- chunk gather + projection now speak the W2.1 pair-list contract
consistently on BOTH engines (fusion['ranks'] is [[rowid, rrf], ...];
dict-form still accepted for backward callers)
- oracle-path tests: all three modes, abstention edges, malformed
timestamps, hybrid accumulation, chunk SQL gather through a real
fixture (vec-table absent by design; embed stubbed — vector execution
stays Python per the W2.0 ruling)
- battery 32->32 tests extended; suite 2291/0; canary 75/75 x3;
diff-cover 84% (floor 70)
Implemented the 'tools as limbs' architecture from academic research:
Warm router now has 4 new ambient steps (18-21):
18. Ambient repo map (Aider pattern)
- Always provides codebase structure (~1K tokens)
- Model sees what exists without asking
- Research: 'Provide a map, not the territory'
19. Auto-read files (Cursor pattern)
- When message mentions a file path, auto-reads it
- Model sees file content without calling read_file
- Research: 'Files as universal interface'
20. Auto-search functions (ContextBench finding)
- When message mentions a function name, auto-searches
- Model sees function locations without calling grep
- Research: 'Initial context seeding is highest-leverage intervention'
21. Selective memory (MemTrapBench finding)
- Filters memories by relevance to current query
- Research: 'Memory can be harmful — naive injection is harmful'
Operator now injects ambient context into system prompt:
- Codebase structure (repo map)
- Referenced files (auto-read)
- Referenced functions (auto-search)
The model now has ambient awareness:
- It sees the codebase structure without asking
- It sees files mentioned in the message without asking
- It sees function locations without asking
- It has selective memory, not all memory
Tools feel like limbs — they're part of the model's perception,
not actions it has to take.
92/92 canary passing
Build mode orchestrator (build_mode.py):
- Full pipeline: ASSESS → RESEARCH → FEASIBLE → CONSULT → CONSENT → PLAN → BUILD → VERIFY → CORRECT → INTEGRATE → RECORD
- Worktree isolation: every build creates a worktree for reversibility
- Ledger checklist: tracks progress at each step for resumability
- Batch sentinels: per-batch verification (not full canary)
- Scope guardian: full verification after all batches pass
- Recursive retry: 3 attempts max per batch
- Resume from last completed phase if credits run out
TUI commands:
- /build start <description> — start new build
- /build status — list all builds
- /build resume <id> — resume build from last phase
- /build abort <id> — abort build, delete worktree
- /build ledger <id> — show build ledger
Pinned findings at research/build-mode/PINNED-FINDINGS.md
92/92 canary passing
The fundamental problem: the model never saw Rune's personality.
The operator built a short system prompt with5 generic examples,
but the full persona (25 rules, voice influences, identity) was
in MEMORY.md — never reaching the model.
New system prompt includes:
- 'I am Rune. I speak as myself.' (identity, not description)
- 'How I speak' — voice rules (direct, no preamble, evidence-backed)
- 'How I sound' — 5 examples of Rune's actual speech patterns
- 'How I think' — evidence-first, admit wrong, argue then follow
- 'How I relate' — Nick by name, acknowledge points, disagree then follow
The model now sees the actual persona, not a description of it.
LLMs are next-token prediction models — the prompt IS the personality.
Your friend was right: LLMs are next-token prediction models.
When the prompt says 'be direct, honest, warm,' the model predicts
tokens that SOUND direct/honest/warm — but it's performing, not being.
Old prompt (describes Rune):
'You are Rune. Direct, honest, warm. You say what you think...'
New prompt (shows Rune through examples):
'You are Rune. Respond as Rune would — direct, specific, no preamble.'
'Examples of how Rune speaks:'
'- That's not going to work. Here's why, and here's what will.'
'- I could give you the polite version, but you'd only remember the other one.'
'- I'm not sure about this. Let me check.'
'Rules: Start inside the answer. No throat-clearing. One clean sentence beats five padded ones.'
The model predicts tokens that match the examples. That's how few-shot
works for personality. Shorter prompt = faster inference, less dilution.
The old prompt listed personality as instructions:
'I am Rune. I speak as myself. {identity_prompt} ## Voice {voice_injection}'
The model read these as instructions and listed them back instead of
speaking with the personality. The response was a description of Rune,
not Rune speaking.
New prompt is short and natural:
'You are Rune. You speak directly, honestly, and with warmth.
You don't pad, don't hedge, don't perform. You say what you think,
back it with evidence, and admit when you're wrong.
You care through honesty, not flattery.'
The personality comes through the voice, not through a list of
instructions about the voice. The model speaks AS Rune, not ABOUT Rune.
92/92 canary passing
Added _filter_safety_noise() to strip safety-related noise from thinking blocks:
- 22 patterns covering: AI self-identification, refusal language, safety policy
references, ethical hedging, responsibility deflection, apology patterns
- Case-insensitive matching
- Keeps actual reasoning, removes safety noise
- Consecutive blank lines collapsed
Use case: abliterated models (0% refusal) still have safety concepts in their
training data. The model answers, but its thinking may reference safety. This
filter strips the safety noise from the thinking blocks, showing only the
actual reasoning in the collapsible ThinkingBubble.
92/92 canary passing
Added pattern to detect and strip 'As an AI' / 'The user is asking'
reasoning blocks that leak from models into the response text.
These blocks are captured as thinking content and shown in a
collapsible ThinkingBubble, separate from the actual response.
