Canonical definitions for paper’s product nouns. Other pages link here rather than redefining terms; if a page seems to use one of these words differently, this page wins.
agent
The AI coding tool whose work paper captures. Supported agents are Claude Code, Codex, and Pi, launched through paperctl start <agent> or shell routing. The ChatGPT app and tools like Conductor and Zed can route through paper persistently, without a terminal launch.
backend
A model-provider route configured within a gateway: the provider host, the request schema it speaks, a model allowlist, and how requests authenticate — transparent or managed. Managed with paperctl tapes backend or from the console’s Gateways page.
cassette
An API extension a tapes deployment serves under /v1/cassettes — search, labels, export, skills, and digests are each cassettes. Day-to-day you use them through their first-class commands and console features; paperctl cassettes is the direct surface.
chain
The hash-linked sequence of turns that makes up a session’s conversation. Each turn references its parent, so the chain records the exact order and ancestry of the exchange.
derived status
A session’s outcome — completed, failed, abandoned, or unknown — computed from its captured events rather than reported by the agent. Shown as the status badge on sessions in the console and CLI.
digest
A written summary of what your organization accomplished in a period — accomplishments, next steps, and topics — shown as a card on the console dashboard. See Dashboard digests.
gateway
Your organization’s named route for model traffic. Agent calls pass through the gateway on their way to the provider, and capture happens there. An organization can have more than one gateway; each holds its own backends. See the Gateway overview.
gateway API key
A per-user credential (it starts with paper_) for calling a gateway directly, without paperd — from CI, scripts, or SDK code. It carries your paper identity only; the backend supplies the provider credential. Created and revoked on the console’s API keys page, and valid on any of your organization’s gateways. See API keys and direct access.
harness
The agent runtime that produced a session — which tool (Claude Code, Codex, Pi) and version was driving the conversation. Sessions record their harness so behavior can be compared across tools.
label
A shared, organization-wide tag you attach to sessions and Skills — a vocabulary for grouping work by release, incident, customer, or experiment. Filters in the console and CLI (--label) match by label. See Organize sessions with labels.
member
A user who belongs to an organization. Members can view the organization’s captured sessions; roles such as admin come from the organization’s directory.
model
The specific LLM that served a request, as reported per call (for example claude-fable-5). A single session can involve several models; the console shows the cost-weighted mix.
organization
The top-level boundary for everything in paper: sessions, search, Insights, Skills, gateways, and membership are all scoped to exactly one organization.
paperd
The paper daemon: a local proxy on your machine (127.0.0.1:51539 by default) that keeps your login token fresh, attaches your paper identity to each request, and forwards your agent’s model traffic to your organization’s gateway. It is not a gateway — nothing is captured or routed on your machine.
provider
The company serving the model API — Anthropic, OpenAI, and so on. The backend determines which provider a call reaches and with what credentials.
search
Semantic search over your organization’s captured sessions, from the CLI (paperctl search "<query>") or the console. Queries match meaning rather than exact keywords, and results are scoped to your organization.
session
One captured agent run, from launch to exit: its turns, tool activity, token usage, cost, duration, and derived status. Subagent runs render inline within the session, and its usage and cost totals include them.
Skill
A reusable unit of knowledge extracted from captured sessions — a markdown document an agent can load to apply what a previous session learned. Skills carry a version and links back to their source sessions. A published Skill is visible to everyone in your organization.
Skill version
An immutable published snapshot of a Skill’s content, with a semantic version number and changelog. Editing a Skill after publication never changes an existing version — corrections ship as a new version.
span
A single call within a trace: one model call, tool invocation, or event. Spans carry their own timing, token usage, and input/output, and link to related spans so causality can be followed.
subagent
An agent run the main agent spawned to work on a subtask — a Claude Code Task agent or a Codex spawn_agent thread. Subagents appear inside their parent session, nested under the tool call that spawned them, at any depth; they are not separate sessions, and the session’s usage and cost rollups include their activity.
success rate
The share of sessions in a period whose derived status is completed. A drop in success rate says something changed — it does not by itself say what caused it.
tapes
The open-source trace engine behind paper’s capture — it stores sessions and powers the console, search, and Skills. You interact with paper; tapes is the layer underneath, also usable on its own.
tool call
An agent’s invocation of a tool — running a command, reading a file, calling an API — together with the arguments it passed and the result it got back. Captured as part of the session’s spans.
trace
One prompt-to-response cycle within a session: the user’s prompt, the model and tool activity it triggered, and the response. A session is a sequence of traces; the conversation-spine view in the console shows one card per trace.
transparent and managed auth
The two ways a backend authenticates to its provider. Transparent: the caller’s own credential (API key or plan sign-in) passes through untouched. Managed: the gateway injects a key your organization stores on the backend. See Backends and providers.
turn
A single message in a session’s chain — one user, assistant, or tool message with its content blocks, model, and token usage.