startup, llm, profiles, todo: five more the prod logs gave up ...

The same 5-day sweep of prod's journal as 5e51f87, continued. Every one of
these failed silently: no error in the UI, just a line in the log or a hole
in the database.

complete()/embed() retry a connection error before falling back.
stream_complete has always given its first chunk two attempts, but the
non-streaming path made exactly one per (server, model), and with a single
configured server (ollama.com) one ReadTimeout raised "All backends
exhausted: ReadTimeout" outright. That is the whole story of 2026-10-08,
when 27 calls died that way in a day — 23 of them memory summarisation, the
rest planning. Attempts now come from settings.llm_complete_retries (2) with
llm_retry_backoff_sec (2.0) between them; a missing model is still not
retried (a 404 is an answer, not a hiccup), and a lone server is still never
blacklisted — blacklisting it would block the next request for _TTL.

Startup waits for Postgres. create_container() opens the pool eagerly and
the retry loop for the DDL tables sits below it, so a host that comes up
before docker takes the whole lifespan with it. On 2026-10-07 15:19:40 the
agent ran `sudo reboot` on itself at the user's request ("Перезагрузи
себя"); the host booted 15:20:03, navi started at 15:20:18, hit
ConnectionRefusedError on the pool and logged "Application startup failed.
Exiting." Connection errors now retry for ~30s; a genuinely broken config
(a missing DATABASE_URL) still fails fast.

The profile loader reports its tally. A profile dropped by an unreadable
config.json is skipped by design so it cannot take the server down, but on
that same 15:20 boot four of them — modeler_3d, navi_code, secretary,
server_admin — were dropped with `Extra data: line 125 column 1`, and the
only trace was one error line each: the UI simply showed four profiles
fewer.

todo recovers a dropped `op`. Of the 3385 todo calls stored on prod, 73
arrived with the discriminator flattened away ({"index": 1, "status":
"in_progress"}, {"action": "view"}, {"": "add", "tasks": [...]}) and got a
bare "Unknown op: None" — a wasted round trip for an intent the remaining
arguments state plainly. The error message now names the five ops and what
each needs.

Message.created_at is stamped at creation. It defaulted to None and only
some construction sites filled it, so 85% of stored rows had no timestamp at
all (every tool message, 14 168 of 14 168) and no query could slice history
by time or measure a pause. Loading a row passes the column through
explicitly, so NULL rows stay NULL instead of being restamped with "now".

Every new test was checked to fail with its fix reverted. Full suite green
(1466 passed, 1 skipped). No frontend or dependency change, so the deploy is
a pull and a restart.

synapse reactions W2: hidden dispatcher profile + synapse_instructions tool

profiles: retire gate flags and retrain prompts for the plan tool ...

- drop planning_enabled / planning_mandatory / planning_phase2_enabled /
  observe_skips_plan_enabled / adaptive_replan_enabled from AgentProfile,
  the loader and admin serialization
- add `plan` to native tools of developer, navi_code, tool_developer,
  secretary, server_admin, modeler_3d — without the gate they would
  otherwise lose planning entirely
- system prompts: call plan for non-trivial multi-step work before
  execution; complex plans get presented for confirmation; stuck or goal
  changed -> plan with a reason

Автономность мелких моделей: OUTPUT DISCIPLINE, milestone-todo, перехват финала хода ...

Два последовательных захода над одной проблемой — мелкие модели (12–30B) в navi_code
теряют автономность на трудных шагах: «остановился поболтать» вместо действия и
слишком большое расстояние между пунктами плана.

