Weekly Report — 2026-W18 (2026-04-27 ~ 2026-05-03)

This week was characterized by significant progress in software stability, data pipeline construction, and deep-dive model evaluation. TokenMonitor saw multiple updates (v0.12.1 to v0.12.3), successfully resolving critical Windows platform bugs, UI/UX inconsistencies, and tray icon display issues. In the robotics domain, substantial progress was made in error recovery data collection and transferring large-scale datasets to the tianhe server, alongside a critical performance analysis of the pi0.5 model v2 which identified significant task-specific regressions in manipulation tasks. Additionally, infrastructure tasks including the batch deployment of reports to GitHub Pages were completed, though development capacity was occasionally constrained by platform-specific CI failures and environment authentication issues.

Weekly Overview

Metric	Value
Date Range	2026-04-27 ~ 2026-05-03
Active Days	6 / 7
Total Conversations	23
Projects	7
Tasks Completed	23
Tasks In Progress	2
Total Tokens	1,828,169,469
Total Cost	$1,439.54
Claude Code Token	1,184,554,421
Claude Code Cost	$892.21
Codex Token	643,615,048
Codex Cost	$547.33
Daily Average Cost	$239.92

Project Progress

TokenMonitor (5 days active) — 🔄 active

Accomplishments:

• Resolved major Windows platform bugs including hover oscillation, tray icon focus issues, and window alignment glitches
• Implemented Cursor IDE auto-reconnect feature with Rust IPC and polling logic
• Released v0.12.1 and updated to v0.12.3 with critical window resize stability (500px cap) and UI polish
• Unified model name formatting, expanded brand color palette, and fixed Cursor FloatBall ‘N/A’ display by implementing a fallback utilization chain in tray.rs

Blockers:

• ⚠️ CI failures due to platform-specific linting (macOS vs Windows)

Robotics ML Pipeline (1 days active) — 🔄 active

Accomplishments:

• Collected 88 NPZ files for error recovery pick_place demonstration
• Successfully transferred 1.9GB of data to the tianhe server via rsync

Blockers:

• ⚠️ Incomplete subtype collection due to script configuration error (missing stuck_no_progress_D1)

Robotics Model Evaluation (2 days active) — 🔄 active

Accomplishments:

• Performed comprehensive statistical analysis of pi0.5 v1 vs v2 performance
• Identified regression patterns in Clean Rollout and Error Recovery tasks
• Created formalized markdown documentation of evaluation results

Blockers:

• ⚠️ Significant performance degradation in pick_place task requires deeper investigation

Gadget Summarization Pipeline (1 days active) — ✅ completed

Accomplishments:

• Successfully batch-deployed 66 accumulated daily reports to GitHub Pages

Key Tasks

• ✅ Root cause diagnosis and fix of Windows hover oscillation bug — Implemented a three-layer defense (state check, resize blocking, and shrink-guard) to break the feedback loop between window repositioning and browser mouse events during chart interaction.
• 🔄 Error recovery pick_place data collection — Systematic collection of robotic error recovery subtypes; currently 23/24 subtypes covered.
• ✅ Analyze pi0.5 model v2 performance vs v1 — Formatted benchmark tables and computed deltas to identify regression patterns in v2 across Clean Rollout and Error Recovery scenarios, discovering critical drops in pick_place tasks (-24.1pp).
• ✅ Implement Cursor IDE auto-reconnect feature — Added Rust-based IPC commands and a frontend polling mechanism to automatically recover connection status without application restarts.
• ✅ Upload error recovery data to tianhe server — Transferred 1.9GB of validated NPZ files and scripts to remote storage for pipeline use.
• ✅ Fix TokenMonitor Cursor FloatBall N/A display — Modified tray.rs to implement a fallback logic (extra_usage -> primary_window_utilization) to ensure Cursor users with plan-based usage can view utilization data.

Problems & Solutions

1. Window height oscillates rapidly when hovering chart bars due to feedback loop between resizing and mouseLeave events. [TokenMonitor] (2026-04-27)

Solution: Implemented a three-layer defense including state-based detection and a ResizeOrchestrator to block observer-driven resizing during hover.

2. SetFocus fails on Windows tray icon clicks because the calling thread lacks foreground ownership. [TokenMonitor] (2026-04-27)

Solution: Switched to SetForegroundWindow which provides implicit rights within the tray icon context.

3. Initial v0.12.1 release failed CI due to cargo fmt and clippy warnings. [TokenMonitor] (2026-04-28)

Solution: Performed local formatting fixes, removed needless returns, and managed a clean git tag deletion/re-creation workflow.

4. Server data upload failed due to missing directories and path mismatches between local and remote expected roots. [Robotics ML Pipeline] (2026-04-29)

Solution: Manually created remote directories and uploaded data to multiple locations to satisfy both local script and downstream tool requirements.

5. pi0.5 v2 showed overall regression in Clean Rollout (-5.6pp) and severe regression in pick_place Error Recovery (-24.1pp). [Robotics Model Evaluation]

Solution: Isolated task-specific performance changes via delta-formatted comparison tables to identify that threading improved (+24pp) while manipulation tasks regressed, suggesting architectural trade-offs.

6. TokenMonitor FloatBall displayed N/A for Cursor utilization due to missing extra_usage fields for certain user tiers. [TokenMonitor]

Solution: Implemented a fallback chain in tray.rs to use primary_window_utilization when extra_usage is None.

Learnings

Architecture (architecture)

• Feedback loops in UI systems (resize → move → event → resize) require multi-layer intervention rather than single-point fixes.
• Git tag management: checking out a tag creates a detached HEAD state, requiring an explicit switch and merge to update a local branch.

Debugging (debugging)

• In async window management, anchoring to stable work area bounds is more reliable than querying the ‘current’ window state during rapid updates.
• Rate limit data structures vary by user tier; UI extraction logic must account for different available fields (plan_usage vs spend_limit_usage).

Tools (tools)

• Pre-release validation must include full CI suites (fmt, clippy with -D warnings) locally to avoid public tag churn.

Domain Knowledge (domain)

• UI color systems for brands should use hue separation for families and brightness for versions, ensuring clear visual distinction between competitors.
• Model version improvements can be task-specific; significant gains in one domain (threading) may result in regressions in another (manipulation/error recovery).
• Error recovery robustness is a distinct evaluation dimension that can decouple from base task success rates.

AI Usage Notes

Effective Patterns:

• ✓ Incremental hypothesis validation (checking logs/state before full implementation)
• ✓ Using human domain knowledge to steer abstract technical implementations (e.g., brand colors)
• ✓ Automated statistical depth: AI effectively added delta columns and aggregate statistics to raw benchmark data without explicit prompting.

Limitations:

• ✗ Tendency to jump to complex technical implementations before confirming user-facing symptoms
• ✗ Failure to proactively validate script targets against manifest ‘ground truth’ in data collection
• ✗ Lack of proactive troubleshooting: AI failed to suggest the ‘/login’ command when encountering repeated authentication failures on the MacBook.

Next Week Outlook

Prioritize investigation into the pi0.5 v2 pick_place task regression. Complete the remaining ‘stuck_no_progress_D1’ subtype for the robotics dataset. Resolve MacBook authentication issues to restore full development capacity. For TokenMonitor, maintain stability following the v0.12.3 update and continue refining the multi-platform CI pipeline.

Token Usage Statistics

Peak Day: 2026-04-30 — $695.11 / 830.5M tokens

Daily Average: $239.92