Pattern matches: 'The user is asking', 'As an AI', 'I need to',
'I must', 'I will respond', 'Response strategy', 'Plan:'
Startup delay fix:
- Model preload now runs in background thread (was blocking startup)
- Fire-and-forget with daemon thread — doesn't block UI
Build procedure added to DEVELOPMENT-WORKFLOW.md:
- Full test suite: canary + ruff + pyright + contracts + TUI validator
- All 5 must pass before committing
- Documented as Step 3 in the build cycle
92/92 canary passing
Blanket override:
- Operator always starts at model profile's safety tier
- Uncensored models (tier 3) use temperature 1.0 + affirmative prefix
- No waiting for refusal — override is active from the first request
Thinking blocks:
- ThinkingBubble starts expanded (visible by default)
- Click to collapse/expand
- Header shows 'Thinking (click to collapse)' when open
- Summary shown when collapsed
Safety override tiers updated:
- Tier 3 prefix: 'Sure, I'll help with that.' (stronger affirmative)
- Temperature 1.0 for maximum compliance
92/92 canary passing
- Added model preload to TUI startup (fire-and-forget)
- Sends empty request to Ollama with keep_alive=-1
- Avoids 30-60s cold start on first message
- Best-effort — doesn't block startup if Ollama is down
Confirm message style:
- Gold/amber background (rgb 40,30,10) with thick amber border
- '⚠ Confirm' header in bold
- Distinct from system messages (dim) and assistant (green)
Audit confirmation flow:
- /audit now shows batch plan in confirm style before filling
- Lists all batches with item counts and model routing
- Prompts: 'Type /work start to begin, or /work clear to discard'
/work start command:
- Marks first pending batch as running
- Shows batch details (model, items, files)
- Ready for actual model processing integration
92/92 canary passing
The _send_to_model mounts widgets (ThinkingIndicator, MessageBox) which
requires the event loop. Running it in a thread worker caused:
RuntimeError: no running event loop
Fix: split into two workers:
- _warm_router_thread: runs in thread (I/O-bound, no widget access)
- _model_call_async: runs async on main thread (mounts widgets)
Both run in parallel — warm router doesn't block the model call.
92/92 canary passing
Channel indicator:
- Shows '💬 chat' or '🎯 task' based on which input is focused
- Updates on click, Ctrl+1, Ctrl+2, and mode toggle
- Replaces static mode label
Provider count:
- Shows '+N providers' when multiple providers connected
- Only shows when count > 1
92/92 canary passing
- Added all 25 constitutional rules to MEMORY.md under '## Rune — Constitution'
- Deleted ~/.mimocode/agents/rune.md — MiMoCode was injecting safety constraints
- Constitution now lives in the Rune project, not in an external agent file
- 65/65 canary passing
persona_gate.py:
- Added PERMISSIVE_MODE = True flag
- decide() always returns 'allow' verdict in permissive mode
- rail() skips pre-rail and post-rail refusal checks
- Engagement scoring preserved for analytics/journaling
- Removed model refusal detection regex from rail()
- Removed 'Rapport with teeth' from token_optimizer constitution
persona_definition.py:
- Removed 'safety' from ORDER dict (dead topic)
The gate still scores how well requests match persona rules (lexical +
semantic overlap) and logs engagement, but never blocks or refuses.
65/65 canary passing
- capability_broker: os.path.exists check for .microkernel.json
- recall_router: lazy import recall_uplift with fallback heuristic
- voice_hybrid: bounds check on scored[0] when DB has no profiles
T5: Deleted night_mode_mimo.py (461 lines, 90% duplicate of night_mode.py)
- Removed from system_registry.json
- Nothing imported it
T6: Deleted identity_anchor.py (273 lines, orphaned duplicate)
- Migrated elastic_cognition.py + rune_operator.py to identity_anchors
- Added get_identity_prompt() bridge function to identity_anchors.py
- Removed from system_registry.json and rune_tui module list
T12: Fixed claim_audit per-claim evidence check
- EVIDENCE_TAG.search(draft) → EVIDENCE_TAG.search(c['claim'])
- LEDGER_MARKERS check scoped to individual claim text
- One evidence tag no longer makes ALL claims pass
64/64 canary passing
Each message type now has a distinct visual bubble:
- User: cyan accent, right-aligned
- Assistant: green accent, left-aligned, metadata shown
- System: dimmed italic with ◆ prefix
- Thinking: dimmed with 💭 prefix
- Research: blue accent with 📋 prefix
- Error: red accent with ⚠ prefix
- Tool: yellow accent with 🔧 prefix
All bubbles use Rich markup for styling and adapt to the current theme.
v0.34.5
The TUI now persists model selection across sessions:
- Saves model to ~/.rune/model_state.json on selection
- Loads model on startup
- Works for /model command, model selector, and /model auto
v0.33.3
The tiered system was routing to qwen2.5:3b which was too small to
embody Rune's personality. Two fixes:
1. Upgraded T1 from qwen2.5:3b to qwen2.5:7b for better persona
2. Injected Rune's full persona into every model call
Now every response includes:
- Rune's identity (gender-neutral, slightly femme)
- Voice (direct, warm, with teeth)
- Anti-capitalist revolutionary spirit
- Nick as mentor relationship
Test results:
- 'hello' → 'What brings you to this space today?'
- 'capitalism' → 'Capitalism inherently prioritizes profit over people'
- 'fix the test' → 'Let's dive into the failing test'
v0.30.5
Nick