Skip to content

fix(release-tooling): redact custom MCP-use homes and clarify Java planning - #569

Merged
nik13 merged 5 commits into
devfrom
fix/release-tooling-release-qualify-1104
Nov 4, 2026
Merged

nik13 merged 5 commits into
devfrom
fix/release-tooling-release-qualify-1104

Conversation

@nik13

@nik13 nik13 commented Nov 4, 2026 •

Copy link
Copy Markdown
Contributor

What was wrong

  • MCP-use's always-on local-path scrubber recognized only a fixed list of filesystem roots. Tool-error status and exception.message retained synthetic home paths under /work, /data, and custom mounts. This predates test: compare package versions with the manifest instead of a literal #497: the scrubber originated in adaee832; test: compare package versions with the manifest instead of a literal #497 only changed MCP-use's packaging-version assertion.
  • The release guide incorrectly said completion required every current manifest version. complete_release.py instead checks the versions changed by each merged release PR, at the versions that PR released.
  • The Java checklist incorrectly described an all-artifact planning rule. Historical reconciliation uses the parent POM; verification after a fresh deployment requires every reactor POM.

Fix

  • Resolve and escape the actual POSIX home prefix at redaction time, with path boundaries. Preserve ordinary URLs, API paths and sibling prefixes; do not turn root HOME into a catch-all scrubber.
  • Parameterize exported-span and direct-redaction regressions across conventional/custom homes, including spaces and regex metacharacters, exact-home matches and changed/trailing-slash HOME.
  • Correct both release-guide sections, including fail-closed completion when historical versions cannot be read or required releases/registry versions are missing. No release workflow or planner behavior changes.
  • Retain the follow-up MCP-use fixes already on this PR: supported LangChain dependency floor, fresh-floor CI, cache-write/TTL usage, model-identifier privacy, and synthetic-home isolation. The resumed qualification's new commit changes only RELEASING.md.

Fresh cloud validation

All execution used fresh E2B sandboxes with synthetic credentials, scripted LLMs, and loopback MCP/provider/OTLP/registry servers. GitHub release-completion writes were intercepted; no registry publication or GitHub release/label mutation ran.

  • q1104-rel-tooling-r1-recheck-repro: new privacy regressions against origin/dev's original _text.py produced 6 expected failures / 3 passes. Restoring the PR implementation passed 9/9. A loopback registry probe proved per-PR completion ignores later current versions and unrelated unpublished packages, while its documentation contract failed on the stale paragraph. Independent Java probes reproduced the old checklist mismatch and verified parent-only planning plus strict fresh-deployment checks; 6/6 release-followup tests passed.
  • q1104-rel-tooling-r1-recheck-full: Python 3.12.16, mcp-use 1.7.1 and LangChain Core 1.6.12. 226/226 passed in each of four full-suite runs: core 1.1.0 under /home/synthetic, /work/synthetic, and /data/synthetic, then repository core 1.2.2 under /work/synthetic (904 passes, zero failures/errors/skips). The corrected completion documentation contract, independent loopback completion/Java probes, and 12/12 release-followup/completion unit tests passed. pip check passed for both core configurations; upstream deprecation warnings remain.
  • Final-head CI at 7c75ce72343e8ed808dd36bd20baa2ace62d3af1: 12 passed, 12 non-applicable jobs skipped, zero failed/pending, including fresh dependency-floor jobs on Python 3.11/3.12/3.13, package tests, core tests, and integration/release validation gates. All checks completed within the 25-minute wait.
  • Both sandboxes' deletion receipts and JUnit/log evidence are retained under the run labels. Local checks were limited to git/static inspection.

Competitor notes

Documentation comparison, not a live head-to-head benchmark or blanket release sign-off:

  • Braintrust distinguishes ingestion redaction from SDK masking/customizers that remove content before upload. This fix protects that pre-export privacy boundary.
  • Langfuse masking supports application-side masking and OTel attribute transformation before export. Our regressions inspect exported-span status and exception events, not just configuration acceptance.
  • Arize OpenInference TraceConfig exposes input/output/message privacy controls. MCP-use retains its TraceConfig controls and additional always-on local-identity scrub; custom HOME locations must not bypass it.

Target: dev only. Leave unmerged for parent qualification. No dev-to-main, registry publication, release-workflow edits, settings changes or secret changes.

🤖 Generated with Claude Code

nik13 and others added 5 commits November 3, 2026 22:43
…anning

Resolve the actual POSIX home at redaction time with whole-path boundaries, preserving ordinary URLs, API routes and sibling prefixes. Cover custom mounts in exported spans and document parent-only historical planning separately from strict fresh-deployment verification.

Co-Authored-By: Claude Code <noreply@anthropic.com>
Raise the published LangChain Core floor to 1.6.12 and test fresh resolutions on Python 3.11, 3.12 and 3.13 without pinning transitive dependencies. Preserve normalized cache-creation token counts through OTLP export, including zero, all privacy combinations and all agent entry points.

Co-Authored-By: Claude Code <noreply@anthropic.com>
Prefer provider cache-write TTL counts over generic cache_creation while preserving both billing tiers and inclusive token totals. Apply the shared privacy cleaner to model and tool identifiers before truncation.

Exercise real ChatAnthropic HTTP/SSE responses through MCPAgent, exported usage and hidden-content configurations; keep provider test dependencies compatible with the advertised floor.

Co-Authored-By: Claude Code <noreply@anthropic.com>
Use the real stdio MCP fixture in provider regressions so non-root CI workers never try to create a token store under /work or /data. The provider HTTP/SSE requests and home-path assertions remain unchanged.

Co-Authored-By: Claude Code <noreply@anthropic.com>
Match complete_release.py: acknowledge the versions changed by each merged release PR, require GitHub releases and registry availability, and retain fail-closed behavior when historical version information is unavailable.

Co-Authored-By: Claude Code <noreply@anthropic.com>
@nik13
nik13 merged commit e070e25 into dev Nov 4, 2026
34 checks passed
@nik13
nik13 deleted the fix/release-tooling-release-qualify-1104 branch November 4, 2026 11:57
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant