RTK Savings Data Not Displaying + Need for RTK Bypass Option

Environment

App version: alma v0.0.756 (latest)
OS: macOS (Apple Silicon)
Model: kimi-k2p5-turbo (via Fireworks)

Bug 1: RTK Savings Stats Always Show "暂无节省数据"

Description

The RTK (Real-time Kompression) panel in the Alma UI consistently shows "暂无节省数据" (No savings data yet) / "When Alma executes commands, RTK will automatically compress tool output. Start a conversation and savings data will appear here."

However, RTK is clearly functioning — tool outputs show [alma compacted XXXXX chars] markers indicating compression is happening successfully. The statistics display just doesn't reflect this.

Evidence

Tool outputs are being compressed: e.g., alma compacted 11292 chars, alma compacted 1183 chars
This has been observed across multiple conversations and commands
The UI stats panel never updates, even after extended sessions with many compressed outputs

Expected Behavior

The RTK savings panel should display:

Total bytes/chars saved
Compression ratio or percentage
Number of compressions performed

Actual Behavior

Panel always shows "暂无节省数据" regardless of how many compressions have occurred.

Possible Cause

The RTK compression engine runs correctly but the statistics/metrics are not being persisted or emitted to the UI layer. The UI component reads from a stats store that is never populated.

Feature Request: Per-Tool or Per-Skill RTK Bypass

Motivation

During a workflow involving ChromeRelay to organize Blackboard course files (thread: mnsr31rnasjxdxoey2), RTK's auto-compaction caused significant issues:

Lost HTML structure: ChromeRelay returns full page HTML/DOM data. RTK compaction strips or truncates this, making it impossible to parse download links, form tokens, or navigation elements.
Lost file metadata: When listing downloaded files, the compacted output dropped file sizes and modification dates that were needed for deduplication logic.
Context loss in multi-step browser operations: A sequence of ChromeRelay commands (navigate → scrape → download) relies on the full output of each step being preserved. Compaction breaks this chain.

Concrete Example

In the Blackboard file organization session:

ChromeRelay returned HTML content from vuws.westernsydney.edu.au containing CDN download URLs for PDFs
RTK compacted the output, losing the URL patterns needed to extract direct download links
The workaround was to use eval to extract cookies/headers from the browser and construct curl commands manually — which shouldn't be necessary if the raw ChromeRelay output were preserved

Proposed Solution

Add a config option to bypass RTK for specific tools or skills:

json

{ "rtk": { "bypassTools": ["ChromeRelay", "ChromeRelayEval"], "bypassThreshold": 0 }}

Or a per-invocation flag:

bash

alma config set rtk.bypassTools '["ChromeRelay", "ChromeRelayEval"]'

Alternative approaches:

Per-tool threshold: Only compress if output exceeds a much larger threshold (e.g., 50KB) for browser-related tools
Smart bypass: Don't compress outputs that contain URLs, HTML, or structured data formats
Per-skill bypass: Allow skills to declare rtk: false in their SKILL.md

Also discussed in https://discord.com/channels/1454390052359503986/1454390459328761857/1484830135365537924

Why This Matters

RTK is great for saving tokens on verbose command output (e.g., ls -la, npm install). But for tools that return structured browser data, the compression destroys the very information the AI needs to act on. A bypass option would let users get the best of both worlds.

Related Config Context

Current RTK-related settings: