Skip to content

Releases: julia-script/effect-inspect

v0.3.1

Choose a tag to compare

@github-actions github-actions released this 01 Oct 22:39
98351e7

Patch Changes

  • 24a8793: Upgrade Effect and its React, Node, and Bun integrations to the stable 4.0.0 release and migrate imports to the stable module paths.

v0.3.0

Choose a tag to compare

@github-actions github-actions released this 26 Sep 14:48
91d3d4f

Minor Changes

  • c5efb18: summary.result.memory places process-wide memory in time: peakHeapAtMs, first/last sample times, sampling cadence (medianIntervalMs, maxGapMs with maxGapFromMs/maxGapToMs) and spansActiveAtPeak, the innermost spans active when the heap peaked. A memorySamplingGap notice reports a gap between samples over 10× the median, with possible causes phrased as possibilities. span gains processMemory: the process-wide samples within the span's interval. Additive; apiVersion stays 1.
  • 3904603: summary now leads with what an interrupted or disconnected run left unfinished. A top-level notices array states easy-to-miss facts (spans without a recorded end, how the session ended, that longest ranks completed spans only, collector eviction). result.unfinished lists the innermost open spans — the last recorded position on each open chain — with their open ancestors, before longest. Every per-session response gains termination (state active/ended/unknown, lastObservedMs, endedAtMs, unobservedTailMs). spans --sort duration|outsideChildren interleaves open spans by a lower bound instead of listing them last. Additive; apiVersion stays 1.

Patch Changes

  • ff10e18: Root --help gains a JSON SHAPE section: the top-level keys of a query response, a reminder that context keys (completeness, termination, ...) are siblings of result, and the most-used key paths. Each per-session command's help points to its field lists and says to read summary's notices first.

v0.2.0

Choose a tag to compare

@github-actions github-actions released this 25 Sep 17:00
a3b8dcd

Minor Changes

  • 27fffee: Let a launcher choose the session ID before starting an instrumented program, through the sessionId layer option or the EFFECT_INSPECT_SESSION_ID environment variable; a random UUID remains the default. The ID survives reconnects, while a different run reusing an ID the collector already holds is refused and counted as a conflict instead of being merged into the original trace. Collision refusal needs the updated collector; older collectors still merge runs that reuse an ID.
  • 1a9a75e: The collector now answers read-only trace queries over HTTP on its port (POST /api/v1/query, GET /api/v1/export?sessionId=ID): session discovery, a run summary, filtered spans, one span with its ancestry, children and events, and correlated logs. Answers are bounded JSON that name the exact session and time reference, report recorded loss or unknown completeness, and refuse a session whose ID another run reused. Saved traces answer the same queries offline, and an exported trace keeps the collector's loss counters in an optional header field that older builds ignore.
  • a2aba7c: New query commands for coding agents: summary, spans, span, logs, export and sessions. Launch an instrumented program with EFFECT_INSPECT_SESSION_ID=my-run-001, then query exactly that run with --session my-run-001, or a saved trace with --file PATH and no collector. Each command prints one bounded JSON document on stdout (--json for one compact line), a diagnostic on stderr on failure, and a distinct exit code per outcome. export writes a lossless .eitrace without overwriting an existing file unless --force. Root and per-command --help document the workflow, flags, JSON fields, errors and exit codes. The interactive --wizard global flag is no longer offered.