feat(orchestrator): add skills field to agent YAML template and CreateAgentRequest API

[geoffjay]

Adds the skills field throughout the agent lifecycle so operators can declare which skills each agent receives.

Changes:

• AgentConfig and CreateAgentRequest gain a skills: Vec<String> field (serde default empty)
• New DB migration m20260513_000017 adds skills TEXT NOT NULL DEFAULT '[]' column
• AgentTemplate in apply.rs gains a SkillsField enum supporting skills: all (expand to all discovered skills) or skills: [name1, name2] (validated against discovered skills)
• schemas/agent.yml updated with oneOf skills property
• main.rs binary gains mod skills so crate::skills::discover_all_skills resolves in API handler

Closes #1210

[geoffjay]

Review: feat(orchestrator): add skills field to agent YAML template and CreateAgentRequest API

Stack Position

Branch issue-1210 is stacked on epic-1209, not main. The parent branch epic-1209 contains the issue-1209 (Skill model) work plus a fix(skills): address review feedback commit.

Parent status: PR #1226 has been re-submitted for review (review-agent label is set, needs-rework has been removed). I verified the two blocking issues from that review are both fixed on epic-1209:

• discover_all_skills() now has a global sort ✓
• Per-entry read_dir errors now use let Ok(entry) = entry else { continue } ✓

This PR is approvable in isolation but cannot merge until #1226 is re-reviewed and cleared. The conductor should treat it as blocked until the parent chain is clean.

---

SkillsField Enum: Correct

#[derive(Debug, Clone, Deserialize)]
#[serde(untagged)]
pub enum SkillsField {
    All(String),      // matches any string — validated at apply time
    Named(Vec<String>), // matches any JSON array
}

#[serde(untagged)] is the right choice here — serde tries All(String) first for string values and Named(Vec<String>) for array values, which correctly handles skills: all vs skills: [a, b] vs skills: []. The Default impl returning Named(vec![]) matches the schema's default: []. ✓

---

Skills Resolution in apply_agent: Correct

All four cases are handled cleanly:

Input	Outcome
skills: all	Expands to all discovered skill names
skills: "typo"	Fails with actionable error: use 'all' or a list
skills: [] / omitted	Assigns no skills (fast path, no discovery call)
skills: [a, b]	Validates each name against discovered skills; fails with missing names and a hint to run agent skill list

The guard SkillsField::Named(names) if names.is_empty() => vec![] correctly avoids calling discover_all_skills() when no skills are needed. ✓

---

Migration: Idempotent and Consistent

let stmt = "ALTER TABLE agents ADD COLUMN skills TEXT NOT NULL DEFAULT '[]'";
if let Err(e) = db.execute_unprepared(stmt).await {
    if !e.to_string().contains("duplicate column name") {
        return Err(e);
    }
}

Follows the same idempotency pattern as other migrations in this codebase (m20260417_000016 etc.). The SQLite "duplicate column name" string check is the established convention here. ✓

The empty down() with an explanatory comment about SQLite < 3.35.0 is the right approach — consistent with other migrations that can't safely DROP COLUMN on older SQLite.

---

Storage / Entity Integration: Consistent

skills is stored as JSON text ("[]" default) in both the entity model and the AgentConfig, following the exact same pattern as rooms. Serialization uses unwrap_or_else(|_| "[]".to_string()) and deserialization uses unwrap_or_default() — matching every other JSON-in-text column in storage.rs. ✓

---

Schema: Correct oneOf

oneOf:
  - type: array
    items: {type: string}
  - type: string
    enum: [all]
default: []

"all" matches only the string variant. ["a", "b"] and [] match only the array variant. No overlap, no ambiguity. ✓

---

Non-Blocking: No Tests for SkillsField

There are no unit tests for SkillsField YAML/JSON deserialization or for the apply_agent skills resolution logic. The #[serde(untagged)] behavior (particularly the fallthrough for unexpected strings reaching SkillsField::All and being caught at apply time) would be valuable to cover. These are tracked in #1214 (the capstone test issue for the epic), so the omission is intentional — just noting it for when #1214 is dispatched.

---

Non-Blocking: discover_all_skills() Called Per-Agent at Apply Time

When skills: all or skills: [named] is used, discover_all_skills() runs once per agent in apply_agent. If an agent YAML defines 20 agents, this means up to 20 filesystem scans. For the typical small .agentd/skills/ directory this is negligible, but it's worth noting as a future optimization candidate (scan once, pass results to all agents).

---

Summary

The code is correct and well-integrated throughout the call chain (apply.rs → types.rs → api.rs → storage.rs → migration). The parent epic-1209 branch has confirmed fixes for both blocking issues from the #1226 review. Two non-blocking notes: tests deferred to #1214 (expected), and per-agent skill discovery at apply time is a minor inefficiency.

Holding merge-queue pending re-review and approval of #1226. Once that clears, this can follow immediately.