml
  - config.yml
  - feature.yml
- - ci.yml
  - release-cli.yml
- pull_request_template.md
- - cases.jsonl
  - README.md
- - cases.jsonl
  - manifest.json
  - README.md
- - dev_manifest.json
  - dev_trace.jsonl
  - ood_manifest.json
  - ood_trace.jsonl
  - sealed_manifest.json
- __init__.py
- analyze_long_knowledge_threshold.py
- build_retrieval_dev_v2.py
- candidate_rerank.py
- long_knowledge_hybrid.py
- long_knowledge_metrics.py
- long_knowledge_schema.py
- query_formation.py
- README.md
- realtime_web_retrieval.jsonl
- realtime_web_retrieval_audited_v2.jsonl
- realtime_web_retrieval_audited_v2_manifest.json
- realtime_web_retrieval_dev_v2.jsonl
- realtime_web_retrieval_dev_v2_manifest.json
- retrieval_failure_attribution.py
- retrieval_metrics.py
- retrieval_schema.py
- run_candidate_admission_bench.py
- run_candidate_rerank_bench.py
- run_long_knowledge_bench.py
- run_long_knowledge_dense_bench.py
- run_long_knowledge_hybrid_bench.py
- run_query_formation_bench.py
- run_realtime_retrieval_bench.py
- run_searxng_engine_bench.py
- run_web_extraction_bench.py
- searxng_engine_bench.py
- sweep_long_knowledge_dense_rrf.py
- sweep_long_knowledge_rerank_rrf.py
- web_extraction.py
- web_extraction_cases.jsonl
- web_extraction_schema.py
- - chat_state_throughput_ab_v1-batched-c1-c4-c8-c16.json
  - chat_state_throughput_ab_v1-batched-health.json
  - chat_state_throughput_ab_v1-c1-c4-c8.json
  - chat_state_throughput_ab_v1-unified-c1-c4-c8-c16.json
  - mixed_scheduler_v1-baseline-c8-c16.json
  - mixed_scheduler_v1-unified-c8-c16.json
  - mixed_scheduler_v1-unified-health-after.json
  - mixed_scheduler_v1-unified-health.json
  - run_g1i_production_matrix.py
  - sidecar_batch_smoke_v1.json
- agent_benchmark_metrics.py
- agent_benchmark_schema.py
- build_wikipedia_title_index.py
- build_wikipedia_xml_index.py
- evaluate_s_level_bench.py
- general_agent_loop_v1.jsonl
- greedy_tool_call_repeat_v1.json
- greedy_tool_call_template_matrix_v1.json
- knowledge_hybrid_shadow_v1.json
- long_horizon_agent_v1.jsonl
- long_text_agent_e2e_v1.json
- long_text_qa_smoke_v1.json
- run_agent_benchmark_metrics.py
- run_command_state_loop_bench.py
- run_general_agent_loop_bench.py
- run_knowledge_hybrid_shadow.py
- run_long_horizon_agent_bench.py
- run_mixed_scheduler_benchmark.py
- run_semantic_tool_gate_bench.py
- run_sidecar_batch_smoke.py
- s_level_targets.v1.json
- semantic_tool_gate_v1.jsonl
- web_recall_5h_v1.json
- web_retrieval_metrics.py
- web_shadow_v1.json
- - cli_smoke.sh
  - install_smoke.sh
  - mock_server.py
- install.sh
- package-release.sh
- README.md
- bench-4080-c-feedback.json
- benchmark.json
- default.json
- production.example.json
- agent-service-v1.openapi.json
- agent-service-v1.schema.json
- debug-trace-v1.schema.json
- execution-plan-v1.schema.json
- long-lived-agent-result-v1.schema.json
- plugin-v1.schema.json
- state-lifecycle-v1.schema.json
- state-reference-v1.schema.json
- stateful-inference-session-v1.schema.json
- statepool-plugin-v1.openapi.json
- usage-record-v1.schema.json
- worker-capability-v1.schema.json
- - Cargo.toml
  - README.md
- - Cargo.toml
  - README.md
- - Cargo.toml
  - README.md
- - Cargo.toml
  - README.md
- - Cargo.toml
- - Cargo.toml
- - app.css
  - app.js
  - index.html
- __init__.py
- generate_website_swarm.py
- run_agent_website_swarm.py
- - compose.yaml
  - README.md
  - settings.yml
- - 0001-rust-only-stateful-runtime.md
  - 0002-statepool-cloud-plugin-boundary.md
- - native-state-memory-scaling.png
  - qwen9-vs-rwkv7-harness.png
- AGENT_BENCHMARK.md
- ARCHITECTURE.md
- BENCHMARK.md
- CODEMAP.md
- CONFIGURATION.md
- DEBUG_TRACE.md
- FRONTEND.md
- HTTP_API.md
- KNOWLEDGE_SETUP.md
- KNOWN_ISSUES.md
- LONG_KNOWLEDGE_BENCHMARK.md
- LONG_TEXT_QA.md
- PROJECT_SPECIFICATION.md
- QUICKSTART.md
- README.md
- RELEASE.md
- REPOSITORY_SURFACE.md
- RUST_MIGRATION.md
- S_LEVEL_BENCHMARK.md
- SCHEDULER.md
