agent_config | The SDK call that reads an agent’s saved prompt and skills from DecimalAI for your code. See the three things called an agent. | Prompt & skills |
| agentversion | The open spec and PyPI package for the manifest, diff and compatibility format DecimalAI builds on. | Manifests |
| Baseline | The manifest a regression check diffs against, typically the last one seen in production. | Regression Check |
| Compatibility Report | Generated when a new manifest is registered: it classifies every existing trace as keep, repair, replay or drop. | Manifests |
| Compatibility Verdict | The per-trace outcome: keep, repair, replay or drop. | Manifests |
| Component | One versioned piece of a manifest: tool, model, prompt, skill, subagent or output_schema. | Manifests |
| Component Verdict | The per-component result of diffing two manifests: COMPATIBLE, REPAIRABLE, INCOMPATIBLE or MISSING. | Manifests |
| Dataset | Curated training data built from filtered, scored production traces. | Datasets |
| Decision Engine | Combines quality and compatibility scores into one keep, repair, replay or drop verdict per trace. | Verdicts |
| Degraded | A parent trace’s status when some, but not all, of its child traces errored. | Multi-agent |
| Delegation | An orchestrator assigns a task to a sub-agent. Control returns when the sub-agent finishes. | Multi-agent |
| Delivered | A skill’s full text reached the model. | Offered, Delivered, Loaded |
| DPO | Direct Preference Optimization: a dataset format of chosen and rejected pairs. | Datasets |
| Drift | A sub-agent’s actual config diverges from what the orchestrator’s manifest expects. | Multi-agent |
| Eval Score | One evaluation result: name, a score of 0.0 to 1.0, passed, source and category, such as quality or compatibility. | Score sources |
| Eval Verdict | The trace-level outcome computed from its eval scores: pass, fail or review. | Verdicts |
| Evaluator | A configured quality check: deterministic, LLM-as-judge or custom. | Evaluations |
| Handoff | A transfer of control between peer agents. The original agent may not regain control. | Multi-agent |
| Impact Report | A regression check’s PR comment: each production trace marked HIGH, MEDIUM or LOW impact. | Regression Check |
| Impact Severity | HIGH, MEDIUM or LOW: how structurally a manifest change touches a trace. A different scale from Severity. | Compatibility Policies |
instrument | Turns on tracing for a framework, as decimalai.init(langchain=True) does. It doesn’t add a skill to your workspace. | Tracing |
| Invocation Mode | Who fires a skill: model (Automatic), user (On-demand, no context cost until called) or any. | Authoring Skills |
| Lift | How much better an agent scores with a skill than without it. Measured by running the same cases with and without it. | Authoring Skills |
load_skill | The tool a model calls to read a skill’s full text. | Skill delivery |
| Loaded | The model itself asked for a skill. Only this step says the model chose it. | Offered, Delivered, Loaded |
| Manifest | A snapshot of an agent’s full configuration at a point in time: tools, models, prompts, skills and sub-agents. | Manifests |
| Manifest Hash | A SHA-256 fingerprint of a manifest’s structure. Registering the same hash for the same agent changes nothing. | Manifests |
| Manifest Status | A manifest’s lifecycle state: active, superseded or draft. | Manifests |
| Offered | A skill’s name was on the list the model saw. | Offered, Delivered, Loaded |
| Orchestrator | An agent that delegates to sub-agents, inferred from manifests or trace links rather than declared. A sub-agent’s trace carries parent_trace_id. | Multi-agent |
| Pack | A curated bundle of skills for a common agent role, such as support or coding. | Skill packs |
| Parent Trace | An orchestrator’s trace, which child traces point to through parent_trace_id. | Multi-agent |
| Progressive Disclosure | A skill’s name and description are always in context, and its full text loads only on demand. | Skill delivery |
| Regression Check | The pre-deploy GitHub Action that diffs a candidate manifest against the baseline. | Regression Check |
| Repair | A deterministic fix of a trace to match a new manifest, at zero LLM cost. | Manifests |
| Replay | Re-running a historical trace against the current agent to compare outputs. | Replay |
| Revert | The same manifest hash reappears: the old manifest is reactivated and the current one superseded. | Manifests |
Runs (--runs) | Repeats the whole eval suite N times and averages each case. More runs narrow the error bars. | skillevaluation |
| Session | A group of traces forming one conversation, linked by session_id. | Trace schema |
| Setup version | A numbered record of an agent’s prompt and skills. You can compare one with another and roll back to it. | Versions |
| Severity | How much a component change matters: none, minor, moderate or major. | Manifests |
| SFT | Supervised Fine-Tuning: a dataset format of input and output pairs. | Datasets |
| Skill | A reusable instruction file (SKILL.md) that changes how an agent behaves. Not a tool. | SKILL.md format |
| Skill Scope | public (a standard the base model may learn over time) or private (house knowledge it never absorbs). | Authoring Skills |
| Skill Type | capability teaches an ability the base model lacks, such as an internal API. preference steers the form of something it can already do, such as a commit convention. | Authoring Skills |
| skillevaluation | The open spec and PyPI package for A/B benchmarking a skill: each case runs with and without it. | skillevaluation |
| Skills Ablation | Running the same eval suite with and without a skill. It’s how lift is measured and retirement detected. | Authoring Skills |
| SkillSafety | The band that says whether a skill is safe to load: Passed, Caution, Blocked or Not yet reviewed. | How skills are vetted |
| SkillScore | The 0 to 100 score of evaluated quality, from benchmark results, task evaluations and AI ratings. Usage isn’t an input. | SkillScore |
| Source Badge | Where a registry skill came from, not an endorsement: community (authored on DecimalAI), imported (synced from GitHub) or featured (a community skill the ranking promoted). | Browse the registry |
| Source Type | Where a trace came from, such as production, test or evaluation. | Trace schema |
| Span | A timed segment within a trace: llm, tool, retriever or other. Spans nest through parent_span_id. | Trace schema |
| Surface | A policy grouping for compatibility rules: tool_registry, model_runtime, prompt_stack, skill_registry, subagents or output_contract. | Manifests |
| Trace | The record of one run, from input to output. | Trace schema |
| Trigger Case | An eval case marked should_trigger: true or false. It grades whether a skill fires at the right times, not its output. | skillevaluation |
| Trigger Health | How reliably an Automatic skill fires at the right times, shown on its registry page. | Browse the registry |
| Version Label | A manifest’s readable name: v1, v2, v3. Auto-incremented, not semantic versioning. | Manifests |
