Spaces:
Sleeping
Sleeping
Commit History
19 days ago
-
Don't auto-route next turn to a knock guest; refresh architecture doc
Knock interjections are now saved with knock_guest=True, and run_chat_app's proactive character shift skips them, since the client's "who do you want to reply to?" choice already decides the next speaker. Architecture doc updated to 2026-09-29: Fly.io hosting, removal of the user-facing model-quality toggle, and a new monetization section. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>8746479
-
Rewrite technical architecture doc to match current app (Cielo)
Replaces the stale April/May .docx (pre-Cielo, pre-dialogue-pipelines, wrong model/table names) with content reflecting the current system: five dialogue pipelines, Cielo coordinator, personalization/axis profile loop, content systems, model routing, and known limitations. Adds a markdown version alongside the regenerated .docx, plus a rewritten generate_tech_doc.py generator to keep them reproducible. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>6c18117
-
Make Medium the default deep tier; pin planning to Sonnet
Acts on the Haiku-vs-Sonnet benchmark: Medium (Haiku) becomes the default for dialogue_deep_tier everywhere (ContextVar default, DB column default, both clients' initial UI state) since it matched Sonnet on 5 of 6 tested calls at a fraction of the cost. Adds get_planning_model() as the one carve-out: pragmatic's build_reasoning_plan gets Sonnet even on Medium tier, since that was the one call where Sonnet showed a real quality edge. Light tier still gets mini there — that tier's promise is minimum cost, not something to silently override. This call fires once per dialogue, not every turn, so the cost of pinning it up is small.71534bb -
Add uniform Light/Deep model tiers per dialogue + Deep/Medium/Light switch
Replaces the global flat/smart llm_mode toggle for the four structured dialogues (maieutic/combative/pragmatic/declarative) and Cielo, which defaulted every request to Fast and silently ran their reply-generation calls on mini instead of the benchmarked-necessary Sonnet tier. Each dialogue's stage-generation calls now always use the deep tier (being inside the dialogue is itself the "this deserves depth" signal, unlike standard chat which needs reply_routing.py's per-turn heuristic); gate/ classifier calls use fixed Light/Lighter tiers. The deep tier is a single user-facing switch, not a hidden toggle: Deep (Sonnet, default) / Medium (Haiku) / Light (mini), threaded per-request in app.py, persisted via a new user_profiles.dialogue_deep_tier column, with matching Settings controls in the web client and mobile app. Requires running migrations/add_dialogue_deep_tier.sql once in Supabase before the persistence layer works. Full inventory and rationale in docs/model_routing_design_2026_07…aae29ff
-
docs: consolidate all model-tier benchmarking into one reference doc
New docs/model_tier_benchmarks_2026_07.md: the classifier split architecture, every benchmark scenario and result across the standard pipeline, perception, and dialogue pipelines, the 3 Anthropic-routing bugs found and fixed, the reply auto-routing feature, a full current-defaults table, and what's deliberately left untested and why. Adds a superseded-by pointer to the top of the older docs/llm_model_strategy.md, whose 3-tier design mostly shipped but whose file references and "not yet built" section are now out of date. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>26bf0bb
-
docs: add internal architecture memory reference
Consolidated file-by-file survey of both repos (backend + mobile) for fast orientation in future sessions. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>199bc14
-
review structure for structured dialogue
6a61c85
-
wire the character voice to the system with py file
79e563e
-
debug chat
cc6625f
-
litellm migration + cross-character exclusion management
- Add llm_client.py: unified LiteLLM client with flat/smart mode routing (flat=all gpt-4.1-mini, smart=Tier A→claude-sonnet-4-6, C→gpt-4.1-nano) - config.py: add ANTHROPIC_API_KEY, OPENAI_QUALITY_MODEL, classifier default → gpt-4.1-nano - All Tier A calls (ask_socrates, confrontation, closing, pragmatic, proactive) now use get_model("A") via shared client - app.py: llm_mode param per request, new GET/DELETE /chat/cross_character_exclusions endpoints - cross_character_utils.py: add remove/clear/get session exclusion helpers - Short-message optimisations: skip past dialogue search and tensions for <10 word messages - chat_repl.py: cross-character knock accept/decline/never handling in terminal REPL - requirements.txt: add litellm>=1.40, anthropic>=0.30 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>b776550
-
create philosophic thread
9077c6a -
start philosophical explanation logic
50f8e3c
-
remove bugs for short responses to exit the conversation
92af567
-
study on model update
56cf5f2