[ROHIT SINGHAL (rohitsinghal4u)]

Description

Adds evidence-based issue investigation after triage, adapted from Azure/azure-sdk-for-net#60514 for Python and PyPI.

Protecting existing workflows

Existing triage label/comment/assignment budgets, owner-routing code/dependencies, permissions, and failure reporting are preserved against main. Investigation is handled by a separate completion-triggered relay, not by a prerequisite of triage writes or completion. The existing Actionlint workflow is restored byte-for-byte; new native contracts run in a separate path-filtered CI workflow.

Automatic handoff is disabled by default. A maintainer can set GH_AW_ENABLE_ISSUE_INVESTIGATION=true after rollout validation. Manual investigation dispatch remains available. The relay verifies a successful same-repository/default-branch triage run, attempt-specific jobs and current artifacts, the explicit safe detection verdict, the target cross-checked from trusted activation-job logs, the applied analysis comment, and owner routing before native dispatch. Its failures cannot gate or roll back the completed triage workflow. The existing triage source, generated workflow, and shared Actionlint are byte-for-byte identical to current upstream main; runtime selection and concurrency are unchanged.

Changes and review fixes

• Use an independent, opt-in completion relay after successful triage. Preserve single-owner assignment and mention-based routing using actual application receipts.
• Validate nonempty terminal outputs before mutations. Reject incomplete/error outputs and conflicting actions rather than returning a misleading successful run.
• Bind investigation comments and closures to the dispatched issue before invoking native handlers; the pinned close handler does not itself enforce its configured target. Recheck open/unlocked status and all clean-triage labels before mutations and Copilot assignment.
• Use the closure handler's comment-first behavior and gate Copilot assignment on the applied recommendation comment's Outcome. Keep authentication/availability skips best-effort.
• Preserve lifecycle-aware compatible-release guidance, specific duplicate evidence, targeted information requests, service-ownership evidence, and human-review exclusions. Read all comment pages before deciding whether an investigation is repeated.
• Generate the workflows with gh-aw v0.88.8 and pin Copilot CLI 1.0.80. Run native handler/serializer and baseline-compatibility contracts in a separate path-filtered CI workflow.

Local validation

• The added investigation compiles with strict/schema validation and zero warnings; its generated output is reproducible. Existing triage source and generated YAML are restored exactly from upstream, not regenerated.
• 21 automated test groups pass, including 370 named deterministic gate, corner-case, and relay integration scenarios. Coverage includes full-file legacy parity, missing/expired/ambiguous artifacts, reruns, actual native dispatcher calls with mocked GitHub writes, every handoff predicate, scope validation, and supported target representations.
• Native actionlint v1.7.12, repository cspell, and whitespace checks pass.
• The real single-owner receipt is rejected by the previously published helper and accepted by the corrected helper.

Live evidence and remaining limitations

Earlier live runs verified a targeted information request and a bounded documentation-fix recommendation on commit 06bc1d0, plus a negative handoff that lacked a valid analysis receipt. Most of the planned live gate matrix was not executed before cleanup. The corrected relay selected the issue and artifact IDs read-only from real upstream triage run 38799256886. Full live gate coverage, automatic completion-triggered dispatch, and persistent coding-agent assignment remain unverified; a native handler success log alone is not proof of assignment. No full-live-validation claim is made for this update.

Copilot assignment remains best-effort with the existing credential chain. Ownership/version decisions remain model-based, and separate API calls are not transactional. Native runtime scripts are pinned and tested rather than copied or modified.

API, compatibility, and security

No SDK implementation, public API, package version, dependency, or release metadata changes. The color-based label taxonomy is unchanged: all service names and all six current category colors remain supported, including future service names using the same color. Positive integer forms, case-insensitive colors, and an optional leading color hash remain accepted. Closed/locked/manual-triage issues and cross-issue output targets are intentionally excluded. The agent stays read-only with shell access disabled; no new secrets or external inference endpoints are introduced.

