Frame Engine (v3-framed):
- Tick-based deterministic pipeline: frames advance on completion, not timers
- FrameRecord/FrameTrace dataclasses for structured per-message tracing
- /api/frames endpoint: queryable frame trace history (last 20 messages)
- frame_trace HUD event with full pipeline visibility
- Reflex=2F, Director=4F, Director+Interpreter=5F deterministic frame counts
Expert Architecture (v4-eras):
- PA node (pa_v1): routes to domain experts, holds user context
- ExpertNode base: stateless executor with plan+execute two-LLM-call pattern
- ErasExpertNode: eras2_production DB specialist with DESCRIBE-first discipline
- Schema caching: DESCRIBE results reused across queries within session
- Progress streaming: PA streams thinking message, expert streams per-tool progress
- PARouting type for structured routing decisions
UI Controls Split:
- Separate thinker_controls from machine controls (current_controls is now a property)
- Machine buttons persist across Thinker responses
- Machine state parser handles both dict and list formats from Director
- Normalized button format with go/payload field mapping
WebSocket Architecture:
- /ws/test: dedicated debug socket for test runner progress
- /ws/trace: dedicated debug socket for HUD/frame trace events
- /ws (chat): cleaned up, only deltas/controls/done/cleared
- WS survives graph switch (re-attaches to new runtime)
- Pipeline result reset on clear
Test Infrastructure:
- Live test streaming: on_result callback fires per check during execution
- Frontend polling fallback (500ms) for proxy-buffered WS
- frame_trace-first trace assertion (fixes stale perceived event bug)
- action_match supports "or" patterns and multi-pattern matching
- Trace window increased to 40 events
- Graph-agnostic assertions (has X or Y)
Test Suites:
- smoketest.md: 12 steps covering all categories (~2min)
- fast.md: 10 quick checks (~1min)
- fast_v4.md: 10 v4-eras specific checks
- expert_eras.md: eras domain tests (routing, DB, schema, errors)
- expert_progress.md: progress streaming tests
Other:
- Shared db.py extracted from thinker_v2 (reused by experts)
- InputNode prompt: few-shot examples, history as context summary
- Director prompt: full tool signatures for add_state/reset_machine/destroy_machine
- nginx no-cache headers for static files during development
- Cache-busted static file references
Scores: v3 smoketest 39/40, v4-eras fast 28/28, expert_eras 23/23
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Wire Interpreter into v2 pipeline (after Thinker tool_output, before Output)
- Rename tool_exec -> tool_call everywhere (consistent convention across v1/v2)
- Switch Director v1+v2 to anthropic/claude-haiku-4.5 (was opus, reserved)
- Fix UI apply_machine_ops crash when states are strings instead of dicts
- Fix runtime_test.py async poll to match on message ID (prevent stale results)
- Add traceback to pipeline error logging
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- send_and_wait: POST /api/send + poll /api/result with timeout
- 5 test cases: greeting, german, DB count, buttons, show tables
- Clears state between tests for predictability
- --graph both: runs v1 + v2 back to back
- Reports live to frontend via /api/test/status
- Both graphs 5/5 green
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Harness reports to /api/test/status with suite_start/step_result/suite_end
- Frontend shows x/44 progress, per-test duration, total elapsed time
- Auto-discovers test count from test modules (no hardcoded number)
- run_all.py --report URL pushes live results to browser
- Fix: suite_start with count only resets on first call
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- OutputSink: collects output, optionally streams to attached WS
- Runtime no longer requires WebSocket — works headless for MCP
- WS connects/disconnects via attach_ws()/detach_ws(), runtime persists
- /api/send/check + /api/send (async) + /api/result (poll with progress)
- Graph switch destroys old runtime, next request creates new one
- Director v2 model: claude-opus-4 (was claude-sonnet-4, reserved)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
RED->GREEN->REFACTOR cycle:
- UI node has state store (key-value), action bindings (op/var), and
local action handlers (inc/dec/set/toggle — no LLM round-trip)
- Thinker self-model: knows its environment, that ACTIONS create real
buttons, that UI handles state locally. Emits var/op payload for
stateful actions.
