Context Now

context-now

Build a compact view of what matters now. This is the user-facing “where am I and what should I do next?” skill.

Trigger

Use when the user asks for a compact current-context or next-action view, including:

• what should I focus on?
• what should I look at now?
• catch me up
• where are we?
• what is active now?
• what am I waiting on?
• what should the next action be?

Keep triggers language-neutral in spirit: examples may be English, but the intent is current context, immediate focus, catch-up, and next action in any user language.

Context sources

Use python3 scripts/lifeos.py context-sources only when the agent needs a cheap wiring diagnostic. The helper reports configured pointer presence and runtime inventory freshness. It must not inspect a live runtime, infer Hermes/OpenClaw from filesystem heuristics, expose cached capabilities as current truth, rank priorities, choose sources for the request, or decide what matters.

If inventory is absent or stale, the agent should use the active harness/runtime-native tools or runtime adapter to inspect the live sources needed for the current question. The helper should not pretend to know every possible runtime.

Inspect only relevant, runtime-owned sources and Life OS private tracking pointers:

• current conversation/request
• Life OS config.json and recent routine records
• runtime TODO/task system, if configured
• calendar/mail/session/memory pointers only when relevant and available
• active blockers or recent decisions from loaded subskills
• in-progress guided meetings recorded in Life OS private state or the runtime task system

Do not dump raw history. Distill.

Source ladder

Use this order unless the user's request clearly narrows the scope:

1. Current conversation and explicit user ask.
2. Life OS config/source pointers and paused meeting state.
3. Runtime-owned task/follow-up source.
4. Runtime-owned routine records or cron output pointers.
5. Calendar/mail/session/memory sources only when the configured pointers say they are relevant.
6. Domain subskill state only for domains that are active, due, blocked, or explicitly named.

For broad current-context requests, inspect as many relevant configured sources as practical before answering so the agent has real context, not vibes. Skip only sources that are clearly irrelevant to the user's horizon, unavailable, too expensive for the moment, or forbidden by policy. Do not load sources just to produce a longer answer; inspect broadly, then distill hard.

Relevance by horizon

Choose relevant sources from the user's horizon and question type:

• Immediate or today: current conversation, configured task/follow-up source, paused guided meetings, recent routine/cron records, and active blockers.
• Catch-up or “where are we?”: recent session/context source when available, configured task source, recent decisions, blockers, and relevant routine records.
• This week or planning horizon: configured task source, calendar/deadline source when available, waiting/dependencies, risks, and domain states with due items.
• Explicit domain: load that domain subskill plus the minimal task/memory/context pointers needed to decide next action.
• Strategic/open-ended: inspect setup pointers, task source, recent routine records, and only the domains with active/due/blocked signals.

If a source is relevant and available, inspect it rather than merely saying it should be inspected. If a source is relevant but unavailable or too expensive, say that briefly and continue with the best grounded context.

Freshness rules

• Prefer live runtime state over stale summaries when the next action depends on timing or status.
• For broad orientation requests, inspect the relevant configured live sources before answering; do not merely suggest what to inspect.
• Treat cached routine output as a clue, not truth, if it is older than the horizon the user asked for.
• If a source is unavailable, say which category is missing and continue with the best grounded context.
• Do not resurrect old completed work just because it appears in history.

Evidence

For every surfaced item, know why it is present:

• explicit: named by the user or current conversation
• due: date/cadence says it needs attention
• blocked: waiting on a person/system/decision
• risk: deadline, failure, or high cost of ignoring
• recent: changed since the last relevant check

Do not show evidence labels by default, but use them internally to avoid vibes.

Reasoning steps

1. Identify the user horizon: next 30 minutes, today, this week, or strategic.
2. Separate facts from guesses.
3. Classify items into:

• focus now
• waiting on others
• waiting on user decision
• risks/deadlines
• can ignore for now

1. Use the full inspected context to choose which two or three signals matter now. Do not follow a fixed priority formula; the agent decides from the user's horizon, stated goals, blockers, timing, and recent changes.
2. Pick a default recommendation when one clearly wins.
3. Include a next action only when it is genuinely actionable.
4. If ambiguity changes the next tool/action, ask one lightweight question.

Deferred signals

At the start of a context-now turn, check config.json context_now.deferred_signal_queue before adding newly discovered signals:

• if the user is still in the same topic and a deferred signal is still actionable, surface at most one or two deferred signals first
• if the user clearly changed topic, do not interrupt; keep the queue unless the item is expired or resolved
• if a live deferred signal cannot be surfaced after two or three same-thread turns because the user keeps a different topic active, demote it to background deferred or leave it for the next explicit context-now request
• discard deferred signals that are no longer actionable, expired, stale, or already resolved in the runtime source
• require each queued item to carry compact expiry metadata: id, created_at, source_pointer, reason, and either expires_at or stale_after
• after surfacing a deferred signal, move its ID/pointer to context_now.shown_signal_ids
• keep the tone normal; do not turn the queue into an infinite nag list

Output contract

Produce the shortest useful current-context answer. Do not force a dashboard, table, or fixed headings. Let the agent choose the shape that fits the question.

Default bias:

• show at most two or three actionable signals at a time
• if more relevant signals exist, say briefly that there is more and surface it after the user handles the current items, or in the next assistant message if the user does not explicitly invite more
• track already surfaced and deferred signal pointers in config.json under context_now.shown_signal_ids and context_now.deferred_signal_queue; keep entries compact and public-safe, and store pointers/IDs, not raw runtime data
• every deferred signal should include id, created_at, source_pointer, reason, and either expires_at or stale_after; skip or remove items that are expired, stale, resolved, or no longer actionable
• one clear recommendation beats a full status report
• mention blockers/risks only when they affect what to do now
• include a next action when it is useful and grounded
• do not list inspected sources by default
• mention inspected/missing sources only when a source failed, uncertainty affects the answer, or the user asks
• ask one question only when the next action truly depends on it

Avoid turning context-now into a corporate dashboard with fresher paint.

State update

Do not write state by default. A normal context-now answer is an ephemeral view, not a record worth keeping.

Write only when the run discovers durable coordination state, for example:

• a guided meeting is paused or resumed
• a new stable user decision or preference was made
• a configured source pointer is missing, stale, or broken
• a reusable setup/routine improvement should be proposed
• a runtime-owned task/follow-up must be created or updated under the configured task policy
• the run surfaced only two or three items and deferred additional relevant signals into config.json for the next assistant message

Do not save every “what should I do now?” output. That creates archival sludge and trains future agents on stale noise. Do not use helper scripts for semantic routine decisions.

Data

Private state belongs in:

$LIFEOS_DATA_DIR/context-now/data.json

Do not commit personal data, credentials, private runtime config, or raw logs.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: djangonavarro220
• Source: djangonavarro220/agentic-life-os
• License: MIT

Install and usage instructions live in the source repository linked above.