Review queue
Merge reviews, low-confidence translations, audit findings and suggested rules, built for agents to work through.
The review queue holds every decision that needs judgment, as small self-contained tasks. Each task carries a payload with everything needed to decide it and the list of actions it accepts. Resolving a task applies the action through the same ops a person would use (split, merge, set a translation, add a rule), so an agent resolving tasks is exactly as safe as a person doing it by hand.
Task kinds
| Kind | Actions | When it appears |
|---|---|---|
| merge_review | keep_merged · split | Keys that share source text were joined into one unit, but their context suggests they might mean different things. |
| merge_suggestion | merge_into_a · merge_into_b · keep_separate | Two units with near-identical text (found by scan_duplicates) that could share one translation. |
| translation_review | approve · edit · retranslate | The model was not confident, or the translation changed a placeholder. |
| audit_finding | apply · edit · dismiss | An audit found a problem in one translation and proposes a fix. |
| rule_suggestion | add · edit · dismiss | An audit proposes a standing rule for a language. |
| glossary_suggestion | add · dismiss | An audit proposes a glossary term: fixed per locale, or never translated. |
| learn_format | keep · retire · retry_with_hint · dismiss | The AI wrote a template for a file format it had not seen. |
Claim, decide, resolve
claim_taskstakes up to 50 open tasks, highest priority first, optionally of one kind. Claimed tasks are hidden from other agents for 30 minutes, so you can run several in parallel.- Read each payload and decide. Use read ops such as
get_unitorsearch_unitswhen you need more context. resolve_taskwith one of the task's actions and a short note. Passtextfor edits,bodyfor a rule edit,key_idsto choose which keys to split,hintforretry_with_hint.
npx @meterapp/glossa claim_tasks --json '{"kind":"merge_review","limit":10}'npx @meterapp/glossa resolve_task --json '{"task":"<id>","action":"split","note":"verb vs adjective"}'Deciding merges
Every key points at a unit found by its normalized source text, across the whole workspace. Usually that is right: "Cancel" is "Cancel" everywhere. Split when the same words mean different things: a verb and a noun, a status and an action, or a word whose translation depends on what it refers to. Keep merged otherwise; fewer units means fewer translations to review.
A prompt that works
Work through the Glossa review queue until it is empty.- Claim 10 tasks at a time.- merge_review: split only when the keys mean different things; say why in the note.- translation_review: approve correct ones; edit with the fix otherwise. Never change placeholders.- rule_suggestion and glossary_suggestion: add the ones that would have prevented a real mistake.Summarize what you changed at the end.The plugin's review-queue skill does this out of the box: see MCP and agent plugins.