Tool schema optimization (PLAN__Tool_Schema_Optimization.md Phases 1-3):
- model_registry.py: ROLE_DEFAULT_TOOLS — distill gets [], research/coder get
narrow tool lists by default; applied in get_role_config() when user hasn't
configured a custom list
- openai_orchestrator.py: keyword routing via narrow_tools_by_keywords() — scans
user message + last assistant turn; narrows active schemas to matched categories
only (e.g. "weather" → 3 web tools instead of 69); zero tools sent for pure chat
- openai_orchestrator.py: _get_cached_tools() — module-level schema cache keyed by
(role, sorted_tool_list, risk_params); eliminates redundant schema rebuilds
- openai_orchestrator.py: _TOOL_SCHEMA_OVERHEAD 3000 → 500 tokens (schemas now
excluded from the per-call fixed estimate since they're cached separately)
- tools/__init__.py: CATEGORY_TOOL_MAP + _KEYWORD_CATEGORY_MAP + classify_tool_categories()
+ narrow_tools_by_keywords() — the classifier logic lives here so both orchestrators
can share it
aider_run tool (cortex/tools/aider.py):
- Invokes Aider as a subprocess with --message --yes-always --no-pretty --no-stream
- Project aliases: cortex / aether_api / aether_frontend / aether_container
- Auto-injects OpenRouter API key from Cortex model registry (no ~/.env needed)
- background=True fires async + registers in agent_manager; notify=True sends push
notification on completion
- admin-only, confirm-required, TOOL_RISK=high
- .gitignore: added .aider.chat.history.md / .aider.input.history / .aider.llm.history
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Model Registry:
- Add per-user custom roles (add/remove via UI); required roles chat/orchestrator/distill
are always present and cannot be removed
- Auto-migrate legacy .env-defined roles to custom_roles on first access
- Role config panel (gear): Remove role button moved inside panel; required badge below name
- Role select: Primary + Backup slots only (was three)
Settings pages — Tailwind CSS migration (CDN, preflight: false):
- local_llm.html, settings.html, help.html, notifications.html, tools_settings.html,
crons.html, integrations.html all migrated; pg-* color tokens; dark/light via data-theme
pg.css fixes:
- input[type=checkbox/radio]: width: auto — prevents pg.css width:100% from stretching checkboxes
- btn-submit: responsive sizing via Tailwind w-full md:w-96 on each button (no longer full-width on desktop)
Documentation:
- MASTER.md, TODO__Agents.md: remove "/ Inara" from titles; description updated to "per-user AI personas"
- HELP.md: persona-agnostic language throughout (NC Talk, Google Chat, push, schedules, HA sections);
roles section restructured to show required vs. custom roles with examples
- notifications.html: subtitle and HA description use "your persona" not "Inara"
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds `anthropic_api` model type so users can authenticate with a direct
Anthropic API key instead of (or alongside) the CLI OAuth session.
- model_registry.py: `anthropic_api` type; `save/get/remove_anthropic_api_key()`
mirroring the Google account pattern; `save_cloud_model()` now picks type
based on credential type (cli → claude_cli, api_key → anthropic_api);
`_resolve_model()` merges api_key from the credential entry
- llm_client.py: `_anthropic_api()` backend (AsyncAnthropic SDK); dispatch
and fallback wiring; usage tracking
- routers/local_llm.py: Anthropic API key management routes
(POST /settings/local/anthropic-key, /anthropic-key/{id}/remove);
`anthropic_api` badge and edit-form credential selector
- static/local_llm.html: Anthropic Cloud Provider block now shows API key
management (add/remove); Add Model → Anthropic tab has credential selector
(CLI vs API key)
- requirements.txt: enable anthropic>=0.40.0
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- reasoning_budget_tokens: optional int field on local_openai models;
when set, injects {"reasoning": {"budget_tokens": N}} via extra_body
into every OpenRouter API call (both tool-loop and confirmation-gate
rounds). Field exposed in the model edit form in Settings.
- session name moved from standalone full-row div between #messages
and #input-area into the persona-switcher block in the header, as a
third dim line under "Cortex · Local". Collapses when empty via
:empty CSS. No JS changes required.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- inject_mode: per-role toggle (parallel to inject_datetime) gates the
"Current mode: Off The Record" line in the system prompt; wired through
model_registry, context_loader, chat router, orchestrator router, and
local_llm settings UI
- OTR orchestrator fix: OrchestrateRequest now carries off_record;
_finalize_job stores it per message and gates log_turn on it; JS
orchestrate payload sends off_record correctly
- Per-message hover metadata: removed always-visible .model-tag; replaced
with .msg-meta strip in the action bar (hover-only); shows model label,
host, fallback indicator, and OTR badge; stored in session JSON
- Send/stop button tooltip: shows role + model and (when tools on)
separate orchestrator model + engine label; live elapsed timer on stop
button via startRunTimer/stopRunTimer
- OrchestratorResult.backend_label: new field; openai_orchestrator fills
it; finalize_job propagates it to job dict and session messages
- GET /backend: exposes orchestrator_model label so the frontend tooltip
can show both models separately
- TODO: session delete confirmation added
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds a synchronous sub-agent spawning tool that lets the orchestrator
delegate tasks to a specific role's model and tool set.
