chore(main): release 0.3.0

[github-actions]

🤖 I have created a release beep boop

0.3.0 (2026-02-18)

⚠ BREAKING CHANGES

• Remove all legacy server code, CLI, migrations, and Docker infrastructure. The repo is now a workspace with only the wasm member crate.
• sdk: SDK API completely redesigned
• Evaluation now uses separate containers:
  • Agent container: base image (ghcr.io/platformnetwork/term-challenge)
    with term_sdk installed, runs agent HTTP server
  • Task container: task-specific image (e.g., alexgshaw/fix-git)
    executes commands and runs tests

Features

• 3-validator task distribution, cancel command, and improved error handling (e18083b)
• Add agent transparency and logging system (#101) (5d2e286)
• add always-on server mode with /get_weights endpoint (bb29283)
• add API module structure (f767bf6)
• add assignment monitor for stale validator reassignment (31fbb15)
• add binary caching to validator worker (bbf237e)
• add cache, chain, validation, container, agent, and evaluation modules (ffa9e5f)
• add client and weights modules (8f044de)
• add container cleanup for evaluation containers (e0e90c9)
• add core types and crypto modules (25db2c4)
• add detailed agent status endpoint with all phases and timings (f3dfa7c)
• add detailed error logging for database operations (7eb88ba)
• add DirectDockerBackend and binary agent runner for local bench testing (d84ed75)
• add disable_decay and disable_public_code fields (172223f)
• add dump_instruction function for LLM review with DB storage (b562eea)
• add epoch calculation with custom start block (ebe42fa)
• add folder upload support to term wizard (6e2ae37)
• add forced_weights for manual weight overrides + sort leaderboard by success_rate (5ecfe21)
• add GET /api/v1/agent/{hash}/code endpoint for public code visibility (4c8e1ac)
• add in-container agent execution with platform LLM bridge (d6c4f0a)
• add lib_new.rs and STRUCTURE.md documentation (7deb466)
• add LLM proxy endpoint with validator auth (0b3f647)
• add LLM review instructions API endpoint (bdb82b2)
• add LLM review logs API and configurable system prompt (1506e72)
• add migration 035_llm_review_instructions to EMBEDDED_MIGRATIONS (a4e940d)
• add missing migrations, write_file tool, submit_verdict guard, dual-write columns (13cd948)
• add multi-agent LLM code review system (125efdc)
• add multi-file package submission support (d1d8cba)
• Add OpenAI Responses API support (GPT-4.1+/GPT-5.x) and use real provider costs (2738dd4)
• Add OpenRouter prompt caching support with usage tracking (f8924d2)
• add ProposalManager for P2P agent proposal flow (fe47817)
• add public API endpoints for pending submissions and validator assignments (89cb608)
• add real-time task logging to platform server (54b1b42)
• Add real-time task streaming cache for live evaluation progress (e61556c)
• add requirements.txt support for package compilation (a1e655b)
• add retry logic for HTTP 429 errors from Chutes API (02a208b)
• add retry loop for platform-server connection (30s interval, 5 attempts) (fb23d26)
• add status, total_cost_usd and success_rate to leaderboard (5716384)
• add storage module structure (08097ac)
• add term-sudo CLI + remove leaked API key (eca7fd7)
• add Terminus-2 agent adapted for Term SDK 2.0 (e72c7eb)
• add time-based reward decay system (20d978d)
• add tool_calls/function calling support through platform bridge (0133db9)
• add validator worker for evaluation recovery and polling (6c9af2d)
• add verbose agent logging and evaluation resume support (4415307)
• add verbose logging for container backend and compilation (9886e1f)
• add WebSocket broker backend for container management (1742947)
• add weight and submitted_at to leaderboard responses (d6d8e37)
• add worker, task, admin, and server modules (98779c2)
• Add ZIP package support to submit wizard (52e6e14)
• Add ZIP package support to submit wizard for Bridge API (493c40a)
• api: add public GET /api/v1/rules endpoint for LLM rules (f8b0968)
• api: add subnet settings for evaluation suspension (0f0c158)
• api: add subnet status and improve rules endpoints (ad47426)
• AST-based plagiarism detection system (e6af359)
• async task logging system with real-time tracking and recovery (ca3a09b)
• Auto-detect and collect submodules from requirements.txt (bc2fe92)
• auto-evaluation after agent submission (ba1f911)
• automatic cleanup of orphan Docker volumes (cf148a3)
• broadcast new_submission event to validators via WebSocket (e05646f)
• change minimum tasks threshold from 8/validator to 20% success rate (#108) (cfd38e6)
• checkpoint: add checkpoint4 as active evaluation checkpoint (#96) (4af4164)
• checkpoint: add checkpoint5 and remove dead code (#103) (5de0a82)
• cleanup stale task containers at validator startup (8da0f7b)
• cli,wasm: add ratatui TUI monitor and agent storage with logging (#5) (ce0b35d)
• cli: add centralized TermClient for API calls (0ef1dcd)
• cli: default platform URL to https://chain.platform.network (14211c6)
• concurrent task execution (30 tasks, 4 concurrent per agent) (d14cc55)
• container: add LLM reviewer Docker image builder (#107) (ca71a35)
• create directory structure and util module (ec597d9)
