glossary
DevAware OS vocabulary
Agent governance
The policy layer that decides which actions an AI coding agent may take, and which require explicit user approval.
Allowlisted execution
A model where an AI agent may directly invoke only a fixed, read-oriented set of commands; anything else requires the user, and destructive patterns are refused.
BYO-key
A model where the tool subscription is separate from cloud AI provider usage — the user supplies their own provider key and the provider bills them directly.
Checkpoint
A Git-independent local file snapshot taken on every chat send, allowing instant rollback of agent-introduced regressions.
Clean prompt
The policy-checked artifact assembled after governance and handed to the model. Explicit scope, constraints, acceptance criteria; no secrets.
Compact final report
An enforced end-of-loop summary stating what changed, what was validated, and what was skipped.
Compact memory
A local distilled-outcome store the next DevAware OS loop can lean on. Embeddings and summaries stay on the user's machine.
Control layer
The governed boundary between human intent and AI execution that runs the DevAware OS pipeline end to end.
Controlled execution
Execution under an allowlist, with a destructive-command spawn-boundary guard and an approval gate on flagged tool actions.
Destructive-command guard
A check that refuses dangerous shell patterns at the spawn boundary before an agent command runs.
Development operating layer
A local-first control plane between human developer intent and AI-driven execution. Bundles governance, prompt control, controlled execution, validation, and compact memory.
Effort meter
A user-visible setting bounding how much work a single DevAware OS loop is permitted to spend on a request.
Execution lane
One of the three user-selected paths a DevAware OS request runs through — local open-source runtime, BYO-key cloud provider, or OAuth CLI session.
Governance block
A reusable policy rule applied to a DevAware OS loop. Sets scope, tool allowlist, and whether actions require explicit user approval.
In-loop verb
A bounded auto-context command (read, grep, symbol, exec, Playwright smoke or steps) an agent turn can invoke to pull evidence, depth-capped to stay controlled.
Intent block
An explicit declaration of what a developer wants this DevAware OS loop to accomplish, captured before governance fires.
Learning Lab
A local, idle-gated training surface that runs user-initiated fine-tuning jobs on the user's machine only when it is idle.
Local-first execution
An approach where the agent loop runs on the user's machine by default, with workspace, memory, and validation local; cloud is opt-in per request.
MCP client
The component that connects DevAware OS to Model Context Protocol servers over stdio, listing and calling tools under a per-project trust grant.
P0 filter
A signal threshold that blocks low-quality intents from entering the pipeline before any work is spent.
Project trust
An explicit user grant, required before DevAware OS runs tools against a repository; repositories are untrusted by default.
Provider independence
The property of not being tied to any single AI vendor — the user connects compatible local or cloud providers and selects which one runs each request.
Secret masking
Hiding secret-shaped values in environment files during project inspection so credentials are not surfaced into prompts or the UI.
Structured timeline
A chat surface that renders each agent turn as explicit lanes — tool request, tool result, injected context, validation result, blocker, next action.
Token mode
A Save / Normal / Spend setting that controls how much context and effort DevAware OS invests in a loop.
Unified command center
The single desktop workspace that consolidates chat, governance, terminals, preview, inspection, file tree, Git, validation, and memory around the pipeline.
Validation gate
An automated check (typecheck, lint, test, build, smoke) a DevAware OS loop must pass before declaring work done.
Workflow noise
Low-signal content in an intent — P0 inflation, prompt bloat, emotional language, vague or missing objectives — that DevAware OS flags or auto-fixes before compiling a clean prompt.