[anttiviljami]

Summary

Fixes PII that reached anonymized callers (MCP read profile, CLI/SDK with anonymize tokens) through entity-api responses — STABLE360-13580. Also makes anonymizeResponse 2.7–5.6× faster on large responses.

PII fixes

Leak	Cause	Fix
Activity caller.EpilotAuth (email, cognito:username, claims)	caller handling only masked top-level string keys; caller is the request security context, i.e. nested	apply embedded-user rules (user strategy) at every depth; add username / cognito:username / preferred_username
getThreadTimeline actor.email	message-api creates activities via POST /v1/activity without caller → not isActivityLike → payload walked unscrubbed	caller no longer required (payload suffices); plus catch-all: emails/phones/IBANs in strings outside recognized shapes are scrubbed (not inside entities, so data_classification: 'public' fields stay untouched)
Opportunity delivery_address / billing_address	type relation_address not classified	relation_address: 'address'; $relation_ref kept in address items
journey_data names/street	fields not in schema only got pattern scrub	apply anonymizeUnknown field heuristics to non-schema object fields (only when schema is known)
message/thread _title (= subject)	no curated rule	message:_title, thread:_title → redact (also covers thread.latest_message)
cc[].name	MESSAGE_PARTY_KEYS was from/to	add cc, bcc, reply_to
file filename	no rule	file:filename, file:_title → value pseudonym (same input → same pseudonym; shape already recognized by epilot-mcp's mask detector)

City / postal code stay visible by existing policy (ADDRESS_KEEP_KEYS).

Performance

Profiled on 0.2.3: ~30% in mapEntries + GC (entries/fromEntries arrays per node), ~20% in scrubText (4 regex passes on every string), 13% in loadSchemas (an await per node), ~12% re-classifying every attribute of every entity.

• loadSchemas: one synchronous pass collects distinct (org, slug), resolved concurrently once each
• mapEntries as a plain loop; walker / leaf mappers return unchanged subtrees as-is instead of copying (result may share objects with the input — read-only; input still never mutated)
• per-call classification cache per (schema, attribute); Set lookups for key lists
• scrubText prefilter: @ for emails, an exact-superset guard for the IBAN/phone patterns (same lookbehinds) — timestamps, numbers, labels skip all regexes. Fuzz test (20k strings) asserts the guard never skips a match
• anonymizeUnknown memoizes per-key rules

Workload (median of 7)	0.2.3	this PR
searchEntities, 1000 contacts (2.4 MB)	210 ms	77 ms
activity feed, 100 × opportunity snapshot + 5 ops (6 MB)	507 ms	102 ms
entity with 200k-item json attribute (12.7 MB)	862 ms	153 ms

Output is byte-identical to the unoptimized build of this branch on all three workloads.

Release as 0.3.0 (curated defaults changed); entity-api bump: https://gitlab.com/e-pilot/product/entities/entity-api/-/merge_requests/1830 (^0.2.2 won't pick it up).

Test plan

• npm test — 45 passed (stable360-13580.test.ts: ticket cases + public fields untouched; performance.test.ts: prefilter fuzz, schema resolved once, input not mutated)
• tsc --noEmit
• entity-api src/anonymization tests (20) pass against this build; new ticket tests fail on 0.2.2
• After publish: lockfile bump in entity-api, re-run ticket repro via MCP read profile

🤖 Generated with Claude Code

https://claude.ai/code/session_01CRtqyNKiBCCdu3JSwr2zGp