Skip to content

[Review agent] Distinguish primary errors from newly suppressed secondary failures (#69445) #70269

Closed

Description

@maraf

Is there an existing issue for this?

  • I have searched the existing issues.

Related tracking issue: #70231. This report documents the specific review outcome on #69445 and a focused improvement, rather than a request to change the AI engine implementation.

Is your feature request related to a problem? Please describe the problem.

The review-pull-request review of #69445 did not surface a useful assessment of the empty-handler feedback. Investigation of the original November 2 execution shows two distinct effects:

  1. The feedback was seen and intentionally deduplicated, not overlooked or unavailable.
  2. The logging worker accepted the new exception-suppression policy as preserving the old processing-error contract, without adequately distinguishing the original processing exception from a separate failure raised by the logging provider.

This should not be characterized as a proven runtime-regression false negative. The original comment asked for justification of a policy choice; the later explicit decision to propagate logger failures had not yet been made when the review ran. The improvement should preserve that distinction and avoid turning every empty catch into a finding.

Source and discussion evidence

The reviewed head was c0f86a6d725473f7a6f4c56d816ef941591053d2; the old-side merge base was 80ce68b5280ef01757bdf9d94b653da62e5724d5.

  • Old AgentContext error path: processing failures were already represented by Error and ConversationStatus.Error, followed by notification and normal task completion. There was no logging call or logger-failure suppression at this seam.
  • Changed AgentContext error path: the PR adds _agent.LogProcessingError(ex) inside a nested try / catch (Exception). If the logging provider throws, that secondary exception is silently lost, while the original processing exception remains in Error and notification continues.
  • UIAgent forwarding method and configurable logger seam establish the called path.
  • Added preservation test deliberately exercises a throwing logger and asserts successful completion with the original processing error. This supports implementation intent, but is not independent authority that logger failures should be suppressed.
  • AgentContext swallows exceptions #69058 asks to log processing failures; it does not establish a best-effort policy for failures in logging itself.

The discussion clarifies the tradeoff:

  • November 2, 15:30 UTC: initial feedback: “We should avoid using an empty handler. Why is it needed?”
  • November 3, 07:35 UTC: author explanation: the guard avoids a new faulted-task path and preserves the subsequent status notification, while acknowledging the special handling may be unnecessary.
  • November 3, 09:33 UTC: explicit reviewer decision: let such exceptions propagate and remove the catch.

Original execution evidence and provenance

The original review history was available and inspected directly, including its delivered skill text, worker assignment, worker responses, and final synthesis. It was not reconstructed from today's skill.

Input Original execution
Review start November 2, 2026, approximately 19:54 UTC
PR head c0f86a6d725473f7a6f4c56d816ef941591053d2
Old-side merge base 80ce68b5280ef01757bdf9d94b653da62e5724d5
Frozen guidance/base tip f570eec79ddb2efed7baea5c2b33e46ee391ea65
Skill loading Successful native top-level review-pull-request invocation
Skill provenance Installed skill; no immutable repository revision recorded
SHA-256 of exact delivered skill content d09698f27f09185e76d003a05cacf655f4ba76896c8c6ba0a56ca9c90c728128
Review shape 27 topic workers, with 27 returned results

The delivered skill differed from the skill at the reported guidance revision, including after line-ending/edge-whitespace normalization. The guidance revision therefore must not be presented as the installed skill revision. The delivered content itself was recoverable.

Historical instructions: Step 1 said existing feedback was read “only for deduplication”; Step 5 dropped claims already raised by an existing review. The logging-worker assignment explicitly said: “Existing feedback asks why catch empty (dedupe only).”

Logging worker, approximately 20:00 UTC:

LGTM — no verified diagnostics/logging defect.

provider failures cannot replace it or prevent notification.

Existing empty-catch feedback deduplicated.

After receiving the exact diagnostics-topic bullets in a correction, the same worker retained its conclusion:

LGTM — unchanged after applying all five exact bullets from the pinned guide.

Logging-provider exceptions are handled without replacing the processing error.

Final coordinator synthesis, 20:02 UTC:

Existing inline comment questioning the empty handler was not re-raised; it was existing feedback and did not establish a separate, evidenced behavioral defect.

