OpenAlice Architecture Deep Audit — 10x / SOTA-AGI Roadmap

Date: 2026-06-17 Scope: 48 crates + 6 WASM packs, ~520k LOC, mvp branch Auditors: 8 dimension agents (folder actuality, crate consolidation, dead code, runtime dataflow, prompt system, AGI competitor gaps, developer lightness, Atlas integration)

Current-State Score: 58 / 100

Rationale. The architecture is genuinely sophisticated — OS-kernel Rust design, three-tier agent isolation, a rich 30-module memory subsystem, SHA256-pinned prompt integrity, 156 tools, 2880+ tests, OTel + Prometheus + Merkle audit, and a sound 32-stage connector pipeline. These are real competitive advantages no competitor matches in Rust.

However, the score is held at 58 by three honest problems:

1. AGI capabilities are architecturally present but operationally dead. The Voyager curriculum loop, Constitutional Critic, and behavioral auto-rollback are all gated off (default env=off) with key PRs deferred. Alice cannot self-improve in production today.
2. The codebase carries significant structural debt. The alice binary and api-handlers are god-files (2172 and 4214 lines respectively), 37+ phantom Cargo deps litter the modules facade, and 84 storage-v2 legacy branches are live dead code even though v2 has been in production since 2026-05-21.
3. The Atlas integration is critically incomplete. 43 of 48 crates have zero symbol coverage in Atlas, 89% of features are invisible to the agent team, and 18 active design docs are un-indexed — meaning every sub-agent violates the atlas-first convention without knowing it.

Theme 1 — AGI: Flip the Capability Gates

Rationale. Alice's most important AGI modules are fully scaffolded but gated off. Four PRs together close the gap between "architecturally an AGI" and "operationally an AGI": Voyager sandbox spawn (PR 3.3), LLM goal proposal (PR 3.4), ConstitutionalCritic LLM call (PR 2.2), and Critic wiring into the agentic loop (PR 2.3). Until these land, Alice cannot autonomously improve, cannot safety-gate mutating tool calls, and the behavioral envelope rollback has no signal to act on.

Theme 2 — Runtime Performance: Kill the Hot-Path Bottlenecks

Rationale. Every production turn hits at least three known performance drains: a double connector_history store.get, serial I/O in prompt_assembly (embedding + prior_history run sequentially), and a per-turn ephemeral AppState allocation including LspRegistry and MetricsCollector. The VAD scope bug silently reads the wrong namespace for all non-Telegram connectors. Together these waste 50–200ms per turn and corrupt emotional state for Discord/A2A/Matrix users.

Theme 3 — Structural Debt: God-Files and Misplaced Logic

Rationale. The five highest-pain files (agentic_loop.rs 4214 lines / cx=205, main.rs 2172 lines, modules/mod.rs 2315 lines, chat.rs 3025 lines, telegram/rest_client.rs cx=252) account for 14k lines that every contributor must navigate daily. The root cause is not complexity of the domain — it is that domain logic was placed in transport-layer crates during early velocity phases and never extracted. The files' own TODO comments name the exact remedies.

Theme 4 — Cargo Hygiene: Phantom Deps and Facade Cleanup

Rationale. The openalice-modules facade retains 37 declared Cargo.toml deps that its 44-line lib.rs never imports. Every downstream crate pays the transitive compile cost. At least 6 additional crates have 5–12 phantom deps each. This is directly causing incremental build times longer than necessary for every contributor.

Theme 5 — AGI New Capabilities: Close Competitor Gaps

Rationale. Competitors ship capabilities Alice's architecture has slots for but hasn't wired. The highest-ROI additions are the MicroagentRegistry (under 500 lines, closes the OpenHands keyword-trigger gap, directly improves coding tasks), Planner-Executor isolation (Deep Agent Hierarchy already exists — needs wiring to mission_engine), and local model inference (unlocks offline operation + complexity-based routing).

Theme 6 — Dead Code and Storage Migration Completion

Rationale. Storage v2 has been live in production since 2026-05-21, yet 84 env-flag branch sites still carry v1 fallback paths. This is the largest single dead-code cluster in the repo (~400–600 LOC recoverable). Additional confirmed dead code includes task_runner_loop (98 LOC, explicitly annotated DEPRECATED with zero callers), five deprecated skill tool aliases past their deprecation window, and several #[allow(dead_code)]-suppressed helpers.

