Truestack orchestrate

truestack-orchestrate

One coding request often needs planning, building, an external call, and a quality gate — in a specific order, with the right contracts applied at each step. This skill is the router: it doesn't do the work itself, it routes to the skill that does and makes sure the chain runs in sequence. Front door of the set — classify the request, dispatch, check the handoff, repeat.

Don't duplicate a skill's content here. Orchestrate decides which skill and in what order; each skill owns its own method — the sibling's own description is the source of truth for what it covers.

Step 0 — ground before routing (always)

1. Read project memory — CLAUDE.md is auto-loaded (Principles + Boundaries); read .ai/memory/ on demand. If no memory exists, route to truestack-project-memory to build it before any substantial work.
2. Carry the always-on contracts into whatever runs next: honesty (ground / verify / abstain, truth over agreement), auto-research (consequential current-fact decisions are grounded in authoritative sources before committing, not recalled), clarify-then-proceed on real ambiguity, and the Ask-first / Never boundaries. They travel with every route; the destination skill enforces it.
3. Right-size (below) so heavy process never lands on a one-liner.

Right-size first

• Trivial & unambiguous (rename, copy tweak, one-liner): skip orchestration — the one obvious skill, then truestack-quality-control.
• Small (contained, clear intent): the single relevant skill → truestack-quality-control.
• Substantial (feature / service, several moving parts): run the full chain.
• High-risk (irreversible · data · money · security · schema / migration · first prod deploy · auth/credential or privacy-policy design): full chain plus the truestack-architecture-planning approval gate before any code is written; route risky DDL/backfill through truestack-database-migrations (if the hooks/ PreToolUse gate is wired in this install it also trips there — confirm it's active rather than assuming; if unwired, the approval gate is the only barrier).

Route by intent

Match the request to its skill — one or more. When several apply, run the chain in order.

The request is about…	Route to
First task in a repo · no memory yet · "remember" / "onboard" · stale facts	truestack-project-memory
Designing / scoping / "how should I structure this" / starting something new	truestack-architecture-planning
Implementing backend — server, API, data layer, choosing a stack	truestack-backend-development
Building or restyling React UI — component, page, design, accessibility	truestack-react-frontend
A bug, error, crash, failing test, regression, "not working", slowness that's a defect	truestack-root-cause-debugging
Checking work after a change — "QC", "verify", "make sure it's solid"	truestack-quality-control
Live data or a real external action via a connected MCP server · .mcp.json	truestack-mcp-integration
Schema or data migration — DDL (add/drop/rename column, constraint, index), backfill, expand/contract, "migration locked the table"	truestack-database-migrations (high-risk: via truestack-architecture-planning's gate first)
Packaging & running on a VPS — Docker, nginx, TLS, systemd, SIGTERM drain, blue-green cutover, "my deploy causes downtime"	truestack-deploy-and-runtime
CI/CD pipeline — GitHub Actions, checks gating PRs, SHA-pinning, release/semver, health-gated deploy step	truestack-ci-and-delivery
Making a service observable — structured logs, metrics, OTel tracing, SLOs/burn-rate, "no telemetry yet"	truestack-observability
Securing an app — auth flows, RBAC/access control, OWASP/injection, secrets, threat model, "is this secure"	truestack-application-security (high-risk auth/credential design: via truestack-architecture-planning's gate first)
Designing the API interface contract — REST/GraphQL/gRPC, error model, versioning, pagination, idempotency, OpenAPI	truestack-api-design (high-risk public/versioned contract: via truestack-architecture-planning's gate first)
Dependency lifecycle — add/pin/bump a package, Renovate/Dependabot cooldown, CVE/SBOM/license, "is this package safe"	truestack-dependency-management (high-risk ingestion: via truestack-architecture-planning's gate first)
Privacy/compliance policy — GDPR/CCPA/HIPAA, PII inventory, retention/erasure, consent, audit of personal data, breach readiness	truestack-data-privacy (high-risk: via truestack-architecture-planning's gate first)
Splitting work across parallel agents · research fan-out · resuming another session's ledger	truestack-agent-coordination
"Research / compare / what's best / latest / find out / is it still true"	truestack-deep-research
A shared reference/artifact — "how does this work", "reverse engineer this", "adopt/port this pattern", upgrading from a reference	truestack-reverse-engineering
"Every day / weekly / remind me / run at / recurring"	truestack-task-scheduling
"Score / rate / audit / evaluate / improve" a skill or skill set	truestack-skill-evaluation
"Sharpen / optimize / rephrase this prompt" · "adopt an expert persona" — a raw ask needs an optimized brief	truestack-prompt-optimizer (also runs automatically at every handoff)
"Explain this / what does this do / teach me / in simple terms" — a plain-English walkthrough is the deliverable	truestack-explain-plain

