v0.97.1 · 2026-07-06

The adversarial review pays — nine fixes in the day's own release

An end-of-day adversarial review of everything 0.97.0 shipped found
real edges in the untested seams; all confirmed on the code, fixed and
pinned the same day:

• Housekeeping can no longer strand an answerable run. The trace
  pruner was blind to what a journal IS: it could delete a PAUSED run's
  journal (the resume substrate) — including the very --resume target
  the click was about to consume, since it pruned BEFORE the spawn.
  Paused journals now survive any ranking (tail check), and the
  imminent spawn's resume target is explicitly protected.
• An answer now targets its own pause. The paused notification
  re-derived the journal at click time from a live map that EVERY run
  overwrites — a mock preview run between pause and answer could
  swallow a real human approval into a mock journal. The paused record
  now carries its journal path, captured at pause time; answering the
  same pause twice warns before re-running gated side effects.
• Typed answers survive. The engine JSON-parses answer values —
  answering 123 to an input gate arrived as a Number and failed the
  gate's string contract; numeric-looking choices could never match.
  Input and choice answers are JSON-encoded (text stays text).
• Preflight's engine contract is alive. parseCheckReport never
  copied the requirements section — the whole engine-stated-contract
  adapter was dead code on the wire. One copy, one wire-level test.
• Diff v2 names the culprit, not a victim. First-divergence ranked
  by clocks mixed across the two runs — a cascade-cancelled task (no
  compare clock) fell back to the OLDER run's epoch and always outranked
  the actual failure. Compare-run clocks only now.
• Drift truth only speaks about its own workflow (the sha check now
  requires the trace to match the active document), the runs-tree cache
  keys on mtime+size (same-tick appends on coarse filesystems), the
  .gitignore nudge re-reads before writing (a stale snapshot could
  revert edits made while the toast waited), and terminal runs (run
  fallback · golden test) honor the spawn-cwd law so their journals
  land beside the workflow.

Time travel, for real

• F5 time-travel debugger (nika ≥ 0.96) — breakpoints in your
  .nika.yaml, F5 replays the newest recorded run of that workflow under
  the real VS Code debugger: step forward AND backward through task
  settles, recorded outputs in the Variables pane, continue runs to the
  next breakpointed task. Replay never re-executes — stepping back is
  free. Every run in the Runs view gains "Debug This Run (Replay · Time
  Travel)" (journal→source matched by workflow name; QuickPick on
  ambiguity). A walkthrough step teaches it.
• Export to OpenTelemetry (nika ≥ 0.96) — a context action on any
  recorded run projects its journal to OTLP/JSON lines: drag into Jaeger
  UI, or POST to Aspire/Grafana/Langfuse (cost rides gen_ai.usage.cost).
  Local file, zero collector. An engine older than the verb is told to
  upgrade instead of a clap parrot.
• Runs view discoverability — Run History and Diff join the view
  title bar; Preflight and Run History join the welcome capabilities grid.

---

📦 VS Marketplace · Open VSX · CHANGELOG · engine: docs.nika.sh/changelog/releases