- SCHEDULER_REVIEW_NOTES.md
- SERVICE_PIPELINE.md
- STATEPOOL_CLOUD_PLUGIN.md
- - gallery-overview.jpg
  - task-wall-complete.jpg
  - task-wall-released.jpg
  - task-wall-start.jpg
  - web-ui-chat.png
- build_finewiki_aliases.py
- build_finewiki_revision_map.py
- check_public_release.py
- check_statepool_contracts.py
- dev
- index_finewiki_candidate.py
- index_finewiki_page_embeddings.py
- prepare_miracl_long_knowledge.py
- verify_release.sh
- - __init__.py
  - hf_scheduler.py
  - scheduler.py
  - state_pool.py
- - __init__.py
  - batching.py
  - chat_prompts.py
  - citations.py
  - claim_verifier.py
  - data_plane.py
  - data_server.py
  - evidence_admission.py
  - model_client.py
  - page_quality.py
  - persistent_state.py
  - prompts.py
  - query_coordinator.py
  - routing.py
  - session_text.py
  - sidecar.py
  - state_answer.py
  - state_batching.py
  - state_evidence.py
  - state_prompts.py
  - state_runtime.py
  - statepool_drain.py
  - statepool_worker.py
  - tool_executor.py
- - __init__.py
  - classification.py
  - decode.py
  - protocols.py
  - state_snapshot.py
- - __init__.py
  - candidate_index.py
  - config.py
  - db.py
  - finewiki.py
  - g1i_tool_call.py
  - g1i_types.py
  - p4_search.py
  - passage_selection.py
  - search.py
  - search_reasoning.py
  - search_request.py
  - semantic_selection.py
  - text.py
  - wikipedia.py
- - analyzer_cases.jsonl
- test_agent_benchmark_metrics.py
- test_agent_benchmark_schema.py
- test_agent_data_plane.py
- test_agent_hybrid_shadow_baseline.py
- test_agent_web_recall_baseline.py
- test_agent_web_shadow_baseline.py
- test_agent_website_swarm.py
- test_analyzer.py
- test_batching.py
- test_candidate_admission.py
- test_candidate_index.py
- test_candidate_rerank.py
- test_claim_verifier.py
- test_db_search.py
- test_decode_protocol.py
- test_evaluate_s_level_bench.py
- test_evidence_admission.py
- test_finewiki.py
- test_g1i_tool_call.py
- test_general_agent_loop_bench.py
- test_hf_scheduler.py
- test_hybrid_knowledge.py
- test_knowledge.py
- test_knowledge_shadow_bench.py
- test_local_discovery.py
- test_long_horizon_agent_bench.py
- test_long_knowledge_bench.py
- test_long_knowledge_dataset.py
- test_long_knowledge_dense_bench.py
- test_long_knowledge_dense_sweep.py
- test_long_knowledge_hybrid.py
- test_long_knowledge_hybrid_baseline.py
- test_long_knowledge_rerank_sweep.py
- test_long_knowledge_threshold.py
- test_long_text.py
- test_mixed_scheduler_benchmark.py
- test_page_embedding_indexer.py
- test_page_quality.py
- test_passage_selection.py
- test_persistent_state_registry.py
- test_pipeline_layers.py
- test_precision_discovery.py
- test_public_release.py
- test_query_coordinator.py
- test_query_formation_bench.py
- test_realtime_retrieval_baseline.py
- test_realtime_retrieval_bench_data.py
- test_realtime_retrieval_gold_audit.py
- test_realtime_retrieval_runner.py
- test_realtime_search.py
- test_retrieval_failure_attribution.py
- test_retrieval_metrics.py
- test_run_command_state_loop_bench.py
- test_runtime_contracts.py
- test_scheduler.py
- test_search_reasoning.py
- test_search_reasoning_baseline.py
- test_search_request.py
- test_searxng_engine_bench.py
- test_semantic_selection.py
- test_session_text.py
- test_state_runtime.py
- test_statepool_worker.py
- test_text.py
- test_web_discovery.py
- test_web_extraction.py
- test_web_extraction_baseline.py
- test_web_extraction_schema.py
- test_web_shadow.py
- test_website_swarm_demo.py
- test_wikipedia.py
- test_wikipedia_local_index.py
- test_wikipedia_xml_index.py
- - api-client.js
  - app.js
- - api-client.ts
  - app.ts
- - api-client.test.mjs
  - mock-controller.mjs
- app.css
- index.html
- tsconfig.json
.env.example
.gitignore
Cargo.lock
Cargo.toml
CHANGELOG.md
CLAUDE.md
COMPATIBILITY.md
CONTRIBUTING.md
DEVELOPING.md
GOVERNANCE.md
LICENSE
pyproject.toml
README.md
ROADMAP.md
SECURITY.md
THIRD_PARTY_NOTICES.md
uv.lock