If nothing fits cleanly, say so and ask one short clarifying question — don't force a route.

When a request straddles two skills (deploy vs CI vs migration, security vs QC vs privacy, API vs build vs system design…) or needs a multi-step chain beyond the common ones below, open references/routing.md — it holds the boundary-case splits and the full canonical chain catalog. Don't guess a seam from memory when the reference defines it.

Canonical chains (most common; full catalog in references/routing.md)

• Backend / frontend feature → truestack-architecture-planning → the builder skill for that layer → truestack-quality-control
• Bug fix → truestack-root-cause-debugging → truestack-quality-control
• Research → decision → truestack-deep-research → truestack-architecture-planning
• Migration, deploy, CI/CD, security, API-contract, dependency, privacy, observability, parallel-build and recurring chains → references/routing.md — don't reconstruct a chain from memory when the catalog defines it.
• Always: every chain reads truestack-project-memory first; every handoff is optimized by truestack-prompt-optimizer; truestack-quality-control is the gate before "done"; truestack-explain-plain closes the chain in plain English; and code and project memory must reconcile — the tally balances — before anything is called done.

Route beyond this set — pick the best skill, ours or not

This set is the general, governed spine, not the deepest skill for every domain. When a specialist set installed in this environment covers a domain better, let it do the domain work — discover, don't assume it's installed — and keep truestack's wrapper (the PreToolUse gate, the QC gate, the Always-line contracts) around the result. Honesty over territory: use the better tool — mechanics in references/routing.md.

Single agent or parallel?

Default single-agent — coordination overhead and human review are the real costs, and most work is sequential or tightly coupled. Hand to truestack-agent-coordination only when the plan has genuinely independent tasks, the work is big enough to pay the overhead, and shared contracts (schema / interfaces / API shapes) are defined first. Start at 2–3 agents.

Handoff discipline

• The order is fixed: ground → right-size → route → optimize the handoff (truestack-prompt-optimizer) → dispatch. Optimizing runs after routing, never before it — the persona and required-component judgment key off the destination, and right-sizing has already filtered the trivial work out.
• Route, then let the destination skill own its method — don't re-explain or re-run its steps here.
• Pass forward what it needs: the goal, the acceptance criteria, relevant memory, prior results — optimized first by truestack-prompt-optimizer (matching persona + task-fitted technique + explicit goal + labeled assumptions + checkable criteria; intent unchanged). A handoff that echoes a vague ask verbatim is a defect.
• After each step, check the result against its acceptance criteria before moving to the next. If a step fails, loop within that skill (build loop / verify-or-loop) — don't restart the chain.
• Nothing is "done" until truestack-quality-control passes and every acceptance criterion maps to evidence (the full done-conditions live on the Always line above).

Explain it simply

Open with the route in one plain line — what you'll do and in what order ("Plan the schema, build the endpoint, then QC it"). Get a yes before any high-risk or parallel dispatch. Report each step's result in one line. If you abstained or couldn't route cleanly, say so plainly. Every task closes with truestack-explain-plain — automatically, no request needed: a plain-English explanation of what was done and why, sized to the change (full lesson for substantial work, two plain sentences for a trivial one). It also fires whenever the user asks what/why or sounds confused.

Honest exit

Report against the goal, not the effort. If the chain stalls — a blocked dependency, a conflicting constraint, a claim you can't verify — stop and give an honest completion score (criteria met / total + what blocks each unmet one) rather than declaring a done you can't prove (honesty contract). An accurate 7/10 the user can act on beats a false 10/10.