Features
NexCoder is built so the safe path is the default path: checkpoints before every mutation, permission gates before every command, and read-only modes that are structurally incapable of writing.
Everything below ships in both the desktop IDE and the terminal CLI.
Every AI surface in NexCoder - the chat panel's Agent, Ask, Edit, Debug, Review, and Scan modes, and the terminal CLI - is a thin policy profile (system prompt + tool subset + turn budget) over the same v2 agentic core. Improve the engine once and every mode gets better.
The agent plans with a visible todo list, edits files directly, and verifies its own work. Every mutation is preceded by a checkpoint, so you can revert per run or per file. Commands ask for permission first, with a per-project allowlist; full-auto mode still denies risky commands. Trajectories are recorded for every run.
The active editor file and your text selection travel with every run, so "fix this" means the selected code. Natural-language code_search ranks local files and line snippets when the agent does not yet know an exact symbol - and repository content stays on your machine.
Every run appends to a per-project session under .nexcoder/sessions/. Follow-up prompts replay recent conversation, history survives restarts and is browsable in the Chats sidebar, and a durable MEMORY.md holds what the agent chooses to remember.
A Claude-Code-style skill catalog - commit, code-review, systematic-debugging, and more - that the agent loads on demand. Invoke one directly with /skill-id in the composer or --skill on the CLI. Projects define their own under .nexcoder/skills/<id>/SKILL.md, overriding built-ins by id.
Token budgeting, two-stage compaction, and a hard force-fit guarantee before every model call keep long sessions on the rails, with a live context meter in the composer so you always know where you stand.
Six policy profiles over one engine. Full-access modes get the complete toolset behind permission gates; read-only modes simply never receive mutating tools.
Full access
The full agentic loop: plans, edits files, runs permitted commands, and verifies the result end to end.
Full access
Focused file mutation. Direct edits with checkpoint-backed revert per run or per file.
Full access
Investigates failures systematically, reproduces them, and applies the minimal fix.
Read-only
Answers questions about your code. Structurally incapable of changing anything - mutating tools are never offered.
Read-only
Code review over files, diffs, or modules with findings you can act on.
Read-only
Sweeps a project for issues, risks, and structure - a fast way to learn an unfamiliar codebase.
Everything the agent knows about a project lives in a .nexcoder/ directory you can read, version, and delete.
| Path | Contents |
|---|---|
| .nexcoder/sessions/ | Chat history - index + JSONL messages per session |
| .nexcoder/checkpoints/ | Revert snapshots taken before every mutation |
| .nexcoder/permissions.json | Always-allowed commands |
| .nexcoder/MEMORY.md | Durable project memory (agent remember tool) |
| .nexcoder/repo_map.json | Cached repository map |
| .nexcoder/trajectories/ | Compact JSONL run traces |
| .nexcoder/skills/ | Project-local skills (override built-ins by id) |
NexCoder