feat(agent): hierarchical executor with path resolve, runtime probe, quiet UI

Make Zed Agent closer to Cursor: deterministic DevOps path index, live Traefik
port probe before blind edits, stop-after-edit, and quieter Russian progress.

config/litellm.base.yaml

@@ -0,0 +1,37 @@
+# Static LiteLLM settings — model_list generated by scripts/gen-litellm-config.py
+
+general_settings:
+  master_key: os.environ/LITELLM_MASTER_KEY
+  store_model_in_db: false
+  max_budget: 50
+  budget_duration: 30d
+
+litellm_settings:
+  drop_params: true
+  set_verbose: false
+  request_timeout: 300
+  num_retries: 2
+  cache: true
+  cache_params:
+    type: redis
+    host: redis
+    port: 6379
+    ttl: 3600
+  callbacks: ["prometheus"]
+  require_auth_for_metrics_endpoint: false
+
+environment_variables:
+  NOVITA_API_KEY: os.environ/NOVITA_API_KEY
+  GROQ_API_KEY: os.environ/GROQ_API_KEY
+  XAI_API_KEY: os.environ/XAI_API_KEY
+  GEMINI_API_KEY: os.environ/GEMINI_API_KEY
+  GIGACHAT_CREDENTIALS: os.environ/GIGACHAT_CREDENTIALS
+  TOGETHER_API_KEY: os.environ/TOGETHER_API_KEY
+  OPENROUTER_API_KEY: os.environ/OPENROUTER_API_KEY
+
+router_settings:
+  routing_strategy: simple-shuffle
+  num_retries: 2
+  timeout: 300
+  allowed_fails: 5
+  cooldown_time: 15

config/model_matrix.yaml

@@ -1,104 +1,76 @@
-# tier × lane (A/B/C) → Novita model id + LiteLLM fallbacks
+# Tier × lane orchestration metadata (model IDs → config/providers.yaml)
 
 lanes:
   A: economy
   B: balanced
   C: max
 
+# LiteLLM per-lane RPM (proxy-side). Zed agent bursts tool calls — keep high for personal IFT.
+# OpenRouter :free upstream may still 429; router fallbacks + cooldown_time apply separately.
 models:
   a-simple:
-    novita: novita/qwen/qwen3-4b-fp8
     fallbacks: [b-simple, groq-llama-8b]
-    rpm: 60
+    rpm: 300
   b-simple:
-    novita: novita/meta-llama/llama-3.1-8b-instruct
     fallbacks: [a-simple, groq-llama-8b]
-    rpm: 60
+    rpm: 300
   c-simple:
-    novita: novita/qwen/qwen3-8b-fp8
     fallbacks: [b-simple]
-    rpm: 60
+    rpm: 300
 
   a-medium-ops:
-    novita: novita/qwen/qwen3-8b-fp8
     fallbacks: [a-medium-code, b-medium-ops]
-    rpm: 40
+    rpm: 240
   b-medium-ops:
-    novita: novita/deepseek/deepseek-v3.2
     fallbacks: [a-medium-code, c-medium-ops]
-    rpm: 40
+    rpm: 240
   c-medium-ops:
-    novita: novita/deepseek/deepseek-v3.2
     fallbacks: [b-medium-ops]
-    rpm: 40
+    rpm: 240
 
   a-medium-code:
-    novita: novita/qwen/qwen3-coder-30b-a3b-instruct
     fallbacks: [a-medium-ops, groq-qwen-coder]
-    rpm: 40
+    rpm: 240
   b-medium-code:
-    novita: novita/qwen/qwen3-coder-30b-a3b-instruct
     fallbacks: [b-medium-ops, c-medium-code]
-    rpm: 40
+    rpm: 240
   c-medium-code:
-    novita: novita/qwen/qwen3-coder-30b-a3b-instruct
     fallbacks: [b-medium-code, c-medium-ops]
-    rpm: 30
+    rpm: 180
 
