Agent Skill Stack v0.3.0

Agent Skill Stack v0.3.0

This release closes the first-hop recall gap without turning the product into a generic auto-router.

Highlights

• expands the Skill trigger description so shaped end-to-end requests can select Agent Skill Stack even when the user never says “Skill”;
• keeps one bounded action with its domain Skill instead of over-triggering the stack curator;
• adds local-index detection for explicit multi-Skill requests and goals that combine a coordination action with sustained or multi-step scope;
• documents explicit $agent-skill-stack invocation as the portable first-use bootstrap across hosts;
• separates first-hop activation from downstream project-profile routing and the runtime dependency gate;
• adds positive recall trials for a one-week Xiaohongshu campaign and explicit Skill-set selection;
• adds a negative trial proving that “write one Xiaohongshu post” does not recall the stack curator;
• verifies that the first-hop activation plan keeps Agent Skill Stack primary while domain and writing-quality Skills remain supporting;
• removes “any goal” claims that an unaware host cannot guarantee.

Why this iteration exists

v0.2.0 could prove that every primary and helper Skill in a selected project route was locally readable before execution. It could not guarantee that the host would invoke Agent Skill Stack before that route existed. v0.3.0 treats these as two different layers:

1. bootstrap Agent Skill Stack explicitly, or let a capable host match a shaped end-to-end goal;
2. derive the workflow and select the smallest verified stack;
3. create project routes for later primary/helper handoffs;
4. use the runtime gate to block incomplete downstream dependencies.

The implementation intentionally avoids a trigger such as “any natural-language request.” A single bounded action should remain with its relevant domain Skill.

Install

npx skills add neilchen2000-pixel/agent-skill-stack --skill agent-skill-stack -g -y

GitHub Copilot / GitHub CLI 2.90.0+:

gh skill install neilchen2000-pixel/agent-skill-stack agent-skill-stack@v0.3.0

First prompt

Use $agent-skill-stack to help me promote this open-source project on Xiaohongshu for one week.

Verification

• Agent Skills structure validation: passed
• Python compilation: passed
• Safe trials: 12/12 passed
• First-hop positive trials: sustained end-to-end goal and explicit multi-Skill request
• First-hop negative trial: one bounded domain action stays out of the stack curator
• Runtime trial: ordered primary-to-helper activation plan resolves completely
• Secret and private-path scan: no matches
• Symlinks: none
• Bundled runtime dependencies: Python standard library only

Read the complete public security audit before installation.

Agent Skill Stack v0.2.0

Agent Skill Stack v0.2.0

This release closes the gap between finding a Skill and actually loading it when the downstream task runs.

Highlights

• adds a deterministic runtime activation gate for project Skill Stack routes;
• resolves the primary Skill, supporting Skills, and named downstream dependencies into one ordered activation plan;
• loads an indexed helper from its local path even when the host did not advertise it in the current Skill list;
• blocks the handoff when a dependency is missing, duplicated, structurally invalid, or unreadable;
• prevents an incomplete Skill handoff from reaching an external write;
• makes project profiles declare their activation and missing-dependency policy;
• upgrades recall checks from search ranking to complete primary-to-helper handoff verification;
• stores no prompt history, routing feedback, usage logs, or credentials.

Why this iteration exists

An index can prove that humanizer is installed and searchable without proving that its instructions were loaded for a real writing task. A workflow can also mention a helper such as dbs-ai-check even when no such installed Skill exists. v0.2.0 treats both conditions as runtime routing problems:

1. the selected route becomes an explicit activation plan;
2. every item must resolve to one readable local SKILL.md;
3. every instruction file must be loaded before its handoff;
4. a missing or ambiguous dependency blocks downstream execution instead of being silently skipped.

Install

npx skills add neilchen2000-pixel/agent-skill-stack --skill agent-skill-stack -g -y

GitHub Copilot / GitHub CLI 2.90.0+:

gh skill install neilchen2000-pixel/agent-skill-stack agent-skill-stack@v0.2.0

Verification

• Agent Skills structure validation: passed
• Python compilation: passed
• Safe trials: 8/8 passed
• Positive runtime trial: an indexed but non-advertised Humanizer is included in the activation plan
• Negative runtime trial: an unavailable named helper blocks the handoff and is reported by name
• Duplicate-install trial: ambiguous same-name Skills block activation
• Secret and private-path scan: no matches
• Symlinks: none
• Bundled runtime dependencies: Python standard library only

Read the complete public security audit before installation.

Agent Skill Stack v0.1.0

Agent Skill Stack v0.1.0

The first public release turns an end goal into a minimal, audited, project-specific stack of AI Agent Skills.

Highlights

• dynamic workflow derivation instead of a fixed domain template;
• local Skill indexing and project-first routing profiles;
• direct, operational, supporting-outcome, and connector discovery lenses;
• 25% community-evidence weight after hard fit and safety gates;
• conflict checks and consent-controlled staged installation;
• shareable SVG Skill Stack recommendation cards;
• no prompt-history, routing-feedback, or usage collection;
• English and Simplified Chinese documentation.

Install

npx skills add neilchen2000-pixel/agent-skill-stack --skill agent-skill-stack -g -y

GitHub Copilot / GitHub CLI 2.90.0+:

gh skill install neilchen2000-pixel/agent-skill-stack agent-skill-stack@v0.1.0

Real examples included in this release

1. Crypto research and backtesting: combines read-only market data, backtesting discipline, risk management, and position sizing while keeping exchange credentials and live orders out.
2. Cited research decision brief: separates source extraction, spreadsheet validation, document drafting, and citation verification while preserving evidence provenance.
3. Support knowledge-base improvement: adds privacy redaction as a hard gate, drafts reviewed help-center updates, and keeps every external publish behind human approval.

Each example includes its derived capability flow, candidate decisions, safety boundary, and a shareable recommendation card.

Verification

• Agent Skills structure validation: passed
• Python compilation: passed
• Safe trials: 5/5 passed
• Secret and private-path scan: no matches
• Symlinks: none
• Bundled runtime dependencies: Python standard library only

Read the complete public security audit before installation.