repost-global-dedupe

name	repost-global-dedupe
description	Cross-pair/global dedupe for Repost-with-agent. Use before any publish-capable run so pairs with different sources/destinations share one content ledger and avoid double-posting the same item through alternate hops such as LinkedIn→X→Bluesky and X→Bluesky.
when_to_trigger	Any time a candidate survived basic per-pair source scraping and before destination compose/publish, especially when multiple pairs can share a destination or source from each other.

Repost Global Dedupe — cross-pair content ledger

This is the cross-pair layer Ethan asked for on 2026-05-04: pairs must not think in isolation. If any pair has already handled the same underlying content for a destination, every other pair should know and skip instead of double- posting.

State file

Global state lives at:

~/.repost-with-agent/global-posted.jsonl

It is append-only NDJSON. It sits beside pairs.json and complements, but does not replace, each pair's local pairs/<id>/posted.jsonl.

Each line should be shaped like:

{
  "ts": "<ISO-8601>",
  "event": "global.publish.success | global.publish.catchup | global.publish.semantic_duplicate | global.publish.remote_duplicate | global.publish.deleted | global.publish.malformed",
  "pairId": "<pair-id that observed/created this record>",
  "contentKey": "<canonical cross-pair content identity>",
  "sourcePlatform": "<site-key, e.g. linkedin | x | bluesky | threads | facebook | your-site>",
  "sourceItemId": "<source platform item id>",
  "canonicalSourceUrl": "<source post URL>",
  "destinationPlatform": "<site-key, e.g. linkedin | x | bluesky | threads | facebook | your-site>",
  "destinationAccountHint": "<configured destination.accountHint if known>",
  "destinationUrl": "<published/matched destination post URL>",
  "destinationId": "<destination platform item id if known>",
  "draftText": "<text that was/would be posted>",
  "status": "posted | caught-up | skipped-duplicate | posted-malformed | deleted-malformed | deleted-runaway | deleted-source-url-leak",
  "note": "<optional human reason>"
}

Missing optional fields are fine. Never rewrite old lines; append corrections as newer lines.

Content key rules

A contentKey represents the underlying idea/item independent of the pair hop. The simplest key is:

<sourcePlatform>:<sourceItemId>

or, if the source ID is unavailable:

url:<canonicalSourceUrl without tracking params / trailing slash normalized>

Before finalizing that key, inherit lineage from the global ledger:

1. Load ~/.repost-with-agent/global-posted.jsonl if it exists.
2. If the candidate's current source URL or source ID matches a previous line's destinationUrl, destinationId, or destination-platform canonical URL, inherit that line's contentKey.
   • Example: LinkedIn→X published contentKey=linkedin:urn:li:activity:123 to https://x.com/.../status/999.
   • Later, x-to-bluesky scrapes that X status as its source. Because the X source URL/ID matches the earlier line's destination, the X candidate inherits contentKey=linkedin:urn:li:activity:123 instead of becoming a new x:999 item.
3. If multiple lines match, use the newest non-empty contentKey and record the inherited-from URL in audit if helpful.

This lineage inheritance is what prevents LinkedIn→X→Bluesky and a direct X→Bluesky pair from both posting the same LinkedIn-origin content to Bluesky.

Crash-recovery derived-source guard

Do not rely only on the global ledger when the current source account is also a configured destination account for another pair. A crashed or interrupted run can create a public destination post before it writes global-posted.jsonl; the next daily sweep may then scrape that public post as a fresh source and cascade it to other platforms.

Before publishing a candidate from an owned destination account such as the X profile used by a LinkedIn→X pair:

1. Search every local pair history (~/.repost-with-agent/pairs/*/posted.jsonl) for a live row whose destinationUrl / destinationId matches the current candidate's canonicalSourceUrl / sourceItemId. If found, inherit that row's contentKey and treat the current source as a derived repost output, not a new organic source.
2. Search source-fanout manifests under ~/.repost-with-agent/source-fanouts/ for destination records matching the candidate URL/id. If found, inherit the manifest source item key (<sourcePlatform>:<sourceItemId>).
3. Search active backfill queues under ~/.repost-with-agent/backfill-queues/ when the candidate text/link is a near-exact match for a queue item's sourceBody/clean draft, even if the failed run never wrote a manifest or pair ledger. Match conservatively: same expanded public URL plus an obvious text overlap, or ≥80-character normalized prefix/quote overlap. When this matches, fail closed with derived-source-shadow rather than publishing.
4. Append a pair.dedupe.derived_source_shadow audit event and a global catch-up line if the destination already exists; otherwise skip the candidate without creating a public post. The audit must name the upstream source item, the matched queue/manifest/ledger proof, and the reason.

