gentle-sdd-continue

Read the SDD workflow FIRST: .claude/skills/_shared/sdd-orchestrator-workflow.md under the workspace when a workspace-scope install wrote it there, otherwise ~/.claude/skills/_shared/sdd-orchestrator-workflow.md. Then treat it as the authoritative SDD workflow instructions for this command. The Claude Code session model is controlled by Claude Code; Gentle AI's always-on Model Assignments policy resolves every Agent tool call, including generic/organic delegation through default.

WORKFLOW:

1. If the gentle-ai binary is available, run gentle-ai sdd-continue [change] --cwd <repo> and treat its dispatcher/status output as authoritative — but only when the session artifact store is openspec or hybrid. When the session artifact store is engram, do NOT invoke the native dispatcher at all — it cannot see the change (it reads only openspec/changes/); resolve status entirely from Engram (mem_search + mem_get_observation on the change's topic keys) using the manual status schema in ~/.claude/skills/_shared/sdd-status-contract.md (the same schema used when the binary is unavailable). The dispatcher is authoritative only for openspec/hybrid. If unavailable, read ~/.claude/skills/_shared/sdd-status-contract.md and produce structured status before acting.
2. Resolve the active change. If $ARGUMENTS is missing and more than one active change exists, ask the user to choose and STOP. Do not guess.
3. Check which artifacts already exist for the active change (proposal, specs, design, tasks)
4. Determine the next phase needed based on the dependency graph: proposal → [specs ∥ design] → tasks → apply → verify → archive
5. Launch the appropriate sub-agent(s) for the next phase only if authoritative status says the dependency is ready. Route only by nextRecommended and dependency states; never infer from free text. If blockedReasons is non-empty, do not proceed to apply, archive, or terminal work. If nextRecommended is verify, verification/remediation may run only to refresh evidence; if nextRecommended is resolve-blockers, report blockedReasons and stop; if nextRecommended is a planning token (propose, spec, design, or tasks), launch the corresponding planning phase. Carry actionContext and allowed edit roots into any sub-agent launch.
6. Present the result and ask the user to proceed

CONTEXT:

• Working directory: Detect agent-side before proceeding by running git rev-parse --show-toplevel with the Bash tool; if that fails, run pwd with the Bash tool.
• Current project: Derive agent-side from the detected working directory basename. Do not use slash-command shell interpolation for this value.
• Change name: $ARGUMENTS
• Execution mode: ask/cache per orchestrator
• Artifact store mode: ask/cache per orchestrator
• Delivery strategy: ask/cache per orchestrator

ENGRAM NOTE: To check which artifacts exist, search: mem_search(query: "sdd/$ARGUMENTS/", project: "{project}") to list all artifacts for this change. Sub-agents handle persistence automatically with topic_key "sdd/$ARGUMENTS/{type}".

Use the lazy workflow instructions to coordinate this workflow. Do NOT execute phase work inline when a native sub-agent is available.

STATUS CONTRACT:

Prefer gentle-ai sdd-continue [change] --cwd <repo> when available — but only when the session artifact store is openspec or hybrid; when the store is engram, do NOT invoke the binary and resolve status from Engram using the manual status schema. Otherwise read ~/.claude/skills/_shared/sdd-status-contract.md and follow it. If status reports workspace-planning with no allowed edit roots, do not launch apply/verify/archive work that would infer repo-local ownership.