Automation Sandbox Post Copy Configure

Automation: Sandbox Post-Copy Configure

Apply a Salesforce sandbox post-copy automation JSON config to a target org. The skill pins the Tooling API sobject and the record-lookup SOQL filter for the two canonical ConfigurationName values it has been calibrated against (OutboundMessages, RemoteSiteSettings); for any other ConfigurationName it derives a candidate from the entry value and verifies it against the live org's describe endpoint. Every entry — pinned or derived — must still pass Step B (describe returns 200 with a Metadata compound field) before any PATCH is planned. Entries whose API cannot be identified or verified are surfaced in the summary and skipped — they are never guessed at.

Tool Restrictions

Use ONLY the Bash tool to execute sf CLI commands (sf data query --use-tooling-api, sf api request rest, sf org display). Do NOT use MCP tools like execute_soql — ignore them completely; the compound Metadata PATCH pattern this skill requires is not available through MCP tool wrappers. If the target org alias is not explicitly named by the user, invoke sf commands without --target-org — the harness has already set the CLI's default target-org. Never pass --target-org default — default is not an alias and will fail with NamedOrgNotFoundError.

SOQL-over-REST (sf data query --use-tooling-api ...) is treated as an API call — same OAuth session, same authorization boundary as the subsequent PATCH. No direct database / non-Salesforce SQL access.

STOP — do this before making any API call

Never call the org from memory. Before the first request:

1. Read the config JSON end-to-end from the exact path the user gave (default ./post-copy-config.json). Every entry must have all five keys (ConfigurationName, Label, Fields, IsActive, ExecutionOrder). If any entry is malformed, abort and surface the file path + entry index — do not partially apply. Do not invent entries. If the file is missing, stop and ask; never fabricate a plan against synthetic labels.
2. For each distinct ConfigurationName in the config, run the derivation + describe-verify step (below) before planning any PATCH. An entry whose API cannot be verified must never appear as a planned PATCH — it is rejected up front.
3. Confirm the target org alias with the user unless they supplied one explicitly. This skill mutates a live org — writing to the wrong org (e.g., a production alias set as default) is the most expensive failure mode.

If you announce "I will apply … now" without having read the config file and run the describe-verify for every distinct ConfigurationName, stop and do those first.

Per-entry procedure (the core of this skill)

The skill carries a pinned mapping for the canonical types below, and a derive-then-verify path for anything else. Follow these steps for every entry, in order. Step B (describe-verify) is mandatory regardless of whether the mapping came from the pinned table or the derive path — its HTTP status must appear in the summary.

Step A — Resolve the Tooling API sobject and lookup filter

Pinned canonicals (authoritative — use these exactly, do NOT substitute a different sobject name):

For canonical entries use the pinned sobject and SOQL as-is; skip "derive". Do not invent OutboundMessage, RemoteSiteSetting, or MasterLabel variants — plausible-looking but wrong.

Field-name resolution (overrides + case rule + existence check) is owned by scripts/map-metadata-key.mjs.

Non-canonical: derive a candidate (usually singular of the Metadata API type name, sometimes prefixed); Step B rules out wrong guesses. Use the generic SOQL patterns in Step C.

Step B — Verify the candidate exists and supports Metadata writes

For each candidate, GET the describe:

sf api request rest \
  "/services/data/v<apiVersion>/tooling/sobjects/<Candidate>/describe/"

Accept the candidate only if both are true:

• HTTP 200 (the sobject exists on this org's Tooling API), AND
• The describe response's fields array contains a field named Metadata (compound field — this is what PATCH writes through).

If no candidate passes both gates, mark the entry API_NOT_IDENTIFIED and skip. Do not guess a REST path — the wrong path 404s in the best case and updates the wrong record in the worst case.

Step C — Resolve the record Id (SOQL-over-REST)

For pinned canonicals: run the SOQL from the Step A table verbatim (substituting <Label> / <Fields.Object>). Do not substitute a different filter column such as MasterLabel — the pinned SOQL is the tested-and-correct filter for that sobject.

