Changelog

Recent changes, summarized from commits and PRs.

2026-06-11

Added

• Personal-context injection. New orchestrator/task_context.py assembles a compact context block from vault notes (semantic search) + active learned-preference rules, injected into the parse prompt via a {{ personal_context }} slot. Degrades gracefully when the vault is empty. PreferenceApplier and memory_store are now wired into the live parser.
• Domain/duration edit pathway. PATCH /tasks/{id} (API) and PATCH /admin/tasks/{id} (dashboard) now accept domain and estimated_duration, with an inline editor in the dashboard task detail panel. These edits fire the CorrectionSubscriber learning loop, which was previously dormant for these fields.
• Skill-subsystem alerting: new skills/alerting.py raises fallback alerts for skill-lifecycle paths that were previously silent — degradation demotions, run-persistence failures, shadow-sample-loss streaks, and evolution park-in-draft. (Skill System)
• Pending-approval surfacing: a Discord ping fires when the nightly auto-draft creates skills awaiting approval, and a standing "⏳ Pending your approval" section in the EOD digest lists every skill parked in draft (auto-drafted or evolution-parked) until acted on. (Skill System)

Changed

• Task parsing is now local-first. parse_task routes to the local model (local_parser, qwen2.5:32b) as primary, with confidence-gated escalation to the cloud reasoner via a new parse_task_cloud route when the local parse confidence is below 0.7. Most tasks now parse at zero marginal cost; ambiguous ones get a cloud second opinion.
• Calibrated duration estimates. prompts/parse_task.md gained explicit duration anchors (quick comms 15 min / errands 30 / focused work 60), fixing the prior "every task is ~1 hour" behavior. The domain rubric was sharpened and now leans on injected personal context to disambiguate work vs personal.
• confidence_threshold re-added with its consumer. The Model-Layer audit (below) removed confidence_threshold from donna_models.yaml / the RoutingEntry model as read-nowhere config; it is re-added in this release together with its consuming logic — ModelRouter.confidence_threshold_for and the confidence-gated parse escalation above. (Model Layer)
• Skill auto-draft now human-gated by default: auto-drafted skills default to requires_human_gate=1, so a single human approval at draft → sandbox is mandatory before a skill leaves draft; the sandbox→shadow_primary→trusted promotions thereafter stay automatic, gated only by the §23.4 run-validity and shadow-agreement thresholds. spec_v3.md §23.5 and lifecycle.md were reconciled to match. (Skill System, spec_v3.md §23.5)
• Serialized placement choke point: Scheduler.schedule_task / schedule_dependency_chain run the read→find-slot→create-event section under an asyncio.Lock, realizing the spec_v3.md §3.7.1 double-booking guard (the earlier "async queue" wording was a design target). (Scheduling, spec_v3.md §3.7.1)
• Budget enforcement: BudgetGuard.check_pre_call now enforces the $100/month hard cap (previously daily-only — the monthly check was dead code) and fires the 90% monthly warning. BudgetPausedError carries a period (daily / monthly). (Cost & Escalation)
• Escalation-gate posture: new config/manual_escalation.yaml → gate.mode. The default shadow consults the gate on every call and logs would-escalate events (escalation_shadow_would_fire) without prompting, persisting, or blocking; enforce runs the interactive decision tree. The router now derives a deterministic cost floor when a caller omits estimate_usd, so the gate is no longer dark. (Cost & Escalation)
• Ledger integrity at the model choke point: ModelRouter.complete() is now the accounting boundary, not just dispatch. Production routers are built via build_model_router(), which requires an invocation_logger, and complete() raises RoutingError rather than make an unlogged billed call — so all spend reaches invocation_log, the table BudgetGuard reads for the $100/month cap. Chat and bot routers are now wired through the factory. (Model Layer)
• Config-driven pricing: per-call cost_usd is computed from the per-alias config rates (input/output_cost_per_token_usd) instead of hardcoded Sonnet $3/$15; the Anthropic provider fails loud on an unpriced model id rather than silently mispricing. (Model Layer)
• Dead-config audit: confidence_threshold was flagged as read-nowhere and briefly removed from donna_models.yaml / the RoutingEntry model — then re-added in this same release with its consuming logic (confidence-gated parse escalation, see Changed above). (Model Layer)
• Tool-validation allowlist can no longer be bypassed: ToolRegistry.execute now requires task_type+agent_name; previously omitting task_type skipped the allowlist entirely (principle #6). (Orchestrator)
• §7.2 sub-agent pipeline documented as dormant: spec_v3.md §7.2, docs/domain/orchestrator.md, and docs/domain/agents.md now state that AgentDispatcher + the PM/Prep/Scheduler/Decomposition agents + the agent-layer ToolRegistry are built-but-unwired, and describe the real live flow (DiscordIntentDispatcher → ChallengerAgent → ClaudeNoveltyJudge → AutoScheduler).

