Research atlas

This appendix keeps discovery machinery separate from the main runtime narrative. Use it when starting from a raw constant, generated module record, prompt-catalog entry, byte offset, minified symbol, native module, or documentation gap.

The atlas is a triage layer, not proof. Promote a finding into the main internals sections only after a focused source read confirms behavior.

It contains three deliberately separate document types:

Type	Owner	Purpose
Source navigation	Bundle module map and generated `data/`	Locate candidate source regions and literal surfaces.
Coverage evidence	Full-system documentation coverage review, Disassembled string-surface review, and domain audit ledgers	Record which source-confirmed mechanisms have an owner and which evidence limits remain.
Documentation architecture	Documentation structure and duplication review	Record canonical ownership, merge/rename decisions, and intentional non-merges.

Source-anchor policy

This page is a research guide. Linked pages and generated artifacts carry concrete anchors.

Semantic alias	Minified anchor	Scope
Research atlas chapter	N/A — navigation page	Groups artifact/bytecode notes, source-anchor methodology, and future watchpoints.
Atlas/research pages	See linked pages and `claude-code-pkg/` artifacts	Concrete bundle anchors and generated inventories live in destination artifacts.

Research workflow

flowchart TD
    Question[Question or raw constant] --> Artifact[claude-code-pkg artifacts]
    Artifact --> Surface[cli.js strings / JS shims / optional temporary graph parse]
    Surface --> Source[Focused source read]
    Source --> Classify{Behavior confirmed?}
    Classify -->|yes| MainDoc[Patch or create internals section doc]
    Classify -->|partial| Backlog[Record opportunity]
    Classify -->|no| Noise[Mark as scan noise]
    MainDoc --> Validate[links and Markdown checks]

Primary reading order

Order	Page or artifact	Research question answered
1	Bundle module map from `cli.renamed.js`	Which Bun module loaders correspond to which Claude Code subsystems, and at what line ranges in the semantically renamed bundle?
2	Full-system documentation coverage review	Which source-confirmed lifecycles were promoted, rejected as false gaps, or left evidence-bounded?
3	Disassembled string-surface review	Which undocumented strings led to real executable mechanisms, aliases, noise, or native evidence limits?
4	Documentation structure and duplication review	Which pages are canonical owners, which duplicate was retired, and why do adjacent architecture/reference/lifecycle pages remain separate?
5	`data/cli-modules.json`	What loader, declaration, and export records did the structural extractor find?
6	`data/cli-module-themes.json`	Which semantic modules matched each populated runtime/vendor theme?
7	`data/prompt-catalog.json`	Which long literal prompts passed the domain/noise filter, with line/offset/hash provenance?

Mechanism-question audits

These ledgers record the question → source trace → documentation update → full reread loop for all 63 canonical mechanism pages in the main wiki. The count follows the final page manifests in the six ledgers; it is not the total Markdown-page count and is not inferred from filename suffixes. Each page was asked at most ten questions per round, and each domain stopped only after a complete pass produced zero new questions answerable from the retained 2.1.215 artifacts. The full-system coverage review first compared the six converged domains and cross-cutting inventories. The later structure review retired one duplicate command page without removing a mechanism, and the independent string-surface review subsequently promoted two executable owners that theme/page inventories had missed.

Domain	Audit ledger	Pages	Final result
Runtime and startup	Runtime mechanism question audit	10	Converged
Context and model loop	Context/model mechanism question audit	8	Converged
Tools, integrations, and security	Tools/security mechanism question audit	15	Converged
Sessions, persistence, and remote	Sessions/remote mechanism question audit	10	Converged + string follow-up
Operations and native support	Operations/native mechanism question audit	10	Converged + string follow-up
Agents and automation	Agents/automation mechanism question audit	10	Converged

“Converged” is artifact-bounded: server, unavailable native/bootstrap, runtime-only, and future-version questions remain explicitly listed as evidence limits rather than guessed answers.

Promotion rules

Treat raw string hits as leads, not behavioral proof.
Anchor every promoted claim with file path, approximate line/byte offset, exact string or symbol, and semantic meaning.
Keep bytecode-only or native-binary-only observations as research notes unless paired with readable JS or safe runtime behavior.
Update docs/SUMMARY.md, section README files, and adjacent cross-links whenever a finding becomes a main documentation page.

Created and maintained by Yingting Huang.