#### Context Part of the EN baseline term extraction (#916+). The hub and sub-pages of the [GitLab Duo Agent Platform](https://docs.gitlab.com/user/duo_agent_platform/) section were scanned for terminology candidates. Pages processed (45 pages total): **Hub:** - `doc/user/duo_agent_platform/_index.md` **Agents:** - `doc/user/duo_agent_platform/agents/_index.md` (summary only) - `doc/user/duo_agent_platform/agents/custom.md` - `doc/user/duo_agent_platform/agents/external.md` - `doc/user/duo_agent_platform/agents/tools.md` - `doc/user/duo_agent_platform/agents/foundational_agents/_index.md` (summary only) - `doc/user/duo_agent_platform/agents/foundational_agents/security_analyst_agent.md` - `doc/user/duo_agent_platform/agents/foundational_agents/ci_expert_agent.md` - `doc/user/duo_agent_platform/agents/foundational_agents/planner.md` (summary only) - `doc/user/duo_agent_platform/agents/foundational_agents/data_analyst.md` (summary only) **Flows:** - `doc/user/duo_agent_platform/flows/_index.md` (summary only) - `doc/user/duo_agent_platform/flows/custom.md` - `doc/user/duo_agent_platform/flows/execution.md` - `doc/user/duo_agent_platform/flows/foundational_flows/_index.md` (summary only) - `doc/user/duo_agent_platform/flows/foundational_flows/developer.md` - `doc/user/duo_agent_platform/flows/foundational_flows/code_review.md` (summary only) - `doc/user/duo_agent_platform/flows/foundational_flows/software_development.md` - `doc/user/duo_agent_platform/flows/foundational_flows/fix_pipeline.md` - `doc/user/duo_agent_platform/flows/foundational_flows/convert_to_gitlab_ci.md` - `doc/user/duo_agent_platform/flows/foundational_flows/sast_false_positive_detection.md` - `doc/user/duo_agent_platform/flows/foundational_flows/agentic_sast_vulnerability_resolution.md` (summary only) - `doc/user/duo_agent_platform/flows/foundational_flows/secret_false_positive_detection.md` (summary only) **Customize:** - `doc/user/duo_agent_platform/customize/_index.md` (summary only) - `doc/user/duo_agent_platform/customize/agent_skills.md` - `doc/user/duo_agent_platform/customize/custom_rules.md` - `doc/user/duo_agent_platform/customize/review_instructions.md` (summary only) - `doc/user/duo_agent_platform/customize/agents_md.md` **Triggers, context, security:** - `doc/user/duo_agent_platform/triggers/_index.md` - `doc/user/duo_agent_platform/context.md` - `doc/user/duo_agent_platform/security_threats.md` - `doc/user/duo_agent_platform/composite_identity.md` - `doc/user/duo_agent_platform/authentication.md` - `doc/user/duo_agent_platform/environment_sandbox.md` - `doc/user/duo_agent_platform/ai_catalog.md` - `doc/user/duo_agent_platform/sessions/_index.md` - `doc/user/duo_agent_platform/model_selection.md` - `doc/user/duo_agent_platform/turn_on_off.md` (summary only) - `doc/user/duo_agent_platform/troubleshooting.md` (summary only) **Model Context Protocol:** - `doc/user/gitlab_duo/model_context_protocol/_index.md` - `doc/user/gitlab_duo/model_context_protocol/mcp_server.md` - `doc/user/gitlab_duo/model_context_protocol/mcp_clients.md` - `doc/user/gitlab_duo/model_context_protocol/mcp_server_tools.md` - `doc/user/gitlab_duo/model_context_protocol/ai_catalog_mcp_servers.md` **Knowledge graph:** - `doc/user/project/repository/knowledge_graph/_index.md` **Agentic chat:** - `doc/user/gitlab_duo_chat/agentic_chat.md` **404 (page does not exist):** `doc/user/duo_agent_platform/customize/memories.md` **Scope note:** This batch covers the hub and all core sub-pages of the Duo Agent Platform section (45 pages total). Pages not covered are low terminology value: `get_started_agent_platform.md` (already in #923); tutorial pages under `tutorials/duo/`, `tutorials/duo_chat_issues/`, `tutorials/connect_claude_desktop_with_gitlab_mcp_server/`; `doc/user/administration/gitlab_duo/configure/access_control/` (admin); chat procedural pages (best_practices, example_prompts, troubleshooting); `agents/external_examples/`, `flows/custom_flows_schema/`, `flows/execution_variables/` (config/schema references); `gitlab_duo/semantic_code_search/`; and `model_context_protocol/mcp_server_troubleshooting/`. Each candidate was scored against the 8-criteria extraction framework in [`terminology_artifacts/term-extraction.md`](https://gitlab.com/gitlab-com/localization/docs-site-localization/-/blob/terminology/term-extraction-workflow/terminology_artifacts/term-extraction.md) (2+ criteria required to include; MR !30) and then filtered by the translation-risk test in [`terminology_artifacts/term-brief.md`](https://gitlab.com/gitlab-com/localization/docs-site-localization/-/blob/main/terminology_artifacts/term-brief.md). Key test: does the term create a real translation difficulty a translator cannot resolve from context alone? Criteria numbers for each term are recorded in the Notes column. Cross-references run: Quickterm CSV (all 20 candidates checked; findings in Notes and FR column). MS Terminology Collection FR (all 20 candidates checked; zero matches -- this section covers AI agent platform terminology not present in the MS collection). Terms already covered in earlier extraction batches are not repeated here. `service account`, `merge request`, `pipeline`, `CI/CD`, `false positive`, `audit trail`, `sandbox` (already in Quickterm), `agentic` / `agentic chat` (already in Quickterm), and `MCP` / `Model Context Protocol` (already in Quickterm) are all excluded. #### Extracted terms <table> <tr> <th>Term</th> <th>File(s)</th> <th>Count</th> <th>In Quickterm</th> <th>FR (from Quickterm)</th> <th>Notes</th> </tr> <tr> <td>AI agent</td> <td>doc/user/duo_agent_platform/_index.md, agents/custom.md, agents/external.md, agents/foundational_agents/*.md</td> <td>~150</td> <td>Yes (<em>agent d'IA</em>, Preferred)</td> <td>agent d'IA</td> <td>An autonomous software entity that perceives its environment, makes decisions, and takes actions to complete tasks in the GitLab Duo Agent Platform. Included despite Quickterm coverage because the Phrase master TBX has this concept (ID UNjYrmRjzXG88l1RYNBwS4, preferred: true) with no FR entry -- Quickterm has not been synced into Phrase and the sync timeline is unknown. Our TBX import is the confirmed path to put FR into translators' working environment. FR is confirmed as <em>agent d'IA</em> from Quickterm (Preferred). Critical disambiguation note: standalone "Agent" in Quickterm and in Phrase maps to the GitLab Agent for Kubernetes -- translators must not apply that entry to Duo Agent Platform content. Criteria: 1,2,4,5,7,8.</td> </tr> <tr> <td>flow</td> <td>doc/user/duo_agent_platform/_index.md, flows/custom.md, flows/foundational_flows/*.md</td> <td>~100</td> <td>No</td> <td></td> <td>GitLab-specific noun: an AI-powered multi-step automated workflow in the Agent Platform. Distinct from "CI/CD pipeline" and from the general English word "flow" (flux). FR has no established equivalent: "flux" is a data-flow term, "workflow" is often kept in EN, "processus" is too generic. High standardization risk: this word appears on nearly every Agent Platform page. Criteria: 1,2,4,5,7,8.</td> </tr> <tr> <td>foundational agent</td> <td>doc/user/duo_agent_platform/_index.md, agents/foundational_agents/*.md</td> <td>~20</td> <td>No</td> <td></td> <td>One of the pre-configured purpose-built agents shipped with GitLab Duo Agent Platform (Planner, Data Analyst, Security Analyst, CI Expert). Contrast with custom agent and external agent. "Foundational" has no single natural FR equivalent: "fondamental" vs "natif" vs "de base" -- the same ambiguity that affects "base role" (#928). Must be consistent across the three agent types. Criteria: 1,2,5,7,8.</td> </tr> <tr> <td>foundational flow</td> <td>doc/user/duo_agent_platform/_index.md, flows/foundational_flows/*.md</td> <td>~15</td> <td>No</td> <td></td> <td>One of the pre-configured flows shipped with GitLab Duo Agent Platform (Developer Flow, Code Review Flow, Software Development Flow, Fix CI/CD Pipeline Flow, etc.). Same taxonomy and same "foundational" translation risk as foundational agent. Must be translated consistently with it. Criteria: 1,2,5,7,8.</td> </tr> <tr> <td>custom agent</td> <td>doc/user/duo_agent_platform/agents/custom.md, _index.md</td> <td>~50</td> <td>Partial (as <code>Custom</code>)</td> <td></td> <td>A user-created agent configured with a system prompt and optional tools, as opposed to a foundational agent. Quickterm has "Custom" with no FR entry. Must be consistent with "custom flow" (below), "custom role" (#928 as <em>rôle personnalisé</em>), and "custom rules". If "personnalisé" is adopted, all four must align. Criteria: 1,2,5,7,8.</td> </tr> <tr> <td>custom flow</td> <td>doc/user/duo_agent_platform/flows/custom.md, _index.md</td> <td>~50</td> <td>Partial (as <code>Custom</code>)</td> <td></td> <td>A user-created AI-powered workflow (flow), as opposed to a foundational flow. Same translation pattern as custom agent. Also appears as "Custom flows" (plural, section heading). Must align with the agent taxonomy above. Criteria: 1,2,5,7,8.