[blink1073]

PYTHON-6178

Changes in this PR

• The nightly fuzz task now harvests its coverage-increasing inputs into the committed fuzz/corpus/ and opens an automated [Fuzz Corpus] PR against main, so later nights build on prior discoveries instead of restarting from the spec seed corpus.
• Harvest dedupes by content digest against a baseline manifest recorded at setup, caps growth at 50 units per run and 64 KiB per unit (overridable from Evergreen), and treats a run that finds nothing new as a clean no-op.
• Crashing inputs are excluded automatically: a crashing target fails the task before the harvest command runs, so reproducers keep the existing fuzz/regression/ triage flow.

Test Plan

• just test test/test_fuzz_harvest.py: 8 unit tests for the harvest module (digest diffing, caps, summary, no-op), written first.
• just setup-tests fuzz fuzz writes a 485-digest baseline manifest, and a unit committed in fuzz/corpus/ reaches both the scratch corpus and the baseline.
• Harvest guards: patch builds skip (patch run), an empty harvest exits 0.
• Evergreen patch on fuzz-regression-rhel8: task green; seeding, baseline manifest, and target replay verified on a real host (patch run).
• Evergreen patch of the nightly fuzz task, using a temporary in-patch guard that runs the harvest without PR creation: all three targets ran 60s crash-free, and the harvest kept 50 units while skipping 17 oversize and dropping 58 over the per-run cap (patch run).
• The GitHub-facing PR creation needs the App secrets, so the first real nightly after merge exercises it.

Checklist

Checklist for Author

• Did you update the changelog (if necessary)? (The repo has no changelog file.)
• Is there test coverage?
• Is any followup work tracked in a JIRA ticket? If so, add link(s). (None filed.)

Checklist for Reviewer

• Does the title of the PR reference a JIRA Ticket?
• Do you fully understand the implementation? (Would you be comfortable explaining how this code works to someone else?)
• Is all relevant documentation (README or docstring) updated?

[Copilot]

🟡 Changes recommended

The nightly harvest’s comment-bot bootstrap lacks assumed-role AWS credentials, preventing it from retrieving GitHub App secrets and creating a PR when changes are harvested.

1 open issue

• Pass assumed-role AWS credentials to the comment-bot subprocess · New

What changed in this PR

Persists coverage-increasing BSON fuzz inputs through automated corpus PRs, allowing future nightly runs to build on earlier discoveries.

Changes:

• Seeds nightly fuzzing from the committed corpus and records baseline digests.
• Adds bounded, digest-based harvesting and automated GitHub PR creation.
• Adds harvest unit tests and documentation, plus temporary Markdown formatter exclusions.

File	Description
test/​test_fuzz_harvest.py	Tests digest filtering, caps, summaries, and no-op behavior.
pyproject.toml	Temporarily excludes two Markdown files from Ruff.
fuzz/​README.md	Documents nightly harvesting, limits, and review guidance.
fuzz/​harvest_corpus.py	Selects new corpus units and writes harvest summaries.
fuzz/​corpus/​.gitkeep	Adds a directory for committed corpus inputs.
CONTRIBUTING.md	Links contributor guidance to the harvest documentation.
.evergreen/​scripts/​setup_tests.py	Seeds committed inputs and records baseline digests.
.evergreen/​scripts/​harvest-fuzz-corpus.sh	Runs harvesting with patch-build and empty-result guards.
.evergreen/​scripts/​generate_config.py	Adds harvesting after nightly fuzz tests.
.evergreen/​scripts/​create-fuzz-corpus-pr.sh	Commits harvested inputs and creates or updates a PR.
.evergreen/​generated_configs/​tasks.yml	Includes the generated nightly harvest command.

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

Harvested fuzz inputs are not replayed across all targets before publication, and spec-resync commits omit the public suffix list, risking untriaged crashes and incomplete or failed updates.

2 open issues

• Stage the public suffix list in spec-resync commits · New
• Pass assumed-role AWS credentials to the comment-bot subprocess

1 previously missed issue

In code that hasn't changed since last review

Replay harvested inputs against all targets before publishing

.evergreen/​scripts/​harvest-fuzz-corpus.sh:26

handle_fuzz() runs the targets sequentially against one shared corpus, so inputs discovered by later targets are not checked by targets that already finished. A green run can therefore publish an input that fails another target, bypassing the intended crash-triage flow. Replay the harvested files through all three targets in fixed-input mode and abort on failure before creating the PR.

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

The authenticated GitHub App push/PR flow remains untested and shares plumbing with spec resync, so credential bootstrap and create/update behavior need confirmation on Evergreen.

No open issues 🎉

2 resolved since last review

• Stage the public suffix list in spec-resync commits
• Pass assumed-role AWS credentials to the comment-bot subprocess

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

Within-run deduplication misses identical candidates, allowing duplicates to consume the harvest cap, exclude distinct inputs, and inflate the reported count.

1 open issue

• Deduplicate candidates within the same run · New

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

The GitHub App-authenticated publication flow needs maintainer validation with production secrets to confirm token acquisition and permissions, branch pushing, and PR creation/update end to end.

No open issues 🎉

1 resolved since last review

• Deduplicate candidates within the same run

_{Review effort: Balanced}