Fixed

• Timezone-correct slot placement: Scheduler.find_next_slot now steps candidates in UTC (DST-safe) but evaluates every time-window against the configured calendar.yaml zone, so the absolute blackout and domain windows are enforced on the user's wall clock instead of UTC — a work task can no longer land at ~4 AM local, and confirmations show the correct local time. (Scheduling, spec_v3.md §6.3)
• Deadline-aware horizon: the search horizon is clamped to the task's deadline / earliest bound (honoring a constrained weekday); an unplaceable dated task now surfaces as needs_scheduling instead of being placed late within a flat 14-day window. (Scheduling)
• Fail-closed calendar reads: placement now builds its busy-set from the union of all configured calendars (personal + work + family) and raises CalendarReadError (with a fallback alert) on any read error, rather than booking blind against an empty calendar. (Scheduling)
• Billed spend dropped on token-limit truncation: a token-capped extension call raised TokenLimitReachedError before the invocation_log write, dropping real spend from budget accounting. The raise now happens after the log + payload writes; auto_drafter / evolution catch it; log-write failures now alert via fallback_alert_fn instead of warning silently. (Cost & Escalation)
• Skill trust-gate evidence loop wired: the production executor factory now constructs SkillRunRepository and injects the bundle's ShadowSampler into SkillExecutor, with a boot invariant that alerts loudly if skills run live without run-persistence/sampler — previously the statistical trust gates ran on data that was never produced, so promotion and auto-demotion were inert in production. (Skill System)
• Skill trust-gate landmines: the requires_human_gate check no longer blocks system-actor demotions (it is scoped to a promotion-destination allowlist); gate evidence is keyed on skill_version_id, so an evolved version no longer inherits its predecessor's track record; and a run counts as valid only if it succeeded with no continued/step_failed/skill_failed step, with a config-driven failure-rate ceiling guarding shadow→trusted. (Skill System)
• Dead evolution transition removed: the contextlib.suppress(IllegalTransitionError) around an always-failing hop in evolution.py was deleted; an evolved version now parks in draft with an explicit alert, and transition() rejects reason="human_approval" from a system actor. (Skill System)
• Discord clarification replies no longer vanish: DiscordIntentDispatcher._resume discarded the pending draft before checking the re-parse status, so a clarification reply that re-parsed to escalate_to_claude (or ready+chat) fell through to no_action and was silently dropped — no judge, no task, no message. _resume now mirrors dispatch(): escalations route to the novelty judge, ready/chat returns chat, an unknown status asks the user to rephrase, and the draft is discarded only on a terminal outcome. (Orchestrator)
• Challenger fail-open is no longer silent: the three fail-open paths (transport error, OSError, execute() exception) now emit dispatch_fallback_alert, and a schema-validation failure degrades to escalate_to_claude instead of proceeding on unvalidated model output. Fail-open is kept (the Challenger must never block task creation). (Agents)
• Atomic task-state transitions: Database.transition_task_state read+validated status before taking the write lock (TOCTOU); read + validate + write now happen inside the lock, honoring the spec §3.7.1 atomicity guarantee.

Notes

• spec_v3.md model-routing and task-parsing sections describe the old cloud-first parsing; reconciliation tracked as S25 in followups.md.

2026-06-06

Added

• Time intent: the parser now emits a structured time_intent classifying when a task happens (exact / window / constrained / recurring / none), persisted as tasks.time_intent_json (Alembic migration). deadline / deadline_type are derived from it. An LLM-free fallback re-extracts common date phrasings when the model omits it. (Task System)
• Routing gate: a deterministic, LLM-free gate routes captured tasks to the scheduler (time-bound), automation (recurring), or backlog (undated). (Scheduling)
• needs_scheduling state: time-bound tasks the scheduler can't place before their deadline surface in needs_scheduling instead of stranding in backlog. (Task System)
• Persona-voice capture confirmations: slot-aware Discord confirmations (template-based, zero-token) replace the static "Scheduled: pending." reply. (Capture a Task)