</td> </tr> <tr> <td>external agent</td> <td>doc/user/duo_agent_platform/agents/external.md, _index.md</td> <td>~30</td> <td>No</td> <td></td> <td>An agent that integrates with a third-party AI model provider (Claude Code, OpenAI Codex, Amazon Q, Gemini) and runs in CI/CD using the external provider's infrastructure. Security considerations differ from built-in agents. FR "agent externe" is the expected calque but must clearly distinguish from custom agent. Note: "Agent" alone in Quickterm maps to the GitLab Agent for Kubernetes (GitLab Agent), not AI agents -- translators must not apply that entry to Agent Platform terminology. Criteria: 1,2,3,5,7.</td> </tr> <tr> <td>AI Catalog</td> <td>doc/user/duo_agent_platform/agents/custom.md, flows/custom.md, agents/external.md</td> <td>~40</td> <td>Partial (as <code>Agent Catalog</code>)</td> <td></td> <td>The GitLab feature for discovering, creating, managing, and sharing custom agents and flows, accessible via Explore > AI Catalog. Quickterm has "Agent Catalog" (the previous name) with no FR entry. The feature has been renamed to "AI Catalog" -- the Quickterm entry is outdated. FR: "catalogue IA" vs "Catalogue IA" vs kept in EN? Decision needed. Criteria: 1,2,5,6,8.</td> </tr> <tr> <td>managing project</td> <td>doc/user/duo_agent_platform/agents/custom.md, flows/custom.md, agents/external.md</td> <td>~15</td> <td>No</td> <td></td> <td>The GitLab project that owns and controls the configuration of an agent or flow: determines visibility (public/private), who can edit it, and the service account naming. Confusability risk: "managing project" sounds like "a project involved in management", not "the project that manages this agent". FR "projet gestionnaire" or "projet de gestion" -- both awkward. Not in Quickterm. Criteria: 1,2,5,7.</td> </tr> <tr> <td>system prompt</td> <td>doc/user/duo_agent_platform/agents/custom.md</td> <td>~5</td> <td>No (closest: <code>prompt</code>)</td> <td></td> <td>The foundational instruction set given to a custom agent that defines its role, persona, behavior, and expertise. Different from a user prompt (the message a user types). Quickterm has "prompt" kept as "prompt" in FR (Preferred). "System prompt" in FR: "prompt système" or "invite système" -- the "système" placement is inconsistent with treating "prompt" as a standalone EN loan word. Decision needed on whether to keep the full compound in EN or translate "system". Criteria: 1,2,3,7.</td> </tr> <tr> <td>composite identity</td> <td>doc/user/duo_agent_platform/composite_identity.md, flows/custom.md, agents/external.md</td> <td>~5</td> <td>No</td> <td></td> <td>GitLab authentication mechanism that combines the user's group memberships with a service account's memberships to control what a flow or external agent can access, preventing privilege escalation across top-level groups. Very specific GitLab security concept. FR "identité composite" is a calque but "composite" is not standard in FR authentication vocabulary: "identité combinée" or "identité croisée" are alternatives. No established FR term. Criteria: 1,2,3,6,7.</td> </tr> <tr> <td>prompt injection</td> <td>doc/user/duo_agent_platform/agents/external.md, gitlab_duo/model_context_protocol/mcp_server.md</td> <td>~6</td> <td>No</td> <td></td> <td>Security attack where malicious instructions are embedded in data processed by an AI agent to override its behavior or extract sensitive information. Prominent in security considerations for external agents and MCP. FR: "injection de prompt" -- consistent with Quickterm keeping "prompt" in EN, but "injection" has multiple meanings in security (SQL injection, code injection). Must not be confused with SQL/code injection. Criteria: 2,3,5,7,8.</td> </tr> <tr> <td>agent skill</td> <td>doc/user/duo_agent_platform/customize/agent_skills.md</td> <td>~20</td> <td>No</td> <td></td> <td>A structured knowledge file (SKILL.md) following the open Agent Skills specification that gives AI agents specialized knowledge and reusable workflows for specific tasks. Agents load the skill automatically when a matching task is detected. Also used as "Agent Skills" (plural, feature name). FR: "compétence" vs "aptitude" vs keeping "skill" in EN -- "skill" is widely kept in EN in FR tech vocabulary. No established FR term in this AI context. Criteria: 1,2,5,6,7,8.