Manual

Every slash command in amber comes from one JSON tree (completions.json). Nothing is hardcoded, nothing is special-cased: the drawer, Tab completion, and the ? help pages are all derived from the same tree. Press / to open the drawer, type? at any depth for the full help page.

Model & Provider

Sessions & Jobs

Policy & Permissions

Files & System

MCP & Plugins

Skills, Learning & Engine

flowchart TB
    ctx["sealed Context, read-only"] --> gate{"compression gate"}
    gate -->|compressed| rebuild["context rebuilt<br/>clear + push"]
    gate -->|kept| copy
    rebuild --> copy["prompt copy<br/>the Context itself is never mutated"]
    copy --> head["head blocks<br/>inserted right after the system prompt"]
    copy --> tail["tail blocks<br/>appended after the conversation"]
    head --> order["within each: ascending prompt_priority<br/>ties keep insertion order"]
    tail --> order
    order --> sources["sources: retrieved memories · skill metadata<br/>session brief · plugin PromptBlockCapability"]
    sources --> omit["a block that renders empty is omitted"]
    sources --> isolated["a plugin block that throws<br/>cannot break the turn"]

flowchart TD
    cycle["compression cycle"] --> response["CompressionResponse<br/>segments · memory_ops · skill_ops"]
    response --> apply["apply_compression_result"]
    apply --> mem["apply_memory_ops<br/>upsert, evidence starts at promote_threshold"]
    apply --> skills["apply_skill_ops<br/>upsert or deprecate, with a trigger phrase"]
    mem --> store
    skills --> store["JsonMemoryStore"]
    store --> dedup["dedup by content hash<br/>a repeat upsert just bumps evidence"]
    store --> recall["MemoryRetriever builds the system-prompt suffix<br/>recall is injected into the prompt copy"]

Windows

Configuration

Project settings live in amber.conf in the workspace root. Provider credentials live outside the project, per provider file:

# amber.conf (project-local, only data paths)
system_prompt=prompts/system.md
tools_prompt=prompts/tools.md
git_prompt=prompts/git.md

# ~/.config/amber/providers/<name>.conf
api_base=http://localhost:8080/v1
model=Qwopus-27B-Fusion

Amber keeps its own files in four places, following the XDG base directories: global config in ~/.config/amber (providers, skills, MCP servers), shareable project config in<project>/.amber (created only when you add something), private per-project state in~/.local/state/amber/projects/<project> (sessions, settings, logs, experience, policy) and the model cache in~/.cache/amber. Running amber from your home directory therefore leaves no project directory behind, and a deleted checkout does not take your conversation history with it./get config paths prints the resolved layout for this machine, and the bundled amber-config skill teaches the agent the same table.

Environment variables

• AMBER_WORKSPACE - override the workspace root (defaults to the working directory)
• AMBER_MODEL - override the default model
• AMBER_PROVIDER - override the default provider
• AMBER_STREAM - set 0 to disable SSE streaming
• AMBER_APPROVAL - approval mode override
• XDG_CONFIG_HOME, XDG_STATE_HOME, XDG_CACHE_HOME - move the config, state and cache roots