This guard covers the exact failure mode where LinkedIn→X created an X post, OpenClaw failed before state was fully written, and the daily x-to-* sweep saw that X post as new source material. In that case the correct action is to repair or catch up the original LinkedIn fanout state, not repost the derived X post to more destinations.

Global duplicate decision

For the current pair and candidate:

1. Resolve the candidate contentKey using the rules above.
2. If pair.policy.globalDedupeEnabled === false, skip this skill. Default is enabled.
3. Run the crash-recovery derived-source guard above. If it matches, do not publish the candidate as a normal source item; either inherit the upstream key and continue dedupe, or skip/catch up with derived-source-shadow when the only safe conclusion is that the source is a repost output from another pair/backfill.
4. Search global ledger lines for the same contentKey where:
   • destinationPlatform === pair.destination.platform, and
   • if both are present, destinationAccountHint / destination profile/account point at the same configured destination identity.
5. Compute the newest live-success verdict for those same-destination rows. A row proves live success only when its event/status is one of global.publish.success, global.publish.catchup, global.publish.remote_duplicate, global.publish.semantic_duplicate, posted, caught-up, or skipped-duplicate, it has a destination URL/ID, and it has no remediation/deletion/malformed flags. Newer rows with global.publish.deleted, deleted-*, posted-malformed, global.publish.malformed, needsRepost, or needsRemediation remove or quarantine the old proof and must not count as duplicates.
6. If the latest same-destination verdict is live success, the candidate is duplicate-global for this destination. Do not publish. Append:
   • a per-pair audit event pair.dedupe.global_duplicate, and
   • a per-pair catch-up line to pairs/<id>/posted.jsonl, and
   • a global catch-up line with event: "global.publish.catchup", status: "skipped-duplicate", and a note naming the matched destination.
7. If no live same-destination verdict exists, the candidate is globally unique for this destination. Continue to destination scrape + Layer 2 semantic dedupe.

Important nuance: the same contentKey may legitimately be posted once per configured destination. LinkedIn→X and LinkedIn→Bluesky are both allowed; two different routes to Bluesky for the same contentKey are not.

When to append global state

Append to the global ledger on every meaningful outcome that proves destination state. Use correction/deletion rows instead of relying on old success rows after cleanup; future dedupe must evaluate the latest verdict, not mere row existence:

• Successful publish: append global.publish.success immediately after the per-pair posted.jsonl success line.
• Remote/fuzzy duplicate catch-up: append global.publish.remote_duplicate when destination scrape proves the item already exists.
• Layer 2 semantic duplicate catch-up: append global.publish.semantic_duplicate when the semantic layer matches an existing destination post.
• Global duplicate skip: append global.publish.catchup so future pairs do not re-reason the same skip.

Do not append global success for failed/uncertain login, CAPTCHA, rate-limit, or compose failures. Those do not prove a destination post exists. If a previously recorded destination post is deleted or found malformed/source-url-leaking, append global.publish.deleted or global.publish.malformed with the old URL, status (deleted-malformed, deleted-runaway, deleted-source-url-leak, or posted-malformed), and needsRepost / needsRemediation as appropriate so later runs do not treat stale proof as live.

Audit event

When this skill skips a candidate, append to pairs/<id>/audit.jsonl:

{
  "ts": "<ISO-8601>",
  "event": "pair.dedupe.global_duplicate",
  "pairId": "<pair-id>",
  "sourceItemId": "<candidate source id>",
  "canonicalSourceUrl": "<candidate source URL>",
  "contentKey": "<resolved contentKey>",
  "matchedPairId": "<pair that already handled it, if known>",
  "matchedDestinationPlatform": "<destination platform>",
  "matchedDestinationUrl": "<existing destination URL>",
  "reason": "same contentKey already posted/caught-up for this destination"
}

For derived-source suppression, append the sibling shape:

{
  "ts": "<ISO-8601>",
  "event": "pair.dedupe.derived_source_shadow",
  "pairId": "<pair-id>",
  "sourceItemId": "<candidate source id>",
  "canonicalSourceUrl": "<candidate source URL>",
  "contentKey": "<inherited upstream contentKey if known>",
  "matchedUpstreamSourcePlatform": "linkedin",
  "matchedUpstreamSourceItemId": "urn:li:activity:...",
  "matchedProof": "global-ledger | pair-ledger | source-fanout-manifest | backfill-queue-text-match",
  "reason": "source post is a public destination output from another pair/backfill; repair upstream state instead of cascading"
}

See also

• skills/repost-run/SKILL.md — calls this before destination publish.
• skills/repost-dedup/SKILL.md — Layer 1 local/destination string dedupe.
• skills/repost-dedup-semantic/SKILL.md — Layer 2 semantic dedupe.
• docs/state-files.md — state schemas.