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

**A local, state-native Agent runtime for RWKV models**

RWKV State Agent is a fully local, general-purpose Agent built around RWKV recurrent state. It reuses computed context across chat turns, executes strict Tool Calls, and advances many isolated Agent States through one AMD Radeon GPU scheduler.

> **One-line pitch:** a private local Agent that remembers through recurrent state, acts through bounded tools, and can advance 100 isolated jobs on one Radeon GPU without sharing their context.

> **Optional cloud story:** the assistant stays resident, but the GPU does not
> have to. The opt-in StatePool Cloud Plugin turns long-lived recurrent State
> into a cloud-native scheduling and lifecycle primitive while leaving the
> default local assistant unchanged.

## Release status

The current public beta is [`v0.3.0-beta.2`](https://github.com/123123213weqw/rwkv-agent/releases/tag/v0.3.0-beta.2). It includes the Rust control plane, local Web UI and CLI, recurrent-State runtime, bounded tools, AMD ROCm configuration, benchmarks, and reviewed evidence summaries.

**New in Beta 2:** the canonical Rust service pipeline now accepts versioned
TaskSpec requests, owns durable task identity, resume/cancel, strict streaming
termination and dependency readiness, and exposes an opt-in local Debug Trace.
Debug capture is release-default `off`; `full` traces contain private bodies and
must remain local. The CUDA model Sidecar and retrieval Data Plane are still
narrow external Python providers, and this release makes no new load-capacity or
production-deployment claim. See
[`docs/SERVICE_PIPELINE.md`](docs/SERVICE_PIPELINE.md) and
[`docs/DEBUG_TRACE.md`](docs/DEBUG_TRACE.md).

The architecture and operating contract are documented in [`docs/PROJECT_SPECIFICATION.md`](docs/PROJECT_SPECIFICATION.md).

## Optional StatePool Cloud integration

This repository contains only the disabled-by-default StatePool **client and
native RWKV Worker interface**. Without `--cloud-plugin`, no plugin client is
constructed and the existing local routing path is unchanged. Wire types,
JSON Schema and OpenAPI stay here so each `rwkv-agent` release can declare an
exact client compatibility window.

The independently released
[`statepool-cloud`](https://github.com/123123213weqw/statepool-cloud) repository
owns the cloud control plane, PostgreSQL/S3 backends, OpenAI/vLLM adapter,
Compose/Helm/KEDA deployment, FinOps benchmarks and competition evidence.
No cloud control-plane source or upstream inference engine is vendored here.

- client configuration: [`docs/STATEPOOL_CLOUD_PLUGIN.md`](docs/STATEPOOL_CLOUD_PLUGIN.md)
- client schemas/OpenAPI: [`contracts/statepool-plugin-v1.openapi.json`](contracts/statepool-plugin-v1.openapi.json)
- cloud product and deployment: [`statepool-cloud`](https://github.com/123123213weqw/statepool-cloud)

