The page claimed a live 422 and 402, async_run stored, and the "hidden"
version not recorded, none of which the stored rows showed at cf52206. A
second live session (2026-10-20 14:59 UTC, traceAI#219 at 18f7f8b) stored
async_run, a stream with its final status and token counts, and an unknown
version answered HTTP 422; this page now says what each session stored, what
is still stand-in only (async_stream, create + wait, cancel, deployments, run
iterators, image and file outputs), and that trainings are not traced.

Also: stream_reload_seconds and the read after done, the "hidden" version,
the CHAIN kind of a failed call, 429 behaviour and troubleshooting (402 is
not claimed), and the unit-suite matrix (3.10-3.13 x replicate 1.0.0/1.0.7).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ne is completion

traceAI#219 ab5f5c4 moves the read after a stream (GET /v1/predictions/{id})
onto a background thread: never retried, no HTTP span, bounded by
stream_reload_seconds, and the span ends when it returns or the budget runs
out. A loop that stops at done and closes the stream is a completed stream.

- New section "The read after a stream": free but counted against the
  account's rate limit for non-create requests, how to turn it off
  (stream_reload_seconds=0), off the caller's path, bounded, no span, when
  the span ends, flush/exit waits, 256 in flight, what is kept.
- Stream row, option text, retry and cancellation bullets, a
  troubleshooting entry for a stream span without tokens.
- Live section: session 3 (one prediction, loop stops at done; 10 in all),
  fi_verify 8 traces, privacy scan of 121 rows.
- Tested versions: the cells the matrix actually ran.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…s timeout

Matches traceAI#219 at ee9cb13 (round-2 review):

- The read uses the package's own connections with the default transport
  (never the client's pool), at most 4 through a transport you pass and
  none on an httpx pool under 8 connections; a hard deadline on the own
  pool; at most 64 reads at once (replicate.stream.read=skipped past
  that); no read for a span that is not exported.
- force_flush(timeout_millis) waits at most half its timeout and never
  longer than asked; shutdown and exit wait at most the budget;
  uninstrument() does not wait; stream_reload_seconds is at most 10.
- The caller's cost stated as measured, not "no latency".
- replicate.stream.read / done_reason / truncated, aborted, a done with
  reason canceled, a stream never iterated has no span, suppressed calls
  are not traced; troubleshooting and limitations updated.
- Live: fourth session, async_stream through the package's own pool
  (0.42 ms for the done event, span ended 104 ms after the loop, 39/17
  tokens); eleven predictions in all; 137 stored rows scanned.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

traceAI#219 round 3 (b5df956, head 86915f8): force_flush() does not wait
for reads by default (replicate.stream.read=flushed; the new
stream_flush_wait_seconds opts in); a plain httpx transport is read through
the package's own connections; an async-only transport's read ends with
its event loop; name lookups are shared and wait for a resolver; at most
64 stream spans wait at once; 4 reads per transport of yours, 16 in all.
The caller-cost figure names its setup; "never longer than asked" is
qualified for inline exporters; the 2 MiB cap for transports that read
whole bodies. Troubleshooting lists every skipped cause, Limitations
carries the README's two bullets, dates are UTC. Live: session 5,
async_stream inside asyncio.run() on httpx.AsyncHTTPTransport(retries=2),
twelve predictions in all.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…he read after a stream is opt-in

traceAI#219 round 4 (d610ede, f1866eb, README 5ca9a1c). stream_reload_seconds
now defaults to 0: a stream span ends at done on the caller's thread, OK,
without status, predict_time or token counts; no request after the stream,
no thread, and flush, shutdown, uninstrument, exit and fork wait for nothing.
A Warning right after the instrument() options says so, how to opt in, and
the opt-in's limits.

"The read after a stream" becomes "The opt-in read after a stream": only
through the package's own pool, for replicate's default transport or a plain
httpx one (any other transport: skipped, custom_transport); flush, shutdown,
exit and uninstrument() do not wait by default (flushed / shutdown /
uninstrumented; stream_flush_wait_seconds and the new
stream_shutdown_wait_seconds opt in); forks: only the package's finisher is
coordinated; a table of every replicate.stream.skip_reason. The measured
cost is from probes at f1866eb. Troubleshooting and Limitations follow.

Live: session 6 (05:15 UTC on 2026-10-21, traceAI f1866eb), sync
replicate.stream with the default settings; thirteen predictions in all;
sessions 2 to 5 are labelled as run with the read on. The fifth session's
"looked up once" is restated as what was measured (one cache entry).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…for HTTP mocks and changed transports; seventh live session

- The opening paragraph says stream spans carry no status or token counts by
  default; the description no longer promises status and tokens for streams.
- The opt-in read's rules as in traceAI#219 round 5: only replicate's own
  RetryTransport over a plain httpx transport without a proxy and unchanged on
  the instance; no read while an HTTP mock patches httpx or httpcore
  (patched_http); the exit cost with an inline exporter.
- Default path: a flush's inline export of dropped streams; URL-like text
  chunks stay in the output; send()/throw().
- Live: the opt-in read in its current form is stand-in only; session 7
  (async_stream, default settings); fourteen predictions.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…aits only for the finisher's export; final-head figures; eighth live session

- A stream stopped before done keeps what it received (LLM for text);
  Ctrl-C recorded alike inside next() and in the loop body.
- uninstrument() never waits for a read; like flush, shutdown and exit
  it waits up to 5 s for the span the finisher is exporting;
  stream_shutdown_wait_seconds also applies at exit; force_flush(0)
  still exports held and dropped spans inline.
- Skip reasons: Sentry HttpxIntegration (patched_http),
  truststore.inject_into_ssl (patched_socket), instrument_client and a
  subclass's _transport_for_url (custom_transport); stub servers see
  the read. Truncated-stream status only with the read on.
- Figures measured from a git archive of traceAI 55c5672 (probe blocks
  named); fork memo and the macOS _scproxy fork hazard; no evidence paths.
- Eighth live session (traceAI f0376ad) and a system.columns-driven scan.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…celled; caller stops alike; figures at traceAI e29fcc6; ninth live session

- Calls still running at a shutdown (including register()'s SIGTERM/SIGINT
  handler) or at exit end as cancelled; Ctrl-C/SystemExit during run(), a
  create, wait() or cancel(), trio cancellations and throw(GeneratorExit)
  are the caller stopping the call; a refused cross-thread close ends nothing.
- "do not wait" / "never waits" now say "for a read" everywhere on the page;
  fork() ends no span.
- Figures re-measured from a git archive of traceAI e29fcc6, with the
  1-minute and 5-minute load and the [D1] skip proportion.
- Ninth live session (traceAI c5b9505); the privacy scan listed views and
  dictionaries without reading them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

A stream still open at uninstrument() ends when closed or completed; once
dropped, at the next traced call (after instrument() again) or at exit.
The troubleshooting entry now carries the 5 s finisher-export clause like
the rest of the page.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ackage version (TH-8320)

The page merges only after traceAI-replicate 0.1.0 is on PyPI, so the note saying it is not there yet would be wrong the day the page goes live. Names traceAI#219 instead of 'the current pull request' / 'this release candidate'.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>