Account Health Check

Audit an Orchestra workspace against the conventions the Orchestra team recommends during onboarding and account reviews, then deliver a prioritised, evidence-backed report. This is a read-only review: it inspects pipelines, runs, and assets through the Orchestra MCP and writes a report file — it never mutates pipelines, settings, or runs. If the user wants something fixed, hand off to create-orchestra-pipeline or fix-orchestra-pipeline.

The workspace is determined by the API key behind the connected Orchestra MCP server, so the review covers whatever workspace that key targets. If the user has multiple workspaces, remind them each one needs its own review (there are no cross-workspace resources).

References

• references/best-practices-checklist.md — the rubric. Every check, what "good" looks like, the MCP signal that detects a violation, its severity, and the doc link to cite. Read this before evaluating — it's the heart of the skill. It has a table of contents; you can read the sections relevant to the data you gathered rather than the whole file at once.
• ../../references/orchestra/mcp/tools-quick-ref.md — Orchestra MCP tool names and arguments.

Workflow

Step 0 — Access

Use the Orchestra MCP if connected (list_pipelines, get_pipeline, list_task_runs, etc.). If it isn't, fall back to the public REST API with $ORCHESTRA_API_KEY (often already in the env); for reading git-backed pipeline YAMLs, $GITHUB_TOKEN. If neither MCP nor a key is available, point the user at the Orchestra MCP setup docs (https://docs.getorchestra.io/docs/mcp) or ask for an API key, then stop — recommend connecting the MCP regardless, since it also powers the fix/triage skills.

REST base: https://app.getorchestra.io/api/engine/public — header Authorization: Bearer $ORCHESTRA_API_KEY.

Step 1 — Confirm scope and probe capabilities

Tell the user which workspace you'll review (the one the MCP key targets) and ask if they want the full review or a subset of areas (e.g. "just alerting and cost"). Default to full. For large workspaces (many pipelines), it's fine to offer to sample or focus on production pipelines first — reading every pipeline definition has a cost, so say so rather than silently truncating.

Check which Orchestra MCP tools are actually connected before you start. Tool coverage varies by server: some deployments expose only the list_*/runtime tools, while a single-pipeline read tool (get_pipeline, or similarly named) may be absent. Every definition-level check — alerts, secrets, concurrency, MetaEngine, set_outputs, branching, unique IDs — needs that read tool. If it isn't present, don't jump straight to metadata-only — fall back to the REST metadata API from Step 0 (GET /pipelines/{alias} with $ORCHESTRA_API_KEY) to pull full definitions directly; it doesn't depend on MCP tool coverage at all, only on having an API key. Only if that's also unavailable (no key, no REST access) does the review actually become metadata-only — say so up front, tell the user those checks will come back Not assessed, and note that connecting a server with definition reads (e.g. the remote Orchestra MCP) removes the need for the fallback. Don't discover this halfway through and silently degrade.

Step 2 — Gather data (read-only)

Pull the evidence before judging anything. Batch the list_* calls first, then read definitions.

1. Inventory — list_pipelines. Capture for each pipeline: name, alias, storageProvider (GITHUB/GITLAB/etc. = Git-backed; ORCHESTRA = UI-only), paused, numTasks, schedule, sensors, webhook, and latest-run metadata. This list alone drives several checks (env-duplication by naming, Git-backing, schedule cadence, pause state).
2. Definitions — read each pipeline's full definition with the single-pipeline read tool (by alias or pipeline_id). This is where most checks live: alerts, concurrency, MetaEngine/matrix, task groups, ${{ ENV. }} refs vs hardcoded secrets, branching conditions, inputs, triggers/ sensors, set_outputs. Default scope for large workspaces: the live + Git-backed set — every unpaused, recently-run pipeline plus every Git-backed (storageProvider != ORCHESTRA) one, since that's what a real audit hinges on. Skip dormant TEST:/demo-named pipelines unless asked, and say what you covered vs skipped. If the MCP read tool is absent, use the REST fallback (Step 1) instead — only skip this step and mark the definition-level checks Not assessed if neither the tool nor the REST fallback can reach pipeline definitions.
3. Runtime signals — list_pipeline_runs over the last 7 days (the metadata window). Use this to spot concurrency skips, over- or under-scheduling, and noisy failure rates. list_task_runs /list_operations add integration-level detail (e.g. source-tool jobs, dbt models) when a check needs it.
4. Assets — list_assets. Assets with no Orchestra operation in 7+ days are a governance signal that work may be running outside Orchestra; confirm with list_operations filtered to the asset's integration before flagging.
5. Python compute (opportunistic, for check 7.6) — if Python tasks are a notable share of the workspace, pull recent PYTHON task runs (list_task_runs / GET /task_runs?integration=PYTHON& page_size=200&time_from=<ISO>&time_to=<ISO>) and sum completedAt - startedAt; note count, frequency, and max memory/cpu params. Skip this if Python isn't materially used.