</td> </tr> <tr> <td>knowledge graph</td> <td>doc/user/project/repository/knowledge_graph/_index.md, duo_agent_platform/_index.md, gitlab_duo/model_context_protocol/mcp_clients.md</td> <td>~5</td> <td>No</td> <td></td> <td>A structured, queryable representation of a code repository that models code entities and their relationships to power AI features. Available as a beta MCP server. FR: "graphe de connaissances" (plural) vs "graphe de connaissance" (singular) -- the singular/plural distinction matters in FR for this term. "Graphe de connaissances" is closer to the semantic web / ontology meaning, which may or may not align. Criteria: 1,3,5,6,8.</td> </tr> <tr> <td>GitLab Credits</td> <td>doc/user/duo_agent_platform/_index.md, flows/custom.md, agents/custom.md</td> <td>~20</td> <td>No</td> <td></td> <td>Unit of consumption for GitLab Duo Agent Platform features: each agent or flow operation consumes credits. Some features are free but require credits in a Monthly Commitment Pool. FR "crédits GitLab" is natural but must be clearly defined as a billing/consumption unit, distinct from other senses of "crédit" (financial credit, course credit). Not in Quickterm. Criteria: 1,2,4,5,8.</td> </tr> <tr> <td>reachability analysis</td> <td>doc/user/duo_agent_platform/agents/foundational_agents/security_analyst_agent.md</td> <td>~3</td> <td>No</td> <td></td> <td>Code analysis that determines whether a vulnerable code path is actually reachable and exploitable by an attacker, used by the Security Analyst Agent to filter out theoretical vulnerabilities. High translation risk: "reachability" in FR security contexts has no standard term. "Accessibilité" is wrong (it means WCAG accessibility). "Atteignabilité" or "joignabilité" are technical calques with no established use. Translator cannot resolve this from context alone. Criteria: 1,2,3,5.</td> </tr> <tr> <td>flow trigger</td> <td>doc/user/duo_agent_platform/triggers/_index.md, flows/custom.md, agents/external.md</td> <td>~15</td> <td>Partial (as <code>trigger</code>)</td> <td></td> <td>The event configuration that determines when a custom flow or external agent runs automatically: mention (service account mentioned in a comment), assign (service account assigned to issue/MR), assign reviewer (service account assigned as MR reviewer), pipeline events. Quickterm has "trigger" as "déclencher" (verb only). The noun form "a trigger" / "a flow trigger" is not covered: "un déclencheur" or "un événement déclencheur" are FR candidates, neither standardized for this context. Must not be confused with CI/CD pipeline triggers. Criteria: 1,2,4,5,7.</td> </tr> <tr> <td>embeddings</td> <td>doc/user/project/repository/knowledge_graph/_index.md</td> <td>~5</td> <td>No</td> <td></td> <td>Dense vector representations used by the knowledge graph to model code entities and their semantic relationships. Identified during second-pass extraction of knowledge graph pages. FR "plongements" is mathematically correct but rarely used in French tech writing; many French AI texts retain "embeddings" unchanged. No established GitLab FR convention. Translator cannot resolve the keep-vs-translate decision from context alone. Not in MS Terminology Collection. Criteria: 1,3,6,8.</td> </tr> <tr> <td>RAG / Retrieval-Augmented Generation</td> <td>doc/user/project/repository/knowledge_graph/_index.md</td> <td>~3</td> <td>No</td> <td></td> <td>AI technique that enhances model responses by retrieving relevant information from an external knowledge source at inference time; used by the GitLab knowledge graph to ground AI features in repository data. Identified during second-pass extraction. FR translation of the full form is structurally different: "génération augmentée par récupération" -- where "récupération" is a false friend (it means retrieval but also recovery). The acronym RAG is widely retained in French AI texts, but GitLab needs a consistent policy. Not in Quickterm or MS Terminology Collection. Criteria: 1,2,3,6,8.</td> </tr> <tr> <td>version pinning</td> <td>doc/user/duo_agent_platform/ai_catalog.md</td> <td>~5</td> <td>No</td> <td></td> <td>Locking a custom agent or flow to a specific version so it does not automatically update when a new version is published in the AI Catalog. Identified during second-pass extraction. "Pinning" is a metaphor with no natural French equivalent: "épinglage de version" is literal and awkward; "verrouillage de version" and "blocage de version" are both used inconsistently in French software documentation. Translator cannot choose without a GitLab-level decision. Not in MS Terminology Collection. Criteria: 1,2,4,7,8.