Fixed

• Strand bug: time-bound tasks are now scheduled immediately by the routing gate and no longer deferred for the Challenger, fixing cases where dated tasks stranded in backlog. (Scheduling)

2026-05-18

Added

• Documentation system: global update-docs skill and docs-updater agent for bootstrapping, updating, and auditing docs across projects (Domain)

Changed

• Documentation cleanup: extracted all inline "not implemented" / "deferred" / "obsolete" callouts from 11 domain docs into open-backlog.md with stable gap IDs (G-1 through G-29)
• Skill System docs refactored: split 769-line skill-system.md into 5 focused subpages (index, setup, lifecycle, evolution, reference) under domain/skill-system/
• Memory Vault docs refactored: split 312-line memory-vault.md into 4 focused subpages (index, semantic, episodic, templates) under domain/memory-vault/
• Management GUI docs refactored: split 495-line management-gui.md into 4 focused subpages (index, api, pages, reference) under domain/management-gui/
• Domain index enhanced: added Mermaid architecture diagram and 7-step "Start Here" reading guide to domain/index.md
• spec_v3.md: added §0 Implementation Status Matrix, removed 5 inline status blocks, moved Phase 6 details to appendix
• followups.md: archived 28 closed items, trimmed to open items only
• properdocs.yml: link validation tightened to warn (CI catches broken links via --strict)
• Clarified distinct purposes of open-backlog.md (feature gaps) vs followups.md (spec questions)

Fixed

• PayloadWriter correctly wired into all ModelRouter instances
• Removed copy-paste error in cost.md (contained skill-system.md content)
• Fixed backup-recovery.md stub with proper intro text
• Updated stale "in flight" marker on slice 15 in slices.md
• Trimmed planned donna_logs.db schema from observability.md (archived to archive/)
• Expanded slices.md with slices 16–24 (escalation, budget, dashboard, chat, tool gaps)

2026-05-17

Added

• Claude Inspector: full forensics UI for browsing LLM calls, comparing payloads, and analyzing cost/performance insights (Insights, Management GUI)
• Payload collection subsystem: PayloadWriter captures full request/response payloads; PayloadEvictor enforces disk budget (Collection)
• Claude Inspector API endpoints for call browsing, payload retrieval, and insights queries
• Deep-link query parameter support on Claude Inspector page
• Claude Code project skills, agents, and hooks for development automation

Changed

• Chat engine now supports session persistence, grouping, and optimistic message rendering

Fixed

• Docker: added DONNA_PAYLOAD_DIR env var for payload storage path
• CI: resolved lint and typecheck failures

2026-05-16

Changed

• Preferences: migrated to event-driven correction pipeline via CorrectionSubscriber and TaskEventBus (Preferences)

Fixed

• UI: migrated drawers to inline expansion (Tasks, Preferences, Candidates) and CenterDialog (Logs, SkillSystem, Shadow)

2026-05-15

Added

• Chat action system: ActionRegistry with handlers for tasks, vault, skills, automations, and debug commands (Chat)
• Quick Chat panel with floating button and Cmd+J toggle
• CorrectionSubscriber for event-driven preference correction logging
• DashboardContext provider and CenterDialog primitive for UI
• Product watch v3 triage cascade with tool_use wiring
• Unit tests for executor tool_use loop and correction event flow e2e test

Fixed

• Automations: atomic success reset in advance_schedule; success un-pauses, failure notifications routed to donna-debug
• Discord: text-based done intent, thread message routing guards
• Skills: on_failure added to claude_with_triage, fallback condition fix
• Calendar: delete Google Calendar event when task is cancelled
• Migration: widen overdue_thread_map snowflake column to BigInteger

2026-05-14

Added

• Automation alert pipeline: defaults, notification channels, multi-channel routing

Fixed

• Discord: text-based done intent and thread message routing
• Scheduler: authenticate calendar client in orchestrator startup

2026-05-13

Added

• Calendar page: week view with time slots, data fetching, week navigation, completed task styling (Management GUI)
• Calendar added to sidebar navigation and routing

Fixed

• Calendar: auth token persistence, nginx proxy, day placement, DST index, overflow clipping
• Auth: allow internal-network requests to user routes without Immich login
• Docker: mount vault volume in API container
• UI: sentinel value for Vault folder select (Radix empty-string fix)
• Chat: use chat_respond template with JSON output instructions