pi-rtk-noemoji

Fork of pi-rtk by Matt Cowger. Identical functionality, all emoji removed from output. Plain text markers only ([OK], [ERROR], [WARN], PASS:, FAIL:, SKIP:, etc.).

Everything else — config, commands, techniques, rtk_configure tool — is unchanged from upstream.

Why

Emoji in tool output wastes tokens (multi-byte), renders inconsistently across terminals, and adds no information over plain text markers.

Installation

pi install npm:pi-rtk-noemoji

Or in ~/.pi/agent/settings.json:

{
  "packages": [
    "npm:pi-rtk-noemoji"
  ]
}

Configuration

Same as upstream. Create ~/.pi/agent/rtk-config.json:

{
  "enabled": true,
  "techniques": {
    "ansiStripping": true,
    "truncation": { "enabled": false },
    "sourceCodeFiltering": { "enabled": false },
    "smartTruncation": { "enabled": false },
    "testOutputAggregation": true,
    "buildOutputFiltering": true,
    "gitCompaction": false,
    "searchResultGrouping": false,
    "linterAggregation": true
  }
}

Recommended safe config above: only techniques that cannot break edit exact-match are enabled. Source code filtering is off — it strips comments/whitespace, causing the model to see different text than what's on disk.

Commands

• /rtk-stats — token savings statistics
• /rtk-on / /rtk-off — enable/disable
• /rtk-what — show current config
• /rtk-toggle-<technique> — toggle individual techniques

Output examples

Build success:

[OK] Build successful (12 units compiled)

Build errors:

[ERROR] 2 error(s):
error[E0308]: mismatched types
  --> src/main.rs:5:12

[WARN] 3 warning(s)

Test results:

Test Results:
   PASS: 42 passed
   FAIL: 2 failed
   SKIP: 1 skipped

   Failures:
   - FAIL test_something
     expected 1, got 2

Known issues fixed in this fork

Command detection was too broad (upstream bug)

Several is*Command functions used command.includes("keyword") to detect whether a bash command was a build/test/search/linter invocation. This caused false positives:

• isSearchCommand: includes("ag") matched "agent", "package", "manage". includes("find") matched "findall", paths containing "find". Any command accidentally classified as a search command had its output reformatted and truncated to 50 results.
• isLinterCommand: includes("black") matched "blacklist". includes("ruff") matched any command with "ruff" as a substring.
• isTestCommand: bare "test" in TEST_COMMANDS matched any command containing the word "test", replacing the entire output with a test summary. This silently destroyed Python script output whenever a line contained the word "ok" (counted as a passed test).

Fix: removed bare "test" from TEST_COMMANDS. Changed isSearchCommand and isLinterCommand to match command names only at command position (start of line or after |, &&, ;), not as substrings.

Source code filtering breaks edit tool

The sourceCodeFiltering technique strips comments and normalizes whitespace when reading files. The agent then sees modified text, but the edit tool requires exact match against the real file content. Edits fail silently or target the wrong text. Recommendation: keep sourceCodeFiltering disabled.

Truncation hides scan/data output

Default truncation.maxChars: 10000 (10KB) and smartTruncation.maxLines: 200 are too aggressive for projects that read large JSON files, scan results, or verbose script output. Findings get silently dropped. Recommendation: disable both for data-heavy projects.

Changes from upstream

All changes are in techniques/*.ts — emoji in generated output strings replaced with plain text. Regex detection patterns (for parsing test/build output) are untouched.

Upstream	This fork
✓	[OK]
❌	[ERROR]
⚠️	[WARN]
📋 Test Results:	Test Results:
✅ N passed	PASS: N passed
❌ N failed	FAIL: N failed
⏭️ N skipped	SKIP: N skipped
• failure	- failure
📄 file	> file
📌 branch	Branch: branch
🔍 N matches	N matches

License

MIT — same as upstream.

Credits

All credit to Matt Cowger for the original pi-rtk.