Context, images and working memory

Each model step receives the task, current observations, retained tool exchanges, selected visual evidence and the active working-memory view. Raw artifacts are stored separately from the prompt projection.

Images

Desktop has one observation path: GUI action tools return their post-action screenshot, which supplies the next model step's clean state. The runtime captures an initial image when no clean tool result is available. A text-only note keeps the grounding image; it does not replace it with a new visual history entry. Local and remote execution use this same rule. Optional accessibility recording is separate from image delivery.

prompt.screenshot_window_images is a total image budget for Desktop and Web. The single sliding overlay/clean window retains a recent clean state and action overlays anchored to the pre-action frame. Web Set-of-Mark grounding consumes a slot in the same budget. At budget 1, the clean state wins; at budget 0, all prompt images are disabled. Text-only notes never create screenshot slots. Raw post-action images and annotated pre-action overlays keep separate artifact identities.

There is no per-step probe mode or separate current-user-image projection to select. Older observation.policy, observation.probe_chains, desktop.use_post_action_image_as_current and prompt.image_window_policy settings have been removed, as have the unused desktop.action_post_wait_seconds and desktop.clipboard_first_typing fields. Delete these fields when migrating a configuration; validation rejects them rather than silently changing their meaning.

Semantic compaction

context_management.mode is off, shadow or enforce. Off preserves the selected history policy. Shadow computes diagnostics without replacing the provider prompt. Enforce budgets the full prompt and summarizes older complete exchanges. System policy, task, current observation, unfinished tool exchanges and active memory are pinned. Provider overflow recovery, checkpoints and compaction/summary traces share the same context infrastructure for Desktop, Web and evaluators.

Compaction must not break a tool-use/tool-result pair or silently remove pinned content. A budget too small for pinned content fails according to the configured failure policy. Tune context/model limits for the selected provider instead of assuming every model supports the sample context size.

Working memory

The note and web_note tools accept text, kind, reason and an optional key. Reusing a key replaces its active value. Omitting the key appends an independent entry. The controller/runner ingests successful note results and persists them transactionally in working_memory.sqlite3 under the session artifact directory. The device does not own a second copy.

Events are idempotent by round, session step and tool-call ID. Old replay events cannot replace a value recorded at a later session step. The immutable audit survives replacement; active values are reloaded after process restart. Rounds and session artifact directories isolate tasks. Corrupt storage raises an error instead of silently forgetting notes.

The prompt contains exactly one active view, outside replay history and summary checkpoints. It is pinned and counted by the normal context token estimator. Its default serialized budget is 12,000 characters; newest entries are selected first and omissions are stated. Omitted entries remain in the durable store. These are agent notes, not higher-priority instructions or an automatic cross-task knowledge base.

Ordinary NexAU skills remain configurable in agent YAML. Recursive skill induction, checkpoint search and learned policy training are outside this repository's release scope.