Digest Bridge

Digest Bridge

Windows Runtime Commands

On native Windows, use the managed Windows runner and the native runtime command target. Set $runtime to the installed runtime root. Multi-agent installs usually use %LOCALAPPDATA%\ai-agents-skills\runtime. Then run:

$runtime = if ($env:AAS_RUNTIME_ROOT) { $env:AAS_RUNTIME_ROOT } else { "$env:LOCALAPPDATA\ai-agents-skills\runtime" }
& "$runtime\run_skill.ps1" "skills/digest-bridge/digest_bridge.py" <args>

POSIX examples below use run_skill.sh and .sh command targets; use the Windows command target above on native Windows.

This uses the managed ai-agents-skills runtime copy of the digest bridge workflow.

When to use

• scan research digests for papers
• scan RSS digests for papers
• create a getscipapers manifest from digest outputs
• request papers mentioned in recent digests

Base path

• $AAS_RUNTIME_WORKSPACE/skills/digest-bridge/

This is a direct Python entry point, so run it from the managed ai-agents-skills runtime workspace with the workspace-local PYTHONPATH.

Core commands

Use functions.exec_command.

cd "$AAS_RUNTIME_WORKSPACE" && PYTHONPATH="$AAS_RUNTIME_WORKSPACE/.local:${PYTHONPATH:-}" python3 skills/digest-bridge/digest_bridge.py scan

cd "$AAS_RUNTIME_WORKSPACE" && PYTHONPATH="$AAS_RUNTIME_WORKSPACE/.local:${PYTHONPATH:-}" python3 skills/digest-bridge/digest_bridge.py scan --source research --min-score 3

cd "$AAS_RUNTIME_WORKSPACE" && PYTHONPATH="$AAS_RUNTIME_WORKSPACE/.local:${PYTHONPATH:-}" python3 skills/digest-bridge/digest_bridge.py request --source research

cd "$AAS_RUNTIME_WORKSPACE" && PYTHONPATH="$AAS_RUNTIME_WORKSPACE/.local:${PYTHONPATH:-}" python3 skills/digest-bridge/digest_bridge.py request --source rss --watch

Operational notes

• Use this after a digest run, not as a replacement for the digest itself.
• Runtime installation includes getscipapers-requester transitively because the bridge's durable manifest/watch handoff invokes that helper.
• The bridge consumes validated digest-items.v1 JSON sidecars only. Markdown is display-only and is never scanned for identifiers.
• Every bridge JSON boundary rejects duplicate object members and non-standard constants such as NaN or Infinity. This applies to producer sidecars, helper acknowledgments, persisted manifest/watch artifacts, and the owned request ledger; malformed owned ledger bytes are preserved and block the helper before any handoff can be banked.
• RSS discovery is limited to the five producer-owned sidecars rss-research, rss-events, rss-jobs, rss-general, and rss-video. Each sidecar's top-level source must match its canonical filename owner; additional or relabeled JSON files cannot enter the handoff. Structured producer outcome fields must also establish that discovery/publication did not fail completely; research status requires exactly arxiv, s2_recommend, and s2_search, each with only status and detail. arXiv is never skippable; either Semantic Scholar source may be skipped when its required seed/topic input is absent. Successful-empty runs remain admissible.
• If a Markdown digest exists without its sidecar, or a sidecar is invalid or oversized, scan and request exit 2 with invalid_digest_sidecar. Rerun the corresponding producer; do not synthesize a sidecar from Markdown. Admission uses directory-entry metadata rather than target existence, so a broken canonical sidecar symlink or a link-like/non-directory RSS digest path fails closed instead of being mistaken for an optional missing producer. On Windows, link-like includes filesystem reparse points such as junctions, not only ordinary symlinks.
• arXiv IDs and DOIs are derived only from validated arxiv.org and doi.org links in the sidecar, never from titles, summaries, or feed text.
• Every handoff is canonicalized to the DOI form accepted by getscipapers-requester; versioned arXiv links and their DataCite DOI form are one request.
• request serializes the manifest/watch transition with the bridge ledger. It records identifiers only after the complete manifest and every requested watch succeed, retains the full bounded producer maximum, and fails closed if the ledger is corrupt, oversized, or symlinked.
• A helper exit code and JSON echo are not sufficient acknowledgment: the bridge bounded-reads the regular persisted manifest and verifies exact DOI coverage, then confirms every strict watch acknowledgment through the helper's locked list-watches ledger before advancing bridge state. A final durable watch must still be active, waiting, or posted, or have reached the successful found state; failed and expired transitions remain retryable and are never banked. A retry after a later item failed may accept and reuse an earlier exact-identity found watch without duplicating it. Durable records are semantically revalidated with the helper-equivalent canonical DOI/ISBN/search identity and both structured/current and legacy watch-key derivations. Creation/update/deadline/check/history timestamps must form a possible chronology; a shape-valid but stale-key, noncanonical, or time-inconsistent ledger record is never banked. Creation, update, last-check, and note-event times more than five minutes ahead of the bridge clock are also invalid; deadlines remain intentionally future-capable. Both the immediate echo and durable record must retain the exact requested services: ["all"] identity. Manifest echo and artifact reads use the helper's 16 MiB output contract, so the full 3,000-item, 500-character ASCII DOI boundary remains admissible. DOI syntax is explicitly ASCII; Unicode IGNORECASE lookalikes cannot be case-folded into a different paper identifier. Helper stdout and combined stdout/stderr are drained concurrently under the endpoint contract; an overproducing or non-terminating child is stopped before any request is banked. One JSON protocol newline is framing rather than artifact content.
• Watch creation stops at the first failed helper call; later papers are not spawned after the transaction is already known to be retryable.
• The request ledger has its own 2 MiB reader/writer ceiling and compact UTF-8 encoding, so the complete admitted maximum batch reloads and deduplicates on the next request instead of becoming unreadable after its first handoff.
• The getscipapers watch store independently locks every helper reader/writer, is byte/count bounded, and fails closed without replacing corrupt state. A watch-store failure leaves the bridge ledger unchanged for a safe retry.
• Explicit invalid helper configuration or unavailable configured manifest/ watch storage fails the handoff; the bridge never banks output silently rerouted to the helper's temporary fallback root.
• Respect --source and --min-score filters instead of broad requests when the user wants a narrower batch.
• scan reports ledger matches as already_requested and never conflates never-requested papers excluded by --min-score, which are counted separately as below_min_score.
• If the user wants actual external retrieval, follow the manifest or request output into getscipapers-requester.
• scan is the dry-run discovery step; request is the transition into manifest/watch creation.
• If the manifest cannot be built, request exits 2 with manifest_failed and records nothing. The papers stay unrequested, so rerun it once getscipapers works rather than rescanning for them.