Continuous Claude
A persistent, learning, multi-agent development environment built on Claude Code that maintains context across sessions.
Evidence shows: setup wizard backs up existing config, uninstall archives and restores, providing rollback; but least privilege, user confirmation, data flow transparency, sensitive data handling, dependency security, external effects, source attribution are only partially addressed. Deductions: no least privilege details, user confirmation only mentioned for wizard, data flow description incomplete, sensitive data (e.g., API keys) handling not explicit, dependency security not mentioned, external effects (e.g., auto-commit) not fully explained, source attribution relies only on MIT license.
Evidence shows: README is consistent, but no dependency availability guarantees or failure message examples. Deductions: self-consistency partially present, dependency availability not mentioned, failure messages not provided.
Evidence shows: README describes multiple usage scenarios and triggers, but capability boundaries and environment fit details are insufficient. Deductions: audience and scenarios well covered, but capability boundaries not explicit, trigger precision partially present, environment fit only mentions Python and Docker.
Evidence shows: README structure clear, installation instructions detailed, license explicit, but naming stability, known limitations, versioning/changelog, maintenance responsibility not fully addressed. Deductions: information architecture and install notes good, but naming stability not mentioned, known limitations not listed, changelog missing, maintenance responsibility not clear.
Evidence shows: README claims token savings and efficiency, but lacks actual data. Deductions: output usability partially present, marginal value claimed but unverified, cost-benefit not quantified.
Evidence shows: claims in README lack traceability, no cross-source corroboration or separation of facts and inferences. Deductions: claim traceability insufficient, cross-source corroboration missing, facts and inferences not separated.
- The setup wizard modifies ~/.claude configuration; ensure you back up and understand changes.
- Auto-commit and auto-execution may have external effects; use with caution.
- Requires Docker and PostgreSQL; ensure environment meets requirements.
- Sensitive information like API keys may be stored in .env; protect accordingly.
What does this agent do, and when should you use it?
Continuous Claude transforms Claude Code into a continuously learning system that maintains context across sessions, orchestrates specialized agents, and eliminates token waste through intelligent code analysis. It introduces continuity ledgers, YAML handoffs, and a memory system powered by PostgreSQL and pgvector for compounding learnings across sessions. The repository includes 109 skills, 32 agents, and 30 hooks, including meta-skill workflows like /workflow, /fix, /build, and integrates the TLDR code analysis tool claiming 95% token savings. The primary framework is the Claude Code CLI, with a setup wizard that installs components automatically and configures PostgreSQL via Docker.
Continuous Claude is a set of hooks, skills, and agents that intercept and augment Claude Code lifecycle events. On SessionStart, it loads continuity ledgers and memory, warms TLDR cache; on UserPromptSubmit, it injects skill activation prompts based on a rule-based system; on PreCompact, it auto-creates YAML handoffs to preserve context; on SessionEnd, a daemon launches a headless Claude to extract learnings from thinking blocks into archival_memory. It provides a /workflow goal router and commands like /fix, /build, /tdd, /refactor that orchestrate specialized agents like sleuth, kraken, plan-agent. Its TLDR tool performs 5-layer code analysis (AST, call graph, CFG, DFG, PDG) and supports natural language semantic search.
- Developers who frequently hit context compaction in long sessions and need to preserve decisions and context across sessions.
- Teams that want to automate multi-step workflows like bug fixing, feature building, and code review via /fix, /build, /review commands.
- Engineers who need to understand large codebases without reading entire files, using /explore and TLDR summaries.
- Users who want to accumulate project knowledge over time, remembering learnings (e.g., auth patterns) and surfacing them automatically in future sessions.
- Researchers or developers doing formal verification with Lean4 via the /prove skill.
What are this agent's strengths and limitations?
- Solves the context compaction problem with handoffs, continuity ledgers, and memory extraction for compounding learning.
- TLDR 5-layer code analysis reduces token consumption by ~95%, making large codebases more efficient.
- Extensive built-in skills (109) and agents (32) covering planning, debugging, review, etc., reducing manual effort.
- Supports Lean4 formal verification via /prove skill, with a proven track record on the Sylvester-Gallai theorem.
- Tightly coupled to Claude Code; not usable with other AI coding assistants or APIs.
- Complex installation requiring Docker for PostgreSQL, uv, Python 3.11+.
- Requires maintaining a database schema and running a daemon for memory extraction, adding operational overhead.
- Includes a large number of agents (32) which README admits will be consolidated in v4, potentially causing complexity and overlap.
How do you install or deploy this agent?
Requires Python 3.11+, uv package manager, Docker, and Claude Code CLI. Clone the repo and run the setup wizard:\n``bash\ngit clone https://github.com/parcadei/Continuous-Claude-v3.git\ncd Continuous-Claude-v3/opc\nuv run python -m scripts.setup.wizard\n`\nThe wizard checks prerequisites, configures PostgreSQL (Docker) and API keys, installs 32 agents, 109 skills, and 30 hooks, and optionally installs math features (SymPy, Z3, Pint) and TLDR. Remote database is supported by setting CONTINUOUS_CLAUDE_DB_URL. Uninstall via uv run python -m scripts.setup.wizard --uninstall`.
How do you use this agent?
After installation, start Claude Code by running claude in your project directory. Trigger workflows via natural language, e.g., \"Fix the login bug\" or /build greenfield \"user dashboard\". Available commands include /workflow (goal routing), /fix, /build, /explore, /premortem, /tdd, /refactor, /review, /security, and /release. See the README for detailed workflow chains and hook triggers.