Anchored Standard
Anchor DeepSeek on a Minimal first request, then unlock Standard tools on demand.
The first request exposes only two tools, heavier tools are unlocked on demand, and CI has contents:read only, showing meaningful privilege reduction. Deductions remain because persistent bash, automatic promotion, and eventual Standard capabilities create a broad action surface without per-action user confirmation. The README thoroughly documents context filtering, catalog changes, durable events, API-key environment variables, and external effects, earning full data-flow transparency. It does not show secret validation, redaction, log-leak prevention, or a sensitive-data policy. There are no runtime npm dependencies and compatibility is identified, but Actions use moving major-version tags and the Harness dependency is a breaking-change-prone preview. Installation avoids overwriting an existing target, while complete uninstall, state rollback, and session recovery instructions are absent. Community status, non-affiliation, copyright, and contribution routes are disclosed, but publisher identity remains unverified and maintenance ownership is only partly established.
Modes, phase transitions, durable events, and verification expectations are mostly internally coherent, with package scripts matching CI. A deduction applies because the package description says “full Standard tools” while the README describes a small resident catalog with heavier tools unlocked on demand; the historical 98/99 results also used a different composition. Zero-dependency tests, Node requirements, a specific Harness revision, and a cross-platform shell fallback support availability, but reliance on a developer-preview Harness limits it. Unknown-key mount failures, duplicate-adapter degradation, and shell-startup failure are discussed, although the source does not demonstrate a comprehensive, consistent set of actionable user-facing failure messages.
The README distinguishes base, zero-tool, whoami, prefab, eternal Minimal, wire-think, and combined modes, including their scenarios and costs. Tool boundaries, context-suppression scope, subagent switches, promotion signals, and configuration defaults are described thoroughly. Trigger behavior precisely distinguishes tool-call, assistant-message, either, and compaction boundaries. Windows, Linux/macOS, and NixOS considerations are covered, but development is explicitly centered on Node 24, Windows, and one preview Harness revision, so broad environment fitness is not fully established.
The terminology, mode table, repository layout, configuration reference, installation, and verification sections form a strong information architecture, reinforced by self-contained-mode invariants. Installation includes PowerShell and Unix commands and protects existing targets. Naming is generally controlled and unknown keys fail, but generated copies, numerous similar modes, and the imprecise “full Standard” wording reduce stability. Examples and verification checklists are substantial, though no standalone FAQ appears. Limitations are unusually explicit: active development has stopped, evaluation became unaffordable, sample sizes are small, score provenance changed, and upstream may break compatibility. LICENSE and package.json consistently specify MIT, so repository licensing is complete despite NOASSERTION external metadata. Only version 0.1.0 and handoff records are shown; there is no formal changelog or release policy. The maintainer limits work to feasible maintenance and accepts Issues/PRs, but identity is unknown and the update commitment is constrained.
The repository offers self-contained, copyable presets with concrete configuration and verification guidance, making its outputs useful to the intended Harness audience. The two-phase tool surface and experimental variants offer plausible marginal value over ordinary Minimal or Standard presets. Deductions apply because the README concedes that the ability advantage is unresolved at small n and that 98/99 cannot be attributed to the current generic composition. Default operation adds no model call and variant costs are disclosed, but halted active development, compatibility exposure, and lack of re-benchmarking for the current composition limit the cost-benefit case.
Core mechanism claims are tied to named directories, plugins, configuration keys, event types, commands, and expected request headers. Performance claims carefully distinguish historical compositions, sample sizes, confidence intervals, and unresolved conclusions, supporting excellent traceability. package.json, LICENSE, and CI corroborate versioning, licensing, test invocation, and CI permissions. However, the supplied material omits plugin implementation, test files, execution logs, and the contents of referenced issues, so the central behavior and results cannot be fully cross-corroborated within this review. Observations, historical facts, hypotheses, and unsettled effect sizes are clearly separated.
