quick-research-campaign
DeepSeek Harness bundle: a bounded, de-confounded hyperparameter sweep ledger with a research-campaign playbook skill and paper-ready report tables
111 results
DeepSeek Harness bundle: a bounded, de-confounded hyperparameter sweep ledger with a research-campaign playbook skill and paper-ready report tables
A generic provider-dispatcher proxy for DeepSeek Harness. Mounts child plugins under a private isolate realm, records every register* call, and fans one operation out to all recorded children with a configurable strategy (parallel merge / race / bail).
Ships a skill that tells the agent when to delegate work to subagents instead of doing it inline, how to fan several out in parallel, and how to check what they returned.
Permission matrix for DeepSeek Harness: 3 sandbox modes x 4 approval strategies = 9 selectable presets, with a global default, a separate LLM-robot default, four-tier risk policies, password-gated approval for high-risk operations, audit log and git checkpoint.
DSH web plugin: per-session API cost monitor. Renders as a dsh-better-sidebar tab, which must be installed separately — this package provides no UI of its own.
动态滚动窗口子代理任务池:一次调用跑完整批独立任务,内部滚动补位,后台运行不阻塞主对话
Offline content-addressed proof for supplied DSH workflow duty separation
A generic provider-dispatcher proxy for DeepSeek Harness. Mounts child plugins under a private isolate realm, records every register* call, and fans one operation out to all recorded children with a configurable strategy (parallel merge / race / bail).
Multi-provider LLM configuration for the DeepSeek Harness LLM seam: declare any number of providers, each with its own endpoint, credential and per-model parameters, across the OpenAI Chat, OpenAI Responses, Anthropic Messages and Google Gemini wire formats.
Parallel side questions and side conversations for DeepSeek Harness without polluting the main context.
Pre-send verification for deliverables: opens the real file to verify row and dedup counts, scans for placeholder residue, diffs incremental files against their stated baseline, and logs the batch in a ledger with candidate vs sendable counts kept separate.
Claim-evidence separation for any document: classifies every checkable claim, traces each to a primary source or a direct measurement, fetches links for real, recounts action-driving numbers, and labels what stays unverified — so a deliverable states its evidence level instead of implying one.
Write changelog entries while the change is in hand: user-facing entries in who/what/how-to-upgrade shape, actions separated from fixes, breaking changes flagged at the top, and a verify pass against the actual diff so the log never claims what the code does not do.
Use the Jev (System One) decision-model paradigm with any OpenAI-compatible LLM — no TypeSafe key required. Registers a jev_decide tool on DeepSeek Harness (dsh).
Save web findings into a durable clip library instead of chat scrollback: each clip records the original URL, fetch date, verbatim excerpt with quotes around any paraphrase, and the project tag — one markdown file per project, greppable, citable after the source page changes or dies.