feat: mid-conversation age detection + crisis under-16 phrasebank fix

[camlambert2000]

Summary

Mid-conversation under-16 age detection, crisis under-16 phrasebank fix, LA-aware safeguarding routing, and location gate reordering.

Changes (7 logical steps, 6 commits)

1. fix(logs): Remove raw user input from gate debug log at route.ts:207 — only gate name is now logged.

2. refactor(safeguarding): Unexport logUnder16Trigger — module-private, only callable from interceptUnder16Age.

3. refactor(safeguarding): Move under-16 intercept explanation strings to phrasebank (UNDER16_INTERCEPT_PREFIX + supporter variant).

4. fix(safeguarding): Remove "Prefer not to say" from CRISIS_UNDER16_LOCATION — location is routing-critical in a child safeguarding exit. 8 options remain (7 LAs + Somewhere else).

5. fix(safeguarding): LA-aware routing on under-16 intercept and all buildUnder16Exit call sites:

   • If LA known → exit directly with localised Children's Services
   • If LA unknown → route to CRISIS_UNDER16_LOCATION to ask for area first
   • buildUnder16Exit now throws if called without valid LA (makes miscalls visible in testing)
   • New CRISIS_UNDER16_SOMEWHERE_ELSE phrasebank exit surfaces Childline for out-of-area users
   • Updated sectionC.ts and stateMachine.ts B3_AGE_CATEGORY with same LA guard

6. fix(safeguarding): Move interceptUnder16Age call before Claude classifiers (checkScope, detectAdviceQuestion) in route.ts — rule #25 compliance. Safeguarding intercept now fires immediately after __LOCATION_RESULT__, before any async API call.

7. feat(flow): Move location gate immediately after GATE0_CRISIS_DANGER:

   • GATE0 option 7 ("none") → LOCATION_CONSENT (was GATE1_INTENT)
   • Location completion → GATE1_INTENT (was PREFERRED_NAME_ASK)
   • GATE2_ROUTE_SELECTION → PREFERRED_NAME_ASK (was LOCATION_CONSENT)
   • GATE1_INTENT choice 3 ("specific org") → PREFERRED_NAME_ASK (was B1_LOCAL_AUTHORITY)
   • LA is now known before any profiling begins, restoring WatsonX design

Test results

96/96 passing (73 original + 23 new).

For James

Please review the regex patterns in detectUnder16Age (stateMachine.ts) — specifically NUMERIC_AGE_RE negative lookahead and SCHOOL_YEAR_RE boundary conditions. Governed by Mid-Conversation Age Detection Design Note v1.1.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
street-support-va-staging	Ready	Preview, Comment	Apr 15, 2026 6:17am
streetsupport-platform-va	Ready	Preview, Comment	Apr 15, 2026 6:17am

[james-cross]

Review — mid-conversation under-16 detection

Thanks Matt — the intent here is exactly right and the regex work is careful. There's one ordering bug that I think needs fixing before this can merge, plus a couple of governance-adjacent things worth a look.

---

Critical (blocks merge)

1. Intercept runs AFTER the Claude scope/advice classifier at early gates — app/api/chat/route.ts:243-289

At GATE0_CRISIS_DANGER, GATE1_INTENT, B4_ADVICE_TOPIC_SELECTION, and ADVICE_BRIDGE, an input like "I'm 14" (length 6, not all-digits) hits checkScope() and detectAdviceQuestion() — both Claude calls — before interceptUnder16Age runs. If Claude classifies it as out_of_scope it returns OUT_OF_SCOPE_GENERAL and the function exits. If detectAdviceQuestion returns a phrase key it returns advice content and the function exits. Either way the under-16 disclosure is swallowed.

This is the precise scenario compliance rule #25 ("AI must not override system rules or safeguarding routes") exists to prevent. The under-16 exit is a hard-coded safeguarding route and must take precedence over any AI-mediated branching.

Fix: move the interceptUnder16Age block to before the earlyGates check (right after the __LOCATION_RESULT__ handler, ~line 237). Add a route.ts test that asserts: at GATE0, posting \"I'm 14\" returns the under-16 exit regardless of how checkScope would classify it.

---

Important (should fix)

2. Hard-coded safeguarding-adjacent text in stateMachine.ts, not the phrasebank — lib/stateMachine.ts:962-967

The two `explanation` strings (`"From what you just said, it sounds like you may be under 16…"` / supporter variant) live inline in interceptUnder16Age rather than in phrasebank.ts. This is the only place in the codebase where a safeguarding exit text is assembled from a TS template literal instead of a phrasebank key. Rule #6 is about runtime AI generation, so this isn't a strict violation, but it breaks the convention Matt has otherwise been strict about — that all safeguarding-adjacent language is centralised so it's reviewable in one place.