   a-complex:
-    novita: novita/deepseek/deepseek-v3.2
     fallbacks: [b-complex, a-reasoning]
-    rpm: 30
+    rpm: 180
   b-complex:
-    novita: novita/deepseek/deepseek-r1-0528
     fallbacks: [a-complex, c-complex]
-    rpm: 20
+    rpm: 120
   c-complex:
-    novita: novita/deepseek/deepseek-r1-turbo
     fallbacks: [b-complex]
-    rpm: 15
+    rpm: 120
 
   a-reasoning:
-    novita: novita/deepseek/deepseek-r1-0528-qwen3-8b
     fallbacks: [b-reasoning, gemini-flash]
-    rpm: 30
+    rpm: 120
   b-reasoning:
-    novita: novita/deepseek/deepseek-r1-0528
     fallbacks: [a-reasoning, c-reasoning]
-    rpm: 20
+    rpm: 120
   c-reasoning:
-    novita: novita/deepseek/deepseek-r1-turbo
     fallbacks: [b-reasoning, gemini-flash]
-    rpm: 15
+    rpm: 120
 
   a-vision-ocr:
-    novita: novita/paddlepaddle/paddleocr-vl
     fallbacks: [a-vision]
-    rpm: 30
+    rpm: 120
     no_escalation: true
   a-vision:
-    novita: novita/qwen/qwen3-vl-30b-a3b-instruct
     fallbacks: [a-vision-ocr]
-    rpm: 20
+    rpm: 120
   b-vision:
-    novita: novita/qwen/qwen2.5-vl-72b-instruct
     fallbacks: [a-vision, c-vision]
-    rpm: 15
+    rpm: 90
   c-vision:
-    novita: novita/qwen/qwen3-vl-235b-a22b-instruct
     fallbacks: [b-vision]
-    rpm: 10
-
-optional_providers:
-  groq-llama-8b:
-    model: groq/llama-3.1-8b-instant
-    api_key: os.environ/GROQ_API_KEY
-  groq-qwen-coder:
-    model: groq/qwen-qwen-2.5-coder-32b
-    api_key: os.environ/GROQ_API_KEY
-  gemini-flash:
-    model: gemini/gemini-2.0-flash
-    api_key: os.environ/GEMINI_API_KEY
+    rpm: 90
 
