#Glossary

One paragraph per term. Read this first if any CodingScaffold output uses a word you have not seen before. Each entry links to the wiki page that goes deeper.

#adapter

A small, native-format integration that teaches a specific coding tool how to behave inside this project. tools adapt --tool opencode writes OpenCode-shaped files; --tool claude-code writes the Claude-Code shape; and so on. Adapters are generated, reviewable, and have no runtime. See Tool-Adapters.

#agentic change

A code change planned or written with help from an AI coding agent (OpenCode, Claude Code, Codex, …). CodingScaffold's session command + agentic-change.md PR template exist so these changes ship with the same shape as any other PR. See Getting-Started.

#artifact

A file or directory CodingScaffold knows about: AGENTS.md, .coding-scaffold/policy/, PR template, eval config, sessions directory, knowledge base. doctor surveys artifacts; the canonical list with rationale lives in src/coding_scaffold/artifacts.py.

#context

The bytes that go into the agent's input window for a single turn: the prompt, project rules, file excerpts, knowledge notes, prior messages. context budget estimates the size; context lint checks the files; context compress writes safer sidecars. See Context-Hygiene.

#doctor

coding-scaffold doctor — the accessibility hub. Surveys what is set up, recommends 1–3 next commands, and explicitly says which advanced surfaces a beginner can ignore. Never installs or writes; always safe to run.

#eval

A small readiness benchmark: eval init creates the config, eval run exercises it, eval report summarizes. It validates that the scaffold is set up — not that the model is good. See Getting-Started.

#knowledge base

A team-shared, reviewable corpus of notes under .coding-scaffold/knowledge/. Multiple backends are supported (markdown, obsidian, foam, mempalace, html). Layered by scope (team/department/unit/company) and maturity (raw → curated wiki). See Knowledge-Base.

#maturity

The lifecycle stage of a knowledge note: raw (captured but unreviewed), wiki (curated for the team), and the layered scopes (team/department/unit/company). knowledge promote moves notes between maturity levels with an audit trail.

#MCP

Model Context Protocol — a way for an agent to talk to external integrations (Slack, Notion, your bug tracker). mcp policy, mcp scan, mcp snapshot, mcp diff govern which integrations the project allows. See Security.

#memory

Reviewable memory entries the agent can recall: memory capture proposes one, memory review accepts/rejects, memory promote moves it between maturity classes, memory audit lists what is in effect. Distinct from knowledge: memory is short runtime hints; knowledge is shared documentation.

#multi-tool project

A project configured with more than one coding tool — e.g., Codex + Claude Code in the same repo. Generated via coding-scaffold setup run --tool codex --tool claude-code (or comma-separated). Each tool gets its native adapter files (e.g., AGENTS.md for Codex, CLAUDE.md for Claude Code) generated side-by-side; setup update keeps them in sync going forward.

#orchestration

A multi-step recipe a tool can drive (tools orchestrate). Distinct from a single prompt: orchestration plans the steps, picks a model per step, and tracks results.

#persona

A target user shape: beginner, control-and-reproducibility, security-review, team-lead. Used by doctor --persona / pilot --persona to surface a focused recipe instead of the full firehose. Personas are documented in Team-Rollout.

#pilot

coding-scaffold pilot --target . --tool <name> — read-only guided wrapper. Probes the local environment and prints the exact 10-minute happy-path commands. Never installs and never writes files. The printed recipe is what the user actually runs.

#policy pack

Files under .coding-scaffold/policy/ that encode the team's non-negotiables: allowed providers, network rules, model selection floors, MCP allowlist. See Policy-Packs.

#provider

A source of model inference: a local runtime (Ollama, LM Studio) or a cloud API (Anthropic, OpenAI, …). probe lists what is available; routing.json describes how the project routes between them.

#routing

The plan for which model handles which class of request: weak_model (fast/cheap), strong_model (high-quality), plus the endpoint and policy. Stored in .coding-scaffold/routing.json.

#scaffold artifact

See artifact.

#scaffold version

A SHA256 snapshot of every generated file written by setup, stored in .coding-scaffold/scaffold-version.json. setup update uses it to tell unchanged files (safe to rewrite) from user-edited files (write a .new sidecar instead). See Upgrading.

#session trace

A reviewable Markdown file under .coding-scaffold/sessions/ recording a single agentic change: task, commands run, diff, follow-ups. Powers reversibility: session checkpoint, session diff, session rollback.

#skill

A reusable scoped instruction the agent loads on demand: a Markdown file under .coding-scaffold/skills/<name>/ with a SKILL.md, optional helpers, and an approved checksum. skills new, skills lint, skills approve, skills export. See Skills-and-Agents.

#stability marker

A label rendered in --help next to each command name: [stable], [preview], [experimental]. Tells experienced teams what they can build infrastructure around. The contract per marker lives in Stability.

#strong model

The model class used for high-quality output (planning, review, hard reasoning). Set in routing.json. The pilot recipe avoids picking one for the user — that is the team's call.

#team manifest

A JSON file (team-onboarding.json) shipped from a team repo that defines shared policy, skills, knowledge sources, and version requirements. team connect / team sync apply it; team push nominates local artifacts upward; team doctor shows the effective merged view. See Team-Sync.

#weak model

The model class used for fast/cheap calls (extraction, classification, format conversion). Set in routing.json.