A command-line coding agent that runs inside your repository. Java 25 + Spring Boot 3.5 + Spring Shell.
It reads the workspace, then acts through a small, explicitly registered set of tools — read, search, patch, run, delegate — and keeps its session state on disk under .coda/, so a run can be inspected or resumed later.
mvn clean package -DskipTests
java --enable-preview -jar target/coda-0.1.0.jarPick a model backend (Ollama is the default and needs no key):
export CODA_DEFAULT_PROVIDER=openai
export CODA_PROVIDER_OPENAI_API_KEY=sk-...
export CODA_PROVIDER_OPENAI_MODEL=gpt-5.4
java --enable-preview -jar target/coda-0.1.0.jar| Command | Alias | Description |
|---|---|---|
ask <message> |
a |
Run one task |
chat |
c |
Interactive loop |
status |
Session / provider / step count | |
session [id] |
Show or switch session | |
provider <name> |
Switch model backend | |
reset |
Reset session memory |
Inside chat: /exit, /status, /reset.
com.fu.coda
├── config/ CodaProperties, EnvLoader (.env)
├── model/ ModelClient + 5 backends (Ollama, OpenAI, Anthropic, DeepSeek, Fake)
├── workspace/ Git context, project-doc pre-scan, fingerprint
├── tool/ Tool, ToolRegistry, ToolValidator, ApprovalGate
│ └── impl/ 7 tools: list_files, read_file, search, run_shell,
│ write_file, patch_file, delegate
├── memory/ WorkingMemory + DurableMemoryStore + LayeredMemory
├── context/ Prompt assembly, section budgets, reduction
├── runtime/ CodaAgent (orchestrator), TaskState, RunStore, SessionStore,
│ SecurityService, ParseService, PromptService,
│ CheckpointService, ToolExecutionService, DurablePromotionService
├── shell/ Spring Shell commands and REPL
├── benchmark/ Runner, fixtures, verifiers
├── metrics/ Run metrics
└── persistence/ CodaPaths (.coda/ layout)
| Tool | Risk | Description |
|---|---|---|
list_files |
safe | List a directory |
read_file |
safe | Read a line range |
search |
safe | Regex search (ripgrep, with a pure-Java fallback) |
delegate |
safe | Spawn a read-only child agent |
run_shell |
risky | Run a shell command in the repo root |
write_file |
risky | Create or overwrite a file |
patch_file |
risky | Replace an exact text block |
Risky tools go through ApprovalGate, configured by coda.approval: ask | auto | never.
Tools are stateless: every execute() call stands alone.
The prompt is split into five sections with per-section character budgets:
| Section | Budget (chars) | Content |
|---|---|---|
prefix |
6,000 | Agent manual: rules, tool signatures, workspace snapshot |
memory |
1,600 | Task, recent files, file summaries, episodic notes |
relevant_memory |
800 | Notes retrieved for the current request |
history |
7,600 | Transcript, semantically compressed |
current_request |
unlimited | Never trimmed |
| total | 16,000 |
Those numbers were measured, not guessed: in this repository the static manual is ~1,728 chars,
7 tool signatures ~1.2k, and the workspace snapshot ~2.5k, so prefix gets 6,000 to stay intact.
When the total is exceeded, sections shrink in the order
relevant_memory → history → memory → prefix — the tool signatures are the last thing to lose,
because a model that cannot see a tool will not use it. Every reduction is logged with
section / before_chars / after_chars / overflow_chars.
History compression runs before length clipping: the last 5 entries are kept whole (up to 1,000 chars each),
repeated read_file calls collapse to the first occurrence, old run_shell calls become one summary line,
and other old entries are clipped to 80 chars.
LayeredMemory
├── WorkingMemory session-scoped: task, recent files (8), file summaries (6), notes (12)
└── DurableMemoryStore .coda/memory/MEMORY.md + topics/*.md
File summaries carry a freshness flag — after a file is edited, summaries built from the old content are invalidated rather than silently reused. Promotion to durable memory is explicit: only notes worth keeping across sessions are promoted, and the call returns both what was promoted and what it replaced.
mvn test22 tests: context budgets and reduction, history compression, prompt contract, response parsing, utils.
benchmark/ holds a task suite (BenchmarkRunner + fixtures + verifiers) that collects metrics/.
- JDK 25 (the build enables preview features)
- Maven 3.9+
- Tool execution and runtime orchestration still have thin test coverage; tests cluster around context and parsing.
delegateis read-only and depth-limited to 1.- Budget numbers come from one measured repository; a second project may need different splits.