Reviewer risks

Likely reviewer	Comment they might leave	Evidence checked	Fix made or reason deliberately kept
Azure SDK engineering owners	Handcrafted receipts do not reflect runtime behavior	Pinned handler and serializer execution	Real runtime regression tests in CI
Copilot reviewer	Scoped closures or gated assignment still affect the wrong issue	Native close-target probe, applied receipts, current issue state	Trusted output binding and clean-triage revalidation
Python owners	Upgrade advice rejects supported versions or previews	Published lifecycle policy and release conventions	Compatible-release guidance, no service/cloud allowlist

All SDK Contribution checklist

• No SDK breaking changes.
• Package CHANGELOG: not applicable to this workflow-only change.
• Contribution guidelines read.
• Title and description reflect the implementation and known validation limits.
• Existing PR history preserved; no merge or force push requested.
• Regression coverage is included in CI.

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).
9 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

Pull request overview

Adds an automated, evidence-gated investigation stage after issue triage.

Changes:

• Adds issue investigation and decision rules.
• Dispatches investigations after clean triage.
• Adds compiled workflow definitions.

Reviewed changes

Copilot reviewed 3 out of 4 changed files in this pull request and generated 6 comments.

File	Description
.github/workflows/issue-investigation.md	Defines investigation behavior and outputs.
.github/workflows/issue-investigation.lock.yml	Compiles the investigation workflow.
.github/workflows/issue-triage.md	Dispatches eligible issues for investigation.
.github/workflows/issue-triage.lock.yml	Compiles updated triage orchestration.

Suppressed comments (1)

.github/workflows/issue-investigation.md:61

• The assignment sink is also repository-wide, so a compromised or mistaken investigation can assign Copilot to an issue other than the dispatched one. Restrict the safe output to the supplied issue number.

    target: "*"

[ROHIT SINGHAL (rohitsinghal4u)]

Thanks for the thorough review — addressing each finding:

1. Fabricated worked example (issue-investigation.md:310) — FIXED. Verified: azure-core/paging.py:79-80 does already guard against the None continuation-token case I claimed was missing. Replaced the example with a real, verified bug I found and confirmed by direct source inspection: an etag operand-order bug in azure-appconfiguration's set_configuration_setting (etag=configuration_setting.etag or etag instead of etag=etag or configuration_setting.etag). See commit 1b600f3.

2 & 5. Stale gh-aw compiler version (v0.81.6) on both lock files. Confirmed valid — this repo's other workflows are on v0.87.1+, and the older compiler's Copilot CLI invocation pattern doesn't copy the binary to a mounted path first. I attempted to upgrade my local gh-aw CLI to match, but the v0.87.10 download was blocked by Windows Defender as a false-positive in my environment, and I'm not going to work around an AV block. This needs a maintainer with a clean v0.87.1+ install to recompile both lock files before merge — flagging clearly rather than shipping something I couldn't verify.

3. Version Currency wording ("only available for the latest package version"). This is intentional and inherited verbatim from the already-merged .NET workflow (Azure/azure-sdk-for-net PR #60514), where it was explicitly requested by that workflow's maintainer as a mandatory decision point rather than soft guidance. I don't think this Python PR should unilaterally soften wording that was deliberately hardened upstream — but happy to revisit if this repo's policy genuinely differs, that's a maintainer call.

4. Exclusion routing ambiguity (noop vs. "Requires a human"). Confirmed this same tension exists verbatim in the current .NET source too — it's not something introduced by this port. Reading intent: well-evidenced excluded cases → "Requires a human" (post the analysis); cases lacking evidence → noop. Open to tightening the wording, but wanted to flag this is a pre-existing ambiguity, not new.

6. Wildcard target: "*" on close-issue/assign-to-agent. Also inherited verbatim from .NET's current production config. Valid architectural question — today the safety boundary relies on the prompt instructing the agent to only act on the dispatched issue_number, not on safe-output-level enforcement. I didn't find a gh-aw option to bind target to a dynamic issue number; if one exists, happy to apply it here and suggest it back upstream to .NET too.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 4 changed files in this pull request and generated 1 comment.

Suppressed comments (3)

.github/workflows/issue-investigation.md:129

• This is stricter than the Azure SDK support policy: an Active major version is fully supported, while customers are only encouraged to use the latest compatible release. The Version Currency rule consequently short-circuits valid reports from any older minor/patch and tells supported customers that support only applies to latest. Reframe currency as diagnostic guidance and reserve support-based deferral for lifecycle/compatibility cases rather than every non-latest version.