## What you can verify

- **Ordinary chat:** greetings remain on the direct path and do not call a tool.
- **State-native memory:** the same session reuses its RWKV recurrent state and only computes new turns.
- **Strict Tool Calls:** `run_command`, `knowledge_search`, `long_text_qa`, and optional `web_search` use one bounded call per model turn.
- **Multi-step execution:** Action → Tool → Observation → repair → verification → final answer stays in one owned State.
- **True streaming:** tokens flow from the ROCm Sidecar through the Rust Controller to the Web UI as NDJSON deltas.
- **Independent parallel States:** jobs have separate Owner, State, Session, Workspace, Trace, and release lifecycle.
- **Live task wall:** [`/tasks`](http://127.0.0.1:18120/tasks) shows real Controller runs and their Route, State, Tool count, status, and elapsed time.
- **100-job proof:** 100 independent website tasks completed with physical decode concurrency 32, without shared context or prebuilt answers.

## Native state-memory scaling

![Native inference-state memory scaling for Qwen3.5-9B and RWKV-7.2B](https://raw.githubusercontent.com/123123213weqw/rwkv-agent/main/docs/assets/native-state-memory-scaling.png)

The current RWKV engine measures **32.5 MiB per recurrent State slot** and reserves **4.06 GiB for 128 slots**, independent of history length. From the frozen [Qwen3.5-9B configuration](https://huggingface.co/Qwen/Qwen3.5-9B/blob/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json) and the [Transformers 5.14.1 cache layout](https://github.com/huggingface/transformers/blob/v5.14.1/src/transformers/models/qwen3_5/modeling_qwen3_5.py), the Qwen FP16 native-cache lower bound is 49.5 MiB of fixed linear-attention state plus 32 MiB per 1K history tokens. At 4K history this is 177.5 MiB per session: the current RWKV slab crosses over near 24 concurrent sessions and uses **4.3× less inference-state memory at 100 sessions**.

This figure isolates native inference State/KV memory; model weights, compute workspace, allocator overhead, and durable transcript storage are excluded. It is an analytical cache-layout comparison backed by measured RWKV State sizes, **not** a measured Qwen native-runtime or whole-process VRAM result. The existing low-concurrency process baseline still records Qwen at 18,339 MiB Ready versus RWKV at 22,393 MiB Ready; see the frozen [Phase 0 comparison](bench/artifacts/long-lived-phase0-live-v1/comparison-and-decision.json).

## Frozen long-horizon comparison

![Qwen3.5-9B NF4 versus RWKV-7.2B frozen Harness comparison](https://raw.githubusercontent.com/123123213weqw/rwkv-agent/main/docs/assets/qwen9-vs-rwkv7-harness.png)

The models use the same frozen dataset and runner, but different precision and runtimes, so this is an observed Harness comparison rather than an architecture-only ranking. See the [source CSV](bench/baselines/long_horizon/qwen9-nf4-vs-rwkv7-fp32io16-v7r4.csv) for full result hashes and preserved metrics.

## Architecture

```mermaid
flowchart LR
    UI["Web UI / Rust CLI"] --> C["Rust Agent Controller"]
    C --> G["Semantic Tool Gate\nprewarmed root State"]
    G -->|direct chat| S["RWKV G1I 13.3B\nROCm Sidecar"]
    G -->|tool task| L["Strict Agent loop"]
    L --> T["Tool data plane"]
    T --> O["Observation"]
    O --> L
    L --> S
    C --> Q["Task wall / Trace"]
    S --> P["Recurrent State pool\nunified ready queue"]
    P --> GPU["AMD Radeon gfx1100\nROCm 7.2.1"]
```

The model never emits a large planner document. Each model turn produces either one strict Tool Call or one final answer. Runtime code owns budgets, state identity, workspace boundaries, validation, metrics, and release.

## Core tools

| Tool | Purpose |
|---|---|
| `run_command` | Controlled file operations, commands, tests, and artifact validation |
| `knowledge_search` | Local long-term knowledge index |
| `long_text_qa` | Evidence-grounded questions over pasted long text |
| `web_search` | Optional retrieval when the user explicitly requests current Web information |

State management, task metrics, artifact verification, and the task wall are runtime features, not model-visible tools.

