conversation-compiler

name	conversation-compiler
description	VCC (View-oriented Conversation Compiler) documentation. Compile Claude Code JSONL logs into adaptive views.

VCC — View-oriented Conversation Compiler

Compile Claude Code JSONL logs into adaptive views for reading, searching, and context recovery.

Usage

python "path/to/VCC.py" <input.jsonl ...> [options]

Option	Description
-o <dir>	Output directory (default: same as input)
-t <N>	Token truncation limit (default 128)
-tu <N>	User message token limit (default 256)
--grep <pattern>	Regex search pattern (Python re — use `a

This tool also supports multi-file processing:

cd "/path/to/target/folder" && python "path/to/VCC.py" *.jsonl --grep "keyword"

Output Files

File	When	Description
.txt	Always	Full transcript, lossless
.min.txt	Always	Brief overview
.view.txt	--grep only	Search-focused view
stdout	--grep only	Search results with block-level line range references

Rules

• Forward slashes only in bash commands. Backslashes in double quotes are escape characters — \U, \l, \. get silently mangled. Use "C:/Users/..." not "C:\Users\...". This applies to ALL bash commands (not just this skill), on ALL platforms (Windows and Linux alike).
• Do NOT use -o unless the user explicitly requests an output directory. By default, compiled files are written next to the input JSONL — this is the intended behavior.
• Do NOT clean up compiled output files after use. Leave them in place unless the user explicitly asks to clean up.

Workflow

1. Compile

python "path/to/VCC.py" "path/to/conversation.jsonl"

Produces .txt + .min.txt next to the input file. Long conversations are automatically split into numbered chunks. The console output lists every produced file with line/word counts — read it to understand the conversation's size and structure before proceeding:

conversation_1.txt  (2808 lines, 34545 words)
conversation_2.txt  (9622 lines, 116015 words)
conversation_3.txt  (3242 lines, 37946 words)
conversation_1.min.txt  (101 lines, 1572 words)
conversation_2.min.txt  (811 lines, 9504 words)
conversation_3.min.txt  (205 lines, 3340 words)

Chunks are numbered chronologically — higher numbers are more recent. (During /recall, the last chunk is the current conversation; the second-to-last is likely where the previous session's context ends.) Use this to gauge how many chunks exist and how large each is, so you can target your .min.txt reads and --grep searches to the right chunks.

2. Read .min.txt

The .min.txt is a scannable outline: user/assistant text is kept (truncated), tool calls collapse to one-line summaries with line references back to .txt.

[user]

Take a look!

[assistant]

Let me take a look.

* Read "code.py" (a_1.txt:19-21,24-34)

Each * line shows the tool name, key parameter, and two .txt line ranges: the call (19-21) and the result (24-34). Read these ranges in .txt for full details.

3. Search with --grep

Always use --grep, never system grep nor your embedded Grep tool. This script's --grep returns important block-level line RANGES that no other grep tools can provide. The output paths are relative to CWD, so cd close to the target first to keep outputs short and save tokens.

cd "/path/to/target/folder" && python "path/to/VCC.py" "path/to/conversation.jsonl" --grep "keyword"

Stdout example (# prefix = shortened filename)：

(#752a87.txt:L67-L81) [assistant]      →  .txt lines 67-81
  77: one matched line ...             →  .txt line 77

Optionally read .view.txt for focused search view.

4. Jump to .txt

All line references point to .txt. Read the referenced range for full context. Remember to read a bit more lines before and after the referenced range to get more context.

SKILLs

/readchat — Read a specific conversation log

Trigger: The user wants to read a specific JSONL conversation log.

Action: Follow the Workflow above with the user-specified filenames or requests.

/searchchat — Search across conversation logs

Trigger: The user wants to search across conversation logs in ~/.claude/projects/.

Action:

1. ls ~/.claude/projects/ — browse project directories, narrow the search area.
2. cd ~/.claude/projects/<project> && python "absolute/path/to/VCC.py" *.jsonl --grep "keyword" — search top-level conversations first.
3. If no results: cd ~/.claude/projects/<project> && python "absolute/path/to/VCC.py" **/*.jsonl --grep "keyword" — expand to subagents.

Critical: Always cd into the target directory first, then use VCC globs — a single cd && python VCC.py <glob> --grep is the correct pattern. Without cd, grep output contains full absolute paths instead of short relative paths, wasting tokens. All content search on JSONL must use VCC's --grep, never system grep or the embedded Grep tool — VCC's --grep returns block-level line ranges with role tags that no other grep can provide.

/recall — Recover context from a previous conversation

Trigger: The conversation opens with a context-continuation summary ("This session is being continued from a previous conversation..."), or the user says /recall.

Action:

1. Find JSONL filename — The continuation message ends with read the full transcript at: <path>.jsonl. Find that JSONL filename. If no JSONL path is present (e.g. the user invoked /recall manually without a continuation header), use the /searchchat method to locate the JSONL before proceeding to steps 2–3.

2. Follow VCC Workflow — Follow the full Workflow above (compile → read .min.txt → --grep → jump to .txt) with the JSONL file found in step 1. The summary is lossy — only the original JSONL is authoritative.

3. Verify against current state — After recovering the conversation, cross-check with reality:

   • REALLY read files — Read all files referenced in the conversation with your Read tool. Pre-loaded content from system-reminders or prior conversation turns does NOT count — it may be truncated, stale, or lossy. You must issue a fresh Read call for each file, even if you believe you already have its content. 99.97% of the time, the user has externally modified nearly every file between sessions. If you skip reading and miss their external changes, your failure rate is effectively 100%.
   • Key details — Identify specific values, paths, configs, and logic mentioned in the conversation. Compare against the actual files to catch drift or errors.
   • Understand the journey — Trace the user's intent, decision sequence, and direction changes. Understand not just what was done, but how and why they got there.