## Support Policy Expectation

By policy, Azure SDK support is only available for the latest package version. Version currency is a mandatory decision point, not just background guidance.

.github/workflows/issue-investigation.md:224

• For confirmed SDK-side issues that hit an exclusion, this rule directs the agent to noop, but the Comment Format explicitly defines Requires a human. Analysis provided below for those same cases. As written, complex but well-evidenced SDK issues can silently disappear instead of giving maintainers the prepared analysis. Emit the human-review outcome for confirmed SDK-side exclusions, while retaining targeted requests/noop only when ownership or required evidence is unresolved.

Do not assign Copilot, even if the above are met, when the issue requires any of the following. Use `noop` or a targeted Insufficient Context request instead:

.github/workflows/issue-triage.lock.yml:1

• This regenerated lockfile is also pinned to gh-aw v0.81.6 and retains the older direct /usr/local/bin/copilot invocation. Recompile issue-triage.md with the current gh-aw release so it uses the repository's mounted Copilot CLI pattern and commit the resulting lockfile; generated YAML should not be patched manually.

# gh-aw-metadata: {"schema_version":"v4","frontmatter_hash":"05b306844fe731bbc51742814306b173e76912201a7e90c5783c248d61a8abd5","body_hash":"dc8a877d119de2565e69737741e0463fd731dbe25d0159ee257cf166e3793683","compiler_version":"v0.81.6","strict":true,"agent_id":"copilot","engine_versions":{"copilot":"1.0.65"}}

[Copilot]

🟡 Changes recommended

The workflows contain unsafe wildcard targets, an incorrect support-policy gate, contradictory outcome instructions, and outdated generated locks.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (5)

.github/workflows/issue-investigation.md:224

• This makes the documented Requires a human outcome unreachable: exclusions are told to emit noop, while lines 260 and 279 say those same exclusions should produce the human-analysis comment. Confirmed SDK-side issues requiring API, security, reliability, or broad design work will therefore receive no investigation result. Route confirmed exclusions to the human outcome and reserve noop for cases lacking enough evidence.

Do not assign Copilot, even if the above are met, when the issue requires any of the following. Use `noop` or a targeted Insufficient Context request instead:

.github/workflows/issue-investigation.md:42

• The workflow already has one trusted target in github.event.inputs.issue_number, but target: "*" allows model output influenced by untrusted issue text to close any issue in the repository. Restrict this destructive safe output to that input (and preferably require customer-reported) so the safe-output boundary contains a misdirected action.

    target: "*"

.github/workflows/issue-triage.lock.yml:1

• This regenerated lock also remains on gh-aw v0.81.6 and invokes /usr/local/bin/copilot rather than the mounted ${RUNNER_TEMP}/gh-aw/bin/copilot pattern emitted by the repository's v0.87.1 workflow. Recompile issue-triage with the repository-standard current gh-aw before merging so the newly chained triage run does not fail at agent startup.

# gh-aw-metadata: {"schema_version":"v4","frontmatter_hash":"05b306844fe731bbc51742814306b173e76912201a7e90c5783c248d61a8abd5","body_hash":"dc8a877d119de2565e69737741e0463fd731dbe25d0159ee257cf166e3793683","compiler_version":"v0.81.6","strict":true,"agent_id":"copilot","engine_versions":{"copilot":"1.0.65"}}

.github/workflows/issue-investigation.md:129

• The Azure SDK support policy does not limit support to the latest point release: Active major versions are fully supported, while the latest compatible release is only encouraged because it receives fixes. Making this a mandatory early exit will incorrectly reject valid reports against supported older minor/patch releases. Base the gate on lifecycle status and known fixes instead of exact version currency.

By policy, Azure SDK support is only available for the latest package version. Version currency is a mandatory decision point, not just background guidance.

.github/workflows/issue-investigation.md:39

• This wildcard lets untrusted issue content redirect the investigation comment to an unrelated issue. Scope add-comment to ${{ github.event.inputs.issue_number }} so output can only affect the issue selected by the dispatcher.

    target: "*"

• Files reviewed: 3/4 changed files
• Comments generated: 1
• Review effort level: Balanced

[ROHIT SINGHAL (rohitsinghal4u)]

Update after a second automated review pass surfaced 3 more real findings — fixed all 3:

1. get_issue → issue_read. Confirmed via my own earlier successful test runs that the model was self-correcting to the real tool name — but that's fragile, not something to rely on. Fixed the prompt to reference the actual tool.
2. Wildcard target: "*" on add-comment/close-issue/assign-to-agent. Took the suggested fix and verified it actually works: scoped all three to ${{ github.event.inputs.issue_number }}, confirmed in the compiled lock output (not silently ignored).
3. Exclusion routing contradiction. You're right that this was a real bug, not just an inherited quirk worth leaving — "noop" and "Requires a human" were both instructed for the same condition, making the human-analysis outcome unreachable. Fixed: well-evidenced exclusions now route to "Requires a human," noop/Insufficient-Context reserved for when the exclusion itself lacks evidence.

Still open, honestly:

• gh-aw compiler version (both lock files). Still couldn't safely upgrade past v0.81.6 in my environment (Defender false-positive on v0.87.10). This is a real, functional risk per your note that triage could fail at agent startup — needs a maintainer with a clean install to recompile before merge.
• Version Currency wording ("support only for latest package version"). I'm intentionally leaving this matching .NET's current, deliberately-reviewed wording rather than unilaterally softening it — but I recognize you've now flagged this twice with specific reasoning (lifecycle/major-version vs. exact-version-currency). This is a genuine policy question, not a text bug, so I don't think it's mine to resolve alone — flagging for whoever owns Python SDK support-policy interpretation.

[Copilot]

🟡 Changes recommended

The support-policy gate is inaccurate, closed issues are not excluded, and both lockfiles require recompilation with current gh-aw.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (3)

Previously missed (1) — in code that hasn't changed since the last review.

.github/workflows/issue-investigation.md:96

• The handoff gate never checks that the issue is still open. Because the investigation is queued after triage, an author or maintainer can close the issue before this run starts; that closed issue still passes these label checks and can receive a new automated comment or Copilot-assignment attempt. Require the retrieved issue state to be open so stale/manual dispatches no-op.

.github/workflows/issue-investigation.md:129

• This version gate is based on an incorrect support-policy premise. The Azure SDK lifecycle policy says an Active major version is fully supported and only encourages customers to use the latest compatible minor/patch release; it does not make every older point release unsupported. As written, the workflow will stop investigation and send supported customers away whenever their point release is not the newest. Please rewrite this section and the Version Currency rule to treat lifecycle status as the support boundary, while recommending an upgrade when the issue may already be fixed.

By policy, Azure SDK support is only available for the latest package version. Version currency is a mandatory decision point, not just background guidance.

.github/workflows/issue-triage.lock.yml:1

• This lockfile was regenerated with gh-aw v0.81.6, so it retains the old hard-coded /usr/local/bin/copilot sandbox invocation (line 938). The repository's v0.87.1-generated workflow instead resolves Copilot from PATH, copies it into ${RUNNER_TEMP}/gh-aw/bin, mounts that directory, and invokes the copied binary (mgmt-sdk-pr-review.lock.yml:772-809); without that pattern, the installed CLI may not be available in the sandbox. Recompile this workflow with the repository's current gh-aw version before merging.

# gh-aw-metadata: {"schema_version":"v4","frontmatter_hash":"05b306844fe731bbc51742814306b173e76912201a7e90c5783c248d61a8abd5","body_hash":"dc8a877d119de2565e69737741e0463fd731dbe25d0159ee257cf166e3793683","compiler_version":"v0.81.6","strict":true,"agent_id":"copilot","engine_versions":{"copilot":"1.0.65"}}

• Files reviewed: 3/4 changed files
• Comments generated: 1
• Review effort level: Balanced

[Copilot]

Copilot review overview

🟡 Changes recommended

Investigation dispatch checks state before the triage safe-output mutations are applied, preventing valid handoffs.

