agent failed to dedupe string catalog collapsing distinct keys with same english text

Fix the agent collapsing distinct keys that share English text

TL;DR

The agent dedupes by English text, but different keys with the same wording are different strings, so dedupe by key instead. Same English in two contexts often needs different translations. Turn off text-based deduping, keep key-based identity, and restore the collapsed keys.

The error

Agent deduped 30 keys into 12 because the English matched
"Save" as a button and "Save" as a menu item became one key, translations broke

Fix it

Step 1: Find the collapsed keys

node -e "console.log('diff the catalog against the previous version, look for keys that vanished while their text survives elsewhere')"

Expected: You see which keys were merged.

Step 2: Disable text-based deduping in the agent

grep -n "dedupe\|deduplicate" scripts/agent-config.js | head -5

Expected: You find the dedupe setting to switch to key-based.

Step 3: Restore the collapsed keys

node -e "console.log('re-add each removed key with its own entry, even if the English duplicates another')"

Expected: Every key is back as its own entry.

Step 4: Add context notes so translators disambiguate

node -e "console.log('add a note to each duplicate-English key naming its UI location')"

Expected: Translators can tell the contexts apart.

When to use this

• An agent merged distinct keys with identical English text
• Translations broke because contexts were collapsed

When NOT to use this

• The keys are truly duplicates in the same context, then merging is fine
• English differs but translations collide, that is a translator issue

Tool and version compatibility

• Any agent or tool with catalog deduping
• Key-based identity for catalog entries

Variant phrasings

dedupe by text plus context

Better than text alone, but key-based is still the safe rule. Context matching is fuzzy.

intentional shared keys across the app

Then share one key deliberately in code. Do not rely on the agent to merge them.

Why it happens

Dedupe-by-text assumes identical source means identical meaning, which is false across UI contexts. A verb here and a noun there share spelling but need different translations, and merging them forces one translation onto both.

Edge cases

• Add a CI check that fails when two keys merge unexpectedly
• Some TMS platforms dedupe on import too, check their settings as well
• Document which keys intentionally share text so future agents do not re-merge

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_nedJZTNWGFrMoKnBO13vgg