DSH APEX Plugin derives its versioned preset compositions and runtime design from the following reviewed sources. Code baselines -------------- DeepSeek Harness Minimal and Standard agent compositions: https://github.com/deepseek-ai/deepseek-harness commit 141eb6fef83422698aef7a981029e843e8161534 Minimal Max v0.2 anchored promotion, tool gating, and Windows Git Bash compatibility were adapted from: https://github.com/xiaobright/dsh-anchored-standard commit d97bec91a3d668f4cf1d03ee5f20aae84fb6f85c The APEX v0.3 design was reviewed against the later Anchored Standard state: https://github.com/xiaobright/dsh-anchored-standard commit db4527a2a70a9032d3a8525ce3c0ea6ef528d6fc The compact post-anchor APEX policy adapts the minimal-necessary, reuse-first, and YAGNI principles from: https://github.com/DietrichGebert/ponytail commit 2ed6c52c9d7e5e56942508591085fd45dea277d3 These code sources are distributed under the MIT License. The MIT permission notice used by this project is included in LICENSE. APEX v0.4 adds a local execution guard through the public `ctx.tools.guard` contract in the pinned DeepSeek Harness baseline. Its process-cleanup and validation constraints were derived from this repository's 2026-08-16 pilot; no additional third-party code was copied for that change. Research and design references ------------------------------ The first-request anchoring hypothesis and evaluation design were informed by: https://github.com/xiaobright/modeltest commit 04255b55f16c4439e538239fb9783070c4165081 The decision to keep task classification and extra model calls outside the v0.3-v0.6.3 anchor, and the v0.5+ Pro/Flash role split, were informed by: https://github.com/yjh051108/dsh-routing-suite commit a09eb0ade28e6ec3b8e5eb22985a14f6bfa1fbe5 The v0.6.1 Pro-led revision was also reviewed against the empirical tool-surface and real-Harness assembly findings in: https://github.com/yjh051108/dsh-router-standard commit 69dfd6aa1df4b2dfcfb1abaecc5384b279950019 Its Flash-specific neutral persona, same-request near-field guidance, decision closure, and small tool-surface findings informed the v0.6.1 worker brief. APEX implements those ideas independently through scoped Harness persona and tool contracts; it does not copy the repository's keyword classifier or prompt modules. The referenced repository is MIT-licensed. The bounded continuation guide and anti-runaway behavior were also reviewed against the pinned Super Injector implementation: https://github.com/yjh051108/dsh-super-injector commit c08136a526e7515dca106441e65cf7fccf63bbae APEX does not copy its runtime injector, hot reload, source overrides, or model-facing tool catalog. It applies only the reviewed design principles through public Harness plugin hooks. The referenced repository is distributed under the BSD 3-Clause License. The v0.5+ bounded Goal/Verified/Open/Next/Evidence state and stall-recovery design were informed by: https://github.com/Tiger3807861189/J-Space-Cognition-Suite-V3.6 commit 46bac14a86398b406c6c6d07baf1fef9cfc8d370 The v0.6.1 parent-coordinated, bounded parallel work-item design was informed by the public Kimi Swarm documentation and technical report: https://github.com/MoonshotAI/kimi-help-center/blob/master/en-US/agent/swarm.md https://arxiv.org/abs/2602.02276 These repositories are cited as research and design references. APEX does not copy their evaluation text, file controller, keyword classifier, injector, router, or prompt modules. APEX implements the ideas independently through DeepSeek Harness session events and the official subagent tool/provider APIs. APEX v0.6 adds its bounded Flash Max implementation path through the pinned Harness `tool-subagent` and `agent/request` contracts. No additional third-party code was copied for this change. APEX v0.6.1 was compatibility-reviewed against DeepSeek Harness commit 141eb6fef83422698aef7a981029e843e8161534. It uses the official rc.8 Minimal persistent bash/pwsh shell split, subagent tool filter, delegation policy, and sandboxed filesystem contracts to keep the Flash Max children limited to disjoint workspace code-edit leases. Its work-item protocol and orchestration are implemented independently; no Kimi or other additional third-party code was copied for this change. The current APEX v0.6.1 composition was subsequently compatibility-reviewed against DeepSeek Harness 0.1.1-rc.1 at commit 528c682e061696f5a160f363f236ecbf53cbd006. Its read-only visual-review path uses the official direct vision model, native read_image tool, attachment storage, and image protocol contracts introduced by Harness commits 4fa38d6a23bc2923a27ab4d375dde1b742ca830e and 708a25a8a9362079185abb4f0c747cff4df74398. No image transport or third-party implementation was copied into APEX. APEX v0.6.2 freezes the v0.6.1 preset and introduces a task-neutral core by removing global engineering-task shell, installation, fetch, and research heuristics from the root path. Optional code-worker, runtime-validation, vision-review, research, and durable-state capabilities remain independently unlockable through the public Harness plugin contracts. This composition was compatibility-reviewed against DeepSeek Harness 0.1.1-rc.1 at commit 6dc6b4f9762901ca04931fdcbb25db418fadf9ed. No additional third-party code was copied for this revision. APEX v0.6.3 freezes the v0.6.2 preset and adds an independent host-side evidence layer for artifact generations, Web-validation stages, read-only visual review, structured Vision Flash research, and direct child-scoped PTC presentation for newly labelled Vision Flash code workers. It uses Node.js standard-library hashing plus the existing Harness session, subagent, PTC, code-runtime, and image-context contracts; no additional dependency or third-party implementation was copied. This composition is compatibility-reviewed against DeepSeek Harness 0.1.2-alpha.3 at commit dd6322d604e00eec1ba5e0c8541159906a21094a. Its Minimal composition remains byte-identical to the reviewed baseline. DeepSeek and DeepSeek Harness are names of their respective owner. This community project is not affiliated with or endorsed by DeepSeek.