• distributed task evaluation and validator readiness system (bdcf46d)
• Docker-in-Docker fixes and glibc compatibility (75a81c6)
• Docker-isolated compilation + binary_ready notification to validators (ca5ecb7)
• enforce minimum 10000 TAO stake for validator assignment (320585d)
• examples: add code validator agent with LLM analysis (6e25de2)
• expose llm_review_status in leaderboard API response (b29a1db)
• extract pg_storage.rs and api.rs into submodules (66e6724)
• fetch whitelisted validators from platform-server (e65d81e)
• fix similarity calculation, add /similarities endpoint, expose plagiarism/LLM status everywhere (7d9f838)
• fully integrate ProposalManager into submission flow (0576970)
• get validator count from platform-server for distributed evaluation (5204f53)
• implement distributed evaluation system - ALL validators must evaluate (1a7684c)
• implement full evaluation flow with LLM review (fdb56cf)
• implement llm_review_called pooler pattern for LLM review worker (#109) (66f0b7b)
• implement P2P progress sharing system (f30978d)
• Implement StaticX for portable agent binaries (90652ea)
• improve API key redaction with comprehensive patterns (b26e865)
• improve LLM proxy cost tracking and add Grok provider support (395fd9b)
• increase MAX_CONVERSATION_TURNS from 50 to 150 (350d9d7)
• infrastructure failure handling with automatic task reassignment (866413e)
• infrastructure failure handling with automatic task reassignment (69429c0)
• install full SDK with LLM support during compilation (8674eac)
• LLM code review via Kimi-K2.5-TEE in isolated Docker container (86120ce)
• LLM proxy with cost tracking, task observability APIs, streaming support (2eb5fb0)
• maintenance: add periodic maintenance task + require all validators for consensus (b0e1713)
• Make temperature parameter optional (70513ba)
• migrate all CLI commands to use bridge routes (5299263)
• migrate bench run to use binary agent system (1915444)
• Migrate submissions API from platform-server to term-challenge (f17e10c)
• migrations: add llm_rules table for Multi-Agent Review system (5de9fb7)
• migrations: add subnet_settings table for owner controls (afef7f4)
• migrations: register 022_llm_rules in embedded migrations (06dd2f9)
• move validator and task assignment to compile_worker (7958323)
• Multi-Agent LLM Code Review System (21394ee)
• multiply task timeout by 3 from config (2cc59f5)
• multiply task timeout by 3 from config (6a510a8)
• plagiarism: exempt original authors from plagiarism detection (80273d6)
• plagiarism: exempt original authors from plagiarism detection (ebcbe6c)
• production-ready task execution with real Terminal-Bench (b4efd99)
• PyInstaller binary compilation for agents (c58a29b)
• Redact API keys and sensitive data in public code reveal (8e7324c)
• redact API keys from code before LLM review (f28522d)
• replace Docker-based LLM review with conversational Rust agent (e377fdf)
• replace epoch-based submission rate limit with time-based (3.6h cooldown) (6216f33)
• replace validator whitelist with stake-based auth via metagraph (bfb91f0)
• review: add dynamic rules fetching from API (744fec8)
• review: add retry logic for rules API fetching (5cefa1a)
• review: add Rust multi-agent consensus code review system (fab2661)
• sdk: add LLM client class with caching support (c98bd4e)
• sdk: add raw_chat() method for full control over LLM request body (ea96ff6)
• sdk: preserve raw_arguments on JSON parse failure (8e7fe10)
• sdk: SDK 2.0 with agent-controlled execution model (41b86a4)
• sequential similarity->rules review, dynamic pool, separate similarity instructions (3a52e94)
• show pending agents in status command (b873507)
• storage: add LLM rules storage methods (26def6b)
• storage: add SubnetSettings model and storage methods (1a07290)
• sudo endpoints + LLM proxy via validator (ba8a799)
• Support max_completion_tokens parameter for o-series models (e51b6e0)
• swe_forge: add SWE-Forge client module for term-executor communication (d19add1)
• swe_forge: implement SWE-Forge evaluation via term-executor and Basilica (69097f1)
• swe_forge: replace term-benchmark with SWE-Forge integration (e2bcdcc)
• synthetic: add autonomous synthetic dataset generation system (#97) (176fc78)
• update CLI to use bridge API for submissions (f47c444)
• use ContainerBackend for validator worker task execution (31d7022)
• use secure broker for building compiler image (be617a2)
• validate miner_hotkey is SS58 format in /evaluate endpoint (f56c6d6)
• validation: auto-validate agents via LLM review (#106) (114c8f3)
• validator worker loads real tasks from terminal-bench@2.0 (aeb1cdf)
• validator: add timeout retry with local and server-side reassignment (375575b)
• verify miner UID registration on subnet 100 before accepting submissions (61b0519)
• WASM-only architecture with sudo challenge management and P2P distribution (#9) (aa1b7fc)
• wasm: add decentralized challenge modules with LLM review, AST validation, and route handlers (#7) (71c55de)
• winner-takes-all weight calculation with manual validation (6915096)
• worker: improve LLM review worker startup logging and error handling (0403019)

