mirror of
https://github.com/EveryInc/compound-engineering-plugin.git
synced 2026-09-19 01:09:55 +08:00
Juan Gomez
6158fc0cb7
Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
75 lines
4.4 KiB
Markdown
75 lines
4.4 KiB
Markdown
---
|
|
title: Codex content transform regexes greedily matched URLs and email-like strings
|
|
date: 2026-07-24
|
|
category: integrations
|
|
module: src/utils/codex-content.ts
|
|
problem_type: logic_error
|
|
component: tooling
|
|
symptoms:
|
|
- "Slash-command rewrite corrupted URLs such as `https://example.com/path` by matching the second `/`"
|
|
- "Slash-command rewrite corrupted route-like values inside URL queries and fragments"
|
|
- "@-agent rewrite corrupted email-like strings such as `user@security-reviewer` by matching the `@` inside a word"
|
|
- "@-agent rewrite partially transformed longer mention-like identifiers"
|
|
root_cause: logic_error
|
|
resolution_type: code_fix
|
|
severity: medium
|
|
tags: [codex, converter, regex, slash-commands, agent-mentions, url-handling, word-boundary]
|
|
related_components: [codex converter, transformContentForCodex]
|
|
---
|
|
|
|
# Codex content transform regexes greedily matched URLs and email-like strings
|
|
|
|
## Problem
|
|
|
|
`transformContentForCodex()` in `src/utils/codex-content.ts` rewrites Claude-style `Task` calls, slash commands, backticked agent names, and `@`-references into Codex-style phrasing. Two of its regexes were too greedy and matched the command/mention markers when they appeared inside unrelated text, producing incorrect output.
|
|
|
|
## Symptoms
|
|
|
|
- URLs such as `https://example.com/path` were rewritten as `https:prompts:example.com/path` because the slash-command regex matched the second `/`.
|
|
- URL queries and fragments such as `https://example.com?next=/unknown-cmd` and `https://example.com/#/ce-plan` had their embedded routes rewritten as commands.
|
|
- Email-like strings such as `user@security-reviewer` were rewritten as `usercustom agent \`security-reviewer\`` because the `@`-agent regex matched the `@` inside a word.
|
|
- Longer handles such as `@security-reviewer_helper` were partially rewritten because `_` was not treated as a continuation character.
|
|
|
|
## What Didn't Work
|
|
|
|
Guarding inside the slash-command replacement callback with `commandName.includes("/")` did not prevent the match; the URL's second slash still entered the callback. Removing the guard alone would not fix the underlying match, and adding a second post-match filter would still have rewritten the URL before the guard could reject it.
|
|
|
|
For `@`-references, there was no boundary check at all, so any `@` followed by a recognized suffix was transformed regardless of context.
|
|
|
|
## Solution
|
|
|
|
Handle each exclusion at its owning layer: preserve HTTP(S) spans once around all free-form rewrites, and keep mention-token boundaries in the `@` regex.
|
|
|
|
`transformOutsideHttpUrls()` splits the content into URL and non-URL spans. Backticked-agent, slash-command, path, and `@`-agent rewrites run only on the non-URL spans:
|
|
|
|
```typescript
|
|
result = transformOutsideHttpUrls(result, (segment) => {
|
|
// Apply the free-form Codex rewrites to segment.
|
|
return transformed
|
|
})
|
|
```
|
|
|
|
This handles ordinary paths, query/fragment variants, and agent-like URL tokens without enumerating every punctuation character that may precede an embedded route or mention.
|
|
|
|
`@`-agent references now require token boundaries before the `@` and after the recognized name:
|
|
|
|
```typescript
|
|
const agentRefPattern = /(?<!\w)@([a-z][a-z0-9-]*-(?:agent|reviewer|researcher|analyst|specialist|oracle|sentinel|guardian|strategist))(?![\w-])/gi
|
|
```
|
|
|
|
## Why This Works
|
|
|
|
The shared wrapper copies each complete non-whitespace HTTP(S) token unchanged and applies every free-form rewrite only to the text between URLs. Commands, paths, and agent mentions elsewhere in the same string still transform normally.
|
|
|
|
A `@` that follows a word character is part of an email handle, username, or other compound token, not a stand-alone agent mention. The `(?<!\w)` lookbehind ensures only boundary `@` symbols are considered, while `(?![\w-])` prevents partial matches inside longer identifiers.
|
|
|
|
## Prevention
|
|
|
|
- Put token-local exclusions in the pattern, but protect shared lexical regions such as URLs once at the layer that owns all affected rewrites.
|
|
- Test common embedding cases early: URLs, markdown links, HTML attributes, quoted strings, email addresses, and parentheses.
|
|
- When one protected context affects multiple sequential transforms, test that every stage preserves it and that equivalent syntax outside the context still transforms.
|
|
|
|
## Related
|
|
|
|
- `docs/solutions/codex-skill-prompt-entrypoints.md` — broader Codex converter architecture and the rule to preserve unknown slash references.
|