AgentsMD: The ToolboxMD project that owns the portable agent operating contract and its installable workflow plugin. Avoid: Matt Skill pack, global prompt repository
Model Router: The separate ToolboxMD router, labelled Prism in the interface, that accepts delegated work and selects its models, effort, and roles. AgentsMD owns workflow, authority, proof, and review. Avoid: AgentsMD model policy, routing matrix
Planner: The top-level agent the user works with. It owns judgment, specs, and acceptance, and delegates the rest. Avoid: main agent
Routing tool: The host-provided means of delegating work with routed model and role selection, such as Model Router. Avoid: model-routing Skill
Project Direction:
The coherent repository-root triad of VISION.md, MISSION.md, and
OBJECTIVE.md that owns the project's confirmed long-range destination,
present purpose, and single current milestone-level outcome.
Avoid: North Star, strategy bundle, project brief
Potentially stale Project Direction:
A committed Project Direction whose known upstream is ahead or diverged and
changes at least one direction file relative to HEAD. The loader reports
potentially_stale and loads those files from the upstream ref; the checkout
copies stay stale until the intended base is reconciled and all three are reread.
Avoid: missing Project Direction, invalid Project Direction
Vision:
The grand and visionary aspirational long-range destination the project wants
to make real, not the current work. It expands ambition and may be directional
rather than measurable. It is recorded in root VISION.md.
Avoid: long-term Objective, roadmap
Mission:
The strategic present-tense reason the project exists, the problem it solves,
and its approach to moving toward the Vision, grounded in what the project does
now. It is recorded in root MISSION.md.
Avoid: Vision, Objective, company slogan
Objective:
The single current milestone-level outcome the project must accomplish now,
with a recognizable completion condition. It is narrower than the Mission but
broader than an individual request, task, Issue, commit, or PR. It normally
organizes and survives several contributing Issues and is recorded in root
OBJECTIVE.md.
Avoid: Current Goal, task outcome, task list, backlog
Delivery Authority: Authorization to deliver an identified outcome or scope in named Issues and repositories through granted operations and targets. It carries through routine repairs; explicit candidate limits still apply. Avoid: standing authority, repository-wide permission
Delivery System: The portable AgentsMD contract that carries qualified work through separate implementation, Proof, review, merge, release, distribution or deployment, installation or activation, loading, Live Verification, and website-current states. Avoid: release pipeline, orchestration engine
Delivery Profile:
The optional root .toolboxmd/delivery.json containing only a Project's real
delivery-command, artifact-build, and website-mapping differences from the
shared Delivery System.
Avoid: Project Record, release policy, delivery state
Merge Unit: One independently reviewable and mergeable change, or one dependent stack that must ship as a single release identity and SemVer transition. Avoid: commit, implementation slice
Final approval PR: The cumulative outcome PR merged under human approval for the task and intended base. Avoid: component PR, task integration branch
Component PR: An internally reviewed change feeding the integration branch and final approval PR. Avoid: final approval PR, independent release
Implementation Slice: One authored part of a Merge Unit with an exclusive writer and an independent exact-SHA review obligation. Avoid: generated promotion pull request, Merge Unit
Delivery Finalization: The post-review lifecycle step for a verified terminal outcome that closes tracker state truthfully, checks every temporary checkout for removal, and retires eligible task-owned transient resources. It preserves required durable information and justified current needs, records exact retained exceptions, and distinguishes unresolved obstacles from completed finalization. Avoid: cleanup sweep, Repository Reconciliation
Temporary checkout: A worktree or equivalent task workspace whose lifecycle ends with verified removal or an exact retained exception during Delivery Finalization. Required persistent local state has a stable home outside it. Avoid: archive, permanent task workspace
Fresh context: A child start seeded with the minimal durable packet and no prior transcript. Do not fork the parent transcript. Outside the routing tool, the default is a host child that honors the requested model and effort and is visible to the user, else a nested child in the coordinator session. A separate host task is the exception when the slice must outlive the parent or continue after it stops, or when the host offers no visible child. Avoid: new sidebar session, new chat as the default, forked parent context
Repository Reconciliation: A bounded repair path triggered when repository orientation detects drift. It refreshes exact evidence, resolves only approved legacy changes, and preserves durable, unsafe, or ambiguous state with an explicit next action. Avoid: cleanup sweep, recurring repository audit
Review-ready pull request stack: A dependency-ordered set of pull request layers whose exact heads, bases, proof, review, revalidation, retarget, and merge states are recorded independently. It begins only from a complete exact-SHA predecessor candidate. Avoid: dependent branch chain, pull request queue
Parent Spec only:
An explicit request that runs to-spec through verified parent Issue
publication and opts out before ticket decomposition or implementation.
