Glossary
Anti-pattern
Section titled “Anti-pattern”A common misuse of a pattern or capability, named with the symptom that reveals it. Principles call theirs a Signal of Violation. See Anti-patterns.
A one-message task description for delegated work: goal, constraints, non-goals, a mirror to follow, a done command, and an ask for assumptions. See Delegation Fit.
Checkpoint Gate
Section titled “Checkpoint Gate”A mechanical block before an irreversible or outward-facing action that parks a one-step decision for a human. See Checkpoint Gates.
Compaction
Section titled “Compaction”Summarizing an agent’s context in place to free space. Keeps continuity but can drop constraints. See Context Handoff.
Context Handoff
Section titled “Context Handoff”A written record of goal, state, evidence, decisions and next step that lets a fresh session resume a run. See Context Handoff.
Context Window
Section titled “Context Window”The amount of text a model can process in a single call. Exceeding it causes truncation or failure.
Correction Diagnosis
Section titled “Correction Diagnosis”Tracing a repeated correction to the input that caused it, then fixing that input. See Correction Diagnosis.
Delegation Fit
Section titled “Delegation Fit”The per-task choice to delegate, pair, or write by hand, made before prompting. See Delegation Fit.
Harness
Section titled “Harness”The interface to the models: the CLI, IDE extension, or application that orchestrates prompts, tools, and context. See Harness Selection.
Keepers
Section titled “Keepers”Repository Maintainers or higher. They endorse proposals, resolve objections, and ratify additions to SHRINE. See Governance.
Orchestration
Section titled “Orchestration”Coordinating multiple models, agents, or pipeline stages so each handles the part of a task it is best suited for.
Progress Breadcrumbs
Section titled “Progress Breadcrumbs”Short, evidence-linked progress notes an agent writes to a shared work board. See Progress Breadcrumbs.
Ratification
Section titled “Ratification”Content becomes part of SHRINE by one of three routes:
- Lifecycle: two keeper endorsements, the objection window, then the draft PR merges (Proposal Lifecycle)
- Expedited: the lifecycle without the objection window, which the CTO or VP of Engineering waives (Expedited Ratification)
- Keeper decision: a majority of keepers ratify or amend directly, and can waive the Discussion; the page links the record (Keeper Decision)
Repository Context
Section titled “Repository Context”Indexed knowledge of a codebase (symbols, references, code graph) exposed to agents. Distinct from memory. See Repository Context.
Reviewable Output
Section titled “Reviewable Output”Agent output shaped to be cheap to review: small diffs, attached evidence, and a review note. See Reviewable Output.
Routing
Section titled “Routing”Selecting which model handles which task based on complexity, cost, and capability requirements.
Scaffolding
Section titled “Scaffolding”Workarounds, wrappers, and patterns built to compensate for model limitations. May become obsolete as models improve. Distinct from Mechanical Scaffolding, which is repeatable structure for token efficiency.
SHRINE
Section titled “SHRINE”Stable-Kernel Hosted Reasoning & Inference Network Environment. The org’s open forum and governance site for how we build with AI.
Skills
Section titled “Skills”Packaged, reusable instructions for accomplishing specific tasks efficiently.
A written agreement before execution: problem, constraints, acceptance criteria with checks, out of scope, checkpoints, abort criteria and stop condition. See Spec, Then Build.
An abstract capability slot: what the capability must do, how to select for it, and its anti-patterns. Not a product list. See FAQ.
Standing Instructions
Section titled “Standing Instructions”A short per-repo instruction file the harness loads every session: commands, idioms, and conventions the agent cannot cheaply derive from the code. See Memory & Context.
Step-Level Routing
Section titled “Step-Level Routing”Choosing model tier and reasoning effort for each step inside an agent trajectory. See Step-Level Routing.
Subagent
Section titled “Subagent”An agent spawned by a primary agent to handle a scoped task, often in parallel with others. Keeps the main context lean and the work isolated.
TTV (Tokens to Value)
Section titled “TTV (Tokens to Value)”Value per win, where a win costs tokens plus human attention (engineer minutes steering, reviewing, babysitting). The north star metric: drive cost per win down, not raw token usage. See TTV.
Unattended Run
Section titled “Unattended Run”An agent run that works for hours without a human watching each step, bounded by a time limit, abort criteria and a stop condition. See Unattended Runs.