MCode Continuous Context, Browser & Local Decision Enhancement Pack

Repository: keys-mcode-continuous-context-browser-decision-enhancement-pack

A pack of validated enhancements for the MiniMax Code (MCode) CLI, built and battle-tested on a 4× NVIDIA DGX Spark (GB10, arm64) running a local GLM-5.3-EXL3 3.0 bpw (abliterated) model through vLLM — no cloud API, $0 token cost.

What's in the pack

Component	What it does	Status
patches/0001	Interactive scrollbar for the fullscreen transcript viewport (three-column grab target, thumb-centering drags)	Validated on fork; upstream issue #217
patches/0002	Opt-in status-line context gauge (Context ▕██████░░▏ 77% left) with thresholds + minimal fallback	Validated on fork; upstream issue #216
patches/0003	Continuous context renewal after 25% utilization, recurring compaction trigger, and regression coverage	Validated locally: 82 tests pass
compaction-fix/	Root cause + repair of the INVALID_CHECKPOINT auto-compaction failures on local vLLM (KV dual-tenancy), now in never-reset mode: 64 K limit → trigger ~48 K → worst-case pair ~104 K vs the 257 K KV cache, enforced on every launch by the local mcode wrapper, plus an interval keeper.sh watchdog	Enforcement + math verified live 2026-09-20; limit is cached per session — restart MCode to activate; first post-restart successful auto-compaction pending
jev-ultrafast/	Structured browser agent on the verified arm64 CDP lane — local djev-spark policy + local GLM-5.3-EXL3 text leg; Syntra client included for adaptive routing/feedback	djev-spark-compatible /v1/systemone wiring added; cloud TypeSafe is opt-in
Syntra + djev-spark	Local decision infrastructure and adaptive policy credits used by the browser/agent stack	Checked out for local integration; model serving remains separately deployable
playwright-mcp/	Microsoft Playwright MCP as MCode MCP tools (~/.minimax/mcp.json / .mcp.json), arm64 browser notes, security gates	Documented; ready to enable
browser-cdp/	The measured CDP lane for browser-harness on arm64: Playwright Chromium 153 headless + --no-sandbox (AppArmor makes it mandatory) + BU_CDP_URL attach; snap Chromium ruled out — confinement blocks CDP entirely	Verified on this machine 2026-09-20; smoke-test.sh PASS
vision-tools/	Eyes for the text-only local lane: analyze_image (door 2, files on disk) + vision_proxy (door 1, chat attachments); local VLM only, the brain sees words never pixels; pattern from tonyd2wild/dsh (MIT)	Ported; failure paths verified live (endpoint-naming errors, degrade-not-throw, byte-identical pass-through); no live VLM on this box yet — validate before use
performance/	Real usage ledger (the runtime's own sqlite accounting) + timed probes + a methodology for validating agent performance claims	Live data + reproducible collector
container/ + Dockerfile	Pre-built arm64 GHCR image: official @minimax-ai/code 0.4.12 overlaid with the patched bundles + arm64 Playwright Chromium + pack tooling	Built & smoke-tested on GitHub's arm64 runner → ghcr.io/drowzeys/keys-mcode-continuous-context-browser-decision-enhancement-pack

Why this pack exists

The TUI patches were built as upstream contributions. The pack also carries the continuous-renewal patch and a maintained reapply hook, so an official MCode update can be rebuilt with the enhancements automatically. This pack makes all of the work — plus the local-stack repairs that upstream will never need (arm64 browsers, local-vLLM compaction, real usage data) — installable in one place, independent of upstream timing.

One-shot apply recipe

MCODE_SOURCE_TREE=/path/to/minimax-code ./install-one-shot.sh

The installer applies all compatible MCode patches, installs the durable launch-time compaction enforcer, configures the 64K local-vLLM ceiling, and installs ~/.local/bin/mcode-enhanced. Launching through that wrapper checks the active MCode release and automatically reapplies/rebuilds the pack after an update. It preserves user config and does not install credentials.

For a nonstandard installation, set the official launcher and source paths:

MCODE_REAL_BIN=/path/to/official/mcode \
MCODE_SOURCE_TREE=/path/to/minimax-code \
  ~/.local/bin/mcode-enhanced

Manual steps, when needed:

# TUI + continuous context patches (against the MiniMax Code source tree, MIT)
git am patches/0001-tui-transcript-scrollbar.patch
git am patches/0002-tui-context-meter-status-item.patch
git apply patches/0003-continuous-context-renewal-25-percent.patch

# Compaction repair + never-reset mode for local vLLM serving
python3 compaction-fix/set-glm53-context-limit.py --limit 64000
bash compaction-fix/check-status.sh

# Playwright MCP (see playwright-mcp/README.md)
npx playwright install chromium   # arm64 real-Chromium path

# Verified browser CDP lane (see browser-cdp/README.md)
./browser-cdp/smoke-test.sh          # launches + verifies end-to-end, PASS/FAIL

# Vision tools for the text-only lane (needs a local VLM endpoint; see
#    vision-tools/README.md — read its Status section first)
./vision-tools/backends/ollama.sh    # then: python3 vision-tools/analyze_image.py <img>

# Performance ledger, reproduced from your own runtime
python3 performance/collect-usage.py --day $(date +%F)

# Or skip all of it — pull the pre-built image (arm64)
docker pull ghcr.io/drowzeys/keys-mcode-continuous-context-browser-decision-enhancement-pack:latest

Upstream MCode itself: MiniMax-AI/minimax-code (source, MIT) / @minimax-ai/code on npm (official releases). The pack's patches and image sit on top of 0.4.12.

Highlights

• The compaction bug that kills long sessions on local vLLM is solved. The compactor's checkpoint request (full history, different system prompt → zero prefix-cache sharing) plus the resident session history exceeded the 257 K-token KV cache — structurally guaranteed to fail at a 200 K configured limit. Two-size adversarial probes proved it; the fix (128 K limit → trigger at 96 K → worst-case pair 200 K < 257 K) is applied and verified. Full analysis in compaction-fix/.
• Real numbers, not claims. 266 calls / 23.8 M input / 177 K output tokens across one day of heavy agent work, straight from the runtime's own ledger — including the 7.7-hour session that died at 7% context before the fix, and the ~606 tok/s prefill measurement for 136 K-token prompts.
• arm64 honesty. No Google Chrome on arm64 Linux; the pack documents the snap-Chromium + Playwright-arm64 reality and wires Jev Ultrafast and Playwright MCP accordingly. Measured further on 2026-09-20: snap Chromium is a CDP dead end (browser-harness --doctor: Snap confinement prevents CDP binding), so the browser lane is now settled by measurement — Playwright's arm64 Chromium, headless, --no-sandbox (mandatory under Ubuntu 24.04+ AppArmor), attached via BU_CDP_URL (browser-cdp/, verified end-to-end).

See ATTRIBUTION.md for component provenance and licenses.

Credits and thanks

With sincere thanks to the original creators and maintainers whose work made this pack possible: the MiniMax Code team, pi-mono, Syntra, djev-spark, Jev Ultrafast, TypeSafe, Microsoft Playwright, Playwright MCP, browser-harness, DeepSeek-Harness-Browser, DeepSeek-Harness-Vision-Tools, the vLLM project, and the maintainers of Node.js, pnpm, Docker, GitHub Actions, and GHCR. Their licenses, source links, and the specific methods used are documented in ATTRIBUTION.md.