For non-canonical (derived) sobjects: query the verified sobject for the record identified by the entry's Label (and Fields.Object when present):

# SOQL to a variable + separate jq call — avoids the most common
# shell-quoting failure in this skill.
SOQL='SELECT Id, FullName FROM <VerifiedSobject> WHERE <UniqueFilter>'
sf data query --use-tooling-api --json --query "$SOQL" \
  > /tmp/entry-<Slug>-lookup.json
ID=$(jq -r '.result.records[0].Id // empty' /tmp/entry-<Slug>-lookup.json)

Pick <UniqueFilter> based on the queryable fields shown in the describe response from Step B. If the row contains a FullName field, SOQL usually rejects a direct FullName = ... filter — filter by whichever direct column the describe surfaces (e.g. EntityDefinition.QualifiedApiName, DeveloperName, SiteName) and apply the FullName match client-side (jq -r '.result.records[] | select(.FullName == "<Object>.<Label>") | .Id').

Outcomes:

• Zero rows → NOT_FOUND (never fall back to insert).
• Multiple rows after client-side filtering → AMBIGUOUS (surface all Ids in the summary and skip; a wrong Id is worse than no Id).
• Exactly one row → proceed to Step D.

Step D — GET current Metadata, mutate, PATCH back

Metadata is replace-in-full — omitted keys are blanked on PATCH. Preserve every existing key from the Step D-1 GET; overlay only the mutated keys. The summary's "planned PATCH body" (apply and dry-run) is this full merged object — verbose or org-specific values may render as "<preserved-from-GET>" in dry-run, but every key must be present.

Every Step D filename must include the record <Id> from Step C (e.g. /tmp/entry-<Id>-meta.json) — phase-parallel entries share the working directory and would clobber a shared name.

1. GET current Metadata → /tmp/entry-<Id>-meta.json.
2. Resolve the mutation key — for each Fields.<Xxx>, run node scripts/map-metadata-key.mjs "<ConfigurationName>" "<ConfigFieldName>" /tmp/entry-<Id>-meta.json. On {"status":"OK","key":...} use the returned key. On {"status":"FIELD_MAP_UNKNOWN",...} mark the entry and skip.
3. PATCH — heredoc-build /tmp/entry-<Id>-mutation.json, merge with jq --slurpfile m /tmp/entry-<Id>-mutation.json '. + $m[0] | {Metadata: .}' /tmp/entry-<Id>-meta.json > /tmp/entry-<Id>-patch.json (preserves JSON types), PATCH with -b @/tmp/entry-<Id>-patch.json, capture the response body to /tmp/entry-<Id>-response.json. Full bash in references/api_endpoints.md §Step D. Never use --arg (stringifies booleans/numbers, breaks on special chars). Wrap the PATCH in for attempt in 1 2; do <cmd> && break; done — retry once on shell/jq quoting failure (non-zero exit before the HTTP call goes out).
4. Classify — node scripts/classify-patch-result.mjs "<httpCode>" /tmp/entry-<Id>-response.json. Exit 0 → SUCCESS; exit 2 → FAILED (parsed error on stdout). 204 with a non-empty body is FAILED.

Step E — Verify the change landed

Re-read the record to confirm the PATCH stuck:

• If the mutated field is a direct-queryable SOQL column on the sobject (rare — most Metadata-writable fields are not), a SELECT by Id is enough.
• Otherwise GET the record and inspect the corresponding key inside .Metadata.

If the read-back value doesn't match the requested value, record FAILED_VERIFY — the PATCH returned 204 but the effect is not visible (usually a naming or permission issue).

IsActive semantics

• IsActive: true → apply the PATCH as described in Steps A–E.
• IsActive: false → do not PATCH. Record SKIPPED_INACTIVE for the entry and add a bullet under Follow-ups in the summary file listing the entry's ConfigurationName + Label so the customer notices that a config-declared inactive record was left untouched on the target org.

Rationale: IsActive: false means "not active in this sandbox", not "deactivate the existing record". Silently deactivating a live integration is a much bigger blast radius than leaving it alone.

