TODO

Last updated: 2026-07-27

Done

v0.2.0 Features

• Streaming tool-call parser — state-machine detects <tool_call> blocks incrementally from streaming LLM output, fires events on </tool_call>
• Tool guardrails — loop detection, idempotent/mutating classification, failure counting with configurable warn/block thresholds
• Pluggable context compaction — Compactor trait + default LLM summarizer
• Self-healing retry — auto re-prompt LLM with error context on tool failures
• Trajectory export — ReAct step serialization (thought→action→observation→response) for RL training data
• Autonomous mode — 24/7 workspace loop: read TODO.md, run agent, validate with tests, git commit/push only on success. Failure pause across restarts
• Skill curator — background task reviews skills, suggests improvements
• Plugin lifecycle hooks — PreToolHook, PostToolHook, LifecycleHook traits
• Skill template preprocessing — ${HERMES_SKILL_DIR} / ${HERMES_SESSION_ID} vars + !\command`` inline shell substitution in SKILL.md
• Cron scheduler — SurrealDB-backed with in-memory noop fallback for testing
• Hermes-style self-healing, loop detection, and compaction integration in loop_runner.rs
• All 85 tests pass, 0 clippy warnings
• v0.2.0 published to crates.io as apollo-agent

Previous (v0.1.x)

• SurrealDB memory schema includes files, chunks, FTS, and sticker cache
• Conversation history is loaded by chat_id before agent calls
• CLAUDE.md is in sync with repo instructions via symlink to AGENTS.md
• cargo clippy --all-targets -- -D warnings passes
• cargo test passes
• cargo build --release passes
• GitHub issue #2 through #6 resolved and closed
• Toolset allowlists, session search, managed skills, Daytona scaffolding

Next

• Wire autonomous mode into CLI entrypoint (apollo autonomous)
• Integrate streaming parser into execution loop — recover_tool_calls runs when the provider returns no native tool calls
• Wire the rx4 bridge into the execution path
• Refresh and persist expired Anthropic OAuth tokens
• Add trajectory collection toggle in config.yaml
• Add tests for Telegram markdown conversion and long-message chunking
• Add tests for audio/sticker handling in Telegram
• Make guardrails state persistable across restarts

Backlog

• Session branching (JSONL tree with parent pointers in SurrealDB)
• Add curator start|stop CLI commands
• Docker runtime adapter to replace Daytona
• WASM-based plugin sandbox (long-term)