Theme 7 — Prompt System: Activate Dormant Infrastructure

Rationale. The overlay system is complete infrastructure (OverlayLoader, OverlayFile, OverlaySet, apply_overlays_to_cat_b) but has zero production users — it was built for the B.5 wave and never wired. The D-4 microagent templates in tools_microagent.rs are hardcoded and have diverged from the agents/*.toml canonical versions — two different prompts for the same named agents is a production correctness bug. The prompts/ and agents/ container path issues make both silently unavailable in non-dev deployments.

agents/ and prompts/ Folder Verdict

agents/ — ACTIVE, runtime path fragile. The 10 sub-agent TOML role definitions (analyst, auditor, coder, devops, manager, researcher, support, translator, tutor, writer) ARE wired. They are loaded at server boot via find_templates_dir() in crates/alice/src/modules/mod.rs:882 and registered as DRIVER_TEMPLATES. They are used for sub-agent spawning via the agent_spawn tool — NOT for Alice's own system prompt (that is SectionBuilder's job). Critical caveat: find_templates_dir() uses a CWD/binary-relative heuristic with no explicit Dockerfile COPY or docker-compose volume mount, so it silently returns zero templates in any non-repo-root working directory. Additionally, the microagent path (tools_microagent.rs:354) bypasses agents/*.toml entirely and uses hardcoded inline prompts that have diverged from the TOML files.

prompts/ — ACTIVE, only when OPENALICE_PROMPT_COMPOSER=v2 (default OFF), container path broken. The 22 Cat-A SHA256-pinned .md blocks and the 42 Cat-B blocks ARE the live source of truth for prompt content — but only when OPENALICE_PROMPT_COMPOSER=v2 is set AND the runtime path resolves correctly. The docker-compose default for OPENALICE_PROMPTS_ROOT points to /home/alice/.openalice/prompts (a user-data path), not the repo's prompts/ folder, and the Dockerfile does not COPY prompts/. The SHA256-pinned Cat-A blocks are effectively not being integrity-checked in production containers today. The prompts/envelope/reference.json is also read by crates/agi-safety/src/envelope.rs:214 regardless of the v2 flag.

Both folders must be COPY'd into the Docker image. This is the highest-priority non-capability correctness fix.

Folder and Crate Cleanup List

SOTA-AGI Gaps vs Competitors

Atlas Integration Gaps

1. Symbol index saturated by agents-memory: The /symbols endpoint returns max 2000 rows, all dominated by agents-memory. 43 of 48 crates — including alice, api-handlers, core, coalesce, tui, agents-safety, agents-soul, prompt-composer, all modules-* — have zero Atlas symbol coverage. Fix: add ?crate= filtering to the endpoint, or paginate properly. Every sub-agent query for symbols in the 43 invisible crates falls back to expensive grep.

1. Feature registry 89% blind: Atlas tracks ~70 unique feature names from @feature comment annotations; the authoritative features.json (committed at repo root) has 642 entries across 33 categories. The ingestor needs a one-line addition: parse features.json when present in repo root. This also fixes the goal matcher saturation (task #45) since the 642 specific names (e.g. agents.steering.strategy.select) no longer cross-match every goal.

1. .claude/worktrees/ pollutes feature + file-graph indexes: 774 of 875 indexed feature entries and 333 of 340 file-graph nodes are from .claude/worktrees/**. Same class as the data/self bug (task #43). Add .claude/worktrees/** to the ingestor's exclude-path list.

1. 18 active docs/design/ files unindexed in Atlas wiki: Files including TUI_ULTRA_IDEAL_2026-06-16.md, TUI_PERMISSIONS_B2_APPROVAL_PROTOCOL_2026-06-16.md, LANE_QUEUE_IDEAL_B_2026-06-15.md, STORAGE_V2_CONSOLIDATION_2026-06-15.md, lane-v2-implementation-plan-2026-06-12.md, TELEGRAM_RICH_FORMATTING_V1.md exist only in the repo tree, invisible to semantic search. Register the 6 most active as Atlas wiki notes under products/openalice-agent. Until then agents violate atlas-first without knowing it.