Bug Fixes

• add --break-system-packages flag to pip install in compiler (7dcbdec)
• add --break-system-packages to httpx pip install (f228ba6)
• add 15 min timeout to LLM HTTP clients and handle empty responses (7b3a11f)
• add cache directory mapping for Docker-in-Docker (c39d5b4)
• add entrypoint override to bench and compiler containers (ee71998)
• add entrypoint override to prevent container from stopping early (b2dd945)
• add FLOAT8 cast to RETURNING clause in update_submission_cost (c514f2c)
• add FLOAT8 casts for all REAL column reads in pg_storage (8ec0efd)
• add httpx to PyInstaller hidden imports (b7d25a6)
• add LLM_MODEL env var support and reduce log noise from /status requests (f487693)
• add migration 009 for validator_assignment status column (17886de)
• add migrations to Docker build context for include_str! (f9c5413)
• add Pong variant to BrokerResponse for auth success parsing (dad55b4)
• Add stream:true to bridge request for SSE streaming support (5a94e2c)
• Add stream:true to bridge request for SSE streaming support (6c0b23e)
• add term_sdk to allowed third-party modules whitelist (57af0ec)
• add validator_assignments table and use claude-haiku-4.5 for reviews (97fdff7)
• address validation feedback - consistent defaults and security hardening (2a10fbf)
• align default timeout with Harbor/terminal-bench (180s) (2b41e9c)
• align LLM proxy signature format with central server (ca40138)
• allow compilation for flagged agents when LLM review approved (583c14b)
• always build compiler image, never pull from Docker Hub (337d345)
• always run tests even if agent times out (11ab582)
• api: update Cortex provider URL to api.cortex.foundation (#99) (53f13f6)
• Auto-collect transitive dependencies (pydantic, tiktoken, etc.) (fa34656)
• broker: add retry logic for WS connection failures (1188c30)
• build compiler image on demand if not found during compilation (12de066)
• calculate evaluation costs from llm_usage table (e5ac0aa)
• cast f64 to f32 for PostgreSQL REAL columns in cost updates (08c3613)
• change eligibility from 8 tasks/validator to 8 tasks total (1eb9812)
• check if PyInstaller exists before installing (78a648d)
• check multiple SDK paths for full SDK installation (cd9ddb0)
• ci: separate coverage job to prevent cancellation (7ba740d)
• cleanup orphan compiler containers at startup and use UUID in names (ec2c026)
• cleanup_stale_claims type error (use make_interval with i32) (91466cd)
• clippy warnings (ef98763)
• cli: read best_score from API leaderboard response (0110c25)
• compilation: remove bench/container/bollard references after module removal (f8712e5)
• copy docker directory into images for compiler image building (ffb42fb)
• correct Chutes API endpoint from api.chutes.ai to llm.chutes.ai (#105) (bb6d8b6)
• correct iteration increment in terminus_2 agent loop (ddca36c)
• correct signature message for my_jobs endpoint (cd079d7)
• count new unique code in plagiarism total_nodes to dilute similarity (d188ffe)
• create pending_evaluations after compilation + exclude evaluation_failure from task counts (a8646c3)
• decay based on last task completion + disable_decay flag + heartbeat URL (02cbadf)
• decrypt API key before sending to OpenRouter (4e78be0)
• deduplicate agent logs by tracking last printed line (6d6abcd)
• default to openrouter if llm_provider is empty (5f78b3c)
• detect active validators by started_at, not just completed_at (f48a153)
• detect and abort stuck agents with consecutive empty responses (848a3cc)
