v4.7.0 - Progressive Capabilities, Context Management & Trusted Access

[siddsachar]

v4.7.0 - Progressive Capabilities, Context Management & Trusted Access

This release builds on v4.6.0 with a prompt-efficiency, long-conversation,
remote-access, and runtime-reliability pass. It discovers enabled external
tools and task-specific skills only when they are relevant, meters the complete
next model input and safely compacts older context, lets owners manage exact
trusted browser addresses without weakening managed deployment policy, and
hardens provider catalogs, OpenAI-compatible streams, chat traces, public
documentation, and deterministic validation around those changes.

Progressive Tool And Skill Discovery

• Auto capability loading - keeps core tools allowed by the active Agent
  Profile directly available while enabled MCP, plugin, Custom Tool, and
  channel capabilities remain searchable until a request needs them.
• Deterministic local search - ranks authorized capability metadata across
  names, aliases, sources, descriptions, tags, and parameter names without a
  provider call, caps returned matches, and keeps stable ordering.
• Exact external invocation - returns the effective target schema, validates
  arguments again before dispatch, preserves the real integration identity in
  live and reopened traces, and never turns search into authorization.
• Unchanged safety boundaries - preserves profile filtering, tool
  enablement, provider compatibility, approvals, prompt-injection handling,
  execution budgets, cancellation, and workspace policy for discovered tools.
• Bounded capability manifests - sanitizes untrusted catalog metadata,
  omits instruction-like descriptions, rejects core, bridge, and ambiguous
  external name collisions, and degrades oversized manifests without flooding
  the model context.
• Progressive skill loading - searches enabled manual and plugin skills,
  activates the selected instructions for the current task, safely confines
  optional UTF-8 references to the skill root, and displays one compact
  Using skill receipt plus an active-skill chip.
• Durable task scope - restores automatically selected skills when a task is
  reopened, retains at most five automatic selections per task, treats a repeat
  load as a no-op, and keeps pinned skills under their existing manual control.
• Parent and child isolation - gives every child Agent its own skill state,
  profile, tool allow-list, approval mode, workspace, and execution budget;
  skill instructions cannot grant a denied tool or enable delegation.
• Tools settings and compatibility - adds Settings → Tools → Capability
  loading, makes Auto-select external tools the recommended mode, keeps
  Load all external tools for eager compatibility, and moves the former
  Search settings surface into Tools.

Context Window Meter And Rolling Compaction

• The desktop composer now shows an event-driven estimate of the complete next
  model input, including system prompts, images, and bound tool schemas. Mobile
  keeps the narrower composer and shows only durable compaction status rows.
• Agent Mode, tool-loop calls, resumes, and Chat Only share token accounting and
  bounded rolling compaction. Older history is summarized as untrusted reference
  data while recent complete turns and tool-call/result groups remain intact.
• Context settings migrate to policy version 2. The historical local 32K value
  becomes Auto (still requesting 32K, capped by native/observed Ollama capacity),
  and the historical provider 128K value becomes Auto. Because an old explicit
  128K selection cannot be distinguished from the former default, restore it via
  Settings → Models → Advanced context → Provider context override if needed.
• Successful compaction produces one durable presentation-only timeline notice
  and, for channel-originated turns, one separately claimed channel notice.
  Context events never enter model prompts, summaries, or token accounting.
• Capacity-aware policy - resolves provider metadata, maintained limits,
  local requested and observed Ollama allocations, and custom-endpoint
  overrides; an unknown remote model uses a disclosed 128K application fallback
  for metering and compaction authority rather than claiming provider metadata.
• Safe headroom - reserves input space according to combined-window versus
  input-only provider semantics and begins compaction at 75 percent of the
  effective context limit.
• Recent-turn integrity - compacts complete older user-led groups while
  retaining at least two recent complete turns and intact tool-call/result
  groups, and marks the rolling summary as untrusted historical reference data
  subordinate to the newest raw user instruction.
• Durable concurrency and recovery - validates checkpoint revisions and
  boundary digests, saves summaries with compare-and-swap, accepts a valid
  concurrent winner, bounds compactor input and output, honors stop requests,
  and fails safely when sufficient context slack cannot be created.
• Overflow diagnostics - remembers when a model exceeds its reported
  capacity during the session and asks for a different model or reviewed
  context override instead of repeatedly sending the same oversized input.

Trusted Remote Addresses And Invitations

• Durable trusted addresses - lets an authenticated owner add one exact
  HTTP or HTTPS origin in Remote Access settings, saves it in
  access_routes.json, applies Host/origin admission immediately, and creates
  an origin-bound invitation without a restart.
• Visible lifecycle controls - lists saved addresses after restart, includes
  them in the invitation route selector, and removes one only after confirmation;
  new HTTP and WebSocket traffic through a removed address is rejected at once.
• Strict origin validation - accepts an exact scheme, host, and optional
  port while rejecting credentials, paths, queries, fragments, wildcards,
  malformed ports, and non-HTTP(S) schemes.
• Managed deployment ownership - shows ROW_BOT_PUBLIC_ORIGINS entries as
  externally managed and makes add/remove controls read-only when
  ROW_BOT_ALLOWED_HOSTS owns Host admission, so UI state cannot override
  explicit operator policy.
• No implied network provisioning - makes clear that trusting an address
  does not perform DNS or reachability checks, provision TLS, configure a proxy
  or firewall, or change Row-Bot's listen address; HTTP remains unencrypted.
• Restored custom routes - repairs invitation creation for exact
  operator-configured browser-facing origins and revalidates the route inventory
  at creation time so stale or unavailable selections are refused.
• Assigned-interface discovery - offers usable unicast addresses actually
  assigned to the computer even when they are outside RFC-private ranges, while
  excluding loopback, link-local, unspecified, and multicast addresses.
• Exposure-specific guidance - gives non-private interface addresses a
  stronger warning about possible external routing, firewalls, plaintext HTTP,
  and the safer HTTPS or Tailscale alternatives.

