The dbn-legal-agent-v3 fine-tune (Track 1 / family) emits a labelled-prose
template — duplicate `answer:` prefixes, markdown-escaped underscores (`\_`),
and a trailing raw JSON blob — rather than the strict JSON the Azure/gpt-4o
path produces via response_format. decodeJsonObject() returned null on that
invalid JSON, so ask() dumped the entire raw blob into `answer`.
Fix at the parse layer (no upstream response_format change, to avoid fighting
the fine-tune's training):
- dbnToolsRepairJsonText(): strip fences, drop only invalid `\_`/`\*` escapes,
then balanced-brace scan collecting every top-level {...} (longest first) to
recover an appended JSON object. Shared by both gateways' decodeJsonObject(),
so all JSON tools benefit.
- dbnToolsParseLabeledFields(): parse labelled-prose into real fields when no
JSON decodes, tolerating escaped key names and collapsing duplicate prefixes.
- ask() null-fallback now builds clean structured fields from the parsed prose
instead of dumping raw; what_remains_uncertain becomes a proper list.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- DeepResearchAgent: engine guard now accepts dbn_legal_v3, claude_sonnet, claude_haiku
(previously stripped these to azure_mini, breaking dbn_legal_v3 selection and
preventing Claude engines from reaching the correct synthesis branch)
- DbnBedrockGateway: remove response_format=json_object from chat payload — LiteLLM
converts this to a tool-use constraint for Bedrock, routing output into tool_calls
instead of content (root cause of the {} empty brief)
- advocate.php: add Claude Sonnet 4.6 (AWS Bedrock) engine option
- account, billing, dashboard, nav, min-sak: pending UI/flow changes from prior sessions
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Routes AI tools across three tiers based on task complexity:
- Azure GPT-4o-mini always: redact, translate, timeline-basic, search-legal (mechanical tasks)
- Claude Haiku 4.5 (Bedrock): ask, summarize, timeline-deep, citations (Norwegian nuance)
- Claude Sonnet 4.6 (Bedrock): korrespond, legal-analysis, deep-research, barnevernet-analyze,
discrepancy-find, advocate (public-facing legal output)
No AWS credentials in app — credentials live in LiteLLM on Colin (same as nova-lite).
Rollback: DBN_BEDROCK_ENABLED=false in .env, no code push needed.
Includes extended thinking support for Pro deep-research via chatWithThinking().
Claude Opus 4.7 constant added for future premium tier (needs litellm_config.yaml entry).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
daveadmin