Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Outside diff comments:
In @.github/workflows/e2e-advisor.yaml:
- Around line 185-189: Guard the trusted analyzer entrypoint before invoking
node: check that "$ADVISOR_DIR/tools/e2e-advisor/analyze.mts" exists and is
readable and only run the node command if it does; if the file is missing, log a
clear message and skip or exit gracefully so CI doesn't show MODULE_NOT_FOUND
and cascade failures. Update the step that runs analyze.mts to perform this
existence check (refer to analyze.mts and the ADVISOR_DIR variable and the node
invocation that writes to "$GITHUB_WORKSPACE/artifacts/e2e-advisor") and ensure
downstream steps handle the skipped case appropriately.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@tools/e2e-advisor/analyze.mts`:
- Around line 349-354: The truncation banner can be lost because
raw.prepend(...) may be immediately trimmed away if raw is already full; change
the pattern to ensure space for the banner before inserting it (e.g., compute
banner = `<...truncated...>` then call a method that trims the buffer from the
opposite end or explicitly remove enough bytes from the end/front to make room,
and only then prepend), or alternatively append the banner to the end of the
buffer instead of prepending so trimming won't remove it. Update the occurrences
that use raw.prepend(...) (and the related text.droppedBytes logic) at the
locations shown (the prepend calls around the variables raw and text) to first
reserve room for the banner or append the banner so the truncation marker cannot
be immediately discarded.
- Around line 104-105: Temporary advisor configDir (variable configDir) and
files like auth.json/models.json and persisted session data must be removed
after the run to avoid leaving API keys on disk; add deterministic cleanup that
removes configDir (use fs.promises.rm or fs.rm with recursive:true, force:true)
in all exit scenarios by invoking deletion in a finally block around the main
async flow and/or registering handlers for process exit signals (exit, SIGINT,
uncaughtException) so configDir is always deleted whether the run succeeds or
errors; ensure deletion occurs after any code that reads/writes those files (the
blocks that write auth.json/models.json and persist session data) to avoid race
conditions.
- Around line 131-137: The credential check (hasLikelyAdvisorCredential()) and
the model materialization (prepareAdvisorConfig()) are inconsistent: preflight
allows provider-specific creds (ANTHROPIC_API_KEY, OPENAI_API_KEY, Bedrock etc.)
but prepareAdvisorConfig() only writes models.json when E2E_ADVISOR_API_KEY is
present, causing "Could not find configured advisor model" later. Fix by making
the flows consistent: either (A) tighten hasLikelyAdvisorCredential() to require
E2E_ADVISOR_API_KEY only, or (B) update prepareAdvisorConfig() to detect
provider-specific credentials (use the same detection logic as
hasLikelyAdvisorCredential()), resolve provider with E2E_ADVISOR_PROVIDER ||
defaultProvider detection, call defaultModelForProvider(provider) and write a
non-empty models.json for that provider even when E2E_ADVISOR_API_KEY is not
set; ensure code paths that reference modelPattern (E2E_ADVISOR_MODEL ||
defaultModelForProvider(provider)) match this logic so prepareAdvisorConfig()
materializes the same model names.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Outside diff comments:
In @.github/workflows/e2e-advisor.yaml:
- Around line 88-93: The workflow step that checks out the trusted advisor code
and then runs analyze.mts is switching the trusted run to a file that doesn't
exist on main and causes MODULE_NOT_FOUND; restore the existing trusted analyzer
path (keep running the analyzer from the PR checkout) or add the same fallback
logic used for dispatch.mts and comment.mts so analyze resolution tries the
legacy path first then analyze.mts (mirror the pattern used around
dispatch.mts/comment.mts); update the checkout/step that references analyze.mts
and the corresponding resolution logic in the workflow (also apply the same
change to the other occurrence referenced in lines 157-175) so the job does not
hard-fail when main lacks tools/e2e-advisor/analyze.mts.