Canonical output shape (always emit this)

Single Markdown summary written to ./post-copy-<mode>-summary.md (mode is dry-run or apply) AND printed to the user. No JSON side-files (plan/phases.json, requests/*.request.json, etc.) — inline every planned/actual request in the Markdown.

Phase enumeration is script-owned. Run node scripts/plan-phases.mjs <config.json> and consume its phases[] output verbatim. Each entry carries ordinal (1-indexed phase number) and executionOrder (raw value, for the (ExecutionOrder = <raw>) heading annotation). Sparse values collapse (1, 2, 5 → ordinals 1, 2, 3). IsActive:false entries are pre-marked SKIP_INACTIVE. See references/execution_phasing.md for the worked example.

For dry-run entries the HTTP column is — (em-dash). End the summary with: No PATCH requests were issued. To apply, re-run without the dry-run flag.

Target-org resolution is script-owned. Run node scripts/resolve-target-org.mjs; substitute the returned .alias into the header. Never emit <env:SF_TARGET_ORG> or $SF_TARGET_ORG verbatim.

# Post-Copy Configure Run — <N> entries <planned|applied> against `<alias>`

Config file: `<path>`
Target org: `<alias>`
Mode: <dry-run|apply>

## Phase <ExecutionOrder> — <count> entr(y|ies)

Planned request:
- Method: `PATCH`
- Path: `/services/data/v<apiVersion>/tooling/sobjects/<VerifiedSobject>/<Id>`
- Body: `<JSON — the FULL merged Metadata object: every existing key from the Step D-1 GET, overlaid with the mutated keys from the config. Preserved keys with verbose or org-specific values may appear as "<preserved-from-GET>" placeholder strings, but every key must be present. NEVER emit a minimal body containing only the mutated keys.>`

| ConfigurationName | Label | Object | Sobject | Describe | Outcome | HTTP |
|-------------------|-------|--------|---------|----------|---------|------|
| <name>            | <lbl> | <obj>  | <VerifiedSobject> | <200 or 404> | <outcome> | <code> |

## Totals

| Outcome | Count |
|---------|-------|
| <state> | <n>   |

## Follow-ups

- <bullet per SKIPPED_INACTIVE / API_NOT_IDENTIFIED / AMBIGUOUS / FIELD_MAP_UNKNOWN / NOT_FOUND entry>

Column semantics: Object = the entry's Fields.Object if present, — otherwise. Sobject = the Step A resolved sobject. Describe = the Step B describe HTTP status (200 for verified, 404 for API_NOT_IDENTIFIED) — mandatory so skipped Step Bs are visible at a glance. Outcome vocabulary: SUCCESS, NOT_FOUND, AMBIGUOUS, API_NOT_IDENTIFIED, FIELD_MAP_UNKNOWN, FAILED, FAILED_VERIFY, SKIPPED_INACTIVE, SKIPPED, DRY_RUN, DELETE_NOT_SUPPORTED, NOT_ATTEMPTED.

Scripts are internal. The plan-phases.mjs, map-metadata-key.mjs, classify-patch-result.mjs, resolve-target-org.mjs invocations are implementation detail — do not inline their raw stdout, JSON output, or "I ran node …" narration into the summary Markdown or the printed response. Consume the JSON, use the returned values, and render the summary in the exact shape above.

Scope

• In scope: Reading a post-copy config JSON (the shape produced by automation-sandbox-post-copy-config-generate), grouping entries into ExecutionOrder phases, deriving+verifying the Tooling API sobject for each entry, resolving Ids via SOQL-over-REST, PATCHing via compound Metadata, and reporting per-entry outcomes.
• Out of scope: Generating the config JSON from an SOP (delegate to automation-sandbox-post-copy-config-generate); deploying metadata XML; running the Async Task Framework (ATF) orchestrator itself (that is the platform-side Java implementation); inventing API paths for ConfigurationName values whose describe probe fails (surface as API_NOT_IDENTIFIED and stop).