## AMD Radeon implementation

### Model and runtime

| Component | Verified configuration |
|---|---|
| Model | RWKV-7 G1I Preview4922 13.3B |
| Parameters | 13,269,245,952 |
| Context | 12,288 tokens |
| Decode | Greedy |
| Precision | FP16 |
| GPU | AMD Radeon `gfx1100`, 51,522,830,336 bytes VRAM |
| ROCm | 7.2.1 |
| PyTorch | 2.9.1 ROCm build |
| Backend | RWKV-7 HF native recurrent backend |
| Stable physical batch | 32 |

The source checkpoint is distributed separately by RWKV and is not committed to Git. The verified checkpoint is [RWKV-7 G1I Preview4922 13.3B](https://huggingface.co/BlinkDL/temp-latest-training-models/blob/main/rwkv7-g1i_preview4922-13.3b-20260720-ctx12288.pth). The ROCm conversion used the RWKV HF adapter's AMD branch at commit `7fd669809cefa97c81f8e8cda6c6a59f9cf04635`.

### Inference optimizations

1. **Recurrent session state** — later turns continue from GPU State instead of prefilling the complete transcript.
2. **B1 zero-copy continuation** — one-row decode reuses its cache directly instead of gathering and scattering about 62 MiB of State every token.
3. **Vectorized chunk continuation** — existing recurrent cache can enter the native chunk-prefill path instead of falling back to token-by-token continuation.
4. **Unified ready queue** — prompt chunks and decode rows are scheduled across independent requests up to physical batch 32.
5. **Prewarmed semantic gate root** — static routing instructions are prefetched once; every request forks the root and appends only dynamic input.
6. **Short direct-chat budget** — ordinary chat uses a 96-token ceiling and a visible-answer stop envelope.
7. **End-to-end streaming** — each Greedy update becomes a Sidecar delta and is forwarded directly by Rust; the Web UI does not simulate typing.

## Verified AMD results

### State reuse and ordinary chat

| Measurement | Before | Gate 5 |
|---|---:|---:|
| First-turn wall time | 63.648 s | **8.240 s** |
| Second-turn wall time | 63.784 s | **7.028 s** |
| First-turn speedup | — | **7.72×** |
| Second-turn speedup | — | **9.08×** |
| Second-turn State reuse | no | **yes** |
| Output tokens, first / second | 78 / 97 | **6 / 6** |

B1 streamed and non-streamed continuation produced identical text, token IDs, and stop reason. B1 gather/scatter workspace was zero.

### Independent State scaling

| Resident jobs | Physical batch | Aggregate output throughput |
|---:|---:|---:|
| 1 | 1 | 1.6661 tok/s |
| 4 | 4 | 6.2571 tok/s |
| 8 | 8 | 11.6955 tok/s |
| 16 | 16 | 19.7453 tok/s |
| 32 | 32 | **29.9838 tok/s** |
| 64 | 32 | 25.7295 tok/s |
| 100 | 32 | 24.0266 tok/s |

At B32, concurrent execution was **10.5792×** faster than the same 32 jobs run serially, while all Greedy outputs remained exact. Decode GPU Busy averaged **88%**, peaked at **98%**, and peak VRAM was **37.87 GB**. A same-protocol Gate 5 warm regression measured 29.3472 tok/s, within 2.123% of the frozen baseline.

### 100 independent Agent jobs

- 100 unique prompts, owners, sessions, States, and workspaces;
- physical decode concurrency 32;
- 100/100 valid final artifacts;
- 0 task failures in the frozen final run;
- 0 protocol leaks, context crossovers, or prebuilt answers;
- average / peak GPU Busy: 86.205% / 99%;
- peak VRAM: 48,093,364,224 bytes;
- final waiting, busy, decoding, and State counters: zero.

The correct claim is **“100 independent tasks · physical concurrency 32”**, not “100-way concurrent decode.”

## Quick evaluation path

With the AMD services already configured:

```bash
# Terminal 1 — model Sidecar
rwkv-g1i-sidecar --host 127.0.0.1 --port 18118

# Terminal 2 — Knowledge / Long Text / optional Web data plane
rwkv-agent-data-plane \
  --host 127.0.0.1 \
  --port 18121 \
  --model-urls http://127.0.0.1:18118

# Terminal 3 — Rust Controller and embedded Web UI
./target/release/rwkv-agent-server-rs \
  --host 127.0.0.1 \
  --port 18120 \
  --model-urls http://127.0.0.1:18118 \
  --data-plane-url http://127.0.0.1:18121 \
  --session-dir ./var/sessions \
  --chat-state-capacity 3 \
  --direct-chat-max-tokens 96
```

Open:

- conversation: <http://127.0.0.1:18120/>
- live task wall: <http://127.0.0.1:18120/tasks>
- health: <http://127.0.0.1:18120/health>

Suggested evaluation:

1. Send `你好，只回复你好。` and observe direct streaming with no Tool Call.
2. Send a second message in the same session and inspect `state reused`.
3. Ask the Agent to create a small project, run tests, repair a failure, and verify the result.
4. Paste a long document and ask for an answer with source evidence.
5. Keep `/tasks` open during the runs and inspect live status transitions.
6. Replay the frozen B32 and 100-State benchmarks on the AMD host.