</td> </tr> <tr> <td>vetted MCP server</td> <td>doc/user/gitlab_duo/model_context_protocol/ai_catalog_mcp_servers.md</td> <td>~5</td> <td>No</td> <td></td> <td>An MCP server that has been reviewed and approved by GitLab or a partner organization before being listed in the AI Catalog, as distinct from community-submitted servers. Identified during second-pass extraction. "Vetted" has no single French equivalent: "vérifié" (verified), "approuvé" (approved), "audité" (audited), and "certifié" (certified) each imply a different level of governance rigor. Translator cannot select without knowing GitLab's governance model for MCP server classification. Not in MS Terminology Collection. Criteria: 1,2,3,6,8.</td> </tr> </table> #### Terms excluded after applying extraction criteria **Already in Quickterm or prior batches:** - `agent` (standalone) -- Quickterm entry maps to the GitLab Agent for Kubernetes, not AI agents. Translators working on Agent Platform pages must not apply this entry to Duo Agent Platform content. - `agentic`, `agentic chat`, `agentic AI` -- already in Quickterm - `AI agent` -- in Quickterm (*agent d'IA*, Preferred) but moved to the candidate table: the Phrase master TBX has this concept with no FR entry and Quickterm→Phrase sync timeline is unknown; our TBX is the reliable path to get FR into Phrase - `sandbox` / `environment sandbox` -- already in Quickterm (*environnement sandbox*) - `MCP` / `Model Context Protocol` -- already in Quickterm - `false positive`, `audit trail` -- already in Quickterm - `service account` -- already covered in #928 - `merge request`, `pipeline`, `CI/CD`, `runner`, `job` -- already covered in prior batches - `prompt` (standalone) -- already in Quickterm as *prompt* (kept in EN) **Clean FR translation, no standardization risk:** - `custom rules` -- *règles personnalisées* is natural and unambiguous - `prompt caching` -- *mise en cache du prompt* is straightforward - `context window` -- *fenêtre de contexte* is natural - `session` (AI session) -- *session* translates directly - `privileged mode` / `privileged runner` -- *mode privilégié* is standard - `network policy` -- generic term used across GitLab and Kubernetes - `parallelization` -- *parallélisation* is standard - `semantic versioning` -- *versionnage sémantique* is established in FR tech documentation - `tool output sanitization` -- *assainissement des sorties d'outil* is derivable from context **Product names, feature names, filenames (not translated):** - `Developer Flow`, `Code Review Flow`, `Software Development Flow`, `Fix CI/CD Pipeline Flow`, `Convert to GitLab CI/CD Flow`, `SAST False Positive Detection Flow` -- named GitLab features - `Planner Agent`, `Security Analyst Agent`, `Data Analyst Agent`, `CI Expert Agent` -- named GitLab features - `semantic code search` -- named GitLab feature - `tanuki-verified` -- UI badge label referencing GitLab's mascot; out of scope for TBX - `AGENTS.md`, `SKILL.md`, `agent-config.yml`, `chat-rules.md` -- filenames **Out of scope for this section (general GitLab/Git terms surfaced by second pass):** - `squash`, `namespace`, `feature flag`, `confidential` (access control flag), `iid`, `scope` (search parameter) -- surfaced from MCP tool descriptions that expose standard GitLab API operations; belong in other extraction batches if not already covered - `work item` -- valid GitLab-specific abstraction for issues/epics/tasks; likely already covered in #925 (Plan and Track batch); to be confirmed before adding **Internal/implementation details (not user-facing):** - `Cloud Connector JWT`, `GLGO exchange JWT`, `CustomersDot service access token` -- AI Gateway authentication token types; not terminology users encounter directly - `workload pipeline` -- ephemeral pipeline created internally during flow execution; mentioned in troubleshooting only, not a user-configured concept - `multi-shot reasoning` -- technical AI inference pattern; low user-facing visibility in this section **Informal or descriptive language (not recurring technical terms):** - `lethal trifecta` -- informal label used once in `security_threats.md` for a three-part threat pattern; not a standardized term - `role-specific reasoning`, `domain-specific expertise`, `context awareness` -- descriptive phrases characterizing foundational agents in overview text; not terms that recur in a standardized way - `composite identity workflow` -- `composite identity` is already in the table; adding `workflow` does not create a separate terminological unit **Data quality note:** - `software development lifecycle` -- EN:Preferred in Quickterm but FR entry (*cycle de livraison logicielle*) is marked Forbidden. Data quality issue for Katerina (#931); not a new TBX candidate.