v0.1.15

Added

• skills package — where every agent discovers Agent Skills, and how to validate one. Agent Skills stopped being a Claude Code feature: the format was released as an open standard in December 2025 and is now stewarded by the Agentic AI Foundation at agentskills.io, with ~45 client products. agentx modeled rules, commands, hooks and sessions but had no concept of skills at all, so a consumer installing skills had to hand-maintain each vendor's directory layout — and quietly install nothing for the agents it had not gotten to.
• skills.RootsFor(agent, scope) returns a Roots{Write, Read} for 12 agents at project and personal scope. Write is the single directory to install into; Read lists every directory that agent discovers from, Write first. Callers must write to Write only. The ecosystem converged on .agents/skills, so a skill installed once is found by nearly every agent; fanning one copy across each vendor's native directory would put the same skill on disk several times, to drift apart and disagree later.
• Two holdouts are modeled explicitly: Claude Code reads only .claude/skills (upstream request) and Kiro only .kiro/skills. Both get their native directory as Write; a test names them, so an agent gaining canonical support upstream surfaces as a failing test rather than as a permanently duplicated install.
• skills.ParseManifest / Manifest.Validate implement the SKILL.md format: name (≤64 chars, lowercase alphanumeric and hyphens, no leading/trailing or consecutive hyphen, must match the skill's directory name), description (≤1024), and the optional license, compatibility (≤500), metadata, and experimental allowed-tools. Validation reports every violation in one pass rather than the first, and unknown frontmatter keys are ignored rather than rejected so a skill valid for a newer agent is not turned away.
• TestEveryRegisteredAgentIsClassified ties the table back to the registry: an agent added to SupportedAgents fails the suite until somebody records where its skills go, or records in deliberatelyUnsupported that it has none. agentx had no such invariant for any capability table, which is how per-agent coverage drifted behind the agent list elsewhere in the library.
• Amp orbs are detected as Amp (#17, #18). An orb's agent shell sets AMP_ORB and AMP_THREAD_ID but not AMP_THREAD_URL, so Amp was never detected there, and a checked-in .codex/ made Codex win instead. The new optional VenueDetector interface marks a host that belongs to an agent; Amp implements it for both variables. Venue signals rank below an agent's runtime signals, so Claude Code run inside an orb is still Claude Code.
• AmpAgent.SessionID rebuilds the thread URL as <AMP_URL>/threads/<AMP_THREAD_ID> (default host https://ampcode.com) when AMP_THREAD_URL is unset. A thread id that isn't a single URL path segment returns an empty session ID.
• Code Puppy declares AgentENVAliases() (code-puppy, codepuppy). Its Detect already accepted codepuppy, which is not its type.

Changed

• Detection order is deterministic. Registry.List() is sorted by type, and Detect walks five phases, strongest evidence first: runtime signals, an explicit AGENT_ENV (type or alias), the own markers of agents without DetectRuntime, venue signals, then everything else, including folder hints such as .codex/. Before this, a tie inside a phase was broken by Go map order and could change from run to run.
• Visible effects: an explicit AGENT_ENV now beats a folder hint every time, where it was random before. When two agents' runtime signals are both present (one agent run inside another), the first by type name wins every time: Claude Code inside an Amp laptop thread reports as Amp, and Codex inside Claude Code reports as Claude Code. Choosing the innermost agent needs process ancestry and is not done yet.

Fixed

• AGENT_ENV aliases count even for agents that map no lifecycle events. The explicit-AGENT_ENV phase reads AgentENVAliases() on its own, so AGENT_ENV=oh-my-pi names OMP even though OMP has no EventPhases.
• Agents without DetectRuntime (Goose, Cursor, Copilot, Cline, …) win on their own marker (GOOSE=1, CURSOR_AGENT=1), inside an orb and over a folder hint such as .codex/. v0.1.14 chose between them at random. A test fails if one of them ever detects from a folder or PATH alone.

Notes

• No SkillsManager, and no Capabilities.Skills flag. Skills are directory bundles, and consumers that install them already own digest tracking, retirement and conflict handling for their whole inventory; a second installer here would be a parallel mechanism drifting from the first. RootsFor returning ok == false is the capability answer, which avoids touching every agent implementation to add a boolean that carries no information the table does not.
• gopkg.in/yaml.v3 is promoted from an indirect to a direct dependency. It was already in the module graph via testify, so no new module is downloaded.