 audit:
   price_drift_threshold_pct: 10

config/orchestration.yaml

@@ -32,9 +32,89 @@ redis:
   ttl_sec: 1800
 
 classifier:
-  # heuristic | hybrid (default) | llm
   mode: hybrid
   llm_model: gigachat-classifier
   low_confidence_threshold: 0.6
   timeout_sec: 15
   max_tokens: 64
+
+# Strong Max plan → cheap coder workers/executor → conditional DeepSeek verify
+hierarchical:
+  enabled: true
+  trigger_tiers: [COMPLEX, REASONING]
+  header: X-AI-Orchestrate  # off | auto | force
+  planner_model: novita-planner
+  verifier_model: novita-verifier
+  planner_fallback: b-complex
+  verifier_fallback: b-complex
+  verify_policy: on_fail_or_hard
+  verify_input: digests
+  synthesize_model: a-medium-code
+  plan_confirm: true
+  agent_executor_model: a-medium-code
+  agent_escalate_model: b-complex
+  agent_escalate_after_edit_failures: 2
+  skip_synthesize_if_single: true
+  max_subtasks: 3
+  max_verify_retries: 1
+  plan_max_tokens: 1024
+  verify_max_tokens: 384
+  synthesize_max_tokens: 1024
+  worker_max_tokens: 768
+  wall_clock_sec: 600
+  digest_chars: 300
+  call_timeout_sec: 300
+  # First turn (kickstart read) can stay longer; after read_file body Coder hangs —
+  # tool_loop uses short timeout + escalate_after_read → DeepSeek.
+  executor_timeout_sec: 90
+  executor_num_retries: 0
+  executor_input_chars: 16000
+  executor_tool_loop_chars: 12000
+  executor_tool_loop_timeout_sec: 45
+  executor_newest_tool_chars: 3500
+  executor_minimal_context: true
+  executor_max_tools: 6
+  executor_tool_desc_chars: 120
+  executor_tools_full_chars: 8000
+  executor_max_tokens: 1024
+  executor_use_stream: false
+  executor_synthetic_on_timeout: true
+  # First turn only: synthetic read_file on plan path if model returns prose.
+  # Mid-loop still stops (no fake edits).
+  executor_synthetic_mode: kickstart_only
+  executor_max_synthetic_continues: 1
+  executor_max_edit_failures: 4
+  executor_midloop_model_retry: true
+  executor_midloop_escalate: true
+  # After successful read_file skip DeepSeek (was 408→400 loop) → edit_file
+  executor_force_edit_after_read: true
+  # After a successful read_file, skip cheap coder — it stalls on fat tool results
+  executor_escalate_after_read: true
+  # Rewrite read_file on already-read paths → edit (breaks re-read loops)
+  executor_rewrite_reread: true
+  # When model returns 200 without tool_calls: one escalate retry before kickstart
+  executor_no_tools_escalate: true
+  # After plan approve: 1) deterministic path index 2) if gaps → cheap find_* model
+  path_resolve_enabled: true
+  path_resolve_model: a-simple
+  path_resolve_timeout_sec: 45
+  path_resolve_max_tools: 3
+  # DevOps (*.calentiq.com / Bad Gateway): probe live ports before edit
+  runtime_probe_enabled: true
+  runtime_probe_timeout_sec: 8
+  # Optional: ssh host for docker inspect (empty = TCP probe from gateway only)
+  runtime_probe_ssh_host: ""
+  # Quiet Zed: milestones only (no ctx bars / LiteLLM waits / subtask dumps)
+  progress_verbose: false
+  show_context_fill: false
+  progress_in_content: true
+  stream_progress: true
+  context_window_tokens: 32768
+  plan_input_chars: 48000
+  llm_cache_ttl_sec: 3600
+  worker_map:
+    simple: a-simple
+    medium_code: a-medium-code
+    hard: b-complex
+  worker_map_max:
+    hard: c-complex