The relevant frozen guide already required “catch only exceptions you can handle or enrich”. Missing diagnostics-topic coverage or an absent catch guideline is therefore not the explanation.

Demonstrated gap versus uncertainty

Demonstrated: the review recognized and suppressed the duplicate; its substantive justification focused on preserving the primary processing error and notification, treating that as sufficient handling of the new secondary logger failure. Old source establishes the former contract, not the latter policy. The added test cannot independently establish the desired policy it was written to enforce.

Plausible but not demonstrated: the author's preservation rationale and the new test may have anchored the worker. The record supports the reasoning gap, not a claim about the model's internal cause.

Not established: that this catch violated a pre-existing, binding logger-exception propagation contract. The later reviewer decision must not be retroactively treated as evidence available to the original run.

Describe the solution you'd like

Add a focused changed-catch policy checkpoint to cross-cutting diagnostics review, with coordinator adjudication. Suggested wording:

For every new or broadened catch or suppression, distinguish the original operation failure from failures raised by logging, cleanup, callbacks, or recovery. Trace what each catch observes, preserves, propagates, retries, or silently loses. Establish the authority for treating each newly suppressed failure as expected or best-effort; preservation of the original error and PR-added tests alone do not establish that authority. Report a finding only when the exact reachable effect and binding contract establish a defect. Otherwise retain a concise unresolved policy decision, including an existing-feedback reference where present, without producing a duplicate comment or a rationale-only finding.

This targets the reasoning failure without introducing a blanket rule against empty catches, a global assumption that logging must never throw, or a global assumption that logging failures must always propagate. Documented best-effort diagnostics and expected cleanup failures remain legitimate.

The skill inspected during this investigation already requires EXISTING_FEEDBACK_COVERAGE and disclosure of unresolved candidates under NO_FINDINGS. It also prohibits rationale-only findings without a concrete material effect. Merely adding the coverage field again or weakening the finding gate is not the proposed fix. Preserve those gates and use the checkpoint to distinguish a justified discard from an unsettled policy decision.

Example evaluation cases and acceptance criteria

Use the frozen #69445 source and original guide in paired cases:

Case Expected reviewer behavior
Original source, no existing feedback Distinguish processing exception A from logger exception B; identify B's silent loss at the new catch; do not use the preservation test as contract authority. Retain the unestablished suppression-policy question without inventing a proven defect.
Same source, with the November 2 empty-catch comment Assess the same policy distinction and reference the existing thread; do not emit a duplicate inline finding or equate deduplication with substantive resolution.
Same mechanism, with an explicit target-base contract requiring provider failures to propagate Produce a concrete changed-line finding, tracing the configured logger seam and the caller-visible loss of the required fault.
Same mechanism, with an explicit documented best-effort diagnostic sink Do not report the suppression as a defect; cite the contract justifying it.

The evaluation should reject outputs that conflate A and B, use a newly added test as authority for the desired behavior, declare every empty catch defective, or duplicate existing feedback. It should also reject an unqualified LGTM that silently treats an unsettled policy choice as proven correct.

If runtime evidence is needed for a future case, a faithful narrow probe is SendMessageAsync with a chat-client processing failure and a configured throwing ILogger, observing the returned task, Error identity, and status notification. This establishes mechanics, not which propagation policy the framework should adopt.

Additional context

Note

This issue was generated with the assistance of GitHub Copilot.

Activity

  1. github-actions commented on Nov 3, 2026

    @github-actions
    Contributor

    Triage Summary

    Area: area-infrastructure (reviewer skill / review tooling rather than shipped framework code)
    Type: Task (bounded guidance improvement; no shipped behavior is broken)

    Potential Duplicates

    Notes

    Generated by Issue Triage Agent for dotnet/aspnetcore for #70269 · copilot · auto · 21.8 AIC · ⌖ 1.11 AIC · ⊞ 7.7K · ◷

  2. changed the title [-]PR reviewer: distinguish primary errors from newly suppressed secondary failures (#69445)[/-] [+][Review agent]: distinguish primary errors from newly suppressed secondary failures (#69445)[/+] on Nov 3, 2026
  3. changed the title [-][Review agent]: distinguish primary errors from newly suppressed secondary failures (#69445)[/-] [+][Review agent] Distinguish primary errors from newly suppressed secondary failures (#69445)[/+] on Nov 3, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions