detect-agent

Several agent environments are currently missed, and Grok Bot's shared marker can be reported as Cursor CLI. This adds explicit environment rules while keeping existing result IDs, existing rules' relative order and the override. Add Crush, Hermes Agent, Kilo Code, Trae, VTCode and Zed terminal context from the public Hugging Face registry; extend Goose with and Kiro with while retaining their previous conditions. Add WorkBuddy using its complete product-name keys, Grok Bot using local or compound cloud-shell markers, QwenWork using non-empty , Doubao Work using non-empty , Cue's Linux desktop-worker context using exact , Muse using its active-agent/tool-call markers, and Manus using its addon-runtime descriptors. Extend the existing Grok Build rule with exact ; other non-empty selector values do not match. Grok Build and Grok Bot remain separate results. Junie detection requires non-empty . User-configurable is ignored, so configuring its data directory alone does not trigger Junie detection or suppress another matching agent. The existing rule order is retained. The new product-specific markers precede inherited Cursor/Claude context; Zed remains a last fallback because also describes human-operated terminals. Configuration paths, packaging flags, and general instance/session/location metadata do not independently identify these new contexts. QwenWork accepts any non-empty product-marker value; site, execution mode, endpoint and old chat/distribution markers do not trigger detection. Doubao Work accepts any non-empty agent-name value without requiring a fixed name. The marker identifies Cue, despite its variable prefix, and does not return . Names and short descriptions for all thirteen new entries are checked against official product sites, documentation and repositories, with links in the evidence notes. Doubao Work and Manus use canonical IDs and and constants and . Their environment-variable conditions are unchanged. Sources and scope: Hugging Face registry at fixed commit e08240742cd64d8f908b22e52358b84c812e882c Hermes harness marker injection Grok Build's terminal-marker contract The included records versions, archive hashes and propagation locations from static inspection of WorkBuddy 5.8.3, Grok Bot 0.71.2 and QwenWorkCN 1.3.1. It contains no proprietary implementation code or real session values and does not claim end-to-end execution tests or coverage of every product variant. QwenWork 1.3.1 sets the product marker to and propagates it to local/host shell paths. The rule intentionally accepts any non-empty value; this observed value is not a matching restriction. VM/QXC guest-shell propagation of the new marker remains unverified; older marker paths are not evidence for that coverage. Doubao Work's comes from a user-provided shell observation on 2026-10-16. The rule accepts any non-empty value and is not an independently reproduced run or a public vendor contract. No real credentials, message/log IDs or runtime values are included; all negative fixtures use synthetic values. Cue's marker and Muse's / non-empty markers come from user-provided runtime observations on 2026-10-16, not locally reproduced runs or inspected application artifacts. Real runtime identifiers are not included. Manus's / descriptor names were reported in an ordinary shell; their values were not supplied. The rule requires non-empty values. This is user-reported shell evidence, not a locally reproduced probe or verification of the inspected desktop distribution. Other addon/group fields do not match independently. A user-reported Grok Bot cloud shell had no ; its rule requires all three of non-empty , exact and exact . The original local branch is retained. Shared infrastructure markers do not prove exclusive brand ownership or resolve nested agents' immediate identity; no tenant/authentication/store/boot/VNC identifiers are used. Validation on Node 26.12.0 and Go 1.27.2: — 250 passed, including 221 shared JSON cases The existing mock-fs dependency fails during test collection on Node 26. The separate test-only commit replaces its internal filesystem patching with precise stubs for present, missing and permission-denied paths. Runtime detection, dependencies and lockfiles are unchanged. Detector cases reset all relevant environment keys and stub filesystem access; TTY state uses the existing controlled fixtures. Cases cover exact and wrong values, empty markers, configuration-only negatives, Grok Bot/Cursor coexistence, each missing or invalid cloud conjunction member, Grok Build separation, Cue-versus-Manus attribution, Muse tool-call context, Manus descriptors, arbitrary non-empty QwenWork product markers and Doubao Work agent names, Zed fallback behavior and retained overrides. Coordinated Python-port contribution: https://github.com/togethercomputer/detect_agent/pull/19. The two shared JSON documents are structurally identical, following the Python port’s upstream synchronization policy.