Заход 1 (З1–З3):
- З1: OUTPUT DISCIPLINE в системном промпте navi_code — «act, don't announce»,
  без few-shot антипримеров. Контракт хода: ответ без tool_calls = конец хода, поэтому
  объявление намерения текстом убивает автономность; правила заставляют вызывать
  инструмент в том же ходе.
- З2: плоский todo + метка группы milestone + декомпозиция. _parse_plan_steps
  возвращает list[tuple[milestone, text]]; milestone — метка группировки (не сущность,
  без статуса), «done» вычисляется при рендеринге; подшаги = больше плоских шагов
  (без вложенности). TUI side-panel группирует по milestone (плоский фолбэк при пустом
  milestone). Plan depth: max 15→20 + правило декомпозиции.
- З3: adaptive re-plan «длинный шаг» — nudge «разбей шаг» при in_progress ≥ порога
  итераций без смены todo (порог 4, раньше общего anti-stall warning на 8).

Заход 2 (шаги 1–3, после cloud-теста 31b vs 12b):
Корневая структурная причина: весь спасательный механизм (anti-stall warning с явным
предложением reflect, adaptive re-plan) живёт только внутри tool-цикла — nudge
инжектируется в pre_turn *следующей* итерации, которой при «остановился поболтать»
нет (ход закрылся по return до post_turn). 31b застревала через «продолжаю
tool-итерации» → дожала до warning → спаслась; 12b — через «замолчала текстом» → мимо
всех nudge.

- Шаг 1: перехват финала хода. Если модель выдала bare-text, но в todo есть открытые
  шаги (pending/in_progress) и лимит не исчерпан — НЕ эмитить StreamEnd, а сохранить
  ассистентский текст в session.context, поставить системный nudge и continue
  (без StreamEnd, без workers — консистентно с multi-iteration tool-турами). Счётчик
  final_interceptions на AgentTurnContext, лимит final_intercept_limit (default 2),
  эскалация жёсткости (мягкий → «second stop»). has_open_steps в todo.py: пустой
  todo → False (защита casual-сообщений), failed/skipped терминальны. Профильные
  флаги final_intercept_enabled/limit в base.py + loader + admin.
- Шаг 2: жёсткий reflect-триггер в промпте — «~3 tool attempts on the same step
  without progress → call reflect IN THIS TURN (tool call, not reasoning aloud)».
- Шаг 3: открыть replan для застревания — «call replan when reflect showed the whole
  approach is dead (not one failed step, but the approach itself)».

Тесты: 874 passed, 1 skipped. Новые — has_open_steps (5), final intercept (5),
milestone-группировка, adaptive long-step nudge, парсер шагов с milestone-маркером.

cloud: navi_code model → gemma4:31b-cloud для тестирования догадки (31b признала
застревание, 12b — нет); .env cloud-host уже gitignored.

Co-Authored-By: Claude <noreply@anthropic.com>

agent: bounded autonomy — scope boundary + observe-vs-act ...

navi_code had unwanted "free flight": an observe request ("look at a
directory") triggered the full Phase 3 plan with milestones + auto-todo,
and goal_anchoring then drove the agent to finish those steps, climbing
into sibling projects and executing milestone docs it found.

Two toggleable, default-off profile flags (on for navi_code):

- scope_boundary_enabled: injects a standing system message keeping the
  agent within the literally requested scope; forbids acting on
  discovered TODO/roadmap/milestone docs (report only).
- observe_skips_plan_enabled: Phase 1 classifies MODE: observe|act; an
  observe request skips Phase 2/3 — no multi-step plan, no auto-todo, no
  "execute step by step" prompt. The agent just gathers info and answers.
  Independent of force_plan (observe on the first message still skips).

Free flight stays reproducible by flipping both flags off.

Co-Authored-By: Claude <noreply@anthropic.com>

compressor: structured summaries, profile-aware compression, adaptive keep_recent ...

- Replace free-form summary with strict Markdown template (Goal, Active Files,
  Decisions, Completed Work, Pending Work/Todo, Errors, Key Values).
- Keep filesystem/code_exec/terminal tool results and messages with
  is_compression_critical=True verbatim during compression instead of 300-char truncation.
- Make compression profile-aware: AgentProfile gains compression_keep_recent,
  compression_max_tokens, compression_prompt_file. navi_code uses dedicated
  compression prompt and larger keep_recent/max_tokens.
- Adaptive partition_messages(): important turns (user corrections, errors,
  critical tools) survive longer; filler/social turns compress sooner.
- Increase default context_summary_max_tokens from 3000 to 4000.
- Propagate active profile changes to ContextCompressor and SubAgentRunner.

