[truefoundry-saas]

Automated cherry-pick for release request b165bf01-e135-417d-9928-61a2f0d3d84a.

Source PRs (merged to main):

• AGE-2642: feat: accept x-tfg-tools so a turn can mount ask_user_secret #1194

Do not merge manually — this PR is merged automatically when the release request is approved. If you do merge it yourself, use "Sync status" on the request so it stops showing as pending.

---

Note

Medium Risk
Changes the turn startup path and tool advertisement for HTTP turns in TrueFoundry mode; malformed headers fail early with 400, and server-side execution is blocked for header tools.

Overview
Adds support for the x-tfg-tools request header in TrueFoundry mode: a JSON array of MCP tool definitions (including optional outputSchema and annotations) that are not persisted—callers must resend the header on resume turns.

Parsing and validation live in inlineTools.ts (size/count limits, duplicate names, MCP ToolSchema). A per-request ICapabilitiesStore is resolved when a turn starts (before a turn id exists) and wired through turns and schedules; scheduled runs omit the header and mount nothing. TurnResourceResolver gains optional rootCapabilities, which attaches extraCapabilities on the main thread only (not sub-agents). Header tools are exposed as client-side system tools (e.g. ask_user_secret) that defer execution to the client rather than running on the server.

LocalToolMCP.listTools now forwards outputSchema and annotations from local definitions so discovery/policy see the same metadata as inline header tools. OpenAPI version bumps to 0.182.0-rc.3.

^{Reviewed by Cursor Bugbot for commit de9f56d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[changeset-bot]

🦋 Changeset detected

Latest commit: de9f56d

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 2 packages

Name	Type
@truefoundry/trueforge	Patch
@truefoundry/trueforge-core	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[qodex-ai]

Qodex review

Reviewed units u1. The Fern specification version advances from 0.182.0-rc.2 to 0.182.0-rc.3.

Effort: ●●●●● (5/5) · No findings worth raising. The diff reads clean.

Since last review: 0 addressed, 0 new.

Scope: the changes since the last review (delta: only what changed since the previously reviewed commit 3f9b02a is reviewed; findings are anchored on the whole change).

Walkthrough

What Qodex checked

Lenses (openqodex review with checked answers, all dimensions)

• Bugs and correctness
• Security (auth, injection, SSRF, secrets, headers)
• Performance regressions
• Maintainability
• Style and consistency

Scanners

• actionlint did not run
• bandit did not run
• brakeman did not run
• cargo-deny did not run
• checkov did not run
• gitleaks 8.21.2
• golangci did not run
• hadolint did not run
• kube-linter did not run
• kubeconform did not run
• osv-scanner 2.6.0
• oxlint did not run
• rubocop did not run
• ruff did not run
• semgrep 1.94.0
• shellcheck did not run
• sqlfluff did not run
• sqllint did not run
• squawk did not run
• tflint did not run
• trivy did not run
• zizmor did not run
• 0 scanner candidates raised or dropped by the reviewer

Quality filters

• PR description: minimal (present: length; missing: checklist, issue link)
• Per-category confidence floor (bug/security/performance 0.7, maintainability 0.71, style 0.75)
• Severity threshold from .qodex.yaml: minor
• Path excludes (none configured in .qodex.yaml)
• Disabled rules (none configured in .qodex.yaml)

Verification (M1b probes against PR preview)

• Preview deployment discovered (no successful deploy for this commit)
• Probes (skipped, no preview to probe)

Cross-review dedup

• Delta review: delta: only what changed since the previously reviewed commit 3f9b02a is reviewed; findings are anchored on the whole change
• Carry-overs (no findings from earlier reviews still present)

Output

• Walkthrough comment posted
• Inline comments (no in-diff findings to attach)

---

Reviewer: shown 2 of 2 changed hunks, opened 2 files, 8 tool calls served, 6 model calls, 0 correction rounds, 21.9s. The library does not report tool calls asked apart from those it served, or a step cap.
Reviewed by Qodex · 46.0s

Every dimension is reviewed. Bugs, security, and performance post inline; low-severity style and convention notes are collected in the group above rather than dropped. Mark a finding resolved or flag a false positive in the Qodex Pull Requests view.