- The current generic composition has not reproduced the advertised Project2 98/99 scores; those results came from an earlier, materially different tool composition.
- The preset can eventually expose bash and heavier Standard tools, and promotion occurs automatically. Review Harness sandbox, approval, and tool permissions before use with untrusted workspaces or prompts.
- Active development has stopped, and the project depends on a DeepSeek Harness preview that permits breaking changes. Pin and reassess compatibility before upgrading Harness.
- The supplied evidence does not show complete secret-redaction, sensitive-data handling, uninstall, or session-rollback procedures.
- This assessment uses only the supplied static files; no tests were executed, and referenced issues, research repositories, and plugin implementations were not inspected.
What does this agent do, and when should you use it?
Anchored Standard is an experimental collection of DeepSeek Harness agent presets, not a standalone model or general-purpose agent runtime. Its base mode exposes only the official Minimal pair, bash and str_replace_editor, on the first model request while context-gate suppresses automatic runtime injections; after a durable tool/call or assistant/message, it promotes the session to a resident discovery-oriented catalog. The repository also ships zero-tool and Whoami warm-up variants, Eternal Minimal, a wire-level think/execute mode, a combined anchoring package, and a prefab seeded with a prior trajectory. The implementation is delivered through agent.cordis.yml compositions, local MJS plugins, durable session events, and dev_tool_search, and is installed inside a user's local DeepSeek Harness preset directory. Three historical Project2 runs scored 98, 99, and 99, but they used an earlier composition; the generic prefab was not re-benchmarked and independent replications have not settled the ability gain. Active development effectively stopped on August 17, 2026, leaving the project in maintenance-only status.
In the base lifecycle, context-gate intercepts both unified DeepSeek Harness injection paths so request #1 retains the Minimal persona, the claimed user batch, and only explicitly allowed message kinds; tool-bootstrap simultaneously limits the visible catalog to bash and str_replace_editor. Once the session records its first durable tool/call or assistant/message, the plugins derive promotion from session.events, restore Standard context on request #2, and expose the bootstrap pair plus dev_tool_search, skill_search, skill_load, and tools explicitly unlocked by the model. dev_tool_search can grow the resident set without dumping the entire Standard catalog at promotion. zero-anchored-standard and whoami-standard insert a zero-tool anchor turn; eternal-minimal permanently presents two tools and routes commands such as dshx list or dshx <tool> '<json>' to registered Standard tools; wire-think-standard uses the deepseek-wire-think route to send tool_choice: none while definitions remain present, then returns to an execute phase. The prefab mode hydrates a blank session with a bundled successful trajectory. The plugins add no telemetry and make no network requests themselves, although model-provider calls and unlocked capabilities such as web_search can require network access.
- DeepSeek Harness developers testing whether the first-request tool schema changes DeepSeek V4 Pro's reasoning trajectory.
- Standard preset users who want a Minimal initial surface but still need web_search, subagent, workflow, or other heavier tools later.
- Agent researchers comparing two-tool, zero-tool, tool_choice: none, permanent-Minimal, and combined depth-gating conditions.
- Evaluation teams that need promotion state to survive resume or reload and want to audit catalog changes through durable request/header events.
- Advanced Harness users willing to review local shell-capable plugins and accept the risks of an experimental, maintenance-only project.
What are this agent's strengths and limitations?
- The first request uses the byte-aligned Minimal tool pair while context-gate controls both runtime-context and pre-step injection paths, creating a concrete and auditable experimental condition.
- Promotion is derived from durable session events, so the phase survives session resume and reload instead of relying only on a transient counter.
- The promoted catalog stays small and grows through dev_tool_search, avoiding an immediate full Standard tool dump.
- Multiple self-contained modes enable direct comparisons among two-tool, zero-tool, permanent-Minimal, wire-level tool_choice: none, and combined mechanisms.
