afloresc/opencode-lite

Purpose and Entry Point

This is a Go CLI implementing an autonomous, project-scoped cognitive agent runtime. It accepts a goal, plans and grounds work into filesystem/code operations within workspace/, monitors execution, and adapts future behavior through persistent memory. Process startup is cmd/cli/main.go; configuration is centralized in internal/config/config.go.

Runtime Orchestration

AgentRuntime is the core control loop: it executes goals, invokes planning and tools, captures results, and feeds observations into monitoring and adaptive subsystems. Agent context and shared contracts define the boundaries among stages.

Paths: internal/agent/runtime.go, internal/agent/context.go, internal/agent/interfaces.go.

Goal Routing and Specialized Agents

A supervisor/classifier selects or biases execution toward specialized analysis, refactoring, or documentation agents. These agents expose a common interface and delegate actual execution back to the shared runtime, preserving one execution path while allowing goal-specific policy.

Paths: internal/agent/supervisor.go, internal/agent/goal_classifier.go, internal/agent/analysis_agent.go, internal/agent/refactor_agent.go, internal/agent/docs_agent.go.

Hybrid Planning and Step Expansion

The planner combines deterministic rules, LLM-derived reasoning, and planner memory to produce executable steps. Step expansion/mapping transforms broad plans into operations, filters historically failing work, and supports planner modes such as coarse, fine, aggressive, and conservative.

Paths: internal/agent/planner.go, internal/agent/master_planner.go, internal/agent/planner_rules.go, internal/agent/step_expander.go, internal/agent/step_mapper.go.

Tool Grounding and Execution Boundary

The contextual grounder maps abstract plan steps to registered concrete tools, using strict or flexible selection modes and project context. Tools are the safety boundary for workspace interaction: filesystem changes, search, patches/refactors, formatting, analysis, and sandbox command dispatch all reside behind the tool package.

Paths: internal/agent/tool_grounder.go, internal/agent/tool_grounder_contextual.go, internal/tools/registry.go, internal/tools/filesystem.go, internal/tools/patches.go, internal/tools/run_command_dispatch.go.

Feedback, Memory, and Self-Optimization

Execution monitoring detects repeats, consecutive failures, and stalled output. Metacognition evaluates those signals and provides confidence/advice; the strategy engine adjusts planner/grounder modes and tool or agent bias. Cognitive and planner memory persist project-specific outcomes, while AOC records longer-term optimization preferences.

Paths: internal/agent/execution_monitor.go, internal/agent/metacognition.go, internal/agent/strategy_engine.go, internal/agent/cognitive_memory.go, internal/agent/planner_memory.go, internal/agent/aoc.go.

LLM and Infrastructure Integration

The runtime has an LLM client/prompt boundary with an Ollama provider implementation, making local Ollama the identified model-service integration. Persistent cognitive state is stored under .opencode/<projectID>/; mutable target content is constrained to workspace/. The system runs as a Go binary and uses local command/file operations rather than a web service.

Paths: internal/agent/llm_client.go, internal/agent/llm_prompt.go, internal/providers/ollama.go, internal/runtime/runtime.go, internal/runtime/system_prompt.go.