Coding agents do not leave behind a neat database of isolated facts. They write files: memory summaries, rollout notes, repository instructions, rules, and documentation. Memory atoms are useful for retrieval, but they are a poor default browsing surface because a person usually needs the surrounding file to understand what a passage means.
Docmancer 0.7.0 adds a file-first terminal explorer. Run two commands:
docmancer setup
docmancer
The CLI and MCP still use memory atoms where precise recall and curation need them. The human interface now starts from complete indexed source files.
Four views of the local corpus
The top row separates four kinds of work:
- Memory shows complete agent-memory files.
- Instructions & Rules keeps user-authored guidance separate from agent memory.
- Docs expands each documentation source into pages and sections.
- Security runs the local audit automatically and shows masked findings.
Memory and instruction files are sorted by their latest update and paginated at 50 files per page. Scope, harness, updated-time, and project filters apply to the complete corpus before pagination. A selected project includes its project and team sources plus global context.
Selecting a row updates the inspector without opening a modal. The right pane lazy-loads the complete privacy-cleaned indexed text with line numbers, including sources larger than 500 KB. Press Enter or V when you want the same viewer full-screen.
Search passages without losing the file
Text search still uses memory atoms under the hood, because passages are the right unit for ranking. Results are grouped by source file, with the best excerpt and the number of matching passages shown once per file. Selecting a result opens the complete file at the matching line range, and [ or ] moves between matches.
Passage-level exclusion and promotion remain available in search mode. Whole harvested source files cannot be forgotten or promoted accidentally.
Readable provenance for generated Codex files
Codex rollout summaries often have timestamps and generated slugs that are useful to software but hard to scan. The TUI now renders those sources as human-readable titles with compact dates and shortened home paths. The exact source path remains attached, so Open original still targets the real file.
Hooks and audit are part of the main loop
The terminal interface helps a person understand the corpus. Recall hooks make that corpus useful across agents:
docmancer install claude-code --hooks
docmancer install codex --hooks
Hooks query the existing local index at session start and before matching prompts, then inject only relevant source-backed memory. They are bounded, fail open, stay silent on weak matches, and do not call a hosted provider.
The Security tab runs the same read-only audit as docmancer memory audit. Likely secrets are masked, and the report also covers stale index state, duplicates, oversized sources, and sources that produce no usable passages.
Version 0.7.0 keeps Docmancer's local CLI and MCP contracts compatible while giving the indexed corpus a human interface that matches how the underlying knowledge was actually written.