config/providers.yaml

@@ -0,0 +1,133 @@
+# Provider profiles — active: PRIMARY_PROVIDER env (default: hybrid)
+# Regenerate: bash scripts/gen-litellm-config.py
+
+profiles:
+  hybrid:
+    label: "Hybrid: Novita workers + Novita plan/verify (no Claude geo-billing)"
+    welcome_credit: false
+    requires: [NOVITA_API_KEY]
+    smart_router:
+      model: novita/meta-llama/llama-3.1-8b-instruct
+      api_key: NOVITA_API_KEY
+    fallbacks:
+      # No groq/gemini here — VPN off on IFT; those hang until client timeout.
+      a-simple: [b-simple, a-medium-ops]
+      a-medium-code: [b-medium-code, a-complex]
+      a-complex: [b-complex, a-reasoning]
+      b-complex: [a-complex, c-complex]
+      a-reasoning: [b-reasoning, a-complex]
+      a-vision: [a-vision-ocr]
+      smart-router-internal: [a-medium-ops, a-medium-code, a-complex]
+      # Planner is Max — fallback to DeepSeek then coder (never Llama)
+      novita-planner: [b-complex, a-medium-code]
+      novita-verifier: [b-complex, a-medium-code]
+    default_fallbacks: [a-medium-code, a-complex, a-medium-ops]
+    lanes:
+      # Prefer Llama for fast lanes — Qwen3 puts answers in reasoning_content
+      a-simple: { model: novita/meta-llama/llama-3.1-8b-instruct, api_key: NOVITA_API_KEY }
+      b-simple: { model: novita/meta-llama/llama-3.1-8b-instruct, api_key: NOVITA_API_KEY }
+      c-simple: { model: novita/meta-llama/llama-3.1-8b-instruct, api_key: NOVITA_API_KEY }
+      a-medium-ops: { model: novita/meta-llama/llama-3.1-8b-instruct, api_key: NOVITA_API_KEY }
+      b-medium-ops: { model: novita/meta-llama/llama-3.1-8b-instruct, api_key: NOVITA_API_KEY }
+      c-medium-ops: { model: novita/meta-llama/llama-3.1-8b-instruct, api_key: NOVITA_API_KEY }
+      # medium_code → coder / Plus-class (disable thinking — else stream hangs)
+      a-medium-code: { model: novita/qwen/qwen3-coder-30b-a3b-instruct, api_key: NOVITA_API_KEY, disable_thinking: true }
+      b-medium-code: { model: novita/qwen/qwen3-coder-30b-a3b-instruct, api_key: NOVITA_API_KEY, disable_thinking: true }
+      c-medium-code: { model: novita/qwen/qwen3-coder-30b-a3b-instruct, api_key: NOVITA_API_KEY, disable_thinking: true }
+      # complex/reasoning A/B → coder/deepseek; C (hard) → Qwen3.8-Max only
+      a-complex: { model: novita/qwen/qwen3-coder-30b-a3b-instruct, api_key: NOVITA_API_KEY, disable_thinking: true }
+      b-complex: { model: novita/deepseek/deepseek-v3.2, api_key: NOVITA_API_KEY }
+      c-complex: { model: novita/qwen/qwen3.8-max, api_key: NOVITA_API_KEY, disable_thinking: true }
+      a-reasoning: { model: novita/qwen/qwen3-coder-30b-a3b-instruct, api_key: NOVITA_API_KEY, disable_thinking: true }
+      b-reasoning: { model: novita/deepseek/deepseek-v3.2, api_key: NOVITA_API_KEY }
+      c-reasoning: { model: novita/qwen/qwen3.8-max, api_key: NOVITA_API_KEY, disable_thinking: true }
+      a-vision-ocr: { model: novita/paddlepaddle/paddleocr-vl, api_key: NOVITA_API_KEY }
+      a-vision: { model: novita/qwen/qwen3-vl-30b-a3b-instruct, api_key: NOVITA_API_KEY }
+      b-vision: { model: novita/qwen/qwen2.5-vl-72b-instruct, api_key: NOVITA_API_KEY }
+      c-vision: { model: novita/qwen/qwen3-vl-235b-a22b-instruct, api_key: NOVITA_API_KEY }
+
+  openrouter:
+    label: OpenRouter reserve (VPN; free :free + pay-as-you-go)
+    welcome_credit: true
+    requires: [OPENROUTER_API_KEY]
+    requires_vpn: true
+    smart_router:
+      model: openrouter/openrouter/free
+      api_key: OPENROUTER_API_KEY
+    fallbacks:
+      a-simple: [b-simple, c-simple]
+      a-medium-code: [b-medium-code, c-medium-code]
+      a-complex: [b-complex, c-complex]
+      smart-router-internal: [a-medium-ops, a-medium-code]
+    default_fallbacks: [a-medium-code, a-medium-ops]
+    lanes:
+      a-simple: { model: openrouter/openai/gpt-oss-20b:free, api_key: OPENROUTER_API_KEY }
+      b-simple: { model: openrouter/openai/gpt-oss-20b:free, api_key: OPENROUTER_API_KEY }
+      c-simple: { model: openrouter/openrouter/free, api_key: OPENROUTER_API_KEY }