- The repository includes zero-dependency tests, composition checks, request/header verification guidance, and explicit provenance limitations for its benchmark claims.
- The core implementation depends on DeepSeek Harness preset, event, and plugin interfaces; no direct ChatGPT, Codex, Claude, or generic API deployment is documented.
- Development and testing target DeepSeek Harness 0.1.0-rc.5 and a specific source commit, while Harness is a developer preview that permits breaking changes.
- The reported 98/99/99 runs used an earlier pwsh + read bootstrap and a full 25-tool promotion, so they cannot be attributed to the current generic template; independent ability gains remain unresolved.
- The project is maintenance-only because evaluation became unaffordable, leaving adopters with additional compatibility and validation work.
- Zero-tool and Whoami modes add one model call per session, while wire think/execute adds one call per turn and disrupts prefix-cache continuity.
- The preset carries shell-level trust, and Eternal Minimal's dshx gateway can execute real registered tools, so its code and permissions require review before installation.
How do you install or deploy this agent?
The documented baseline is DeepSeek Harness 0.1.0-rc.5, tested against source commit 47f9438 with Node.js 24. After cloning the repository, install the base mode on Linux or macOS with:
dsh_home="${DSH_HOME:-$HOME/.dsh}"
mkdir -p "$dsh_home/.agent-presets"
test ! -e "$dsh_home/.agent-presets/anchored-standard"
cp -R preset "$dsh_home/.agent-presets/anchored-standard"On PowerShell, use:
$target = Join-Path $env:USERPROFILE '.dsh\.agent-presets\anchored-standard'
if (Test-Path -LiteralPath $target) { throw "Preset already exists: $target" }
New-Item -ItemType Directory -Force -Path (Split-Path -Parent $target) | Out-Null
Copy-Item -Recurse -LiteralPath '.\preset' -Destination $targetEach alternative mode can likewise be copied by itself into a same-named directory under .dsh/.agent-presets. Fully restart DeepSeek Harness, create a blank session, and select Anchored Standard (experimental); do not switch an existing session from another preset. Provider-backed modes need DeepSeek connection configuration. wire-think-standard resolves it from row configuration, the llm-deepseek settings section, or DEEPSEEK_BASE_URL and DEEPSEEK_API_KEY. For prefab, the repository recommends following prefab/AGENT_INSTALL.md's installation contract.
How do you use this agent?
Create a blank session, select the installed preset, and send the real task. Under base Anchored Standard, the first request sees only bash and str_replace_editor; after the first durable tool call or text response, the next request automatically enters the resident phase. Ask the model to use dev_tool_search when a heavier capability is needed, because promotion does not reveal the full Standard catalog. To verify the setup, export session JSONL and inspect request/header events: the first tools array should be exactly ["bash", "str_replace_editor"], the first messages should omit AGENTS.md/CLAUDE.md digests and the skill-catalog reminder, and the next changed header should add dev_tool_search, skill_search, and skill_load. With Eternal Minimal, run dshx list through bash to enumerate gateway tools and commands such as dshx web_search '{"query":"..."}' to invoke a real registered tool. Run npm test for the repository's zero-dependency test suite. bootstrapMaxTokens is unset by default and should be configured explicitly in the tool-bootstrap row only if a first-request cap is desired.
How does this agent compare with similar options?
Compared with the official Minimal preset, base Anchored Standard keeps Minimal's real two-tool schema only for the initial request, then restores context and expands capabilities through discovery tools; remaining permanently on Minimal gives up Standard's broader tool set. Compared with exposing the full Standard catalog immediately after promotion, this repository keeps a small resident set because its experiments found that a full dump pulled the trajectory back toward standard-like behavior. The README also names dsh-routing-suite as a task-aware thinking-mode router and J-Space Cognition Suite V3.6 as a model-agnostic inference-time cognitive-control Skill. The maintainer reports that users found those projects better in some scenarios, but provides no unified head-to-head benchmark.