Provider, Streaming And Chat Reliability

• Provider-scoped catalog credentials - refreshes OpenAI, Ollama Cloud,
  OpenRouter, Requesty, Anthropic, Google, xAI, MiniMax, OpenCode, and Atlas
  Cloud catalogs through the provider auth store, so a key saved only in
  Settings immediately populates provider-qualified model rows without a
  legacy environment-key mirror.
• Long-response timeout policy - gives OpenAI-compatible requests separate
  10-second connect/pool, 120-second write, and 900-second read-inactivity
  limits. ROW_BOT_OPENAI_COMPATIBLE_READ_TIMEOUT accepts a positive seconds
  override and falls back safely when invalid.
• Safe pre-stream retry - retries a timeout or remote-protocol failure once
  only when no stream event has been emitted, never replays partial output, and
  checks cancellation before retrying.
• Quieter tool traces - makes collapsed live and reopened tool rows smaller,
  lighter, and less visually dominant while preserving full result content and
  success/failure state inside their expansions.

Public Website And Documentation

• Stable release alignment - finishes the public 4.6.0 metadata, downloads,
  Linux command, cache revisions, structured data, fallbacks, and cross-page
  navigation that landed immediately after the 4.6.0 source tag.
• Workbench-first explanations - updates the public docs and landing copy
  around parent-led agents, durable documents, authenticated remote access,
  Docker/VPS operation, compact presentation, and the complete models, tools,
  memory, workflows, code, design, messaging, and voice surface.
• Clearer public imagery - leads the documentation with the Knowledge
  workspace, revision-tags screenshot assets, handles query strings in local
  image validation, and keeps architecture diagrams full width.
• Public-site conversion contracts - centralizes the marketing-page Google
  tag bootstrap and measures desktop download, Linux install-view, and
  installation-guide actions; this is public-website behavior, not Row-Bot
  application telemetry.
• Progressive capability guide - documents Auto and eager modes, tool and
  skill discovery, chat receipts, parent/child boundaries, troubleshooting,
  and the related Tools, Skills, extension, mobile, and request-lifecycle pages.
• Remote-access guide - documents trusted-address persistence and removal,
  exact-origin rules, assigned-interface warnings, externally managed settings,
  and the DNS, TLS, proxy, firewall, and reachability responsibilities Row-Bot
  deliberately does not assume.

Tests And Release Validation

• Capability-discovery coverage - verifies ranking, collisions, manifests,
  schema validation, approvals, cancellation, execution budgets, child
  boundaries, persistence, transcripts, status, plugins, MCP, channels, mobile,
  and public documentation.
• Context-management coverage - verifies policy migration and resolution,
  complete-input accounting, compaction boundaries, summary validation,
  persistence, concurrency, failures, Agent and Chat Only parity, channels,
  SMS, mobile, and the desktop meter.
• Remote-access coverage - verifies route inventory, non-private interface
  classification, invitation revalidation, trusted-address mutation, runtime
  HTTP/WebSocket admission, managed-policy precedence, UI controls, and docs.
• Provider and chat regressions - verifies timeout phases, retry/no-replay,
  cancellation, provider-only credentials, catalog caching and qualification,
  first-class providers, Settings dialogs, and live/reopened trace styling.
• Public-site contracts - verifies fallbacks, internal links, versioned
  images, responsive calls to action, architecture lightboxes, generated pages,
  tag initialization, and download/install event payloads.

Breaking Changes And Caveats

• Auto-select is the recommended external capability mode. Use Settings →
  Tools → Capability loading → Load all external tools when an older provider
  or integration requires every enabled external schema to be bound eagerly.
• Context policy version 2 interprets the historical local 32K and provider
  128K values as Auto because old settings did not distinguish defaults from
  explicit choices. Restore a deliberate provider 128K cap in Advanced context
  settings if required.
• Rolling compaction uses the selected model/provider to summarize older
  history. It does not create a new provider route, but cloud-selected tasks
  send the bounded aged range to that cloud provider as part of normal model
  use.
• Every trusted remote address admits the same single Row-Bot owner after
  invitation authentication. Add exact origins only when you control the route,
  and remove them when they are no longer needed.
• Direct HTTP remains unencrypted, and a non-private assigned interface may be
  externally routable. Prefer private Tailscale HTTPS or an operator-managed
  HTTPS reverse proxy outside a trusted LAN.
• Trusting an HTTPS origin does not create DNS, certificates, proxy trust,
  firewall rules, or a listener. Those remain explicit operator responsibilities.
• The public marketing pages load Google tag services and record the listed
  install/download interactions. The Row-Bot application itself adds no
  first-party telemetry or hidden phone-home behavior.
• The default OpenAI-compatible read-inactivity limit is now 900 seconds. Set a
  reviewed positive ROW_BOT_OPENAI_COMPATIBLE_READ_TIMEOUT value if an
  endpoint needs a different ceiling.

---

This discussion was created from the release v4.7.0 - Progressive Capabilities, Context Management & Trusted Access.