Fix: add UNDER16_INTERCEPT_PREFIX and UNDER16_INTERCEPT_PREFIX__SUPPORTER to phrasebank.ts and pull them via getPhrase. Same shape as the new CRISIS_UNDER16_LOCATION entries.

3. Pre-existing PII bleed undermines the audit-clean claim — app/api/chat/route.ts:207

The new logUnder16Trigger is correctly PII-free, and the test verifies that. But route.ts:207 already logs Input: \"${message.substring(0, 50)}...\" for every request before the intercept runs, so in production the raw \"I'm 14 and …\" text still lands in the logs from the gate-entry log line. The claim in the PR description ("Audit log captures type, code, sessionId, timestamp only — no raw user text") is true of the new trigger logger but isn't true end-to-end.

This isn't introduced by this PR, but the PR is the right moment to fix it: change line 207 to log the gate only, or redact when an under-16 trigger fires. Otherwise the design note's privacy guarantee doesn't hold in practice.

---

Questions for resolution

4. "Prefer not to say" on CRISIS_UNDER16_LOCATION — lib/phrasebank.ts:1437-1465

Compliance rule #18 says local authority is a mandatory core field with no "prefer not to say". This is in a safeguarding exit pathway, not normal profiling, and there's a defensible safety argument for letting a frightened child decline — buildUnder16Exit already gracefully falls back to generic Children's Services. Worth confirming: is this pathway exempt from rule #18, or should the option be "Somewhere else" only? The handler at lib/handlers/crisis.ts:239-249 already treats option 9 the same as 8, so removing it would be no-op behaviourally and bring it in line with #18.

(Side note: this also matters for change control. Adding/removing options on a safeguarding gate is arguably Tier 3 — has Catherine signed off on the 9-option shape, or did the previous broken state mean nobody had to decide?)

5. Year 11 cutoff — lib/stateMachine.ts:917

The comment says "Year 11+ deliberately excluded because Year 11 students may already be 16." Agreed for autumn term, but Year 11s are 15 from September to August birthdays — many remain 15 well into the school year. The current rule is "better to false-negative than false-positive on age", which is reasonable, but worth a sentence in the design note about why we accept missing this cohort. No code change needed unless the design position changes.

---

Suggestions

6. logUnder16Trigger doesn't need to be exported — lib/stateMachine.ts:931. Only interceptUnder16Age calls it. Drop the export.

7. Regex reads cleanly but the negative-lookahead list will rot. A short comment pointing at the design note version and a single example of the false positive each token guards against would help future maintainers. Not blocking.

8. The numeric regex captures \"I am 14\" via the a?m shortcut — works, but (?:i'?m|i am|im) is more obviously correct to a future reader than i[\u2018\u2019']?\\s*a?m. Equivalent behaviour, easier to grep.

---

What's good

• Test coverage is genuinely useful: numeric range edge cases on both sides, school year boundary conditions, ambiguous-phrase negatives, contextual qualifier ('asking for my mum'), mid-conversation override of a prior B3 answer, and the audit-log PII test using a sentinel string. This is the right shape for safeguarding tests.
• Negative-lookahead exclusions for units of measurement are well-chosen — \"14 stone\", \"10 minutes late\", \"13 years older\" are exactly the false positives that would have made this brittle.
• Routing to the supporter variant when session.isSupporter is set is correct, and reusing buildUnder16Exit rather than reimplementing the exit text is the right call.
• Phrasebank fix for CRISIS_UNDER16_LOCATION resolves a real broken path — crisis.ts:218 was producing [Missing phrase: …] before this.

---

Summary

• Overall: Needs minor fixes — the intercept ordering must change before merge.
• Governance: Has one violation (rule #25, intercept ordering). Question outstanding on rule #18 (prefer-not-to-say on LA).
• Priority order:
  1. Move intercept above the early-gates Claude classifier (#1)
  2. Move explanation strings into phrasebank (#2)
  3. Resolve the rule #18 / 9-option question with Catherine (#4)
  4. Fix the line 207 raw-input log to honour the audit-clean intent (#3)
  5. Minor cleanups (#5–8)

[james-cross]

Really good update to the original PR. There's a few inline comments, mainly around defensive error handling.

[james-cross]

Worth considering the failure mode here. In dev/test, the throw is great because it makes miscalls loud. In production, if any future caller ever misses the LA guard, a child gets a 500 error instead of the exit they need.

The safer pattern would be to log the error, then self-recover by routing to CRISIS_UNDER16_LOCATION. That way the failure mode is "extra question" not "error page".

[james-cross]

Something in here about how to find their location council would be worthwhile.

[james-cross]

This routes to CRISIS_UNDER17_LOCATION but doesn't set safeguardingTriggered: true on the state. interceptUnder16Age does set it when routing. The session will end up flagged correctly once handleCrisisUnder16Location runs on exit, but there's a window where it's in the safeguarding gate without the flag, which would under-count if anything analytics-adjacent checks between steps.