Harden Hermes Agency security, lifecycle, packaging, and Fabric

[DeployFaith]

Thinking Path

• Hermes Agency is the execution and coordination substrate, with Hermes Fabric providing its operator-facing control plane.
• The repository had a large mixed worktree spanning transport lifecycle, remote-task trust, model routing, skill governance, packaging, CI, Fabric reliability, policy, and documentation.
• Security-sensitive mutations and recovered queued tasks needed fail-closed authorization and current-trust revalidation.
• Keryx lifecycle ownership, Fabric filesystem/process handling, deterministic test partitioning, and package resource validation needed explicit regression coverage.
• The worktree also needed preservation, ownership classification, generated-artifact quarantine, focused commits, and reproducible evidence before publication.
• This pull request keeps the integrated baseline reviewable through nine focused commits while validating the final combined tree.
• The benefit is a safer, reproducible Agency/Fabric release candidate with deterministic CI, isolated package smoke tests, documented policy, and rollback evidence.

Linked Issues or Issue Description

No public issue exists. This PR addresses a repository-wide hardening and release-readiness gap:

• Observed behavior: remote task/pool trust checks, Keryx channel cleanup, profile filesystem writes, local process supervision, global Fabric routes, packaging gates, and test orchestration had incomplete failure and regression coverage.
• Expected behavior: security-sensitive operations fail closed; recovered work is revalidated; owned resources are cleaned up precisely; profile writes are contained and atomic; global routes require instance-admin authority; builds and tests run through deterministic bounded gates; release artifacts prove packaged resources work from an isolated wheel.
• Reproduction/evidence: the retained validation logs and ownership manifests are documented in docs/hermes-agency-remaining-work-plan.md.
• Scope: code, tests, CI/release workflows, model-set consolidation, governed shared-skill lifecycle, Fabric reliability/policy, and operator documentation. No deployment is included.

What Changed

• Consolidated Agency model routing into the canonical OpenAI Codex model set and aligned status tests/docs.
• Added governed shared-skill proposal, validation, promotion, storage, migration, CLI, and profile-scope lifecycle with focused tests.
• Made pool mutation authentication fail closed, propagated CLI tokens, and revalidated recovered queued tasks against current trust.
• Corrected Keryx channel ownership and failed-connection cleanup with lifecycle tests.
• Hardened Fabric profile path containment/atomic writes, local process supervision fallback, global route authorization, OpenAPI coverage, UI behavior, and environment-isolated tests.
• Repaired deterministic Fabric test partitioning and moved CI to explicit serialized/server/workspace lanes.
• Added frozen package installation, package/release validation, isolated wheel resource smoke tests, and workflow/Compose checks.
• Added tested UI JavaScript/CSS artifact budgets and excluded generated Storybook output from source-branding scans.
• Updated security, compatibility, operations, policy, QA, typing-baseline, and rollback documentation.
• Preserved the unrelated local desktop attachment and quarantined generated dependency/runtime backups rather than committing or deleting them.

Verification

Local validation completed on 2026-07-14:

• ruff check . and ruff format --check .
• mypy src/ --exclude '_generated'
• SDK: 406 passed, 33 deselected
• Agency canonical suite: 242 passed
• Full non-integration Python suite: 631 passed, 33 deselected
• Governance/default-staff focused suite: 64 passed
• Bridge contract suite: 11 passed
• Fabric frozen install, workspace/catalog preflights, full workspace typecheck, stable-runner partition tests, policy tests, and source scans
• Production build and artifact budgets: JavaScript 4,914,705 / 5,250,000 bytes; CSS 374,279 / 400,000 bytes
• Branding scan: 20,989 matches across 1,945 files; zero violations
• Isolated wheel build/install/import/CLI/resource smoke and pip check
• GitHub Actions YAML parsing, Docker Compose validation, committed-path hygiene, credential scan review, and git diff --check

Detailed execution evidence and rollback notes: docs/hermes-agency-remaining-work-plan.md.

Risks

• The PR is broad because the retained mixed baseline contained coupled Agency/Fabric/model/governance work. Focused commits preserve subsystem review boundaries.
• Canonical model-set consolidation removes legacy preset files; consumers using those filenames must migrate to openai-codex-only.
• Skill governance adds a new lifecycle and storage surface; it is covered by 33 focused tests plus the complete Python suite.
• Global Fabric operational routes now require instance-admin authorization and may reject callers that previously relied on permissive behavior.
• The UI bundle remains large; an explicit failing budget now prevents further unreviewed growth.
• The plugin directory is not claimed to pass strict Mypy. Its current 824-error diagnostic backlog is documented rather than suppressed.
• No deployment was performed. Runtime rollout and rollback remain separately approval-gated.

For core feature work, check ROADMAP.md first and discuss it in #dev before opening the PR. Feature PRs that overlap with planned core work may need to be redirected — check the roadmap first. See CONTRIBUTING.md.

Model Used

• OpenAI Codex gpt-5.6-sol, high-reasoning engineering mode, with repository/file tools, shell execution, test/build execution, Git, and GitHub CLI. Work was performed directly without Hermes Agency delegation.

Checklist

• I have included a thinking path that traces from project context to this change
• I have specified the model used (with version and capability details)
• I have checked ROADMAP.md and confirmed this PR does not duplicate planned core work
• I have searched GitHub for duplicate or related PRs and linked them above
• I have either (a) linked existing issues with Fixes: # / Closes # / Refs # OR (b) described the issue in-PR following the relevant issue template
• I have not referenced internal/instance-local Paperclip issues or links (only public GitHub #NNN / github.com/paperclipai/paperclip URLs)
• My branch name describes the change and contains no internal ticket id or instance-derived details
• I have run tests locally and they pass
• I have added or updated tests where applicable
• I have updated relevant documentation to reflect my changes
• I have considered and documented risks above
• All Paperclip CI gates are green
• Greptile is 5/5 with no open P2s, recommendations, or follow-ups
• I will address all Greptile and reviewer comments before requesting merge