Review effort: Balanced
Findings: 2 · 2

Open (4)

• Dispatch check cannot observe buffered triage labels · New
• This generated workflow still invokes /usr/local/bin/copilot directly inside the AWF sandbox. The…
• Recompiling triage with v0.81.6 regresses it to the old sandbox launch path that invokes…
• This generated lock uses gh-aw v0.81.6 and invokes /usr/local/bin/copilot inside the sandbox…

Resolved since last review (6)

• The pinned GitHub MCP server exposes issue_read, not get_issue, and its v1.4.0 deprecated-alias…
• These wildcard targets allow issue-derived agent output to comment on or close any issue in the…
• The assignment output is also authorized for every issue, although this run is intended for exactly…
• This exclusion path says to noop, but the comment-format contract below says the `Requires a…
• This overstates Azure SDK policy and makes the workflow defer every report on an older minor/patch…
• This concrete example asserts a bug that the current code disproves: azure/core/paging.py:79-80…

[Copilot]

Copilot review overview

🟡 Changes recommended

Independent safe-output processing can close or assign issues without the required analysis comment, and assignment receipts are not fully bound to the target issue.

Review effort: Balanced
Findings: 1 · 2

Open (3)

• Assignment receipt may belong to another issue · New
• Issue may close without required explanation · New
• Agent may be assigned without investigation context · New

Resolved since last review (4)

• Dispatch check cannot observe buffered triage labels
• This generated workflow still invokes /usr/local/bin/copilot directly inside the AWF sandbox. The…
• Recompiling triage with v0.81.6 regresses it to the old sandbox launch path that invokes…
• This generated lock uses gh-aw v0.81.6 and invokes /usr/local/bin/copilot inside the sandbox…

[Copilot]

Copilot review overview

🔵 Needs a closer look

The workflow can close customer issues and initiate coding-agent work, so final human review is warranted despite the comprehensive safeguards and tests.

Review effort: Balanced
Findings: 1 · 2

Open (3)

• Assignment receipt may belong to another issue
• Agent may be assigned without investigation context
• Issue may close without required explanation

[Copilot]

Copilot review overview

🟡 Changes recommended

Best-effort assignment currently fails on expected credential errors, and comment pagination can miss prior investigations.

Review effort: Balanced
Findings: 3

Open (3)

• Accept completed_with_skips for best-effort authentication failures · New
• Paginate issue comments before checking for prior investigations · New
• Use the hyphenated ignore-if-error configuration key · New

Resolved since last review (3)

• Assignment receipt may belong to another issue
• Agent may be assigned without investigation context
• Issue may close without required explanation

[Copilot]

Copilot review overview

🟡 Changes recommended

Runtime assignment receipts lack the number field required by the handoff validator, blocking valid single-owner investigations.

Review effort: Balanced
Findings: 1

Open (1)

• Assignment receipt field mismatch blocks single-owner handoffs · New

Resolved since last review (3)

• Use the hyphenated ignore-if-error configuration key
• Paginate issue comments before checking for prior investigations
• Accept completed_with_skips for best-effort authentication failures

[Copilot]

Copilot review overview

🔵 Needs a closer look

The elevated assignment credential can cause preflight reads to fail before best-effort error handling applies.

Review effort: Balanced
Findings: None

Resolved since last review (1)

• Assignment receipt field mismatch blocks single-owner handoffs

Previously missed (1)

In code that hasn't changed since last review

Use workflow token for preflight reads to preserve best-effort auth

.github/​workflows/​issue-investigation.md:164

The assignment credential is also being used for the helper's issue/comment reads. If GH_AW_AGENT_TOKEN is expired or lacks repository read access, prepareAssignment fails before the native handler can apply ignore-if-error, so the documented best-effort authentication behavior becomes a hard job failure. Use the normal workflow token for preflight reads and pass the elevated credential separately through the handler's supported GH_AW_ASSIGN_TO_AGENT_TOKEN variable, matching the generated safe-outputs job.

[Copilot]

Copilot review overview

🔵 Needs a closer look

It introduces issue-closing and agent-assignment automation whose full live gate matrix remains unverified.

Review effort: Balanced
Findings: None

[Copilot]

🔵 Needs a closer look

Privileged workflow chaining and model-driven closures need human review, especially with unresolved correctness issues and incomplete live validation.

3 open issues

• Preserve prototype getters in GitHub context overrides · New
• Canonicalize issue numbers before concurrency grouping · New
• Disable failed-job issue reporting in generated workflows · New

_{Review effort: Balanced}