- Thinker's context includes UI state so it can report current values
- /api/clear resets UI state, bindings, and controls
- Test runner: action_match for fuzzy action names, persistent actions
across steps, _stream_text restored
- Counter test: 16/16 passed (create, read, inc, inc, dec, verify)
- Pub test: 20/20 passed (conversation, language switch, tool use, mood)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- testcases/*.md: declarative test definitions (send, expect_response,
expect_state, expect_actions, action)
- runtime_test.py: standalone runner + pytest integration via conftest.py
- /tests route: web UI showing last run results from results.json
- /api/tests: serves results JSON
- Two initial testcases: counter_state (UI actions) and pub_conversation
(multi-turn, language switch, tool use, memorizer state)
- pub_conversation: 19/20 passed on first run
- Fix nm-text vertical overflow in node metrics bar
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Thinker emits ACTIONS: JSON line with follow-up buttons
- UI node is now pure code (no LLM call) — renders actions as buttons,
extracts tables from pipe-separated tool output, labels for single values
- Controls only in workspace panel, not duplicated in chat
- Process cards only in awareness panel, failed auto-remove after 30s
- Auth expiry detection: 403/1006 shows login button, stops reconnect loop
- Sensor timezone fix: zoneinfo.ZoneInfo("Europe/Berlin") for proper DST
- Cache-Control: no-cache on index.html
- Markdown rendering in chat messages
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
6-node pipeline: Input -> Thinker -> Output (voice) + UI (screen) in parallel
- Output: text only (markdown, emoji). Never emits HTML or controls.
- UI: dedicated node for labels, buttons, tables. Tracks workspace state.
Replaces entire workspace on each update. Runs parallel with Output.
- Input: strict one-sentence perception. No more hallucinating responses.
- Thinker: controls removed from prompt, focuses on reasoning + tools.
- Frontend: markdown rendered in chat (bold, italic, code blocks, lists).
Label control type added. UI node meter in top bar.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Python code blocks containing SQL keywords (SELECT, CREATE) were
incorrectly re-wrapped in the SQL template. Now only blocks explicitly
tagged as sql/sqlite get wrapped.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Thinker tool results stream directly to user, skipping Output node (halves latency)
- ProcessManager process_start/process_done events render as live cards in chat
- UI controls sent before response text, not after
- Button clicks route to handle_action(), skip Input, go straight to Thinker
- Fix Thinker model: gemini-2.5-flash-preview -> gemini-2.5-flash (old ID expired)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- ThinkerNode: reasons about perception, decides tool use vs direct answer
- Python tool: subprocess execution with 10s timeout
- Auto-detects python code blocks in LLM output and executes them
- Tool call/result visible in trace + HUD
- Thinker meter in frontend (token budget: 4K)
- Flow: Input (perceive) -> Thinker (reason + tools) -> Output (speak)
- Tested: math (42*137=5754), SQLite (create+query), time, greetings
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Per-node context fill bars (input/output/memorizer/sensor)
- Color-coded: green <50%, amber 50-80%, red >80%
- Sensor meter shows tick count + latest deltas
- Token info in trace context events
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- SensorNode: 5s tick loop with delta-only emissions (clock, idle, memo changes)
- Input reframed as perceiver (describes what it heard, not commands)
- Output reframed as voice (acts on perception, never echoes it)
- Per-node token budgets: Input 2K, Output 4K, Memorizer 3K
- fit_context() trims oldest messages to stay within budget
- History sliding window: 40 messages max
- Facts capped at 20, trace file rotates at 500KB
- /api/send + /api/clear endpoints for programmatic testing
- test_cog.py test suite
- Listener context: physical/social/security awareness
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