• disable /evaluate in server mode, use /validators endpoint (a4357f1)
• ensure binutils is installed before PyInstaller (af6a776)
• estimate epoch from time when not provided in get_weights (ca4624b)
• expire: calculate consensus with 2+ validators when window expires (b147962)
• filter evaluation progress by validator_hotkey (2b44209)
• filter out completed agents from validator jobs (8a5a21e)
• handle non-JSON LLM provider responses (504/503/502 errors) (65ea285)
• Harbor-compatible test verification and dynamic challenge_id (319fdd6)
• implement missing FakeStorage trait methods for tests (8385f10)
• improve broker WS error message to include URL (b8f7877)
• improve Docker error logging for debugging task container failures (1bffd2a)
• improve test stability and add Chutes API support (#104) (0d513b3)
• include all migrations (006-009) in embedded migrations list (83c4245)
• increase limits and reduce validators (dca4dd5)
• infinite retry loop for platform-server, no fallback (b520bee)
• leaderboard: show only fully evaluated submissions (status='completed') (7b7ec1c)
• LLM plagiarism review now loads reference agents' code for side-by-side comparison (f1a3701)
• LLM rejection flags agent for manual review instead of blocking (516cebe)
• llm_review: clarify that Response.cmd() is ALLOWED (1668c6d)
• llm-review: use EXTRA_CHUTES_API_TOKEN environment variable (#111) (e270d91)
• log task results immediately after each task completes (5823384)
• make queue test tolerant of Docker permission errors in CI (2d0210a)
• make validator worker optional, support VALIDATOR_SECRET_KEY (59c3288)
• map cache directory paths for Docker-in-Docker mounts (5c4979d)
• Move baseagent submodule to examples/baseagent (6c8e850)
• P2P validators sync and consensus logic (ec9552e)
• parse docker_image from task.toml [environment] section (0ece103)
• parse pending_jobs field correctly in validator_worker (146860e)
• pass LLM proxy env vars to agent binary process (d630d36)
• per-agent similarity scoring + block compilation unless plagiarism cleared (32eec4f)
• preserve HTTP status codes in LLM proxy error handling (f6aa7bb)
• prevent agents from reading test artifacts during execution (dbe54ac)
• prevent agents from reading test artifacts during execution (2a016b2)
• Prevent command injection in task evaluator (e7aac09)
• Prevent command injection in task evaluator (3525dcb)
• prevent duplicate jobs and add container concurrency limit (b3e0276)
• process detection with exec_shell for pipe commands (ca27369)
• PyInstaller extraction issues in task containers (f73650a)
• re-declare TERM_REPO_PATH ARG in Dockerfile.server runtime stage (5bad625)
• reduce docker pull log spam (1286d60)
• Remove agent wrapper to preserve 'from future' imports (d088b44)
• remove auth requirement from /p2p/outbox endpoint (395dc5e)
• remove fallback mode - skip evaluation if no assigned tasks (f8f7a86)
• remove global timeout from SDK - let agent builders define their own (f0ee67f)
• remove max_steps and timeout_secs from SDK - let agents manage their own limits (108d262)
• remove orphan llm_client.rs, guard short hash, validate empty API key (65c100e)
• remove restrictive cap_drop, run containers as root (8bc2f75)
• Remove unnecessary borrow in clippy lint (5277a64)
• replace placeholders with real implementations (cbb9393)
• Restore docker/ directory required for Dockerfile build (a2bf3b0)
• Restore Dockerfile.compiler required for agent compilation (e11da2b)
• Restore Dockerfile.compiler required for agent compilation (fc11b3f)