## Install and dependencies

### Required

- Ubuntu 24.04 or another ROCm-compatible Linux distribution;
- ROCm 7.2.1;
- Python 3.10+;
- ROCm PyTorch, Transformers, FastAPI, and Uvicorn;
- Rust 1.97+ for the Controller and CLI;
- RWKV G1I Preview4922 13.3B weights converted to the verified HF native format;
- approximately 48 GB GPU memory for the verified 13.3B / 100-resident-State profile.

### Python environment

Install ROCm PyTorch using the package source appropriate for the target ROCm version, then install the project:

```bash
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
# Install the official ROCm PyTorch wheel before the project extras.
python -m pip install -e '.[realtime,agent,dev]'
```

### Rust Controller

```bash
cargo build --release --workspace
cargo test --workspace
```

### Radeon Sidecar environment

```bash
export G1I_BACKEND=hf_recurrent
export G1I_HF_MODEL_PATH=/absolute/path/to/preview4922-13.3b/hf-fp16
export G1I_HF_DTYPE=fp16
export G1I_MODEL_ID=rwkv7-g1i-preview4922-13.3b
export G1I_CONTEXT=12288
export G1I_STATE_CAPACITY=132
export G1I_PERSISTENT_STATE_CAPACITY=101
export G1I_MAX_BATCH_SIZE=32
export G1I_PREFILL_CHUNK_SIZE=32
export G1I_BATCH_WINDOW_MS=10
```

Do not commit model weights, private prompts, credentials, `.env` files, or runtime State.

## Repository map

```text
crates/agent-cli/       Rust terminal client
crates/agent-core/      strict Tool protocol, budgets, lifecycle, events
crates/agent-runtime/   sessions, recurrent State, tools, sandbox, research
crates/agent-server/    Rust HTTP Controller, streaming, embedded Web UI
src/rwkv7_scheduler/    ROCm recurrent cache scheduler and State pool
src/rwkv_agent/         Sidecar, data plane, routing, long text, retrieval
demos/                  100-independent-State Agent demonstration
benchmarks/             State scaling, reuse, routing, and Agent regressions
web/                    RWKV conversation workspace, task wall and runtime status UI
docs/                   architecture, setup, benchmark, and specification
```

The main development map is [`docs/CODEMAP.md`](docs/CODEMAP.md). The canonical split-stack startup is documented in [`docs/QUICKSTART.md`](docs/QUICKSTART.md); the verified Radeon configuration is documented here.

## Privacy and execution boundaries

- Model inference, recurrent State, transcript, workspace, and benchmark artifacts stay on the local host.
- Automatic cross-session preference extraction is intentionally disabled for this release.
- `run_command` is opt-in and restricted to a configured workspace. The verified AMD environment uses an unprivileged user, user/network namespaces, PRoot, bounded output, a timeout, and no unsafe fallback.
- The HTTP Controller has no public authentication. Keep it on loopback, use SSH forwarding, or place it behind an authenticated private gateway.
- Optional `web_search` is the only feature that intentionally uses external network sources.

## Known limitations

- The Radeon path uses the HF native recurrent PyTorch backend; the NVIDIA Albatross MMA extension is not claimed to run on ROCm.
- The 40-case semantic routing set scored 95% with zero missed Tool requests and two false Tool routes.
- The live task wall stores only the latest 100 runs in Controller memory and resets on restart; it is an operational view, not a task database.
- Model weights are external to GitHub and must be converted separately.
- Public authentication, distributed scheduling, and automatic long-term personal memory are outside this release.

## Verification and evidence

The project keeps frozen input hashes, runner hashes, raw JSONL, aggregate metrics, ROCm time series, failure traces, videos, and artifact checksums. Run:

```bash
bash scripts/verify_release.sh
```

Reviewed benchmark summaries and screenshots are available under [`evidence/`](evidence/). Raw traces, private prompts, model assets, and machine-local configuration remain outside the repository. Override `EVIDENCE_ROOT` when validating another environment.

## License

[MIT](LICENSE)

rwkv-agent

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Download README