fix(astra): stop dropping and collapsing habits when listing

[thomasluizon]

The bug

Asked "quais meus habitos de hoje" (what are my habits today), Astra dropped scattered habits and collapsed identical siblings — e.g. a user with six identical Water - 710ml sub-habits saw Water listed once, and mid-list items (Dinner, an overdue posthog, etc.) vanished entirely.

Root cause

The prompt has gpt-4.1-mini free-text reproduce the full nested habit index. Two well-documented small-LLM failure modes bite:

• Duplicate-collapse — identical adjacent lines collapse under induction (the six Water - 710ml lines are byte-identical except for a GUID the model is told never to emit).
• Lost-in-the-middle — items in the middle of a long list get dropped.

The index data itself is complete and uncapped; this is an enumeration/relay failure, not missing data or a token cap (MaxOutputTokenCount = 8192, the answer ended cleanly).

What this PR does (Stage 1 — backend hardening)

Server-side only, so it improves every client on the next deploy with no app update:

1. SiblingTitleDisambiguator numbers exact-duplicate-title siblings (n of N) in both the prompt index (ActiveHabitsSection) and query_habits output. The six waters become Water - 710ml (1 of 6) … (6 of 6) — distinct, countable lines the model cannot silently merge.
2. One index instruction line: repeated habits are marked (n of N); list every numbered instance.
3. finish_reason == length guard in both AI round paths (buffered + streaming): a truncated response is logged (Warning, EventId 8), never silently rendered as a complete list.

Scope / what's deferred

This kills the duplicate-collapse half fully and hardens against silent truncation. The lost-in-the-middle half is only fully solved by having the app render the habit list so the model stops relaying it as free text — that's the next stage (cross-platform, needs the web + mobile chat surfaces). Deliberately not routing listing through a forced query_habits call here: it would add a round-trip per "what's today" without fixing enumeration (the model still relays), and that routing only pays off once the app renders the payload.

Compatibility

Purely additive prompt/text formatting — no DTO or shared-contract change, no client change required. Safe to deploy any time.

Tests

SiblingTitleDisambiguator unit tests, index + tool duplicate-numbering tests, and buffered/streaming truncation-guard tests. Full suite green: 3564 passed, 0 failed.

🤖 Generated with Claude Code

[claude]

Solid fix. The SiblingTitleDisambiguator correctly scopes numbering per-sibling-group (applied at every recursion level in both ActiveHabitsSection and QueryHabitsTool), the ordinal case-sensitive matching is appropriate and tested, and the early-exit on no duplicates keeps the hot path clean. The finish_reason == length guard is a sensible observability improvement — logs at Warning with a stable EventId rather than silently swallowing truncated output. Test coverage is thorough: unit tests for the disambiguator (empty, no-dups, case-sensitive, multi-group), integration-level section/tool tests verifying the numbered suffix in rendered output, and buffered + streaming truncation tests with a recording logger. No DTO or client contract change, safe to deploy immediately.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
100.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud