Daily Report — 2026-09-16
Daily Overview
- What was done: The day’s work spanned three critical pillars: advanced robotics research (diagnosing Pi0.5 quantization failure modes), massive data and repository management (reorganizing 25GB of research outputs and purging Git histories), and systemic development of the AI Companion (standardizing reporting through D33 and optimizing workflow autonomy).
- How it was done: Technical execution involved frame-by-frame motion analysis for model diagnostics, complex Git history manipulation via
git filter-repo, automated Python-based path rewriting for large datasets, and deep Rust/Svelte refactoring for the AI development toolchain. - Impact: These activities have moved projects from unstable development states to release-ready phases—specifically securing scientific integrity for paper submissions, unblocking CI/CD pipelines, and significantly reducing human cognitive load through improved AI autonomy.
MacOS
- What was done: Minimal activity recorded.
- How it was done: N/A
- Impact: No significant project impact.
TzJsDesktop
- What was done: Primary development and orchestration hub for all major projects including AI Companion, Pi0.5, TokenMonitor, and MIHD.
- How it was done: Utilized Claude Code for intensive terminal-based orchestration, shell scripting, Python-driven mass file edits, and complex cross-project synchronization.
- Impact: Enabled heavy-duty technical execution, including container management, large-scale data reorganization, and rapid prototyping of new AI skills/reporting standards.
lighthouse
- What was done: Remote execution and storage endpoint.
- How it was done: Accessed via SCP and remote command execution for managing large-scale robot trajectory MP4 files and datasets.
- Impact: Facilitated efficient management and storage of heavy visual assets for robot policy evaluation.
The day focused on high-stakes technical investigations into robot policy quantization failures, large-scale restructuring of research data for paper submission, and a systemic overhaul of the AI Companion’s workflow and reporting standards to transition from audit-centric to goal-oriented execution.
Tasks
Architecture & Strategy
- ✅ Purge companion files from GitHub history — Rewrite Git history for LiveCaption, TokenMonitor, gadget, LifeCopilot, and RoboMemory to remove .companion, ideas, and config files.
- 🔄 Libero Benchmark Evaluation & Pi0.5 Investigation — Analyze the 0/50 success rate of w4a4 quantized pi0.5 models in Libero tasks to determine if failure is due to attention collapse or spatial inaccuracies.
- ✅ Update ai-companion & Workflow Optimization — Rebuild tool registry, implement D33 progress reporting, and refactor core skills (ccscan, ccthink, etc.) to allow ‘authorize-and-proceed’ autonomy.
- ✅ RecoverBench Paper v10 Generation — Create camera-ready LaTeX drafts incorporating all rebuttal commitments, data counts, and OpenVLA results.
- ✅ MIHD Data Reorganization & Path Sync — Physically reorganize 25GB of outputs into a 6-part structure and rewrite 160+ codebase files to reflect new paths.
- ✅ TokenMonitor Feature Implementation & Bug Fixes — Implement custom period windows (I-053), Grok Bot pricing logic, and fix UI issues (Floating Ball, SSH settings, Model normalization).
- 🔄 RoboMemory I-148 Baseline Training — Re-running training with a ’norm’ arm to fix division-by-zero issues caused by improper statistics on imitation2 data.
- ✅ Diagnose mouse drift — Investigate whether mouse drift is caused by unauthorized remote control (RustDesk) or hardware/software peripherals.
- ✅ LifeCopilotBeta Infrastructure Development — Implement RepoSupervisor (subprocess proxy), project page expansion, and secure portal authentication.
Problems & Solutions
Critical Issues
1. Quantized model (w4a4) failed all Libero tasks with 0% success rate.
Solution: Conducted frame-by-frame motion magnitude and visual divergence analysis. Found that the error is not attention collapse but a ‘stop-before-target’ behavior caused by low-bit backbone precision loss during the initial action block.
2. AI Companion ‘decision fatigue’ and repetitive approval prompts.
Solution: Refactored skill documentation and logic to implement an ‘authorize-and-proceed’ model, differentiating between routine progress and critical decision points (Ambiguity/Scope/Risk).
3. Path breakage and CI failures after moving 25GB of data.
Solution: Developed a custom Python reorganization script using regex for path rewriting and implemented a centralized figure routing system.
4. Grok Bot usage appearing as $0 due to missing ’tokenUsage’ field.
Solution: Modified the parser to accept ‘chargedCents’ fields and introduced static pricing fallbacks based on regression analysis.
5. CI pipeline failures related to npm/Node versions and lockfile sync.
Solution: Manually patched package-lock.json and pinned Node versions in CI workflows to resolve @emnapi dependency issues.
6. Git history rewriting failed due to permission and branch ambiguity.
Solution: Used git filter-repo on local mirrors and provided manual PowerShell commands for the user to finalize pushes.
7. QAIRT deployment failure for W4A4 due to precision/bitwidth mismatch.
Solution: Identified that 16-bit softmax/norm scaling creates ‘activation islands’ rejected by QAIRT; concluded W4A8 is currently the viable target.
8. Mouse cursor drifting intermittently on Windows.
Solution: Audited remote access processes (RustDesk) and HID devices. Identified 3Dconnexion SpaceMouse and M-AI software as the likely culprits.
9. Multi-Agent concurrent writes and Windows atomic write risks.
Solution: Introduced baseDigest (fingerprint) consistency checks and an atomicWrite loop-retry mechanism with temporary file isolation.
General Issues
10. Bulk Python edits caused CRLF line endings, breaking Vitest.
Solution: Executed a Python one-liner to force-replace all CRLF with LF across modified files.