Two data-shape traps, learned the hard way:

• Big list_* outputs. list_pipelines and list_assets can exceed the tool-output token limit on real workspaces and get spilled to a file instead of returned inline. When that happens, don't try to read the whole file into context — jq it. Probe shape first (jq '.result | type, length'), then compute the signals you need with jq/scripts (counts by storageProvider, paused, cron cadence, staleness) and only surface the aggregates. list_assets is paginated (page/page_size/total) — note when you've only seen page 1 and treat it as a sample.
• Nested fields arrive as JSON-encoded strings. In list_pipelines, schedule, sensors, webhook, and triggerEvents come back as strings like "[]" or "{\"enabled\":true}", not parsed arrays/objects. A naive "is it empty?" test counts "[]" as non-empty and reports every pipeline as scheduled/sensored — a false positive. Parse them first (jq 'fromjson') before judging. This is exactly the kind of silent miscount to guard against.

Field names also vary between the YAML form and the stored API form (e.g. max_active vs maxActive, set_outputs vs setOutputs). Match on the concept, not one exact spelling.

Step 3 — Evaluate against the checklist

Work through references/best-practices-checklist.md against the data you gathered. For each violation, record: the check, the specific offending pipeline/task/field (evidence by location — pipeline → task → field/key name, not vague claims), the severity, and the recommendation with its doc link. A clean check is worth noting too — the report's value is partly the reassurance that the basics are covered.

Never reproduce a secret value. For credential findings (e.g. checks 4.3/4.4 — a literal secret in a task param, or a sensitive value in a plain-text env var), cite only the location and the field/env-var name and state that it holds a literal secret. Do not copy the value, or any fragment of it, into your working notes or the report — not even a masked prefix. The location and name are enough to fix it; the value must never enter the context or conversation history. The same applies to any other field whose contents are themselves sensitive.

Be precise and avoid false positives. If you didn't gather the data a check needs (e.g. you sampled pipelines, or a 7-day window hid something), mark that check Not assessed rather than passing or failing it. A confident "I couldn't see this" beats a guess.

Step 4 — Note what MCP can't see

Several conventions live in areas the MCP doesn't expose — RBAC roles and groups, IP restrictions, API-key rotation, the secrets-manager backend, deployment model, and whether source tools (Fivetran, Airbyte, etc.) are actually paused at the source. These can't be auto-checked. List them in the report under Manual verification as a short checklist the user runs in the UI, rather than asserting pass/fail. The checklist file marks these items explicitly.

Step 5 — Score account health

Turn the evaluation into a single Account health score out of 100 plus per-area sub-scores, so the user has a headline number to track over time and a leaderboard of where to improve. Compute it with the Scoring method below — it's a weighted pass-rate over the checks you could actually assess, not a vibe. Report a coverage figure alongside it (how many assessable checks you evaluated) so a metadata-only review is honest about the score being provisional. Never inflate the score by counting [MANUAL] or Not assessed checks as passes — they're excluded, not free points.

Step 6 — Write the report and summarise

Write the report to orchestra-account-review.md (in the working directory unless the user names a path) using the template below, then give a short chat summary: the health score and band, the top 3–5 fixes, and the path to the file.

Write for a customer who wants to act, not read. Keep it tight and scannable — the value is the ranked fix list, not exhaustive prose. Concretely:

• Lead with the score and the Fix first list — that's the part they'll act on.
• One findings table, real findings only, sorted High → Low. Fold evidence into the finding line; don't give it its own column. Never add a row per "Not assessed" check — collapse all of them into the single coverage line at the end.
• Cut anything that doesn't change what the reader does. No methodology dump, no per-area table sprawl, no restating the same point in three sections. Aim for a page or so.