+      a-medium-ops: { model: openrouter/openrouter/free, api_key: OPENROUTER_API_KEY }
+      b-medium-ops: { model: openrouter/openrouter/free, api_key: OPENROUTER_API_KEY }
+      c-medium-ops: { model: openrouter/openrouter/free, api_key: OPENROUTER_API_KEY }
+      a-medium-code: { model: openrouter/cohere/north-mini-code:free, api_key: OPENROUTER_API_KEY }
+      b-medium-code: { model: openrouter/cohere/north-mini-code:free, api_key: OPENROUTER_API_KEY }
+      c-medium-code: { model: openrouter/cohere/north-mini-code:free, api_key: OPENROUTER_API_KEY }
+      a-complex: { model: openrouter/google/gemma-4-26b-a4b-it:free, api_key: OPENROUTER_API_KEY }
+      b-complex: { model: openrouter/nvidia/nemotron-3-nano-30b-a3b:free, api_key: OPENROUTER_API_KEY }
+      c-complex: { model: openrouter/nvidia/nemotron-3-nano-30b-a3b:free, api_key: OPENROUTER_API_KEY }
+      a-reasoning: { model: openrouter/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free, api_key: OPENROUTER_API_KEY }
+      b-reasoning: { model: openrouter/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free, api_key: OPENROUTER_API_KEY }
+      c-reasoning: { model: openrouter/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free, api_key: OPENROUTER_API_KEY }
+      a-vision-ocr: { model: openrouter/openrouter/free, api_key: OPENROUTER_API_KEY }
+      a-vision: { model: openrouter/openrouter/free, api_key: OPENROUTER_API_KEY }
+      b-vision: { model: openrouter/openrouter/free, api_key: OPENROUTER_API_KEY }
+      c-vision: { model: openrouter/openrouter/free, api_key: OPENROUTER_API_KEY }
+
+# Always registered when keys present (all profiles)
+fixed_models:
+  # Hierarchical: strong Max plan (rare); DeepSeek verify (conditional)
+  novita-planner:
+    model: novita/qwen/qwen3.8-max
+    api_key: NOVITA_API_KEY
+    max_tokens: 1024
+    temperature: 0
+    disable_thinking: true
+  novita-verifier:
+    model: novita/deepseek/deepseek-v3.2
+    api_key: NOVITA_API_KEY
+    max_tokens: 1024
+    temperature: 0
+  # Optional Claude — only if ANTHROPIC_API_KEY works later (VPN + billing)
+  claude-haiku-planner:
+    model: anthropic/claude-haiku-4-5-20251001
+    api_key: ANTHROPIC_API_KEY
+    requires_vpn: true
+    max_tokens: 1024
+    temperature: 0
+  claude-sonnet-verifier:
+    model: anthropic/claude-sonnet-5
+    api_key: ANTHROPIC_API_KEY
+    requires_vpn: true
+    max_tokens: 2048
+    temperature: 0
+  grok-3:
+    model: xai/grok-3
+    api_key: XAI_API_KEY
+    requires_vpn: true
+  groq-llama-8b:
+    model: groq/llama-3.1-8b-instant
+    api_key: GROQ_API_KEY
+  groq-qwen-coder:
+    model: groq/qwen-qwen-2.5-coder-32b
+    api_key: GROQ_API_KEY
+  gemini-flash:
+    model: gemini/gemini-2.0-flash
+    api_key: GEMINI_API_KEY
+  gigachat-classifier:
+    model: gigachat/GigaChat-2-Lite
+    api_key: GIGACHAT_CREDENTIALS
+    ssl_verify: false
+    max_tokens: 64
+    temperature: 0
+  gigachat-pro:
+    model: gigachat/GigaChat-2-Pro
+    api_key: GIGACHAT_CREDENTIALS
+    ssl_verify: false

config/routing_rules.yaml

@@ -38,6 +38,18 @@ gateway:
     - novita
     - traefik
     - mnesia
+  medium_code_keywords:
+    - поправь
+    - исправь
+    - fix
+    - edit
+    - patch
+    - bugfix
+    - typo
+    - rename
+    - add test
+    - unit test
+    - lint
   ocr_keywords:
     - ocr
     - прочитай текст
@@ -64,6 +76,8 @@ litellm:
       tier: SIMPLE
     - keywords: ["bash", "docker", "swarm", "ci/cd", "gitea", "rebar", "erlang", "devops", "kubectl", "terraform", "playwright", "npm", "wsl"]
       tier: MEDIUM
+    - keywords: ["поправь", "исправь", "fix", "edit", "patch", "bugfix", "typo", "rename", "add test", "lint"]
+      tier: MEDIUM
     - keywords: ["рефакторинг", "refactor", "архитектура", "architecture", "спроектируй", "design system", "microservice", "distributed", "migration plan"]
       tier: COMPLEX
     - keywords: ["step by step", "reasoning", "prove", "analyze deeply", "think through"]