Arkiv 0.3.1: hardened context hook and import check

Plugin 0.3.1 hardens the local SDK context hook and the installed-import check added in 0.3.0. On ordinary projects the agent sees the same SDK context as before.

Fixes

Each fix began as a failing test.

• Context text is fixed. The hook injects only text made entirely of its own fact lines, whatever answers on the session's local IPC name. Installed versions are measured only up to six digits per part.
• Import check never echoes unusual imports. It reports unmeasured for any SDK import that is not a plain package subpath with short names. Re-export reads are bounded: each declaration file is read once, at most 128 per check. Canonical skills Arkiv-Network/skills@cd8bb37 (#63).
• Four edge cases are now handled correctly:
  • A hook request retried after a lost reply still delivers its change.
  • A declaration-only change re-states the declaration count.
  • A workspaces root (a workspaces field, pnpm-workspace.yaml or lerna.json) with no SDK dependency of its own is unmeasured instead of absent.
  • An SDK found above the lockfile's folder is no longer confirmed by that lock.
• The Codex manifest now carries the shared keywords.

Evidence

• 88/88 tests on Linux ARM64. 87/88 on Windows, where one Linux-only test is skipped. claude plugin validate --strict passes.
• Comparison with 0.3.0 on ordinary inputs:
  • 320 comparisons on both hosts, with and without the session process, using six SDK fixtures and eight import-check files.
  • The injected text is identical in all of them, so the October 16 ablation still applies.
• Public installs of 0.3.1 in isolated Windows profiles match the repository packages exactly: 347 Claude files and 47 Codex files. Both updated from 0.3.0.
• Native sessions from those installs, run on Claude Code 2.1.302 and Codex CLI 0.163.0:
  • Claude: 54/54 checks. They cover every category, file and directory changes, compaction, /clear, resume, the import check, 8 concurrent sessions, missing Node and the connection header.
  • Codex: 20/20 checks. They cover every category, resume, the untrusted-hook gate, cleanup and the header.
  • Long session: one Claude session ran for 11 minutes, past the 600-second timeout of the asynchronous hook that starts its session process. That process stayed up and delivered a later SDK downgrade correctly.
  • No production traffic: Claude ran with no MCP servers, and Codex used a local mock.

Update

• Claude Code: claude plugin marketplace update arkiv, then claude plugin update arkiv@arkiv, then restart.
• Codex CLI: codex plugin marketplace upgrade arkiv, then codex plugin add arkiv@arkiv, then start a new session. If you have not trusted the Arkiv hooks yet, do it once with /hooks.

Arkiv 0.3.0: local SDK context

Plugin 0.3.0 adds local SDK context for Claude Code and Codex CLI, plus the canonical installed-import check.

What's new

• Local SDK context hooks (both hosts). At session start and on later prompts, the plugin reads the workspace's package.json, npm lockfile and installed @arkiv-network/sdk declarations, then tells the agent one category: compatible, out_of_range, not_installed, mismatch or absent. It names the supported range >=0.8.1 <0.9. It never imports project code. It makes no HTTP, TCP or DNS requests and writes no cache or identifiers. Supported on Windows and on Linux with Node 20.8+. macOS stays unmeasured.
• Codex native manifest. The Codex package now uses .codex-plugin/plugin.json, so Codex loads its hooks. Review and trust them once with /hooks.
• Canonical skills Arkiv-Network/skills@49c0e06 (#62). In Claude Code, the router's installed-import check runs after Write/Edit on files that import the SDK. For an SDK outside the supported range, the router proposes an explicit upgrade plan instead of adapting newer examples.
• Skill-usage collection stays disabled.