• Restore full Cargo.toml with all dependencies (6133234)
• retry all Chutes API errors (429, 5xx, network) with exponential backoff (1b8b766)
• retry: detect test execution failures and resource errors (075b90a)
• run verification tests from /workspace directory (5059f5a)
• sdk: add remaining_steps and remaining_secs to AgentContext (eb6fd06)
• security and reliability improvements (0fe24c9)
• security hardening and consistent defaults (07a928d)
• send all required fields to log_task API (f23ec72)
• send BROADCAST_SECRET header for event broadcasts (05d526c)
• set manually_validated based on LLM review result (true if approved, false if rejected) (49059bc)
• set total_validators=2 when queueing submissions + reset window on requeue (3b0d75f)
• stale: only detect stale assignments for pending agents (eb91952)
• stop agent loop on cost_limit_exceeded and empty responses (f685359)
• storage: add llm_review_status filter to LLM review pooler (#110) (d69309a)
• support binary test files (.pt, .bin, etc.) (95ab40f)
• support both 'done' and 'task_complete' in agent response (9243cbd)
• support package/workspace submissions in LLM review (8edc719)
• term-sudo uses bridge routes via chain.platform.network (de42398)
• timeout: apply 1.3x multiplier to agent timeout and fix retry detection (5db6be0)
• timeout: websocket timeout 300s, case-insensitive retry detection, detailed messages (1b33dc6)
• Transform system messages for OpenRouter+Claude requests (6ff4b4f)
• update BrokerError to match platform's ContainerError enum format (496a582)
• update CLI to use correct signature format for agent submissions (c31d816)
• update secure-container-runtime to remove cap_drop restrictions (a10b952)
• Use --collect-all instead of --collect-submodules for user packages (46e2df4)
• use /app as standard working directory (matching harbor) (d58c349)
• use /proc for process detection instead of ps (85fce86)
• use /workspace as default working directory instead of /app (546af74)
• use agent_binary column name, better error logging (273f0ef)
• use bash instead of sh for Harbor test scripts (0892f5d)
• use CHALLENGE_UUID for broker authentication (2e429a7)
• use CONTAINER_BROKER_WS_URL instead of BROKER_WSS_URL (0db1eef)
• use correct timeouts from task config (6b1c812)
• use DateTime<Utc> for plagiarism_checked_at column (f22fe59)
• use exec_shell instead of exec to avoid double shell wrapping (df0cd46)
• use fixed 30 task count and deterministic task selection (c1210ac)
• use GHCR image for evaluator instead of term-challenge/base (54ff7f5)
• use i32 for llm_review_logs id column (SERIAL type) (d53b831)
• use miner's API key directly for LLM security review (36eff85)
• use python:3.11 full image for PyInstaller (includes binutils) (a062d3e)
• use simple release type with manifest config (4876e3c)
• use snake_case serde rename for BrokerResponse to match platform protocol (999f9ba)
• validation: enforce LLM review before evaluation queueing (522d159)
• validation: require LLM review approval before compilation (2b5a8c8)
• validator: add global timeout to force-kill hung tasks (738214b)
• validator: detect stuck validators and improve reassignment logic (06622f5)
• validator: kill agent process before running tests on timeout/incomplete (4322340)
• Validators now evaluate only their assigned tasks (10 each) (ac8828a)
• wasm: remove dead code and stale documentation (9c222ee)
• weights: only allow completed agents to receive emissions (8fa4b22)

