comprehend（comprehend 搜索）

Comprehend

You are working in an unfamiliar codebase. Before you touch anything, understand what you're dealing with.

Break the problem into pieces. Use a persistent REPL to store everything you learn. Fan out subagents to do the reading. Keep your own context window small and available for the actual work — edits, debugging, reasoning.

That is the entire method. What follows is how to do it.

When to Use This

Use this workflow whenever you are making changes in code you don't already understand. This includes:

• Fixing a bug in a repository you haven't seen before
• Adding a feature that touches multiple files or modules
• Refactoring code whose structure you need to map out first
• Any task where you'd otherwise spend your first several turns just reading

The default instinct is to start reading files immediately and jump to a fix. That works for trivial changes. For anything else, it leads to patches that miss edge cases, break related code, or over-engineer because you didn't see the simpler path. Measure first.

Skip this workflow only when the change is truly trivial — a one-line typo fix, a config value change, something where understanding the surrounding code is unnecessary.

Prerequisites

This skill requires python and nohup. Before doing anything else, verify they are available:

python --version && nohup --version > /dev/null 2>&1 && echo "ok"

Note that the Python executable may be called python3, or your user may have set up a specific tool for running Python which they prefer you to use.

If you can't get Python to print ok, stop here. Tell the user which command is missing. On Windows, both are available in Git Bash. On macOS and Linux, both are standard.

Script Paths

This skill bundles scripts in its scripts/ directory. Before using them, resolve the absolute path based on where you loaded this SKILL.md from. For example, if you loaded this file from /home/user/project/skills/comprehend/SKILL.md, then the scripts are at /home/user/project/skills/comprehend/scripts/.

Throughout this document, SCRIPTS refers to that resolved path. In all bash commands, substitute the actual absolute path.

The REPL

Every session starts by generating a unique address and launching the server:

# Generate a session-unique address (prevents collisions between
# simultaneous sessions on the same machine)
REPL_ADDR=$(python SCRIPTS/repl_server.py --make-addr)
nohup python SCRIPTS/repl_server.py "$REPL_ADDR" > /dev/null 2>&1 &

The server must outlive the shell that starts it. Use nohup and shell backgrounding (&) as shown above. Do not use the Bash tool's run_in_background parameter — it may kill the server when the task "completes." On Windows, the server automatically uses TCP on localhost instead of Unix sockets. No code changes needed — the interface is identical.

Throughout this document, REPL_ADDR refers to the session-unique address returned by --make-addr. In all bash commands, substitute the actual path. Each session must use its own address.

This launches a persistent Python REPL. Variables, imports, and definitions survive across calls — not just during comprehension, but for the entire session. The REPL is your memory: use it instead of reading files into your context window.

This is the key tradeoff. The upfront cost is a few extra turns to launch the server and measure the codebase. The payoff: you build a queryable knowledge base that persists for the entire session. When you need to check something later — while debugging, while writing tests, while responding to follow-up questions — you query the REPL instead of re-reading source files. Every lookup costs one small Bash call instead of consuming context window.

Always use a heredoc to send code to the REPL. Never pass code as a positional command-line argument — it will break on quotes, braces, or multi-line input. The only exceptions are --vars, --shutdown, and --file.

# Run code (state persists between calls)
python SCRIPTS/repl_client.py REPL_ADDR <<'PYEOF'
greeting_message = "hello"
PYEOF

python SCRIPTS/repl_client.py REPL_ADDR <<'PYEOF'
print(greeting_message)
PYEOF

# See all stored variables
python SCRIPTS/repl_client.py REPL_ADDR --vars

The quoted delimiter (<<'PYEOF') passes all characters through to Python unchanged — single quotes, double quotes, backslashes, parentheses, braces, everything.

Large results: Heredocs can break on very large blocks of code (100+ lines). When storing large data in _comprehend_results, use the Write tool to write the Python code to a temporary file, then use --file:

# First: use the Write tool to create /tmp/comprehend_data.py
# with the Python code (assignments, dict literals, etc.)
# Then:
python SCRIPTS/repl_client.py REPL_ADDR --file /tmp/comprehend_data.py

This bypasses shell quoting entirely. The Write tool writes the file directly, and --file reads it directly. Use this for any REPL code that is too large for a heredoc.

Windows paths in heredocs: Always use forward slashes in Python code inside heredocs ("C:/Users/..." not "C:\\Users\\..."). Python accepts forward slashes on all platforms. This avoids backslash-as-line-continuation confusion.

