dsh-skill-always-apply

@firefly0621/dsh-always-apply

English | 中文

Opt-in Cordis consumer that contributes skill bodies marked alwaysApply: true (Cursor-aligned frontmatter) to the system prompt of every model request, without requiring a skill tool load.

This package does not register a skill provider. Mount it beside dsh-skill / dsh-skill-filesystem (and usually dsh-tool-skill). It is not part of the default dsh-base composition; install with:

dsh plugin --profile web add @firefly0621/dsh-always-apply

Or insert the shipped cordis.patch.yml into a custom base.

The package also runs against upstream @deepseek-ai/dsh-skill releases that predate the typed alwaysApply field. When a catalog summary does not carry the field, this consumer loads the candidate and reads alwaysApply from the definition or the skill file's frontmatter, so no host-code change is required.

Skill frontmatter

In a local SKILL.md (or flat .md skill) parsed by @deepseek-ai/dsh-skill-filesystem:

---
name: my-standing-rules
description: Standing session rules
alwaysApply: true
---

Follow these rules for the whole session.

alwaysApply uses the same boolean spellings as disable-model-invocation / user-invocable. An invalid value is treated as not opted in, so this consumer skips injection while the skill stays in the discovery catalog.

Routing copy (description / whenToUse) stays on the skill provider that registers the skill; this consumer only selects and injects bodies. Mount skill providers that publish clear routing text so discovery and the model catalog stay useful.

The shipped @firefly0621/dsh-skill-karpathy-guidelines provider marks karpathy-guidelines with alwaysApply: true. Mount this consumer next to that provider to inject the Karpathy body without a skill tool load.

Behavior

On every system-prompt/assemble (the assembly that runs before each model step):

1. Skip agentless assemblies and sessions whose session.header.origin === 'subagent' (default; skipSubagent: false overrides).
2. snapshot() the viewing agent's skills. Incomplete observations reuse the last complete always-apply text for that agent (or inject nothing when none exists yet) and do not write the warm cache, so the next assembly retries.
3. Select Config.names plus skills whose summary, loaded definition, or skill-file frontmatter carries alwaysApply: true, minus disabledNames. Candidate loads run concurrently; byte-budget skips still apply in name order.
4. Honor section maxTotalBytes, and contribute one skill:always-apply system-prompt section that lists names and embeds each renderSkillContent block.

Rendering is memoized per agent and invalidated by skills/change, so membership and bodies refresh on the next assembly after a catalog change, and the prompt prefix stays byte-stable between changes for KV reuse.

Config

Always-apply injection is a host standing-instructions path: frontmatter alwaysApply: true and Config.names do not require modelInvocable. A skill with disable-model-invocation: true can still be injected here while staying out of the model-facing skill catalog.

Model Experience

Always-apply standing instructions

What the model sees

A skill:always-apply section at the front of the system prompt when at least one selected skill fits the complete-section budget: a short <system-reminder> naming the always-apply set, then each skill's canonical <skill_content> block. Because the section is part of the system prompt, every step receives it, compaction never shadows it, and a catalog change lands on the next complete assembly (incomplete rediscovery keeps the previous complete text until then).

Token effect

One section whose size is the reminder envelope plus every rendered skill body that fit under maxTotalBytes, re-sent in every request's system prompt.

KV Cache effect

The section text sits in the request prefix. While the always-apply set is unchanged the rendered text is byte-stable, so the warm prefix cache is reused; a skills/change refresh recomposes the section, invalidating reuse from that token forward.

Known Limitations and Deferred Work

• Opt-in only — product defaults do not mount this plugin; operators add it explicitly.
• Bodies must avoid {{...}} prompt-variable syntax — the section is interpolated by the prompt renderer; a body containing a complete {{...}} group is skipped with a warning (an unclosed {{ is kept as literal prose).
• A complete persona replaces every section — an agent whose composition registers a complete persona (agent presets) suppresses all prompt sections, including this one, for that agent.
• Catalog still lists model-invocable always-apply skills — those entries remain in the skill catalog; the reminder tells the model not to re-load them when the body is already present in the system prompt.
• Bypasses model invocation policy — always-apply and Config.names inject regardless of modelInvocable; use disabledNames to exclude.