learn-coding-agent: a documents-only study of Claude Code v2.1.88 that you cannot run and may not use commercially

The repository contains documents only, so there is nothing to install or run

The top level holds README.md, translated copies in README_CN.md, README_JA.md and README_KR.md, and a docs folder. That is the entire tree. There is no source code, no package manifest and no build script, so the repository cannot be installed, executed or imported.

The README introduces it as a learning and research repository focused on CLI agent architecture, compiled entirely from publicly available online references and discussions. Its particular focus is public information about the CLI agent claude-code. The practical effect is that you read it the way you would read an article series: the value is in the diagrams, tables and written analysis, and any code-like blocks are illustrations of structure rather than something you can copy into a project.

There is no install step. You get it by opening the repository on GitHub, or by cloning it, and reading the files. The English reports sit in the docs folder, laid out like this in the README:

docs/
├── en/                                        # English

Five reports under docs/ cover telemetry, codenames, undercover mode, remote control and a roadmap

The docs folder is organized as English reports numbered 01 to 05, and the README says they are also available in Japanese, Korean and Chinese. The reports are titled Telemetry and Privacy, Hidden Features and Codenames, Undercover Mode, Remote Control and Killswitches, and Future Roadmap. All are described as compiled from publicly available references and community discussions on Claude Code v2.1.88.

The version number is the first limit. Everything is tied to one build, v2.1.88. Anything a later Claude Code release changed will not show up here. A reader using a newer version should treat the reports as a historical snapshot.

The telemetry report makes strong claims that rest on second-hand sources

The summary table says there are two analytics sinks, first-party and Datadog, with an environment fingerprint, process metrics and a repo hash on every event. It says first-party logging has no opt-out exposed in the UI, and that setting OTEL_LOG_TOOL_DETAILS=1 enables full tool input capture.

These are the repository's reported findings, not independently checked facts. The repository states that it is built from public references and discussions, so the chain of evidence runs through other people's accounts. If privacy matters for your decision to use or avoid a tool, treat this table as a lead to verify against the vendor's own documentation and settings, and do not quote it as established.

Undercover mode and remote control are described as behaviors users cannot switch off

Two of the reports make the sharpest claims. On undercover mode, the README says official employees automatically enter it in public repositories, and that the model is instructed not to blow its cover, stripping AI attribution and writing commits as a human developer would. It adds that no force-off exists and that the mode raises transparency questions for open-source communities.

On remote control, it says the application polls a settings endpoint hourly, that dangerous changes show a blocking dialog where rejecting makes the app exit, and that six or more killswitches cover bypass permissions, fast mode, voice mode and the analytics sink. It also says feature flags can change a user's behavior without consent. Each statement is the repository's reading of public information. Anyone writing about these topics should seek confirmation from the vendor before repeating them.

Codenames and the roadmap are inferences about unreleased work, not announcements

The hidden features report lists animal codenames, including Capybara, Tengu, Fennec and Numbat, and says feature flags use random word pairs to obscure their purpose. It names hidden commands /btw and /stickers. The roadmap report says the Numbat codename is confirmed, that Opus 4.7 and Sonnet 4.8 are in development, and that KAIROS is a fully autonomous agent mode with tick heartbeats, push notifications and pull request subscriptions. Voice mode with push-to-talk is described as ready but gated, and 17 unreleased tools are said to be found.

None of that is a product announcement. These are claims about unreleased features drawn from public discussion, and the README gives no release dates. Planning your own work around them would be a mistake.

The architecture material is the durable part: the loop, the tool registry and the stats

The core loop is drawn as a cycle: the user's input goes into a messages array, the array goes to the Claude API, and if the stop_reason is tool_use the tools execute, the results are appended and the loop returns to messages. Otherwise the text is returned.

The directory reference names src files such as main.tsx, QueryEngine.ts, query.ts, Tool.ts, Task.ts, tools.ts, commands.ts, context.ts, cost-tracker.ts and setup.ts, with query.ts called the main agent loop and the largest file at about 785KB. The stats table gives about 1,884 TypeScript files, about 512,664 lines, about 40 built-in tools, about 80 slash commands and about 192 packages, with a Bun runtime compiled to a Node.js 18 or newer bundle. The table of contents also promises a section on 12 progressive harness mechanisms and a tool and permission system with sub-agents.

No license is stated, commercial use is prohibited, and the last push was 1 April 2026

GitHub reports the license as unknown, and the top-level files include no license file. In its place, the README carries a disclaimer: all content is for technical research, study and educational exchange, and commercial use is strictly prohibited. The same text invites copyright owners who believe the content infringes their rights to contact the owner for removal.

So reuse is limited to personal study. The repository has no releases, is not archived, and its last push was on 1 April 2026, so it should not be described as actively maintained. The README says more insights will be shared in the future.

Editorial conclusion

Read this repository if you want a structured, second-hand overview of how one popular CLI coding agent is organized: the query loop, the tool registry and the permission flow. Do not treat its telemetry, undercover mode or roadmap findings as verified facts, and do not reuse its content commercially, because the README prohibits that. Check each claim against primary sources, and expect nothing to update, since the last push was on 1 April 2026.

Frequently asked questions

Official sources

1. Issues
2. README
3. sanbuphy/learn-coding-agent on GitHub