Evidence

• Context-only ablation, October 16. Each host ran 73 cases × 3 repetitions per arm in installed-SDK fixture workspaces, graded by a blind Opus 5.5 judge with 3 votes per answer:
  • Claude Code: 218/219 with context vs 211/219 without; 0 pairs worse, 7 better.
  • Codex CLI: 207/219 vs 204/219; 2 pairs worse, 5 better. Neither worse pair is attributable to the context.
• 78/78 tests on Linux ARM64 and 77/77 on Windows. claude plugin validate --strict passes.
• Installation was checked in isolated profiles:
  • Fresh public installs match the repository packages: 49 Claude files and 47 Codex files.
  • Updates from 0.2.1 succeed on both hosts.
  • Native sessions from the installed 0.3.0 packages received compatible for a 0.8.1 fixture and out_of_range for a 0.6.8 fixture on both hosts. The MCP was suppressed or mocked, so no production traffic was sent.

Update

• Claude Code: claude plugin marketplace update arkiv, then claude plugin update arkiv@arkiv, then restart.
• Codex CLI: codex plugin marketplace upgrade arkiv, then codex plugin add arkiv@arkiv, then start a new session and trust the Arkiv hooks once with /hooks.

Arkiv 0.2.1: anonymous connection attribution

Adds a shared package-version and host header to the existing Arkiv MCP connection. Claude uses a local fixed-header helper and supports ARKIV_TELEMETRY=off; Codex sends a documented static header. No environment value, installation/session ID, local path, prompt, SDK state or skill-use counter is transmitted by this attribution.

Skills, project rules and context hooks are unchanged. The helper reads no files, performs no network request and adds no agent context. Node 20+ is required for the Claude rules hook and header helper. Existing MCP operation telemetry and hosting logs have their own privacy limits.

Verified: 39 serial tests, generated/import/static checks, strict Claude validation, actual Claude/Codex header connections against a local fixture, and anonymous public branch installation with exact bundle-byte parity in fresh isolated profiles. The required user plugin configurations and Codex config are unchanged; an additional global Claude metadata hash changed during the window and remains unattributed. The companion MCP passes classified preview response comparisons and uploaded-source verification; production promotion follows separately.

New context hooks, a fresh complete hook ablation, production skill-use collection and dashboard panels remain pending. See the README privacy section and the preserved historical qualifications in docs/verification.md.

Arkiv plugin 0.2.0 — Claude Code and Codex CLI

Version 0.2.0 adds a Codex CLI package beside the Claude Code package. Both share the arkiv identity and version, install from this repository and bundle the existing Devfolio (Arkiv: Global Tour Stop) MCP profile at https://mcp.arkiv.network/devfolio. No connection key is required.

Claude Code

New installation:

claude plugin marketplace add Arkiv-Network/claude-plugin
claude plugin install arkiv@arkiv

Update from 0.1.6:

claude plugin marketplace update arkiv
claude plugin update arkiv@arkiv

Restart Claude Code. Confirm claude plugin list shows arkiv@arkiv enabled at 0.2.0, then run the README check prompt: a successful server_status call should identify devfolio.

Codex CLI

codex plugin marketplace add Arkiv-Network/claude-plugin
codex plugin add arkiv@arkiv

Start a new Codex session and confirm the version with codex plugin list --marketplace arkiv --json. Invoke the router with $arkiv. The Codex package uses streamable-http, has no Claude hooks and does not add project rules unless you ask for them.

Source

Both packages import the same skills as real files from merged skills revision f87c2f5 (Arkiv-Network/skills#57). The source lock binds both destinations by Git blob, SHA-256 and size; license declarations are preserved.

Verification and limits

Public installation from main (b99f18f) passed on Windows in isolated profiles: Claude Code 2.1.296 and Codex CLI 0.162.1 each install 0.2.0, enabled. Both installed caches are byte-identical to plugins/arkiv and plugins/arkiv-codex, and both MCP configurations point to /devfolio. Habitual settings and plugin registries stayed unchanged. This public check opened no session and made no MCP call.

The pre-publication candidate passed build, import, static and serial tests, strict validation, an actual 0.1.6 → 0.2.0 Claude update, native sessions for both clients against a loopback MCP fixture, and fresh installs on Linux ARM64. Details and retained failures are in the verification scope.

Live MCP connectivity is unverified. Codex desktop, IDE and web, macOS and Linux x86-64 are unverified; native Codex sessions used --no-daemon. The existing Claude evals were not rerun as a new complete ablation, and there is no Codex eval certification.