Co-Authored-By: Claude <noreply@anthropic.com>

Refactor profile tool config to explicit agent/subagent structure ...

Replaces the confusing mix of enabled_tools + mcp_servers + subagent_tools
with a single explicit structure:

    tools: {
      agent:   {native: [...], mcp: {server: [groups]}},
      subagent:{native: [...], mcp: {server: [groups]}}
    }

Why:
- Old fields mixed native and MCP names (mcp__server__tool) in one list,
  making it impossible to tell at a glance what a subagent actually gets.
- subagent_runner.py had 25 lines of runtime MCP filtering logic that
  was hard to follow and error-prone.

Changes:
- AgentProfile: add ToolConfig / ToolScopeConfig pydantic models,
  keep old fields (enabled_tools, mcp_servers, subagent_tools) for
  auto-migration via _migrate_tools validator.
- loader.py: read new "tools" key, auto-migrate legacy configs.
- agent.py: _tool_list now accepts ToolScopeConfig.
- subagent_runner.py: simplified — profile.get_subagent_tools() returns
  the exact scope, no runtime filtering needed.
- context_builder.py, list_tools.py, spawn_agent.py: updated to use
  profile.get_agent_tools() / get_subagent_tools().
- All 6 profile config.json files migrated to new schema.
- Secretary subagent now explicitly gets navi-web MCP tools for web search.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Add inherit_system_prompt and is_subagent_only mechanisms ...

- inherit_system_prompt: subagent parameter to prepend parent's system
  prompt as a base layer before subagent specialisation
- is_subagent_only: profile flag blocking switch_profile, allowing
  spawn_agent only; shown with [subagent only] tag in list_profiles
- Document both in docs/profiles.md and docs/tools.md

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Refactor MCP integration: server groups in profiles ...

- mcp_servers.json: add 'groups' (read/write/admin) for gnexus-book
- AgentProfile: new 'mcp_servers' field (server_name -> group list)
- Profile loader: parse and persist 'mcp_servers' in config.json
- Agent._tool_list(): expands mcp_servers into concrete tool names
  via McpManager.resolve_group(), wildcard '*' supported
- /agents/profiles API: includes 'mcp_servers' in response
- Profiles no longer list individual mcp_ tools in 'enabled_tools'
- discuss: gnexus-book read group
- server_admin: gnexus-book read+write+admin groups

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Add profile editing to admin panel ...

Backend:
- navi/profiles/loader.py: add save_profile_to_dir() to write config.json
  and system_prompt.txt back to disk
- navi/core/registry.py: add ProfileRegistry.update() for in-memory updates
- navi/api/routes/admin.py: new endpoints GET /admin/profiles/{id} and
  PUT /admin/profiles/{id} for reading and saving full profile config

Admin panel:
- Profiles table: add Edit button per row
- Drawer form with sections: Basic, Model & Generation, Thinking,
  Planning, Sub-agent, Tools, System Prompt
- All fields editable inline (text inputs, checkboxes, textareas for lists)
- Save via PUT request, updates both disk and in-memory registry without
  server restart

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Disable thinking stalls for 3D subagents

changed llm & new ollama param

Add context providers: dynamic system message injection per LLM call ...

- navi/context_providers/ registry + built-in public_url provider (global, always injected)
- context_providers/ user directory, hot-reloaded via reload_tools
- AgentProfile.context_providers field for per-profile opt-in providers
- Agent._collect_context_injections() called before every tool-calling loop
- reload_tools now reloads both user tools and user context providers
- manuals/write_context_provider.md for Navi, docs/context_providers.md reference

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Set temperature=1.0, top_k=64, top_p=0.95 for all profiles (Google recommended for gemma4) ...

