gsd-build/get-shit-done v1.38.4 on GitHub

Fixed

SDK uses full installed agent/workflow prompts — The SDK was bundling stripped-down copies of agent definitions (~17% of the real content), missing critical instructions like plan file naming conventions, scope reduction rules, and discovery protocols. The SDK now loads the complete installed agents at runtime and resolves @-file references instead of stripping them.
SDK executor receives actual plan content — executeSinglePlan was passing null to the prompt builder instead of the parsed plan file. The executor now loads, parses, and passes the full plan (tasks, objectives, verification criteria) to the prompt.
SDK verification checks VERIFICATION.md, not just session exit code — A verify session that wrote status: gaps_found to VERIFICATION.md was treated as "passed" because the session itself didn't crash. The gap-closure retry loop now reads the actual verification status from disk.
SDK plan ID derivation for bare PLAN.md files — Plans named PLAN.md (instead of 01-01-PLAN.md) produced an empty-string ID, causing downstream execution issues.
SDK headless discuss mode prevents interactive tool calls — The self-discuss step loaded the full interactive workflow prompt, causing the agent to invoke AskUserQuestion and Skill() in headless mode. A mandatory headless override is now prepended to prevent interactive tool usage.

Deleted 13 bundled SDK prompt files (sdk/prompts/agents/, sdk/prompts/workflows/) that were maintained as stripped-down copies and had drifted from the real agents.

Install/upgrade: npx get-shit-done-cc@latest