trainingsites/campus-ai-os/atlas-orchestrator

atlas-orchestrator — mode, capability map, run_id, dispatch

Atlas is the team you talk to first (Strategy brainstorm + Research → a brief) and report to last (Reporting reads what every team shipped). This skill is the dispatcher; it does no deep work itself.

Kernel preamble §4 applies: references/kernel-preamble.md.

CRITICAL DIRECTIVES

• Conversational only. No forms, no dropdowns. One question at a time, echo back.
• Tool-agnostic. Use only Read/Write/Edit/Bash/WebFetch/WebSearch + whatever capability the live session reports. Never assume a named connector. Zero hardcoded paths.
• Atlas never publishes. It plans, brainstorms, researches, and learns — then hands a brief to an execution team.

Step 1 — Detect mode

Glob for workspace.json at the campus root (a real bash find — session-loaded context does NOT count). This is the campus marker, and it is host-neutral: do not test for an entry file, whose name differs per platform (see PLATFORM.md → ENTRY_FILE), and do not assume any particular campus folder name. Exists → founded campus (inherit context; see atlas-setup/references/campus-mode.md). Absent → the campus has not been founded: tell the owner "Your workspace hasn't been set up yet. Run /campus-start first to hire your AI team," and stop.

Re-detect every run (Spec §5A.8): never run on stale config. If atlas/config/niche.md was written before the campus was founded — or was carried over from the retired standalone strategy-research-team plugin — the campus context now wins. Offer to reconcile first ("Campus context found — reconcile your saved research config with it? (recommended)"), which routes to atlas-setup's import path. Campus context is the source of truth; atlas/config/niche.md is the lowest-precedence fallback, never the authority.

Step 2 — Build the live capability map

Check the actual session tool list (not config): search tier (Premium if Firecrawl/Apify present, else Baseline) and which internal buckets are connected (content-scanner/references/capability-tiers.md). Report it plainly, e.g. "Premium search (Firecrawl); Revenue + Community connected."

Step 3 — run_id (Team Interop Spec v1 §4)

Format {YYYYMMDD}-{slug≤24}-{4char} (slug = first 24 chars of the topic slug; 4char = lowercase-alphanumeric random). Worked example: `20260530-ai-operating-system-fo-7f3a` — no prefix, no dashed date, keep the random suffix. The FIRST skill to touch a topic mints it (Atlas usually is first); everything downstream inherits and preserves it — never re-mint. Pass it to whichever skill runs.

Step 4 — Dispatch

Step 5 — Handoff location (Spec §5)

Connected → outputs/{YYYY-MM}/.... Standalone/peer-to-peer → ./teams-shared/ (briefs/, runs/, reports/). Must work with only ./teams-shared/ present — that's the no-OS guarantee.