Glossary: Project Learnings

Mirrored from docs/design/project-learnings/references/glossary.md. Edit the source document in the repository, not this generated page.

Project-specific vocabulary used in 1_overview.md, 2_design.md, 3_vision.md, and 4_decisions.md. Standard industry terms (glob, YAML, SQLite, MCP) are excluded unless this feature gives them a specific meaning.

Term	Meaning
Active	Learning lifecycle status indicating the record is eligible for injection. Opposite: superseded. See 2_design.md §7.2.
Body	The multi-line markdown content of a learning record — the rule, the reason, and the application guidance. Loaded on demand via orbit.learning.show, never injected directly. See 2_design.md §2.1.
Dedup set	Per-session, agent-local set of injected learning IDs. Each push layer consults it before emitting a <system-reminder> to prevent duplicate injections across layers. See 2_design.md §4.4.
Engine pre-prompt injection	Layer 1 of the push pipeline. orbit-engine queries matching learnings before spawning an agent and prepends summaries to the agent prompt. Universal across agents. See 2_design.md §4.1.
Evidence	Provenance attached to a learning record — commit SHAs, task IDs, or external refs that produced or substantiate the learning. See 2_design.md §2.1.
Learning record	The first-class Orbit resource representing one piece of project knowledge. YAML on disk, indexed in SQLite, mutated through orbit.learning.* tools. See 2_design.md §2.
MCP-sidecar injection	Layer 2 of the push pipeline. orbit-mcp attaches a learnings field to MCP tool responses whose arguments or output reference file paths. Cross-agent. See 2_design.md §4.2.
Pre-tool-use hook	Layer 3 of the push pipeline. A Claude Code PreToolUse hook on `Edit
Pull surface	The orbit.search tool (with kind: "learning") and the orbit-learnings skill. Used for active query at task start or when an agent has time to ask. Complement to push, not substitute. See 2_design.md §6.
Push layer	One of the three injection sites that surface learnings into agent context without an agent query. The three layers are engine pre-prompt, MCP-sidecar, and Claude Code hook. See 2_design.md §4.
Scope	The trigger condition for a learning. Phase 1: path globs and tags evaluated as logical OR. Phase 2: adds symbol IDs and semantic seeds. See 2_design.md §3.
Semantic seed	Reserved field (scope.semantic_seed) for phase 2. Short text describing what a learning is “about”; used as the embedding source for semantic-similarity ranking. See 3_vision.md §1.2.
Stale	A learning whose referenced files, commits, or tasks no longer exist. Detected opportunistically via orbit learning prune --stale-only. See 2_design.md §7.3.
Summary	The one-line rule of thumb for a learning, designed to fit in a <system-reminder>. Always injected; never substituted with the body. See 2_design.md §2.1, §4.5.
Supersede	Lifecycle transition where a newer learning replaces an older one. Both records persist; the old one’s status flips to superseded and gains a superseded_by back-reference. See 2_design.md §7.2.
Symbol-aware scope	Reserved field (scope.symbols) for phase 2. Matches against knowledge-graph symbol IDs rather than file paths, surviving renames. See 3_vision.md §1.1.
Tag	Free-form string label on a learning record. Survives file renames where path globs don’t. Matched as exact strings in phase 1. See 2_design.md §3.2.