platform-apex-anonymous-run
Run anonymous Apex against the connected Salesforce org via
, capture the debug log, and narrate compile-time and runtime outcomes back to the developer.
This is the agent-side equivalent of VS Code's Execute Anonymous Apex (document and selection) commands.
This skill is
runtime, not generation — for authoring
/
files use
; for running Apex unit tests use
; for deep debug-log analysis (governor breakdowns, SOQL-in-loop detection) hand off to
.
Tool Restrictions
Use ONLY the Bash tool to execute
, and the
tool to stage snippet temp files. Do NOT use MCP tools for execution.
Anonymous Apex is NOT read-only
Anonymous Apex executes with the running user's permissions and can perform DML, callouts, and platform events. Treat every invocation as a write unless the developer has stated otherwise.
-
Verification-style scripts (preferred for "test this"): wrap the body in a savepoint + rollback so org state is untouched:
apex
Savepoint sp = Database.setSavepoint();
try {
// ... code under test ...
} finally {
Database.rollback(sp);
}
-
Production org heads-up: if the resolved
points at a production org (no scratch/sandbox markers in
), surface a clear warning before running. This is informational only — there is no automated block. Always wait for an explicit "yes, run it" before executing destructive scripts in prod.
-
Never run anonymous Apex you did not generate or have not been shown — if the developer pastes a snippet, echo it back and confirm before executing.
Workflow
Step 1 — Identify the target org
Resolve the active org alias from configuration. If
is set, the
flag may be omitted from the command, but always log which alias was used in the report.
bash
sf config get target-org --json
Throughout this skill,
is the resolved alias or username. If no
is set, ask the developer; do not silently default. If the org is not authenticated, re-authenticate with
or switch orgs with the
skill.
Step 2 — Resolve the input mode
| Mode | When | Action |
|---|
| File mode | Developer points at an existing path ending in (or any path they specify) | Run sf apex run --file <path>
directly |
| Snippet mode | Developer pastes Apex code into the conversation | Write to .sfdx/tmp/anon-<unix-ts>.apex
first, then run sf apex run --file <tmp-path>
|
Why a temp file for snippets, instead of an inline flag? The current
CLI only supports
(and interactive stdin). It does
not expose an
flag. Even where inline code is supported by other tooling, multi-line Apex passed inline runs into shell-escaping pitfalls (single quotes in string literals, backslashes, embedded
). Writing to a temp file is the only reliable path for arbitrary snippets.
Verify CLI flags before deviating:
Supported flags (as of writing):
,
,
,
,
.
Do not invent flags — if the task asks for something not listed, surface that to the developer rather than guessing.
Step 3 — Set up trace flags (for useful logs)
returns a debug log only if a
is active for the running user (or a streaming tail is attached). Recommended path — let the developer tail logs in another terminal:
bash
sf apex tail log --target-org <alias> --color
This auto-creates a short-lived TraceFlag for the running user and streams logs as anonymous Apex executes. Mention this in the report so the developer can copy/paste it.
If no trace flag is set up,
will still execute the code and return compile/runtime status — only the
debug log body will be missing or sparse.
Step 4 — Snippet mode: write the temp file
Only applies when the input is a pasted snippet:
bash
mkdir -p .sfdx/tmp
TS=$(date +%s)
# write the snippet content to .sfdx/tmp/anon-${TS}.apex via the Write tool, NOT via shell heredoc
Use the agent's
tool (not a heredoc) so the snippet is preserved verbatim — heredocs subject the content to additional shell expansion. Echo the resolved temp path to the developer in the report. Do not auto-clean the temp file after execution — leave it under
for inspection. The
directory is conventionally gitignored.
Step 5 — Execute
bash
sf apex run --file <path> --target-org <alias> --json
- Always pass . Human-format output conflates compile vs runtime errors.
- If is already configured, may be omitted, but log the alias used.
- The command exits non-zero on compile errors. Capture both stdout and the parsed JSON.
Step 6 — Parse the JSON response
The
response shape (relevant fields):
json
{
"status": 0,
"result": {
"compiled": true,
"success": true,
"compileProblem": "",
"exceptionMessage": "",
"exceptionStackTrace": "",
"line": -1,
"column": -1,
"logs": "...full debug log text..."
}
}
Decision tree:
| | Meaning | Surface |
|---|
| — | Compile failure | , , , the offending source line |
| | Runtime exception | , , plus log tail |
| | Success | Whatever the script printed via (extracted from ) |
(top-level) means the CLI itself failed (not authenticated, file not found, network). Surface the raw error and stop.
