tamash-playwright 0.13.0-beta.2 → 0.13.0

package/.env.example

@@ -95,3 +95,23 @@ OLLAMA_API_KEY=
 # no vision fallback, no action-recovery tactics. A fast, free, deterministic first line of
 # defense, not a full replacement for an AI provider. Just set:
 # HEALER_PROVIDER=tamash
+
+# --- Failure analysis: classify a genuinely-failed expect() assertion ---
+# A separate capability from healing above — never fixes anything, only explains a test failure
+# that already happened after every retry was exhausted: likely-defect vs. likely-wrong-locator
+# vs. likely-timing-or-environment vs. inconclusive, with a short explanation. Covers any final
+# failure, not just expect() -- an action (click/fill/etc.) healing already tried and reported on
+# is analyzed too, with healing's own diagnosis folded in as extra context rather than skipped.
+# Reuses whichever HEALER_PROVIDER is already configured above (tamash always declines — no
+# reasoning capability). No reporter to add to playwright.config.ts, runs automatically the same
+# way healing does.
+#
+# On by default, same polarity as HEALER_ENABLED above (unset or anything other than false/0
+# leaves it on) -- independent of HEALER_ENABLED, though: turning healing off does not also turn
+# this off, they check two separate flags. Costs a real AI call on every genuinely-failed test
+# (unlike healing, which only spends when an action fails and is opportunistic about it), so a
+# suite with several broken tests means several calls, every run, until each is fixed. Each
+# call's token usage shows up as its own `failure-analysis-tokens-used` annotation in the report
+# -- deliberately a different annotation type from healing's own `llm-tokens-used`, so the two
+# costs can be weighed against each other, not just healing's. Set to false to turn it off:
+# FAILURE_ANALYSIS_ENABLED=false