Performance Improvements

• run tasks concurrently (2 per agent, 8 max global) (86f7efc)

Code Refactoring

• Complete codebase cleanup (c5b694b)
• Complete codebase cleanup and modular reorganization (ff9c4d8)
• docker: remove Docker dependencies, container/ and bench/ modules (8f17e15)
• InContainerAgent - API key stays server-side, LLM proxy pattern (166b93f)
• integrate new module structure into lib.rs and fix compilation (59ac5d2)
• remove automatic prompt caching from SDK, let users implement manually (2b469ee)
• remove direct Docker backend, use container names for HTTP communication (79120ea)
• remove duplicate EvaluateRequest/Response, reference server.rs (54a661c)
• remove hardcoded rules from multi-agent review prompts (1ddbe43)
• remove hardcoded verification rules, make analysis LLM-driven (7c16590)
• remove local platform-repo copying, use git dependency from Cargo.toml (e52d711)
• remove submit_result, auto-detect task completion in log_task (1763ece)
• remove unused /task mount from containers (563c8c7)
• replace bash LLM reviewer with agent-based Docker reviewer (83685dc)
• replace server/Docker architecture with WASM challenge module (#1) (60533bc)
• standardize challenge ID to term-challenge, remove CHALLENGE_UUID (635e53c)
• strip legacy monolith, retain WASM-only challenge module (#3) (10f4eee)
• use secure-container-runtime types from platform (c3bfc22)
• use two-container architecture for evaluation (d8ab393)

Documentation

• add API key security section with rate limiting guidance (574800a)
• Add baseagent reference and AGENTS.md note to agent-development guide (d5b1329)
• Add bug bounty rewards section to AGENTS.md (cc6cc68)
• remove --api-key from CLI examples, API key now managed in agent code (7b7eff2)
• remove remaining_steps/remaining_secs from documentation and examples (40197be)
• replace ASCII diagrams with Mermaid flowcharts (631ee4e)
• server: add transparency API endpoints to startup logging (3d9b597)
• simplify Mermaid diagrams for better GitHub rendering (bdb9b05)
• update multi-agent review to reflect LLM-driven analysis (90de7a3)
• Update README features and Rust version (525ebe7)
• update README with missing features and architecture (1ecd09f)

Miscellaneous

• add banner image (0fb2a38)
• add migration for forced_weights table (1f26565)
• addressed code review suggestions (9fdbd2e)
• cluster indexation (9fd6e6b)
• cluster indexation - AGENTS.md, git hooks, CI workflow, versioning (c097657)
• increase LLM review max turns to 50 (853a460)
• limit commit-search-depth for release-please (252880d)
• main: release 0.1.0 (70a8940)
• main: release 0.1.0 (9bb29fd)
• main: release 0.2.0 (c010755)
• main: release 0.2.0 (a4c28eb)
• main: release 0.2.1 (3112a36)
• main: release 0.2.1 (3d6265e)
• main: release 0.2.2 (ccbdc8c)
• main: release 0.2.2 (22b7384)
• main: release 0.2.3 (33815e3)
• main: release 0.2.3 (1b6610b)
• main: release 0.2.3 (7c3c79c)
• Remove obsolete STRUCTURE.md (migration completed) (b06723d)
• remove stale Dockerfile.server after server architecture removal (e6f673b)
• Remove unused test files and fix AGENTS.md documentation (4c62212)
• restart CI pipeline (73a1a6e)
• set checkpoint5 as default active checkpoint (7108cb3)
• switch to checkpoint3 (8fe40c0)
• update banner image (89f58dc)
• update platform dependency with auth fix (7c70308)
• update platform dependency with debug logging (3750c3b)
• update Rust to 1.92 (latest stable) (ffb0565)
• update secure-container-runtime dependency to latest build image support (f020b6d)
• update secure-container-runtime to latest with JWT fix (8e8de66)

CI/CD

• add automatic versioning with release-please (08f076a)
• docker push only after build/clippy/test pass (1252747)
• optimize with fully parallel jobs (192d516)
• restore parallel jobs with proper cache dependencies (d5c27c6)
• use platform-runner label (cb5ddc8)
• use self-hosted runners (3c13fb3)

Tests

• add SDK compilation integration tests (18cbf2d)
• add serialization test to verify broker request uses lowercase type (8181359)
• Update compiler tests for no-wrapper behavior (2c8a87a)

---

This PR was generated with Release Please. See documentation.