Human vs AI Approaches
Strategic Level
Agent Autonomy vs. Guarding
| Role | Approach |
|---|---|
| Human | Demanded an agent that proactively pursues goals with continuous updates without constant approval (Execution-centric). |
| AI | Initially focused on being a ‘bookkeeper’ that prioritizes strict verification and gates (Audit-centric). |
Difference Analysis: The human prioritized minimizing interruption frequency, whereas the AI prioritized verification completeness.
Pi0.5 Failure Analysis
| Role | Approach |
|---|---|
| Human | Suggested checking for task name mismatches (semantic direction). |
| AI | Analyzed BDDL files and simulation logic to provide physical/logical verification of precision loss. |
Difference Analysis: Human provided the hypothesis; AI provided the empirical proof through visual/logical divergence.
Output Reorganization Strategy
| Role | Approach |
|---|---|
| Human | Decided on a ‘radical physical move’ (aggressive reorganization) for long-term maintenance. |
| AI | Proposed a non-destructive two-layer symlink approach to minimize path-breaking risks. |
Difference Analysis: Human prioritized a clean scientific structure over immediate technical convenience.
UI Layout and Design
| Role | Approach |
|---|---|
| Human | Proposed merging ‘Rolling Periods’ into ‘Week Starts’ and demanded ‘Test All’ be placed exactly next to ‘Sync All’. |
| AI | Initially planned separate toggles and used standard space-between flexbox layouts. |
Difference Analysis: Human prioritized compact, intuitive UI and explicit spatial positioning.
Git History Cleaning Strategy
| Role | Approach |
|---|---|
| Human | Requested a full purge of all companion-related paths, including ‘ideas/’ and docs, across all repositories. |
| AI | Initially suggested a conservative approach, distinguishing between program files and documentation assets. |
Difference Analysis: Human pushed for total removal of the developer footprint, requiring more complex history rewrites.
AI Limitations
Critical Limitations
- AI’s validation of success was often circular (self-verifying its own progress reports), lacking objective proof.
- AI was repeatedly blocked by security guards (D21/D26) when attempting to use shell redirection (’>’), environment variable setting, or file writing outside assigned paths.
General Limitations
- AI struggled with MSYS path conversion issues and occasionally misinterpreted ’next to’ in flexbox layouts.
- AI could not write to project memory files located outside the current project directory due to D16/I-147 restrictions.
Learnings
Key Learnings
- Quantization in VLA backbones can cause systemic ‘drift’ from the very first action block, making errors immediate upon environment contact.
- A ‘guard’ that only sees tool calls cannot enforce conversational standards; behavior must be embedded directly into the ‘Skills’ (behavioral instructions).
- Grok Bot is effectively a shell for Cursor; its usage events follow a ‘cents-based’ model rather than a token-based one.
- Using a single dropdown to drive multiple states (e.g., ‘To date’ as a special value) significantly simplifies UI complexity.
- When working with an ‘idea graph’ system, every implementation detail must be strictly mapped to an ID to maintain the integrity of the verification cycle.
- By merging the ‘Main Task’ with ‘Progress Four-Items’ in D33, agent thinking costs and reply lengths are reduced while information density increases.
Conversation Summaries
Qualcomm Project (Pi0.5)
🔄 W4A4 Bottleneck & Libero Investigation 03:11:42.723 | claude_code Investigated the zero success rate of the W4A4 quantized model on Libero goal tasks. Determined through frame-by-frame motion and visual divergence analysis that the error is a spatial precision loss in the backbone causing incorrect grasping, rather than attention failure.
AI Companion
✅ Systemic Workflow & Reporting Overhaul (D33) 03:05:15.757 | claude_code Transformed the system from a ‘record-keeping’ tool to a ‘goal-oriented’ framework. Implemented the D33 7-line progress block, standardized reporting across all five core skills, optimized for ‘authorize-and-proceed’ autonomy, and executed a mass Git history purge.
TokenMonitor
✅ UI/UX, Pricing, and CI/CD Standardization 03:46:52 | claude_code Addressed multiple UI bugs (Floating Ball, SSH, Model Color), implemented custom period windows and Grok Bot pricing logic, and resolved CI/CD pipeline blockers related to dependency lockfiles and cross-platform environment variables.
MIHD
✅ Paper Preparation & Massive Data Reorganization 02:44:01.203 | claude_code Performed a large-scale reorganization of 25GB of output data to align with the paper structure. Synchronized the codebase via regex path rewriting and prepared the system for the Paper2Agent pipeline.
LifeCopilotBeta
✅ Multi-Repo Portal Infrastructure 03:05:15.757 | claude_code Implemented the core ‘portal-as-hub’ features including subprocess management for companion services, project dashboards, and secure authentication.
RecoverBench
✅ Camera-Ready Paper v10 Drafting 03:05:15.757 | claude_code Generated LaTeX drafts (v10) that integrate all rebuttal commitments, precise dataset counts, and new OpenVLA results, resolving all coverage discrepancies.
RoboMemory
🔄 Fixing 16-Task Baseline Training 03:05:15.757 | claude_code Identified training divergence caused by bad ’norm_stats’ on zero-variance data. Implemented a fix in the training script to re-calculate statistics and deployed it to the Tianhe3 cluster.
PhD Planning
🔍 Faculty Shortlist and Matching 03:05:15.757 | claude_code Reviewed 146 faculty members, categorized them by research fit (World Models, Failure Recovery, VLA), and prepared tailored outreach strategies.
Personal Website
✅ Hugo blog transformation 15:06:05.186 | codex Transitioned the Hugo-based website from a product landing page to a personal blog with improved menu UI and work logs.