AWA v2

Today: $0.00 0 calls

ORCH

No tasks yet

Type a task and hit "Just do it" or "Let's talk"

Auto-Calendar — Awaiting Your Decision

Events extracted from email + WhatsApp. KEEP/ACCEPT pushes to calendar; X drops.

AWA Intuitions Awaiting Confirmation

Judgment patterns the pattern detector flagged for your review.

Threads — semantic conversation threads

Read-only view of the detangled threading on real data. Pick a thread on the left to see its messages on the right; with no selection the right pane shows the recent tagged stream.

AWA Data Flow — every layer, every action

How a piece of information becomes graph truth. Core Knowledge Contract: source_log is the single entry; enrichment is automatic; agent-origin data never auto-promotes — you approve. Node labels name the real tables / modules.

1 · INGESTION — capture, tag origin

gmail

signal

whatsapp / WAD

imessage

sms

calendar

web fetch

task thinking

integrator decision

autocal / resolver

audit finding

ORCH "remember X"

action: poll / webhook / agent-write → log_source(content, origin)

external = ground truthagent = LLM-derived (review-only)user = you asserted

▼ stamped with metadata_origin

2 · SINGLE ENTRY

source_log origin · source_name · source_ref · content · metadata

action: INSERT-OR-IGNORE dedup on (source_name, source_ref) → triggers maybe_refresh_db_flow_for_source

▼

3 · DB FLOW — enrichment pipeline (rebuild_universe)

a · mechanical dedupdedup_key, content_fingerprint, evidence_id

→

b · entity recognitiondeterministic_key → entities (status=candidate)

→

c · enrichmentnormalize metadata, scope, artifact_kind, sender_display

→

d · conversation groupingsource_conversations, conversation_segments

→

e · semantic threadingsemantic_threads, thread_sources, thread_outputs, event_candidates

raw_sourcesenriched evidence

entitiescandidate

source_entitieslinks

semantic_threads

▼ enrichment feeds BOTH branches

4A · KNOWLEDGE — patterns

observationsdated datapoints: mention · measurement · transaction · relation_start/end · reference

action: classifier routes into the temporal model

▼ accumulate over time

pattern_detectorscores forming hypotheses, decides promotion

▼

Intuitions knowledge table

status: forming → promoted / discarded / dormant
type: hard_fact · preference · judgment
idempotent via intuition_key

review: /api/intuitions/proposed → confirm / dismiss

4B · PROPOSALS — transient corrections

proposals unified table

issue_type: identity_mismatch · autocal_ambiguity · autocal_policy · entity_link · spam_candidate

+ specialised proposal tables

identity_resolution_actionsmerge X=Y

entity_link_proposalsrelations

thread_hub_proposals

source_blacklist_proposals

review: /api/proposals → apply / dismiss
dismiss → rejected_proposals (tombstone, dedup)

▼ rule 8 — agent-origin NEVER auto-promotes · you approve · safe hard-id links auto only for external/user

5 · PROMOTION (on your approval)

promotion.pywrites the approved attribute / relation onto the canonical entity

▼

6 · CANONICAL GRAPH (truth)

contacts

places

organizations

things

services

comm_channels

events

entity_linksrelations

entity_featurestags

knowledgepromoted attributes

Agent-world (separate, NOT in this flow): after a task, the retro (7-question post-mortem) → post_mortems.md + activity_log. That is the agent reflecting on its own work, not data about your world — it never enters the pipeline.

Known gaps (do not yet flow): file / image attachments live on disk (/var/lib/awa/attachments), not source_log. Calendar also has a direct events-table route alongside source_log.

Alerts

Provider Status

Tool Test Results

Contacts

By Model

Model	Provider	$/1M in	$/1M out	Calls	Tokens	Cost	Avg Latency	API Fail	Task Fail

By Purpose

Purpose	Calls	Tokens	Cost

Recent API Calls

Db flow

Evidence → recognize → catalog (thread→hub) → structure (autocal + graph review) → extract (findings + knowledge) → act. Hub = label on entity (event/project/topic). Thread = cross-channel timeline.

Hack ticket