session-workspace

name	session-workspace
version	1.0.0
description	Router for agent session lifecycle: bootstrap, protocol, state/context-window management, compaction preservation, notifications, end-cleanup. Topics: context-window-management, notification-handler, pre-compact-context, session-end-cleanup, session-manager, session-protocol, session-start-bootstrap.
params	{"topic":{"enum":["context-window-management","notification-handler","pre-compact-context","session-end-cleanup","session-manager","session-protocol","session-start-bootstrap"],"required":true,"infer":"from user request; ask only if ambiguous"},"depth":{"enum":["quick","deep"],"default":"quick"}}
routes	{"context-window-management":"references/context-window-management.md","notification-handler":"references/notification-handler.md","pre-compact-context":"references/pre-compact-context.md","session-end-cleanup":"references/session-end-cleanup.md","session-manager":"references/session-manager.md","session-protocol":"references/session-protocol.md","session-start-bootstrap":"references/session-start-bootstrap.md"}

session-workspace

Router skill. Dispatch to EXACTLY ONE playbook; never load the cluster. [DOC]

When to use

Any session-lifecycle moment: starting/resuming a session, persisting or priming state, approaching the context limit, handling a notification, or ending cleanly. If the task is content work (not lifecycle), this is the wrong skill — anti-scope below. [INFERENCE]

Routing (topic → playbook)

• session-start-bootstrap — cold start: load context, prime, announce stage.
• session-protocol — in-session rules of engagement / spine discipline.
• session-manager — read/write .specify/context.json: stages, artifacts. [CONFIG]
• context-window-management — budget tokens; decide trim vs. compact.
• pre-compact-context — snapshot state BEFORE compaction so nothing is lost.
• notification-handler — route/triage an inbound notification mid-session.
• session-end-cleanup — flush state, write handoff, leave a clean tree. Pick by intent, not keyword overlap; ambiguous → ask one question, don't guess. [INFERENCE]

Inputs / Outputs

• In: topic (required); depth (quick=essentials | deep=exhaustive + per-step verification). Out: whatever the routed playbook produces — this file emits no artifacts. [CONFIG]

Spine & gates

Discover → Analyze → Execute → Validate. Enforce constitution v6.0.0, evidence tags (Alfa set, see references/verification-tags.md), and the script-first rule (prefer existing scripts over ad-hoc commands). [DOC]

Acceptance criteria

• Exactly one route Read; no sibling playbooks loaded. [INFERENCE]
• topic resolved (or one clarifying question asked), depth honored.
• Output carries Alfa tags, one family, consistent spelling.
• Gate each dispatch with assets/quality-rubric.json + assets/dispatch-checklist.md. [CONFIG]

Anti-scope

No multi-topic merges, no content authoring, no .specify/context.json writes outside the session-manager route. If unsure which topic, ask — do not fan out across playbooks. [ASSUMPTION]