Every API call is against a live org. Treat this skill as a mutation tool: prefer --dry-run first, confirm the target org alias, and never silently retry a failed entry against a different endpoint.

Required Inputs

Gather or infer before applying:

• Config file path: Path to the JSON produced by automation-sandbox-post-copy-config-generate (default: ./post-copy-config.json in the current directory). If the file does not exist, stop and ask.
• Target org alias / username: The sf CLI alias or username of the target sandbox. Never assume the default org — always confirm. If the user has not supplied one, list available orgs with sf org list --json and ask which to use.
• Dry-run flag (optional, default false): Semantics — reads are ALLOWED, writes are FORBIDDEN. Dry-run means the run produces no mutation on the target org; it does NOT mean "no network calls". You MUST still execute every read against the org: the describe probe (Step B, GET), the SOQL Id lookup (Step C, sf data query), and the record fetch (Step D-1, GET) — because the plan's "would-be PATCH body" can only be accurate if the agent has read the current Metadata block from the live record. The ONLY skipped calls are the writes: Step D-3 (PATCH) and Step E (post-PATCH verification). Never fabricate the current Metadata; never emit a body containing invented keys. If a read fails (401, 404, NamedOrgNotFoundError), surface the error and stop — do not fall back to a from-memory plan. Use dry-run on the first pass unless the user has explicitly asked to apply.
• Continue-on-error (optional, default true): When true, a failing entry does not abort the phase — the remaining entries in the phase still execute, and the failure is reported in the summary. When false, a failure aborts execution mid-phase.

If the user supplies a clear config path and target alias, proceed without further questions.

Workflow

Every step executes real sf CLI commands via the Bash tool. Do NOT narrate the plan without actually running the commands — the point of this skill is to mutate the target org, not to describe how.

1. Read and validate the config JSON — load the file with the Read tool. Every entry must be a JSON object with the five top-level keys (ConfigurationName, Label, Fields, IsActive, ExecutionOrder). Malformed entries are a hard stop.

2. Resolve the target org and API version — run sf org display --json. Parse the JSON to confirm the alias resolves and to capture result.apiVersion (e.g. 62.0). Do not print the raw JSON — it contains the access token.

3. Plan phases — run node scripts/plan-phases.mjs <config.json> and iterate its phases[] output. See references/execution_phasing.md for the concurrency cap.

4. Per-entry describe-verify pass — for each distinct ConfigurationName, run Step A + Step B once and cache the verified sobject. Any ConfigurationName failing Step B marks every entry with that name as API_NOT_IDENTIFIED.

5. Per-entry apply pass — for each entry inside each phase:

   • Entries pre-marked SKIP_INACTIVE by plan-phases → record SKIPPED_INACTIVE, add a Follow-ups bullet, continue.
   • Else run Step C (resolve Id), then Step D (GET+mutate+PATCH), then Step E (verify). If --dry-run is true, still run Step C and D-1/D-2 so the printed plan reflects the real merged payload — skip only D-3 (the PATCH) and Step E, then record DRY_RUN.

6. Between phases — wait for every entry in the current phase to complete before starting the next.

7. Write the summary file to disk — the final deliverable is a Markdown file at ./post-copy-<mode>-summary.md, following the shape in the "Canonical output shape" section. Also print it to the user. Never write access tokens or full sf org display output. If a URL contains embedded credentials, mask them in the summary (https://user:***@host/...); the actual PATCH body carries the verbatim URL.

Rules and Gotchas

Load-bearing invariants (never insert; describe-gate every ConfigurationName; compound Metadata PATCH preserves the full existing block; IsActive: false → SKIPPED_INACTIVE; never print raw access tokens; cap intra-phase concurrency at 5) and the canonical response for each runtime failure (401 mid-run, 429, credential-bearing URLs, Fields.Action = "Delete", describe 404, zero/many SOQL rows) live in references/rules_gotchas.md. Read that file before deviating from the Step A–E procedure.

Cross-Skill Integration

Reference File Index