FORM NOT VOID, MIND NO CORE

Appendix: Glossary and ADR Index

2026.08.10

A. Glossary

ChineseEnglishDescription
AgentAgentAn LLM-driven software entity that can autonomously invoke tools to complete tasks
身份声明Identity StatementThe file in AGENTS.md that defines the agent's role and principles, serving as the source of the system prompt
文件系统即自我Filesystem as SelfCodeCoder's core principle: the agent's identity, knowledge, and capabilities are defined by files on disk, not hardcoded
工具ToolNative primitives compiled into the binary, cannot be added or removed at runtime (26 total)
技能SkillProcedural knowledge stored as .md files, changing the agent's reasoning path
草稿Prompt/DraftThe draft state of a Skill, stored in prompts/, can be promoted to a formal Skill
能力CapabilityExecutable artifacts written by the agent, with a manifest declaring environment and lifecycle
环境EnvironmentThe execution backend for a Capability: Shell / Wasm / Docker
生命周期LifecycleThe operating mode of a Capability: OneShot / OnDemand / Persistent
权限键Permission KeyFine-grained permission identifier, formatted like run_command:git
会话白名单Session AllowlistRuntime in-memory permission cache, cleared when the session ends
项目白名单Project AllowlistPermission configuration persisted in codecoder.json, effective across sessions
天花板规则Ceiling Rule@shell environments can only be granted up to AlwaysThisSession trust level
静默降级Silent DegradationAutomatically falling back to a lower security level when an environment is unavailable — CodeCoder prohibits this
turnTurnThe complete processing cycle from receiving a user message to autonomously deciding output completion
命令通道Command Channel (cmd_tx)The channel carrying the user's active intent
事件通道Event Channel (event_rx)The channel carrying streaming increments and structured state
协作式取消Cooperative CancellationGracefully terminating agent operations via a shared CancelToken, rather than forcibly killing threads
子代理Sub-agentA read-only child instance created via the agent tool, with nesting depth locked to 1
只读工具集Read-only Toolset9 read-only tools available to sub-agents
工作图Work GraphA milestone dependency graph (DAG) constructed in advance
会话SessionThe message tree recorded after the fact (JSON format), the carrier of facts
里程碑MilestoneA single node in the Work Graph, with state and dependencies
验收门Acceptance GateThe levels of verification for milestone completion
命令门Command GateLevel-one acceptance: agent self-reports completion
检查门Check GateLevel-two acceptance: deterministic automated checks (compilation, templates, file count, etc.)
审查门Review GateLevel-three acceptance: structured architectural review by an independent read-only sub-agent
自恢复循环Self-recovery LoopBounded automatic retry under needs_fix state
无用户模式Headless ModeAutonomous run mode with no user present
预授权Pre-authorizationPermission configuration in codecoder.json
熔断Circuit BreakingAutomatic termination after K consecutive failures
上下文压缩CompactionTwo-tier compression strategy: tier-1 discards + placeholderization, tier-2 structured summarization
持久化PersistenceSession file JSON format and migration chain
结构化摘要Structured SummaryFive-field summary template for tier-2 compaction
可观测性ObservabilityEvent streams, NDJSON real-time logs, structured tracing
常驻目录RegistryThe file index table formed by scanning skills/, prompts/, and capabilities/ at startup
持久化记忆MemoryKey-value memory actively written by the agent, shared across sessions
离线单元测试L1 TestUnit test with no LLM calls
黑盒行为验证L2 TestBehavioral test using ScriptedProvider
真实 LLM 冒烟L3 TestEnd-to-end test connecting to a real LLM

B. ADR Index

CodeCoder's complete Architecture Decision Records are stored in docs/adr/, totaling 31 records (data collected: July 28, 2026).

ADRTitleCorresponding Chapter
0015Unified Message ModelPart 1 Ch. 1, Part 2 Ch. 1
0016Channel Topology and Event ModelPart 1 Ch. 1, Part 2 Ch. 1
0017Provider Neutral Message ModelPart 1 Ch. 1, Part 2 Ch. 1
0018Tool Trait and Permission KeysPart 2 Ch. 2
0019Sub-agent Capability BoundaryPart 2 Ch. 3
0020Skills and Capabilities RegistryPart 3 Ch. 1, 2
0021Capability Environments and LifecyclePart 3 Ch. 3
0022Self-authoring Safety LoopPart 3 Ch. 4
0023Context CompactionPart 4 Ch. 4
0025Prompt as Skill Draft TierPart 3 Ch. 1, 2
0026Background Agent Headless RunnerPart 4 Ch. 3
0030BG Objective Acceptance GatePart 4 Ch. 2
0032Client-Server ArchitecturePart 5 Ch. 1
0033BG Ledger and Exit CodesPart 4 Ch. 3
0034Persistent Supervisor Cross-restartPart 3 Ch. 3, Part 4 Ch. 3
0036Compound Command KeyingPart 2 Ch. 2
0039BG Review Gate and ObservabilityPart 4 Ch. 2, Part 5 Ch. 2

C. Project Data Sheet

Data collected: July 28, 2026

MetricValue
Source files31 (24 under src/ + daemon/ + client/)
Test files13
Test cases481 (including 1 #[ignore])
Built-in tools26
Skills6
Capabilities1
Architecture Decision Records (ADRs)31
Programming languageRust
LLM protocolOpenAI-compatible (Provider-neutral)
Run modesInteractive (TUI / Socket) and Headless (environment-variable-triggered)
Communication protocolUnix socket, daemon-client architecture

Note: Glossary terms appear in English in the main text, with Chinese annotations in parentheses on first occurrence.