gastownhall/gascity v0.15.1 on GitHub

Highlights

This is a hotfix release for v0.15.0, shipping the skill-materialization groundwork that v0.15.0 noted as planned follow-on work. If you maintain a city that uses Claude Code agents and wanted pack-defined skills to reach your agents without hand-plumbing symlinks, this release is for you.

Scope is intentionally narrow. This release does not include the session model unification, shared domain layer refactor, or other in-flight main-branch work. It is branched directly from v0.15.0 and only carries targeted fixes.

Upgrade Notes

Safe upgrade from v0.15.0. No config migration required.
Skill materialization is automatic for Claude Code agents. Drop a skill at <pack>/skills/<name>/SKILL.md and it becomes visible to agents in that pack's scope the next time the agent starts or the supervisor ticks.
Agent-local skills are supported: drop them at agents/<name>/skills/<skill>/SKILL.md.
Name collisions between agent-local and shared skills are surfaced by gc doctor.
The recommended post-upgrade flow is:
- gc doctor
- gc doctor --fix (repairs stale implicit-import caches, registers core bootstrap pack)
- gc doctor

Skill Materialization

Added the skill materializer: pack-declared skills at <pack>/skills/<name>/SKILL.md are materialized into each agent's Claude .claude/skills/ sink via symlink at start time and on supervisor tick.
Added two-stage materialization: stage 1 writes into the scope root during supervisor reconciliation (matches the existing agent workdir == scope root case); stage 2 writes into per-session worktrees via a PreStart hook for agents whose sessions live outside the scope root.
Added agent-local skills at agents/<name>/skills/<skill>/SKILL.md, partitioned separately from shared pack skills so one agent's specialization cannot bleed into peers.
Added collision detection in gc doctor: when two agents in the same pack would receive a skill with the same name from different sources, doctor flags it with agent names and source paths.
Added the gc skill list command: shows the effective skill set an agent sees, including bootstrap implicit-import pack skills (gc-work etc.) alongside city-pack skills.
Added a core bootstrap pack that ships operator skills (gc-work and friends) via implicit import, so new cities pick them up automatically.
Added gc internal materialize-skills for operator introspection and scripted use.

Assigned-Skills Prompt Appendix

Added a prompt appendix that tells each agent exactly which skills are visible to them at session start. Skills are partitioned into "assigned to you" (agent-local) and "shared" (pack-level) sections so the agent knows what is theirs vs. shared.
Runtime-gated: the appendix only renders on paths where skill materialization actually delivers (stage-1 at scope root or stage-2-eligible sessions). No false claims of "you have these skills" when materialization would not fire.
Opt-out via inject_assigned_skills = false at the agent level for packs that manage their own skill prompting.

Runtime and Operator Improvements

Fixed SkillsDir and MCPDir population for agents declared in city.toml (previously only convention-discovered agents/<name>/ directories populated these fields).
Fixed a class of skill rendering regressions when SKILL.md frontmatter began with a UTF-8 BOM.
Added a post-compose pass that normalizes agent-local asset directories across both convention-discovered and city.toml-declared agents.

Reliability and Release Quality

Added unit, integration, and acceptance-tier tests spanning the full skill lifecycle (add/edit/delete/rename), CLI surfaces, stage-1 and stage-2 materialization paths, and collision detection.
Cleared 45 pre-existing lint issues that had accumulated on the hotfix branch during development.
Validated end-to-end via multi-scenario manual smoke: city-scoped skills, rig-scoped skills, agent-local skills, worktree materialization, and long-lived named_session mayor drain behavior.