Step 7 — Surface the debug log
- Short logs (< ~200 lines): inline the log body in the report between fenced code blocks.
- Large logs: write the log to and report the path. Include the last 30 lines inline as a tail summary.
- Empty / missing log: likely no active TraceFlag. Surface the Step 3 setup hint and proceed with whatever compile/runtime status was returned.
Highlight these patterns when present in the log:
| Pattern | Why it matters |
|---|
| lines | Governor consumption snapshot — flag SOQL/DML/CPU near-limit |
| Unhandled exception within the anonymous block |
| Unrecoverable error — show the full trailing block |
| count > 1 inside a loop | SOQL-in-loop hint (hand off to ) |
| count high | Unbatched DML hint |
Do not attempt full log parsing here — surface signals only, then hand off to
for deep analysis.
Step 8 — Report
text
Anonymous Apex run: <one-line summary — file or snippet, success or failure>
Org: <alias> (mode: scratch | sandbox | production)
Source: <file path or temp path for snippet>
Compile: success | <error + line:column>
Runtime: success | <exception type + message>
Limits: <CPU=x/10000ms, SOQL=y/100, DML=z/150> (only when log includes LIMIT_USAGE_FOR_NS)
Log: <inline | path .sfdx/tmp/anon-<ts>.log>
Rollback: applied | not applied | n/a
Next: <suggested follow-up>
Examples
Example 1 — File mode
"Run
scripts/seed-test-data.apex
against my default org."
- Resolve from
sf config get target-org --json
.
- Confirm the file exists; if not, stop and surface .
- Run
sf apex run --file scripts/seed-test-data.apex --target-org <alias> --json
.
- Parse JSON. Report compile/runtime status, log tail, and org mode.
- Suggest: "If this seeded real data and you'd like to verify without persisting, re-run with the rollback wrapper (snippet mode)."
Example 2 — Snippet mode (read query)
"Execute
System.debug([SELECT count() FROM Account]);
and tell me the count."
- Resolve .
- Echo the snippet back; confirm.
- Write the snippet to (Write tool).
- Run
sf apex run --file .sfdx/tmp/anon-<ts>.apex --target-org <alias> --json
.
- Parse ; extract the line for the count.
- Report: "Account count = N. Source: (kept for reference)."
Example 3 — Verification with rollback
"Test that this Apex correctly upserts a Contact, then rollback."
-
Resolve
. If prod, surface a heads-up before running.
-
Wrap the developer's snippet:
apex
Savepoint sp = Database.setSavepoint();
try {
// ---- developer snippet begins ----
Contact c = new Contact(LastName = 'Smoke', Email = 'smoke@example.com');
upsert c Email;
System.debug('Upserted: ' + c.Id);
// ---- developer snippet ends ----
} finally {
Database.rollback(sp);
System.debug('Rolled back savepoint.');
}
-
Write to
, execute, parse JSON.
-
Report compile/runtime status, the upserted Id from the log, and
.
Failure Modes
| Symptom | Cause | Recovery |
|---|
No authorization information found for ...
| Org not authenticated, or alias is wrong | Run ; re-auth with or use |
ENOENT: no such file or directory, open '<path>'
| file path is wrong or relative to the wrong cwd | Confirm absolute path; re-run |
| non-empty in JSON | Apex compile error | Surface , , ; show that line; suggest a fix |
| with | Runtime exception inside the anonymous block | Surface exception type + message + stack; show governor counts if present |
| field is empty even on success | No active for running user | Tell developer to run sf apex tail log --target-org <alias>
in another terminal, then re-run |
| with no | CLI / network / auth failure before execution | Surface raw stderr; do not retry blindly |
| Unrecognized flag error | Spec drift with the installed CLI | Re-check ; do not invent flags |
Rules
- Always pass .
- Always resolve from configuration or the developer; never hardcode.
- Never use -style inline flags — they are not supported by the current CLI and are escape-hostile. Always go through .
- Always echo a pasted snippet back to the developer for confirmation before executing.
- For verification-style scripts, default to wrapping in + .
- For prod orgs, surface a heads-up but do not auto-block — the developer is in charge.
- Do not auto-delete temp files under .
- This skill executes anonymous Apex; it does not author, deploy, or test / files. For those, hand off to , the deploy skills, or
platform-apex-test-generate
.
- For deep log analysis, hand off to .