Reads the atmosphere from sources that stay up when Bluesky PLC's AppView does not. On 2026-08-16 the AppView timed out and Jetstream returned 503 while PDSes and Constellation served every request — this skill is built around that split.
Invoke
python3 <skill>/scripts/atproto.py <command> [--json]
Digest output is the default and is what belongs in a transcript. Reach for
--json only when the result will be transformed programmatically; raw records
are large and re-reading them into context defeats the purpose.
| Command | Answers |
|---|---|
feed [actor] [--hours 3] [--html PATH] |
The Following timeline, threaded |
status [--actor X] |
Which layer is broken right now |
interactions <actor> [--hours 8] [--scan 100] |
Who liked/replied/reposted/quoted recent posts |
thread <at-uri> |
Replies to a post |
posts <actor> [--limit] |
An actor's posts, from their own PDS |
records <actor> <collection> |
Any collection in any repo |
resolve <actor> |
DID, handle, PDS host |
backlinks <target> [--collection C --path P] |
What references a record or account, in ANY lexicon |
identity <actor> |
Handle renames and PDS migrations, from the PLC audit log |
lexicons [query] [--hours N] [--others] [--schema NSID] |
Which collections are active network-wide |
Source model
Each read targets the cheapest source that survives independently. Prefer the PDS for anything a repo owns — it is authoritative and had no outage.
| Need | Source |
|---|---|
| Records of a known repo | That repo's PDS |
| Who interacted with a URI | Constellation (constellation.microcosm.blue) |
| handle ↔ DID ↔ PDS, and its history | plc.directory, entryway resolveHandle |
| Which collections are active network-wide | UFOs (ufos-api.microcosm.blue) |
| Search, feed generators, chat | AppView only — no substitute; say so |
feed is the one composite read: there is no AppView-free getTimeline, so it
rebuilds the Following timeline from follows → PLC → each followee's PDS, then
hydrates what the AppView would have inlined — facet byte-ranges to character
ranges, blob CIDs to cdn.bsky.app URLs, quote and parent URIs to fetched
records — and threads the result by reply root. --html writes a self-contained
Preact reader (same components as austegard.com/bsky/thread-reader.html), so a
feed request is one call, not a fan-out the caller re-derives by hand.
Ranking, mutes, and blocks live in the AppView and are absent by construction:
this is the raw follow graph.
Constellation is the persistent index. Do not rebuild one locally: it already indexes the whole network, is operated independently of Bluesky PLC, and is reachable when the AppView is not.
interactions walks the five Bluesky paths in LINK_PATHS. backlinks walks
whatever Constellation has indexed — tangled issues, vouches, list membership,
lexicons nobody here has heard of — so reach for it when the question is "who
references this, anywhere" rather than "who liked this post". Both are
two-phase for the same reason: links/all reports which sources are non-empty,
and only then does enumeration cost anything.
lexicons answers a different question from sample_firehose in
browsing-bluesky: UFOs consumes Jetstream and rolls it up per NSID, so a
single call reports what is running across the whole network without sampling
anything here. --others drops app.bsky., chat.bsky. and com.atproto.,
which is what turns the leaderboard from "Bluesky is large" into a list of the
other apps. --schema NSID fetches the published lexicon document from the
publisher's own repo, deriving the publisher by the reverse-DNS convention
(com.whtwnd.blog.entry → whtwnd.com); publishing that record is optional and
plenty of busy collections skip it, so a miss is a fact about the publisher, not
a failure.
Cost shape
State is a /tmp SQLite scratch (ATPROTO_CACHE to relocate), session-scoped
by design. Cold start is the only start, and that is fine because wall-clock
and token cost are decoupled — the script absorbs the HTTP fan-out and returns
a digest. Measured: interactions --hours 8 over 100 posts ≈ 18s and ~30 output
lines, against ~15 tool calls doing it by hand.
interactions is two-phase on purpose: one cheap links/all per post reports
which link paths are non-empty, so per-path enumeration only runs where
something exists. Preserve that when extending.
Handling failure
Unavailable is raised for egress blocks, 4xx, and exhausted retries; the CLI
exits 2 with the reason on stderr. Report which layer failed rather than
retrying blindly — "the AppView is down" and "atproto is down" are different
facts and the distinction is usually the answer the user wants.
Two known limits worth stating plainly when they bite:
- Third-party PDSes may be egress-blocked. ~3% of a typical follow graph self-hosts (eurosky.social, blacksky.app, personal PDSes). Those reads fail with a clear allowlist message; add the host to project egress settings.
- Times come from TIDs, decoded from the record key rather than fetched. Client-generated, so treat them as approximate ordering, not attestation.
Overlap with the Atmosphere MCP
aturi.to/api/mcp (hosted, keyless, read-only, beta) covers the same three
areas and about 30 more. It was measured on 2026-08-28 at 38 tools / ~8.6k
tokens of schema resident per session, returning pretty-printed JSON rather
than digests, with every read routed through one host. These commands go
straight to plc.directory, Constellation and UFOs, which is the point of this
skill — the 2026-08-16 split is exactly the case a single upstream reintroduces.
Use the MCP for what it alone has: the waypoint catalog and the atproto docs
search. Do not route these three through it.
Writing
This skill reads. For posting, use muninn_utils.bsky_card — it already writes
straight to the PDS (facets, blobs, embeds, create_session) and therefore
works during an AppView outage. muninn_utils.bsky_limit enforces the
300-grapheme cap, which len() gets wrong on emoji and combining marks.