[callumflack]

Summary

Moksha builders can obtain free test credit automatically and resume setup after an interruption. Interrupted data-connection work can lose the context needed to continue safely: which owner and server are selected, whether a payment was already sent, or which consent request was created. Preserve that context and provide recovery commands that reuse the original operation.

Owner inspection now distinguishes Account sign-in from a Personal Server session, shows other accounts' servers separately, and retains network/environment/home context in server-start guidance. Paid reads persist the exact signed payment header before transmission and stop if recovery state cannot be saved. Consent creation saves a frozen intent and idempotency key before the request, exposes unfinished intents for inspection, and supports --resume across test-credit setup and consent with current grant checks.

The Research Shelf example also gains opt-in, metadata-only observations and a maintained browser/Motel verifier for durable save, render, acknowledgement, same-grant iteration and recovery. Missing observations remain unknown and fail verification.

Public Builder instructions now include a provider-independent delivery/evidence procedure in the installation manifest. Ordinary collection and app integration require no private Unity checkout, reference app or Motel. The optional Research Shelf contributor verifier uses those public rules; its doctor now selects the same SHELF_MOTEL_BINARY as launch. The doctor regression failed before the fix and passes afterward. The connector skill work from CLI #105 is included here; #105 is closed as superseded so this is the single CLI review and merge target. Private local history notes are excluded.

Agents can read the bundled workflows directly with vana skills show connect-data --json (or create-connector / builder). The response retains skill metadata and installed status and adds the full instructions plus required reference content. Reading does not install skills or write the skill cache or agent configuration. Public docs and CLI help expose this entry point, so the running agent can continue from an ordinary data request without choosing a harness or restarting a session.

Bare interactive vana offers four goals: Collect or manage my data, Use or share data, Build or test a connector, and Manage my Personal Server. The source picker shows 12 rows with an always-visible Ctrl+O shortcut; Enter on no results retains the source name for the connector-guide handoff. Discovery requires no sign-in. Before source credentials or provider access, collection establishes a valid Vana owner and a ready, approved server. Automated modes return structured reasons and recovery commands preserving source/path/streams and account/server context. Token cancellation now reports cancellation. Skill selection reads instructions without installation; persistent installation stays optional and confirmed. Connect Data and Create Connector are version 1.3.0; Builder remains 1.2.0. Connector creation/Factory integration remains a separate handoff.

Onboarding verification for a4e337b: pnpm validate passes 1,027 tests across 71 files; build and package checks pass. Actual terminal source selection and masked Readwise-token cancellation pass. Local HTTP fixtures verify Account recovery and independent readback of saved synthetic records. Structured Codex review reports no actionable P0 findings (0.95 confidence). Live Readwise collection and production storage remain unverified.

Moksha setup calls Account's app-key-signed test-credit API using existing SDK 4.3.1 signing support. It saves its funding key before submission, retains the returned operation, and uses GET-only recovery once that operation is known. Claim links, caps, disabled funding and pending operations have explicit outcomes; consent starts after credit is usable. New apps still initialize their key with vana app register. Existing saved consent intents and mainnet requests keep their existing flow.

Moksha request → saved funding → usable credit → explicit consent
resume        → same funding operation and consent request

Review fixed two regressions: Account's balance_sufficient response can include unfinalized deposits, so the CLI now reconciles existing escrow before treating credit as ready; and dead-writer cleanup now removes only the observed owner's file in an atomically published, nonempty lock directory, preserving any successor's lock.

Deployment gate: deploy and enable the service before merging/releasing this CLI change. Maciej must confirm the funded Account deployment, apply migration 029, configure/fund the Moksha funder and reconciliation cron, and deploy gateway #170 and Account #1963. Moksha setup stops resumably while that API is unavailable. A real empty-escrow setup and interrupted-operation check are still needed. The CLI does not depend on the unreleased SDK #221 helper. BUI-936 tracks the remaining acceptance; BUI-945 owns the service rollout.