- cortex/tools/agents.py: spawn_agent(task, role, tier, timeout, max_rounds)
- Supports local_openai and gemini_api model types
- Per-host asyncio semaphore (keyed by host_id or model type)
- asyncio.wait_for() enforces timeout; admin-only tool
- cortex/model_registry.py: max_concurrent field in host schema (default 3,
clamped 1-20); backfilled on _normalize() for existing hosts
- cortex/routers/local_llm.py + local_llm.html: "Max parallel" number input
in host add/edit forms
- cortex/tools/__init__.py: spawn_agent registered in TOOL_CATEGORIES["Agents"],
_CALLABLES, TOOL_ROLES (admin), and _ALL_DECLARATIONS
- Docs: TOOLS.md count 44→45, spawn_agent section; HELP.md tool table updated;
ARCH__FUTURE.md Round 2 completed items; TODO__Agents.md spawn_agent checked;
CLAUDE.md tool count and list updated
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Each role can now disable the current date/time header injected into the
system prompt. Default is true (all existing roles unchanged). Useful for
pure processing roles (summarizer, classifier, translator) where temporal
context is irrelevant or could cause unexpected model behavior.
Changes:
- model_registry: set_role_config/get_role_config gain inject_datetime field
- context_loader: load_context gains inject_datetime param (default True)
- orchestrator router: passes inject_datetime from role_cfg to load_context
- local_llm router: reads inject_datetime from POST body, passes to registry;
role_config_data_js includes the field
- local_llm.html: checkbox in role config panel; populate on open, save on submit
Session logs still timestamp every turn (HH:MM header in YYYY-MM-DD.md files)
regardless of this setting — the toggle only affects the system prompt header.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
tools/__init__.py shrinks from 1,137 → 250 lines. Each domain file now
owns both its callables and its FunctionDeclarations (DECLARATIONS list),
so adding a new tool only touches one file.
New TOOL_CATEGORIES dict exported from __init__ — used by the UI for
grouped tool checkboxes.
Role config UI (Settings → Model Registry → Role Assignments):
- ⚙ button per role expands an inline configure panel
- Textarea for system_append (injected into system prompt for this role)
- Grouped checkboxes for tool allow-list (all checked = no restriction)
- POST /api/models/role-config saves both fields; updates ROLE_CONFIG_DATA
in-page so re-open reflects current state without a page reload
Backend:
- model_registry.set_role_config() writes system_append + tools to registry
- TOOL_CATEGORIES exported from tools/__init__ for UI rendering
- TOOLS.md header updated: 30 → 39 tools (ae_journal_* and cortex_* additions)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Each role in model_registry.json can now carry two optional keys:
system_append — injected into the system prompt at position 7 (after
memory, closest to the turn) for the active chat_role
tools — explicit tool allow-list; intersected with the user's
access-level filter so it can only restrict, never elevate
No changes needed for existing users — missing keys fall back to current
behavior. Add keys to a role to give it a specialty focus:
"coder": {
"primary": "claude_cli",
"system_append": "You are in code-specialist mode...",
"tools": ["web_search", "file_read", "shell_exec", "scratch_write"]
}
Changes:
- model_registry.py: get_role_config() returns system_append + tools
- context_loader.py: role_append param appended as "--- Role Context ---"
- tools/__init__.py: get_tools_for_role/get_openai_tools_for_role accept
optional tool_list and intersect with access-level filter
- orchestrator_engine.py: tool_list threaded through run/resume/checkpoint
- openai_orchestrator.py: tool_list threaded through run/resume/checkpoint;
_build_client now calls get_openai_tools_for_role instead of returning
unfiltered OPENAI_TOOL_SCHEMAS
- routers/orchestrator.py: pulls role_cfg for chat_role, passes both
role_append and tool_list to context loader and engine
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Backend toggle now cycles through chat role models by label instead of
cycling service type strings (auto/claude/gemini/local).
- model_registry: get_model_for_slot() — resolves a specific priority
slot without walking the fallback chain
- llm_client: complete() gains slot param; explicit slot selection
dispatches directly to that model with no silent fallback
- routers/chat.py: ChatRequest.slot; GET /backend returns chat_models
[{slot, label, type}] for the UI; _stream_chat uses resolved model
label for the response tag when a slot is pinned
- app.js: toggle loads chat_models from /backend, cycles by label,
sends slot in chat payload; legacy model field removed from payload
- app.js: fix Gap B — agent mode placeholder no longer says "Gemini
tool loop"; now says "orchestrator"
- DESIGN doc: updated to reflect phases 1+2 complete, catalog-as-code
decision, Gap A/B documented, Phase 3 implementation details
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add Claude Opus 4.6 and Sonnet 4.5 (previous versions, still available)
- Fix context_k for Opus 4.7 and Sonnet 4.6: both have 1M context (was 200)
- Haiku 4.5 context_k 200 is correct
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Remove outdated gemini-2.0-flash and gemini-1.5-pro.
