fix(data-layer): redact loop event/status/decision outside their allowed sets

[diazMelgarejo]

The design promise this closes a gap in

docs/specs/v0.20-agentic-turn.md (this repo's own current spec, currently at master HEAD) states the loop event journal is meant to be content-free by construction: task, prompt, command, and output are excluded from runtime/loops/events.jsonl for exactly this reason, on the theory that two append-only journals under .agent/ with opposite privacy guarantees means the weaker one becomes the leak. .agent/tools/data_layer_export.py already honors that promise for the fields it was built to exclude — test_exports_privacy_safe_loop_events_and_quality_counts asserts a "task" value never makes it into the export.

But the journal's remaining three fields — event, status, decision — were never given the same treatment. They are supposed to be a closed, small vocabulary: harness_manager/loops/runner.py writes exactly 10 distinct event names, 10 status names, and 3 decision names, full stop — there's no code path that emits anything else. normalize_loop_event(), though, treats them as open text: str(entry.get("event") or "loop_event") copies whatever string is there straight into the exported action field, and the status/decision union does the same for result. The write side is a closed enum; the read side trusted it as free-form input. That gap is the one place the content-free-journal guarantee this repo explicitly designed for was left unenforced.

Concretely, this means any row with an event/status/decision value outside the real set — from a future bug in the supervisor, a hand-edited events.jsonl, or a loop kind that doesn't exist yet — flows verbatim into the dashboard/analytics export with zero validation. Nothing catastrophic on its own, but it's exactly the kind of small, structural gap that turns into a real leak the moment something upstream of it goes wrong, which is the whole reason the journal was designed to be content-free in the first place.

The fix

Mirror the actual literals runner.py writes into three module-level sets — VALID_LOOP_EVENTS, VALID_LOOP_STATUSES, VALID_LOOP_DECISIONS — and a small _allowed_or_unknown() helper that redacts anything outside the allowed set to "unknown". That's not a new convention: "unknown" is already the fallback this file uses everywhere else for missing or unrecognized values (pii_level, result, harness, and half a dozen other fields all do the same thing already). This fix just extends the pattern to the two fields that had skipped it.

Tests

Added test_redacts_loop_event_status_and_decision_outside_the_allowed_sets next to the existing loop-event privacy test. Verified in both directions before opening this PR: it fails against the pre-fix code (a <script> tag and a free-text "leaked prompt" string both land in the exported JSONL verbatim), and passes with the fix applied. Full suite: pytest tests/test_data_layer_export.py — 8 passed.

Scope

One file, one function, plus its test. No schema changes, no new dependencies, no behavior change for any well-formed event.

Note

Redact unrecognized loop event, status, and decision values in data layer export

• Adds allowlists (VALID_LOOP_EVENTS, VALID_LOOP_STATUSES, VALID_LOOP_DECISIONS) in data_layer_export.py to enumerate permissible values for loop fields.
• normalize_loop_event now passes event, status, and decision values through a new _allowed_or_unknown helper, replacing unrecognized values with 'unknown' instead of copying arbitrary text into the export.
• Behavioral Change: any loop event/status/decision value not in the allowlists will now appear as 'unknown' in agent-events.jsonl rather than the original string.

^{Macroscope summarized 5a2724e.}

[diazMelgarejo]

Related: #65 (two smaller, unrelated fixes found in the same repo while working on this one).