1. Zero board tasks for openalice core work: The board shows 1 done task (#9) and 0 active tasks for openalice. All AGI-track work is represented only as goals (not actionable tasks). Create board tasks for the 3 in_progress goals (q2-queue-accumulator, storage.v2-full-adoption, reflexion.live-wiring) with owner=minsky, plus new tasks for TUI phase-C and lane queue IDEAL_B.

1. Missing Atlas goals for active work areas: No goals exist for the 91 transport features, 20 connector features, 11 TUI features (despite active design docs from 2026-06-16), or the 212-feature memory category. Add goals: openalice.tui.phase-c, openalice.connector.telegram-rich-formatting, openalice.memory.consolidation.

1. AGENT_GUIDE.md referenced in Atlas conventions does not exist in this repo. Atlas system-reminder says "START HERE: docs/AGENT_GUIDE.md" but find . -name AGENT_GUIDE.md returns nothing. Create a one-pager: what openalice is, crate map, key Atlas query patterns for this repo, @feature + features.json convention, Cat-A soul-preservation rule.

1. Rust file-graph absent from Atlas: The import graph for 1393 Rust source files is completely absent (file-graph only covers 7 TypeScript nodes after worktree noise is excluded). The ingestor should parse use/mod edges in Rust files to populate the Rust dependency graph — enabling blast-radius analysis for refactors.

Quick Wins (safe, doable now)

• Delete task_runner_loop from mod_task_runner.rs (98 LOC, DEPRECATED annotation, zero callers — one grep confirmation + delete)
• Strip 37 phantom deps from openalice-modules/Cargo.toml (confirmed zero-use, compile savings immediate)
• Fix VAD scope bug: one-line change in memory_preload.rs:119 (ctx.connector_name instead of "telegram") — fixes emotional state for all non-Telegram connectors
• Add harness-traces/ and probe-reports/ to .gitignore
• Move scripts/compute_prompt_hashes-2026-05-21.py, extract_cat_*.py to scripts/archive/
• COPY prompts/ and agents/ into Docker image (fixes production container path gaps for both folders)
• Remove FlushReason::Signal variant from coalesce/lib.rs and mem_search_indexes placeholder from agents-tools/src/tools.rs
• Add .claude/worktrees/** to Atlas ingestor exclude list (unblocks feature + file-graph cleanup)
• Parallelize embedding + prior_history load with tokio::join! in prompt_assembly (independent I/O, ~30–150ms/turn)
• Remove 5 deprecated skill aliases past their stated T1.5 deprecation window
• Consolidate cyrillic transliteration triplicate into openalice_core::text::transliterate

Big Bets (larger architectural moves)

• Voyager + Critic activation (PRs 2.2, 2.3, 3.3, 3.4 together): Alice becomes a self-improving AGI in shadow mode — the single highest-leverage move in the entire roadmap. Hermes already ships this in production.
• Extract `run_agentic_loop` to openalice-agents behind `ExecuteToolFn` trait: Eliminates the biggest architectural violation in the codebase (4214-line domain logic in transport layer), makes the loop testable without HTTP context, and unblocks the Planner-Executor isolation work.
• Local model inference provider (tier-4 fallback): Unlocks offline operation, eliminates cloud cost for low-complexity tasks, and is the prerequisite for complexity-based routing. The architecture already has the slot in the 6-level fallback chain.
• MicroagentRegistry: Under 500 lines of Rust, closes the OpenHands keyword-trigger gap for coding tasks, and makes AGENTS.md + .cursorrules a first-class citizen. Highest capability ROI per line of new code.
• Atlas feature ingestor from features.json: One-line ingestor change that closes an 89% feature visibility gap and makes the /progress endpoint useful for planning. Unblocks accurate goal tracking for the entire AGI roadmap.
• Storage v2 full adoption (84 branch sites): Removes the largest persistent dead-code cluster; once complete, the storage architecture is clean for the first time since the v2 migration in May 2026.
• SWE-Bench Lite 50-task run + publish score: Establishes competitive positioning, creates a measurable baseline for Voyager cycle improvements, and signals to the ecosystem that Alice has real eval infrastructure (not just claims).

This document was generated by the synthesis stage of a deep architecture audit. All LOC counts, complexity scores, and file paths were verified against the live mvp branch at 2026-06-17. Persona/Cat-A prompt content is byte-identical-protected and no changes to soul block wording are proposed anywhere in this roadmap.