No, you should not use the claimed 5x savings or 95–97% hit rate as a budgeting guarantee; they are project targets or health indicators whose results depend on the model, cache, and session content.

• The README describes “5× fewer tokens”; its cache section defines 95–97% as a healthy-session prefix-cache hit rate and says compression itself costs no more than 2%.
• It also lists upstream cache TTL expiry, model switches, and bili bugs as causes of lower hit rates, so the figures are not an environment-independent SLA.
• The project supports derived child sessions inheriting the parent’s compressed context, which suits subagent collaboration, but information completeness still depends on summarization.
• Long sessions rely on incremental, layered, reversible compression; the project also explicitly cannot exceed the model or provider’s per-request context limit.

Use these figures to understand the design goals, not to calculate your bill or debugging quality directly. Complete logs, exact tool output, and non-summarizable state may not remain fully visible after compression.

Yes, because this is the project’s primary long-session use case, but OpenCode requires special handling to avoid competing compression layers.

• The README explicitly targets “small context windows (100K is enough)” and “month-long single sessions (billions of tokens)”.
• Compression is incremental, reversible, and layered: summaries are written in small ranges and can be decompressed on demand, which fits large refactors better than one full summary at the context limit.
• pi, omp, and OpenCode 1.x/2.x all have native plugin installation paths; the OpenCode installer also disables native auto-compaction.
• The README warns that enabling OpenCode’s native auto-compaction together with billion-context causes double compression.

It still cannot exceed the provider’s per-request context limit, and retention of exact refactoring constraints depends on summary quality.

Yes, because the README explicitly addresses user-level npm installation and port isolation across multiple lanes, although you still need to verify PATH and the active instance state.

• On Linux/macOS, the recommended --prefix=~/.local avoids sudo and EACCES caused by an old root-owned prefix; bili is installed under ~/.local/bin.
• Manual bili start owns port 8787, while native hooks and launcher lanes use a self-managed zone beginning at 18787; collisions increment automatically.
• During an upgrade restart, the proxy waits up to five seconds for the previous build to drain and reuses the same lane port when possible; it hops and logs only when the port is genuinely occupied.
• This fits parallel client use, but stale processes, a manually started 8787 service, or abnormal exits can still cause a client to connect to the wrong instance.

It addresses common installation and port-lifecycle problems, but does not remove the operational cost of proxy processes, logs, and upgrades.

It depends: choose billion-context when you need long sessions, stable cache prefixes, and reversible recovery; keep OpenCode’s native compression when it is sufficient and sessions do not approach the context limit.

• Native installation writes the plugin into OpenCode’s real configuration and disables compaction.auto, showing that the two compression mechanisms should not run together.
• The README requires a manual configuration backup before installation; removal stores configuration snapshots in .bili-bak.
• The project claims month-long single sessions and a 95–97% prefix-cache hit rate, but the result depends on model switches and upstream cache TTL.
• With a git checkout rather than the npm package, the README says OpenCode has no published entry and may not load through its native Npm.add path.

The deciding factor is therefore not the OpenCode version alone, but whether you need a long-lived, recoverable context lifecycle and can accept configuration changes.

No, not for direct deployment, unless enterprise security and compliance explicitly permit a local proxy that rewrites model requests and accept the licensing and session-storage implications.

• The project operates as a local proxy between the client and model provider, observing, rewriting, or routing requests; that may conflict with network isolation, endpoint controls, or sensitive-code policies.
• Codex integration usually involves starting a proxy, setting HTTPS_PROXY, or using bili codex, increasing enterprise networking and troubleshooting surface area.
• The project data marks the license as Other, and the README mentions attribution requirements beyond MIT; closed-source integration, redistribution, and commercial deployment cannot be reviewed as MIT-only.
• The proxy creates processes, ports, logs, configuration, and session files; the project insights call for review of request routing, log storage, and data compliance.

Claude and Codex may be technically integrable, but compliance prerequisites outweigh compression benefits here. The README provides no guarantee of enterprise auditing, data isolation, or zero-logging mode.

It depends: all three clients are covered, but the unified element is the context-management goal, not an identical installation or routing procedure.

• Claude can be installed through the marketplace and then configured with /billion-context:bili-setup; the Hermes installer copies a Python plugin into ~/.hermes/plugins/billion-context/.
• The launcher list includes bili codex, bili hermes, and bili claude, so each has a project-defined integration path.
• Codex cannot obtain complete proxy routing from plugin installation alone; the project insights state that it usually requires starting bili, setting HTTPS_PROXY, or using bili codex.
• Client configuration, authentication, session formats, and subagent behavior can differ, and the README does not promise feature equivalence.

You can standardize the operational entry point and compression layer, but not assume one plugin configuration covers all three clients; Codex proxy reachability is the critical deployment dependency.