Also fixes discuss profile memory tools: use combined "memory" tool name, not nonexistent split variants.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add per-phase planning flags and planning_mandatory ...

- planning_mandatory: disables DIRECT shortcut, forces all phases to run
- planning_phase1_enabled / phase2_enabled / phase3_enabled: per-phase toggles
- planning_phase2_enabled replaces planning_reflect_enabled (migrated in loader with backward compat)
- Migrate all profile configs; rewrite docs/profiles.md as full config reference

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add Ollama multi-server fallback with in-memory blacklisting ...

- New FallbackOllamaBackend (navi/llm/fallback.py): tries servers and
  models in priority order; on LLMConnectionError blacklists the server
  for the process lifetime, on LLMModelNotFoundError blacklists the
  (server, model) pair — eliminates latency from repeated failed probes
- OllamaBackend now raises typed LLMConnectionError / LLMModelNotFoundError
  instead of bare LLMBackendError; accepts list[str] | str | None for model
- AgentProfile.model changed from str to list[str] (str auto-normalised);
  all profiles updated to ["gemma4:31b-cloud", "gemma4:26b-a4b-it-q4_K_M"]
- New config field OLLAMA_BACKENDS_FILE: path to [{host, api_key?}] JSON;
  when set, registry creates FallbackOllamaBackend instead of OllamaBackend
- ollama_backends.json template added (gitignored — contains API key)
- current_model ContextVar type widened to list[str] | str | None

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Use gemma4 cloud model by default

Improve subagent system: isolated tools, custom prompts, context transfer, timeout ...

AgentProfile:
- New fields: subagent_tools, subagent_planning_enabled, subagent_system_prompt
- loader.py: loads subagent_tools/subagent_planning_enabled from config.json,
  reads optional subagent_system_prompt.txt per profile

Profiles:
- Each profile now has a dedicated subagent_tools list (focused subset, no admin tools)
- subagent_planning_enabled: false (configurable per profile)
- New subagent_system_prompt.txt per profile with executor-focused instructions

run_ephemeral:
- Uses profile.subagent_tools instead of enabled_tools
- Builds subagent context without persona or profiles block (focused executor)
- Injects subagent_system_prompt after profile.system_prompt
- Accepts context_transfer: priming exchange injected before task message
- Wall-clock timeout (default 5 min) checked per iteration
- Returns (result_text, completed: bool) instead of bare string
- Optionally runs planning phase if profile.subagent_planning_enabled

spawn_agent:
- Removed briefing param; task is now fully self-contained
- Added system_prompt param: custom injected prompt for this specific task
- Auto-reads parent scratchpad context_transfer section via get_section()
- Result prefixed with [STATUS: completed|limit_reached]
- Timeout 300s

scratchpad:
- Added get_section(session_id, section) helper for cross-session reads

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add profile discoverability: list_profiles tool + system prompt injection ...

- AgentProfile: new short_description (1-line) and full_description (dict
  with specialization / when_to_use / key_tools) fields
- All 3 profile configs: structured descriptions added; list_profiles added
  to enabled_tools
- _build_system_prompt: now accepts full AgentProfile; injects compact
  "Available profiles" block into every system prompt so Navi always knows
  what other profiles exist and when to switch — dynamically, no hardcoding
- ListProfilesTool: new built-in; returns structured per-profile details
  (specialization, when_to_use, key_tools); accepts optional profile_id
  for single-profile lookup
- registry: register list_profiles_tool after profiles registry is built

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Restructure profiles: directory-based format with config.json + system_prompt.txt ...

Each profile is now a subdirectory under navi/profiles/ containing:
  config.json       — model, temperature, enabled_tools, and other settings
  system_prompt.txt — raw system prompt, editable without touching Python

Added navi/profiles/loader.py for auto-discovery of profile directories.
Removed individual profile .py files (secretary, server_admin, smart_home, developer).
profiles/__init__.py now simply calls load_profiles_from_dir() at import time.

New profiles can be added by creating a directory with the two required files —
no Python changes needed.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>