Report structure

Use this template:

# Orchestra account review — <workspace name or "current workspace">

_Reviewed <YYYY-MM-DD>. Read-only audit against Orchestra best practices._

## Health score: <NN>/100 — <band emoji + label, e.g. 🟡 Needs attention>

| Area | | Score | |
|------|--|------:|--|
| Pipeline design & structure | 🟢/🟡/�� | <NN>/100 | <bar> |
| Environments & promotion | 🟢/🟡/🔴 | <NN>/100 | <bar> |
| Version control, Git & CI/CD | 🟢/🟡/🔴 | <NN>/100 | <bar> |
| Connections & credentials | 🟢/🟡/🔴 | <NN>/100 | <bar> |
| Alerting & observability | 🟢/🟡/🔴 | <NN>/100 | <bar> |
| Performance & cost | 🟢/🟡/🔴 | <NN>/100 | <bar> |

_<a> of <b> checks assessed (<c>%). <Add "Provisional — metadata-only; neither the MCP nor the REST
fallback could reach pipeline definitions." if applicable.>_

<Render `<bar>` as a 10-cell meter, e.g. `██████████` filled to the score. Status dot is a quick-glance
read of the same score: 🟢 ≥75, 🟡 60–74, 🔴 <60 — no extra computation, just a color on the number that's
already there. Drop an area row with no assessable checks rather than scoring it 0. Pipelines reviewed:
<n of m>; note any sampling here.>

## Fix first
1. **<headline fix>** — <impact in a few words; the concrete action>. ([docs](<link>))
2. ...
<The 3–5 highest-impact items. This is the report's payload — make each one do-able.>

## Findings
| Severity | Finding (with evidence) | Fix |
|----------|-------------------------|-----|
| High | <what's wrong + the specific pipeline/task/field; concrete value only if non-sensitive> | <what to do> ([docs](<link>)) |
<Real findings only, sorted High → Low. Skip clean and Not-assessed checks here. For secret/
credential findings, name the location and field only — never put a secret value (or fragment) in
this table.>

## Working well
- <2–4 clean checks worth the reassurance>

## Check manually (not visible to the MCP)
- [ ] <RBAC · IP restrictions · secrets backend · API-key rotation · source-tool pausing>

_Not assessed (<n> checks): <areas, e.g. alerting, secrets, concurrency> — needs pipeline
definitions. <Drop this line entirely if coverage was full.>_

Severity guide

• High — security/data-loss risk or a clear reliability hazard: hardcoded secrets or plain-text sensitive env vars, production pipelines not Git-backed, no failure alerting on production, no concurrency limit where overlapping runs would corrupt data.
• Medium — meaningful cost, maintainability, or governance drift: one-pipeline-per-environment, repeated tasks that should be a MetaEngine matrix, over-scheduling instead of sensors, alerts defined in the UI rather than YAML, webhook alert endpoints with no auth.
• Low — polish and best-practice nudges: heartbeat/failure alert-channel split, LIMIT on SQL sensor checks, stale assets to bring under orchestration, set_outputs enabled where unused.

Calibrate to impact and the user's stage — don't bury a High finding under a pile of Lows.

Scoring method

The health score is a severity-weighted pass-rate over the checks you actually assessed — same inputs as the quick checklist, turned into a number. Compute it deterministically so the same workspace always scores the same:

1. Take the assessable checks. Every non-[MANUAL] check in the rubric that you evaluated. Exclude [MANUAL] checks and any you marked Not assessed — they're neither numerator nor denominator. ([MANUAL] checks never count; a check is only assessable if you had the data.)
2. Weight each by severity: High = 5, Medium = 3, Low = 1.
3. Credit each by outcome: pass (✅) = full weight, partial (⚠️) = half weight, fail (❌) = 0. A check fails once if it has any finding of its severity — don't multiply the penalty by the number of offending pipelines (the evidence column already conveys breadth).
4. Score = round( 100 × Σ(credit) ÷ Σ(weight of assessed checks) ).
5. Per-area sub-scores: the same formula scoped to one area's assessed checks. Drop an area with no assessable checks rather than scoring it 0.
6. Coverage = assessed assessable checks ÷ total assessable checks, as a %. If coverage is low (e.g. a metadata-only review where definition checks were all Not assessed), label the score Provisional — the number reflects only what was visible.

Bands for the headline label:

• 90–100 🟢 Excellent · 75–89 🟢 Healthy · 60–74 🟡 Needs attention · 40–59 🟠 At risk · 0–39 🔴 Critical

Keep the score honest: it measures the assessed surface, not the whole account. A 95/100 at 40% coverage means "what I could see looks great, but I couldn't see most of it" — say exactly that.