Finish Agent Skill installs from trusted runtime receipts

[Open-Squilla]

An installation-only Agent turn now finishes from trusted built-in runtime receipts after all requested Skills succeed. The Agent keeps working on mixed requests, retains instructions added during installation, and accurately describes installed Skills that are unavailable or degraded.

Scope

Scope boundary: early installation-tool visibility, turn-local receipt reuse, batch completion, pending user input, and pinned-catalog descriptions. New user input allows failed installs to be retried while successful receipts remain reusable. Disabled, shadowed, needs_setup, and degraded results retain their availability explanations.

Branch

Base branch: main
Target exception: N/A

Updated against main after #1592 and #1593 merged.

Issue

Linked issue: None
If None, reason: finish scoped Skill installation turns reliably without losing remaining work or user instructions.

Release Note

Release note: Agent Skill installation ends once all requested installs succeed, preserves mixed-task work and new user instructions, and explains unavailable Skills accurately. Newly installed Skills become visible on the next turn. No version metadata change.

Tests

Ruff: full source and tests passed on the reviewed stack.
Pytest: 390 tests passed after merging current main, covering Agent installation completion, settlement, ToolContext compatibility, CI planning and workflows. Earlier reviewed-stack checks covered 80 installation and mid-turn input regressions.
Build: WebUI build and full-package mypy passed on the reviewed stack.
Regression tests: added for same-batch and sequential installs, partial failure, mixed tasks, input during fetch/result delivery, user-selected directory retries, risk-token validation, unavailable/degraded results, cancellation, and positional compatibility.
Notes: turn logic is platform-neutral. Updated-head CI provides platform acceptance; previous PR-head CI does not validate this update. Default tests remain offline, deterministic, credential-free, and fork-safe. No new live-model validation was performed for this update.

Maintainer Live Check

Maintainer live check: no
Surface: N/A

Safety

Finalization trusts only built-in runtime receipts. Skill document text and model-supplied retry parameters cannot authorize retries or finalization. Existing permission checks, risk approval, cancellation, Plan/goal constraints, and pinned-catalog behavior remain enforced. ToolContext appends the receipt field after existing published fields. No secrets, private transcripts, local artifacts, or machine-specific paths are included.

Third-Party Origin

Third-party origin: none
Original implementation; no third-party code copied.

[github-actions]

Merge queue CI: success — run 35134181420, attempt 1.
Tested queue candidate: b73694112a8b5e9efac652b8d15841ece2f6f6fc. This result belongs to that combined tree, not necessarily the PR's current head.

The run summary lists reused suites and executed checks. A green PR check alone does not mean queue validation has passed.