Add gemini-2.5-flash-lite (GA) and the three Gemini 3.x preview
models (gemini-3.1-pro-preview, gemini-3-flash-preview,
gemini-3.1-flash-lite-preview). All have 1M token context windows.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds cloud provider management to /settings/models:
- Google Accounts section: add/remove Gemini API keys with labels
- Add Model form: provider tabs (Local / Google / Anthropic) with
catalog dropdowns that auto-fill label and context_k
- Provider badges on model rows (Anthropic / Google / Local)
- /settings/local now redirects to /settings/models (canonical URL)
- save_cloud_model() in model_registry for Anthropic/Google entries
- Distill role migration restored in _migrate_from_local_llm
- Test fixes: version assertions updated to V2
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds a providers section to the per-user model registry for Anthropic and
Google as first-class providers alongside local hosts. Google accounts
(API keys) are now stored as a list so multiple Google accounts can coexist.
Changes:
- model_registry.py: V2 schema, auto migration V1→V2 (pulls gemini_api_key
from auth.json into providers.google.accounts), _resolve_model() merges
account API key for gemini_api type models
- routers/orchestrator.py: uses model-resolved api_key when orchestrator
role resolves to a gemini_api model with account_id
- ANTHROPIC_CATALOG and GOOGLE_CATALOG constants for model picker (Phase 2)
- New functions: get_google_api_key(), save/remove_google_account(), get_catalog()
- Documentation: ARCH__BACKENDS.md updated to V2 schema, DESIGN doc added
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- openai_orchestrator.py — new ReAct tool loop engine for any
OpenAI-compatible endpoint (OpenRouter, Open WebUI, Ollama, LiteLLM);
model handles both tool loop and final response, no Claude handoff needed
- tools/__init__.py — auto-derive OpenAI JSON Schema from existing Gemini
FunctionDeclarations so tool definitions have a single source of truth
- routers/orchestrator.py — route to openai_orchestrator when model registry
"orchestrator" role resolves to a local_openai type host
- routers/chat.py — pass role to _backend_label(); fix fallback_used logic
(only meaningful for explicit backend overrides, not auto-routing)
- static/app.js — add null/"auto" to backend cycle; fetch local model hint
without overriding the auto default on page load
- model_registry.py — _normalize() back-fills host_type on old registry files
- requirements.txt — add openai>=1.0.0
- ARCH__BACKENDS.md — document OpenAI-compat backend and routing logic
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds host_type ("openwebui" | "openai") to the host schema so Cortex can
talk to both Open WebUI/Ollama and OpenRouter/standard-OpenAI endpoints.
Path differences per type:
openwebui (default): /api/chat/completions, /api/models
openai: /chat/completions, /models
model_registry.py:
- host_type added to host schema (default "openwebui", backward compat)
- save_host() accepts host_type parameter
- _resolve_model() passes host_type through with the merged host fields
llm_client._local():
- Reads host_type from resolved model_cfg
- Selects correct chat completions path accordingly
routers/local_llm.py:
- save_host route accepts host_type form field
- fetch-models uses /models for openai type, /api/models for openwebui
- Existing host rows show type selector pre-filled from stored value
local_llm.html:
- "Add host" form includes type selector
To use OpenRouter:
- Add host: URL = https://openrouter.ai/api/v1, Type = OpenAI-compatible
- API key from openrouter.ai (store in .env or model_registry.json only)
- Fetch models or add manually (e.g. anthropic/claude-sonnet-4-5-20251022)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Introduces model_registry.py as the single source of truth for all LLM
backend configuration. Replaces scattered backend settings across user_settings,
config distill_backend_*, and the UI toggle.
model_registry.py:
- Per-user home/{user}/model_registry.json with version, hosts, models, roles
- Models have: type (local_openai|claude_cli|gemini_cli|gemini_api), label,
model_name, host_id, context_k (tokens), tags (capability labels)
- Roles map to priority chains: primary, backup_1..backup_4
- Built-in IDs (claude_cli, gemini_cli, gemini_api) always resolvable
- Auto-migrates existing local_llm.json on first access
- CRUD: save_host, remove_host, save_model, remove_model, set_role
- get_model_for_role(): registry → .env default → hardcoded fallback
config.py:
- role_chat/orchestrator/distill/coder/research .env defaults
- defined_roles: comma-separated standard role list (extensible)
- get_defined_roles() and get_role_default() helper methods
llm_client.complete():
- New role= parameter (default "chat") for registry-based routing
- model= still accepted for explicit UI toggle override
- _claude() and _local() accept model_cfg dict instead of raw string
- _local() uses pre-resolved config from registry
memory_distiller.py:
- distill_mid/long now use role="distill" (no more distill_backend_* .env vars needed)
cron_runner.py:
- brief jobs use role="chat"
routers/chat.py + auth.py:
- Use model_registry instead of user_settings for local model info
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