Avoid: parent-Spec-only, planning-only
Skill:
An Agent Skills-compatible package discovered through SKILL.md, with a named
invocation boundary and metadata describing when to load its body.
Avoid: command, prompt file
Skill Catalogue: The authoritative AgentsMD inventory of active Skills, retained procedures, and deferred, retired or upstream-reference material, including provenance. Avoid: Product Registry, Plugin Registry
Active Skill: A Skill that AgentsMD owns and exposes through active plugin discovery. Avoid: installed Skill, referenced Skill
Adapted Skill: A Skill or retained procedure whose current AgentsMD behavior derives from an external Skill and records that origin and local change. Avoid: copied Skill, Matt Skill
Native Skill: A Skill or retained procedure created and maintained inside AgentsMD. Avoid: adapted Skill
ToolboxMD-native Skill: A Skill or retained procedure created in another ToolboxMD project and intentionally bundled by AgentsMD with its original lineage intact. Avoid: third-party Skill, AgentsMD-native Skill
Upstream Reference: An external Skill recorded for later evaluation but absent from active AgentsMD discovery. Avoid: supported Skill, bundled Skill
Deferred Skill: A Skill kept inactive until its recorded reconsideration trigger occurs. Avoid: retired Skill
Retired Skill: A previously considered Skill whose behavior is intentionally absent from active AgentsMD discovery. Avoid: deleted Skill, deferred Skill
Product Registry:
The toolbox.md discovery surface for ToolboxMD products and their roles.
Avoid: Skill Catalogue, Plugin Registry
Plugin Registry:
The toolboxmd/marketplace distribution surface for installable ToolboxMD
plugins.
Avoid: Product Registry, Skill Catalogue
ContextMD: The future ToolboxMD Agent Knowledge and Learning System. Avoid: glossary, context file
World Model: ContextMD's structured representation of entities, relationships, observations, provenance, and learned experience. Avoid: glossary, project documentation
Persistent Host Automation: Agent-created or agent-maintained host services, scheduled jobs, checks, and recovery automation that operate beyond the current task. It excludes project-local build tooling and one-off diagnostics. Avoid: machine runbook, persistent script
Scoped proof: Complete required coverage composed from authenticated unaffected baseline results and current affected and artifact checks under a trusted Project policy. Avoid: skipped proof, fast check, complete test rerun
Complete-proof baseline: An exact commit with direct successful proof for every declared check, used as the cumulative comparison anchor for scoped proof. Avoid: previous scoped candidate, latest commit
Operations module: The existing name for an Operations-owned Reference containing procedure detail. It is not a separate package or discovery type. Avoid: custom loader, orchestration service
Elon method: The AgentsMD method combining first principles, idiot index, current constraint, and the ordered Algorithm to choose means within confirmed direction and authority.
Algorithm: The fixed sequence of questioning requirements, deleting unnecessary work, simplifying survivors, accelerating, and automating last.
Current constraint: The evidence-supported limiter whose relief enables the next useful progress toward the authorized outcome under confirmed Project Direction. Avoid: busiest component, permanent bottleneck
Idiot index: The ratio of quoted finished cost to constituent cost for comparable units and scope, used as a diagnostic hypothesis about possible overhead. Avoid: guaranteed savings, delivery estimate
Canonical instruction source:
The stable global/AGENTS.md shared through each configured host's native global link.
Avoid: plugin-cache instructions, project instructions
Private preferences:
Personal defaults in PREFERENCES.md at the canonical AgentsMD clone root,
subject to explicit task instructions, required project constraints and authority.
Avoid: shared policy, project preferences
Procedure: Instructions for a bounded action, selected through Operations and stored in ordinary linked files without a separate host Skill discovery entry. Avoid: hidden Skill, mode command
Reference: A linked supporting document containing detail, examples, or constraints loaded when its caller's stated condition applies. It has no separate discovery entry. Avoid: hidden Skill, package type
Workflow: An ordered sequence of actions toward an outcome, which may use several Procedures and References; it is not another registration type. Avoid: Skill, package type
Playbook: pstack's upstream term for an operating guide. AgentsMD retains its useful content in Procedures and References, without a separate playbook package type. Avoid: fourth package kind
Human summary: The three lines that open every agent-written Issue and PR, What, Why, and So what, so a human can tell what it is about and what happens next without reading further. Avoid: TL;DR, executive summary
Elon record:
The four lines every Issue and PR carries: Requirements and who asked, Deleted,
Bottleneck, Checked myself. The Project Direction hook blocks gh creation
without them.
Avoid: Algorithm record, design rationale