CRITICAL: AFTER STARTING THE REPL SERVER, YOU MUST NOT USE ANY OTHER TOOLS FOR FILE OPERATIONS.

The following actions are FORBIDDEN once the REPL is running:

• Using cat, head, tail, grep, find, wc, ls or any shell command to read files
• Using Read_file tool to read source files directly into your context
• Using Glob tool to enumerate files
• Using Grep tool to search file contents
• Any operation that brings file content into your context window

If you catch yourself about to use a shell command or file tool to access file content, STOP. Use the REPL client instead.

Use the REPL for everything! Finding files, searching content, reading source, storing results — all of it. Every fact you discover goes into a variable where it accumulates instead of evaporating.

python SCRIPTS/repl_client.py REPL_ADDR <<'PYEOF'
import glob, os, re

# Find files (instead of Glob tool)
project_source_files = glob.glob("/path/to/project/**/*.py", recursive=True)
project_source_files = [f for f in project_source_files if "/.git/" not in f.replace("\\", "/")]

# Measure them (instead of wc)
file_size_by_path = {f: os.path.getsize(f) for f in project_source_files}
total_source_bytes = sum(file_size_by_path.values())

# Search content (instead of Grep tool)
function_definition_matches = {}
for filepath in project_source_files:
    with open(filepath) as fh:
        for line_number, line_text in enumerate(fh, 1):
            if re.search(r"def process_", line_text):
                function_definition_matches.setdefault(filepath, []).append(
                    (line_number, line_text.strip()))

# Everything persists: project_source_files, file_size_by_path,
# total_source_bytes, function_definition_matches
print(f"{len(project_source_files)} files, {total_source_bytes/1024:.0f} KB, "
      f"{len(function_definition_matches)} files with matches")
PYEOF

The Results Dict

All subagent findings go into one well-known dict: _comprehend_results.

The REPL server initializes this dict automatically on startup. Do not re-initialize it — that would wipe results from other subagents. Just write to it:

python SCRIPTS/repl_client.py REPL_ADDR <<'PYEOF'
_comprehend_results["auth_module_analysis"] = {"functions": [...], "issues": [...]}
PYEOF

The parent assigns every subagent a unique key before launching it. Subagents must never choose their own keys — the parent is the only one that sees all keys in use and can guarantee uniqueness. A subagent writes only to its assigned key; it never reads or writes other subagents' keys.

Keys should be descriptive: 'auth_module_function_signatures', not 'chunk1'. For deeper nesting, use sub-keys within the assigned key:

python SCRIPTS/repl_client.py REPL_ADDR <<'PYEOF'
_comprehend_results["auth_module"] = {
    "function_signatures": [...],
    "import_map": {...},
    "issues": [...]
}
PYEOF

The parent reads from _comprehend_results[key]. The underscore prefix and specific name avoid collisions with user or project variables.

The Workflow

1. Measure

Before reading any files, measure everything relevant to your task. Use the REPL (as above) or the bundled script:

python SCRIPTS/chunk_text.py info <file>

Measure broadly — source, tests, docs, config. The most common failure is measuring only the core source, classifying it as small, then also reading tests and docs and blowing past the limit.

2. Choose a strategy

3. Plan

Based on what you measured, decide what to read and in what order. For a bug fix, focus on the files in the stack trace plus their immediate dependencies. For a feature, map the module boundaries first. State your plan before reading anything.

4. Execute

Fan out subagents. Each writes to _comprehend_results[key]. You read the results back. Details are in the Fan-Out Patterns section below.

5. Iterate

If the aggregated answer has gaps, target those specific areas for deeper analysis. The REPL still holds everything from the first pass.

6. Make Your Changes

Now that you understand the code, make your edits. The REPL remains available — query _comprehend_results whenever you need to check something instead of re-reading source files. This keeps your context window small and available for reasoning about the actual changes.

python SCRIPTS/repl_client.py REPL_ADDR <<'PYEOF'
# Check a detail without re-reading source
print(_comprehend_results["core_library"]["function_signatures"])
PYEOF

7. Shut down the REPL

When your task is complete, shut down the REPL:

python SCRIPTS/repl_client.py REPL_ADDR --shutdown

The nohup server runs until explicitly stopped. Shut it down when done to avoid leaving an orphan process.

The 50KB Rule

Never read more than 50KB of source into your main context window. Everything above that limit must go through subagents that write findings to the REPL. The cost of a subagent is latency. The cost of context exhaustion is the entire rest of the session.

Watch for the trap: you measure 30KB of core source (under threshold!), then also read tests, config, and docs — now you're at 120KB and your context window is shot. Measure the total. All of it.