Evidence

• pnpm validate: 1,008 tests across 69 files, TypeScript, ESLint and formatting pass on the combined tree. The two credit/lock regressions failed before their fixes. The existing idle-wait test now advances a controlled clock after the real interaction starts; removing the idle guard still fails it.
• pnpm pack:check passes on the combined tree with 458 files and all required CLI/runtime/skill entries.
• A controlled built-CLI smoke check uses four separate processes and recovers a lost funding response with one funding key, one operation and one consent request. No live funding or paid read was performed.
• The 221 focused CLI preflight tests, build and formatting passed. The old transcript generator was deliberately not rerun for this revision because it can launch a browser under a replaced HOME; the dedicated browser-isolation chat owns that repair.
• Verifier helper tests: 15 passed for the earlier boundary fix. Read-only closeout reviews of those earlier changes found no P0 blockers. Structured Codex review of installer commit eba0582 completed with no P0 findings (0.97 confidence). This was a read-only source review at the repository’s P0 threshold; the runtime and packaging proofs are recorded separately above/below. The subsequent CLI-first bootstrap / welcome commit 57a3665 also received structured Codex review with no P0 findings; this was source inspection, with runtime proofs recorded separately.
• Funding changes were reviewed in this session; a separate Codex helper was blocked by automatic approval review because it would export the unpublished funding delta. No separate-model pass is claimed for those changes.
• Fresh-agent public-only trial: installed skills and Builder reference match source; all 22 required local links resolve. A sandbox denies private Unity reads. Exactly one file-import collection uploaded all three synthetic records unchanged. Separately, controlled Research Shelf success, intended missing-save failure and restored success passed with actual Motel query evidence and full cleanup. After the doctor fix, a fresh ordinary run passed again. These separate synthetic trials do not establish delivery of that file-import source into Research Shelf, live provider collection, real consent/payment or native delivery.
• The prior public-boundary revision was downloaded without GitHub authentication and installed byte-identically. The installer revision (eba0582) was exercised in real terminals with network fetch blocked: first-use yes, install-hook yes/no, and JSON/non-interactive skips passed; installed skills and the reference match source. Build, package-content checks, native Darwin ARM64 SEA smoke/archive checks and Unix installer smoke passed. Windows execution is left to CI. Standard published installs receive this revision after merge and release.

Related Unity work: breadboard #1957, ownership #1938, approval #1940, closed duplicate backport #1961, and verification #1933. The ownership, approval and verification PRs are already merged; #1961 was closed without a production release. #1957 is the related open Unity PR, currently waiting for GitHub Actions billing/spending eligibility to be restored before checks can run. Automatic funding is implemented on the CLI side in this PR and awaits Maciej's service rollout and live acceptance. Private native connector delivery remains in BUI-939/BUI-940, and the complete fresh-builder trial remains BUI-937.

• CLI-first bootstrap proof: the actual npm tarball ran outside the checkout with network requests and child-process launches blocked. All three workflows and Builder's required reference matched source without skill installation. Real terminal welcome/finish and builder handoff passed. Native Darwin ARM64 SEA build/smoke/archive checks and Unix installer smoke passed. These checks do not establish live authentication, Readwise collection or native Windows execution.

Merge Danger

Door: two-way for application code; retain the request journal when rolling back because older CLI versions cannot resume funding setup.

Blast Radius: builder setup, local recovery state and opt-in skill installation.

Receipt and consent stores now reject corrupt state and write atomically with protected permissions. Existing retained requests remain supported; new consent intents add recovery state without storing credentials. Server write destinations and explicit consent requirements remain intact. Observation output is disabled unless explicitly configured and excludes payloads, credentials and payment headers.

Created with Codex.

[github-actions]

🎉 This PR is included in version 0.48.0 🎉

The release is available on GitHub release

Your semantic-release bot 📦🚀