Anchored Standard Progressive Preset
August 23, 2026 · View on GitHub
Current conclusion
The mechanism is feasible on the pinned maintained @deepseek-ai/dsh@0.1.1-rc.2.code.1 as a per-session Agent Preset. It does not export a warmed model state and does not inspect or replay reasoning text. The useful state is durable Harness session history: the first request is assembled with a two-tool schema, and subsequent schemas are derived from persisted events.
Pinned source and local integration
- Upstream:
xiaobright/dsh-anchored-standard, commitdb4527a2a70a9032d3a8525ce3c0ea6ef528d6fc, MIT. - Bundled package:
packages/anchored-standard-pluginversion0.2.0. - Product-owned preset copy uses the locale-aware 渐进式标准模式 / Progressive Standard Mode fallback, while the maintained Harness source owns the zh/en picker dictionaries for
anchored-standard,cordis-with-products,router-spec, androuter-standard. Unknown user presets keep their own metadata; this desktop repository carries no DSH pnpm patch. - Installer provenance includes
LICENSE,NOTICE,UPSTREAM.json,LOCAL-PATCHES.md, package version, and an app-managed SHA-256 marker. - Installation: atomic sync to
<DSH_HOME>/.agent-presets/anchored-standardwith ownership marker; user-owned same-name presets are never overwritten; invalid packaged resources produce anunavailablenotice and Standard remains operational.
Runtime contract
- A new top-level session exposes only
bashandstr_replace_editor, keeps the Minimal persona, and removes automatic workspace-instruction and skill-catalog messages from that request. - The first persisted
tool/callorassistant/messagepromotes the current compaction epoch (promoteOn: either). - Promoted requests expose the two bootstrap tools plus
dev_tool_search,skill_search, andskill_load. Other Standard capabilities require an explicitdev_tool_searchunlock and are reconstructed from durable events after resume. - A
compaction/endboundary returns to a controlled bootstrap-plus-work set until a new durable signal. Subagents start in the resident phase. - Missing required tools or an invalid composition rejects the selected preset. Standard is not modified and remains the default.
Variants zero-anchored-standard and whoami-standard use a zero-tool anchor turn (fixed test notice or “你是谁” self-introduction) as the first request, promoting on the anchor reply (assistant/message).
Security and privacy boundary
The plugin logs no prompts, responses, credentials, or reasoning. Diagnostics are limited to session identifiers, phase/tool names, and promotion event type inside Harness-owned logging paths. The managed installer never overwrites an unknown or locally modified preset directory. The preset uses the official Harness tool implementations and their existing sandbox/approval policies, except for the upstream Minimal-compatible str_replace_editor filesystem realm documented in the composition.
Evidence
| Retrieved at | Source | Version/date | Key evidence | Confidence |
|---|---|---|---|---|
| 2026-08-16 | Community preset repository | commit db4527a2... | Preset composition, progressive bootstrap hooks, tests, and MIT license | high for mechanism |
| 2026-08-20 | Local pinned rc.8 packages | @deepseek-ai/dsh@0.1.0-rc.8 | Official roster discovers the preset as healthy; Standard remains default; a loopback mock provider captures the two-tool first request and five-tool second request | high |
| 2026-08-16 | DeepSeek thinking mode and context caching | current at plan approval | Ordinary history/cache is not a portable one-time model warm state | medium; revalidate before changing product claims |
| 2026-08-16 | Community modeltest | community benchmark | Reported 98/99 scores are experiment motivation, not a universal product guarantee | medium |
Experimental boundary
The request-schema sequence, persistence, isolation, compaction, strict failure, and installer behavior are locally tested (108 vendored tests). No new V4 Pro credential was available in this implementation run, so the claimed capability or score improvement is not yet field-verified. A future paired experiment must run fixed tasks at least ten times per preset and record schema hashes, completion rate, score, and variance without retaining reasoning bodies. The mechanism never intercepts private model-request fields or captures/replays hidden reasoning.
Related documents
- Parent: Knowledge index
- Architecture: System architecture
- Testing: Testing strategy
- Local patches: LOCAL-PATCHES.md