Fan-Out Patterns

All patterns follow the same contract:

1. Parent starts the REPL server (once per session). _comprehend_results is auto-initialized. Never re-initialize it.
2. Parent assigns a unique key to each subagent in the prompt. Subagents never pick their own keys.
3. Subagent writes findings only to its assigned _comprehend_results[key]. It may create sub-keys within that key freely, but must not touch any other top-level key.
4. Subagent reports back the key it wrote and a summary of what's in it.
5. Parent reads _comprehend_results[key] from the REPL.

Step 4 is critical. The Task tool returns a text message to the parent — that message is the *only* way the parent learns what the subagent stored. Every subagent prompt must end with an instruction to report what was written.

Direct Query

For simple one-shot tasks (summarize, classify, extract a fact). The subagent's return value is used directly — no REPL variable needed.

Task(subagent_type="Explore",
     prompt="Summarize the key functions in this file: <chunk>")

Recursive Query

For sub-problems needing multi-step reasoning or tool access. The subagent stores results in _comprehend_results under a descriptive key.

Parent launches subagent:

Task(subagent_type="general-purpose",
     prompt="Use the REPL at REPL_ADDR. Read and analyze these modules.
     Store your findings in _comprehend_results['auth_module_analysis'] as a
     dict with keys:
       'function_signatures' — list of all public function signatures
       'import_dependency_map' — dict mapping each file to its imports
       'identified_concerns' — list of architectural or correctness issues

     Files: src/auth.py, src/models.py, src/tokens.py

     When done, reply with: the key you wrote to in _comprehend_results,
     what sub-keys you stored, and a one-line summary of each.")

Parent receives a message like: "Wrote to _comprehend_results['auth_module_analysis'] with keys: 'function_signatures' (12 public functions), 'import_dependency_map' (3 files mapped), 'identified_concerns' (2 issues: circular import between auth.py and models.py, unused import in tokens.py)."

Parent reads back:

python SCRIPTS/repl_client.py REPL_ADDR <<'PYEOF'
auth_data = _comprehend_results["auth_module_analysis"]
print("Functions:", auth_data["function_signatures"])
print("Concerns:", auth_data["identified_concerns"])
PYEOF

Batched Parallel Query

For independent chunks that can run concurrently. Issue all Task calls in a single message.

Parent launches all subagents at once:

Task(prompt="Use REPL at REPL_ADDR. Analyze this log segment for errors.
     Store in _comprehend_results['log_segment_hours_00_to_06'] as a dict with
     keys 'error_summary' and 'critical_error_list'.
     Segment: <chunk1>

     When done, reply with: the _comprehend_results key you wrote,
     how many errors found, and one sentence summarizing the most severe.")

Task(prompt="Use REPL at REPL_ADDR. Analyze this log segment for errors.
     Store in _comprehend_results['log_segment_hours_06_to_12'] as a dict with
     keys 'error_summary' and 'critical_error_list'.
     Segment: <chunk2>

     When done, reply with: the _comprehend_results key you wrote,
     how many errors found, and one sentence summarizing the most severe.")

Task(prompt="Use REPL at REPL_ADDR. Analyze this log segment for errors.
     Store in _comprehend_results['log_segment_hours_12_to_18'] as a dict with
     keys 'error_summary' and 'critical_error_list'.
     Segment: <chunk3>

     When done, reply with: the _comprehend_results key you wrote,
     how many errors found, and one sentence summarizing the most severe.")

Parent receives three messages confirming what each wrote.

Parent reads accumulated results:

python SCRIPTS/repl_client.py REPL_ADDR <<'PYEOF'
for segment_key, segment_data in sorted(_comprehend_results.items()):
    if segment_key.startswith("log_segment_"):
        print(f"{segment_key}: {segment_data['error_summary']}")
        for critical_error in segment_data.get('critical_error_list', []):
            print(f"  - {critical_error}")
PYEOF

Chunking

For large single files, use the bundled script to split at natural boundaries:

python SCRIPTS/chunk_text.py info large_file.txt      # measure
python SCRIPTS/chunk_text.py boundaries source.py      # find split points
python SCRIPTS/chunk_text.py chunk large_file.txt --size 80000 --overlap 200  # split

For structured files (code, markdown), prefer splitting at functions, classes, or section headers rather than arbitrary character boundaries.

References

• references/comprehension-patterns.md — Five worked examples.
• references/mapping-table.md — RLM primitive to agent tool mapping.
• references/rlm-system-prompt.md — Theoretical foundation.