Arkiv Claude Code plugin 0.1.6 — Devfolio profile

The Arkiv plugin now bundles the Devfolio (Arkiv: Global Tour Stop) MCP profile at https://mcp.arkiv.network/devfolio. It includes Arkiv building guidance and the event sources. Plugin users do not need a second connection.

Update from 0.1.5

Run these commands in order:

claude plugin marketplace update arkiv
claude plugin update arkiv@arkiv

Restart Claude Code. Run claude plugin list and confirm arkiv@arkiv is enabled at 0.1.6. Use the README check prompt: a successful server_status call should identify devfolio. Review and remove separately installed Arkiv skills and a separately added Arkiv Devfolio connection to avoid duplicates; agents must not remove or switch connections on their own.

New installation

claude plugin marketplace add Arkiv-Network/claude-plugin
claude plugin install arkiv@arkiv

Canonical guidance is imported from merged skills revision 6c27e1d. The URL guard rejects credentials, query strings, fragments and other profile paths. The standalone MCP pointer, synthetic evals and consumer instructions match the bundled connection. The reference-check command now supports the nested plugin layout without editing imported source bytes.

Verification and limits

Generated-source, canonical-import, static and reference checks pass. All 34 serial package tests and both strict CLI validations pass. An actual anonymous public-main 0.1.5 → 0.1.6 update passed with the two commands above in an isolated Claude configuration; the updated cache is enabled at 0.1.6 and its MCP URL is /devfolio. Habitual settings and plugin inventories remain unchanged. Management commands did not start a production MCP session.

Six affected behavioral cases ran three times per arm with read-only mocks and --no-publish. Raw outcomes are 17/18 WITH and 3/18 WITHOUT. One grounded retrieval answer has retained false failing grader votes; direct review accepts it against every unchanged clause and the successful tool trace. All WITH process checks pass; no unmocked call occurred. This is bounded, qualified evaluation, not a perfect raw score or a fresh whole-suite claim. Earlier failures remain in the verification history.

Tested on Windows with Claude Code 2.1.296. Synthetic checks do not establish production MCP health, signing, live-chain correctness or macOS/Linux behavior. No MCP service or Hub deployment is included. The Hub guide remains a separate local preview.

Arkiv Claude Code plugin v0.1.5

Arkiv Claude Code plugin v0.1.5

Release commit: 9269f21; accepted candidate: 516c397; plugin tree: 5113e8278b57eefaa8df90669445bd954cb4ae1f.

Qualified source coverage: 73 cases, 219 usable WITH and 219 complete WITHOUT observations; 21 final-source cases and 52 cases retained through exact consumed-source equivalence. 0 baselines are fresh and 73 are historical exact-input observations. This is evidence across revisions, not a fresh whole-bundle causal ablation. Native raw FAIL votes and separate Root adjudications are preserved. Earlier product and provider failures retain their original results.

Public default-branch installation completed all 15 management commands with three exact installed-cache inventories. Installed-cache onboarding completed exactly two membership sessions with the verbatim README prompts, the pinned model and synthetic loopback MCP. Habitual configuration and source-auth hashes remain unchanged.

Raw outcome passes: 210/219 WITH versus 128/219 WITHOUT; raw outcome delta +37.44 percentage points. Ten raw grader failures (nine outcome and one process) were separately accepted by Root against the unchanged criteria and actual answers/tool traces. Their original scores, votes and FAIL results remain preserved. Separately accepted outcome delta: +41.55 percentage points. All 16 negative cases show zero Arkiv Skill triggers. These results do not guarantee every future response.

The 34 deterministic tests and 31 SDK checks are retained through exact executable-fence, oracle and dependency equivalence; they were not repeated on this version. Source-specific import, strict manifests and package checks passed on the accepted final source. Scope: Windows, Claude Code 2.1.296 and claude-opus-5-5. No General production-health, macOS/Linux or live-chain correctness claim. No funded signing or Hub deployment; the Hub remains localhost only.

Verification scope.