epic: deliver zero-to-first-memory desktop onboarding and isolated local channels

[bradtaylorsf]

Goal

Deliver a zero-to-first-memory Answer Engine experience that a non-technical user can install from one copy-pasted prompt, launch from a desktop icon, connect to the AI agents they already use, import their history with informed consent, and prove memory recall in a fresh chat—while keeping a stable personal-memory instance completely isolated from local staging development.

Non-Technical Summary

A user should be able to paste a short prompt into Codex, ChatGPT Work/Desktop, Claude Code, or Claude Cowork and be guided through the complete setup. The installer welcomes them, asks which agents they use, checks their computer, installs the supported Answer Engine skills and MCP/CLI integrations, imports approved history, proposes an initial organization, and walks them through remembering and recalling one fact in a new chat.

A lightweight desktop application provides one-click access and health controls. The stable personal-memory channel keeps running in the background and cannot be affected by a staging channel used by local coding agents.

Supported Baseline

• macOS on Apple Silicon with at least 16 GB RAM.
• Windows 11 with WSL2 and a supported GPU with at least 8 GB VRAM.
• Capability detection recommends an appropriate local model/runtime and reports reduced-function or cloud-backed fallbacks honestly.
• Unsupported hardware or agent surfaces must fail clearly; the product must not claim local connectivity where the client cannot provide it.

Product Principles

• Local-first and private by default.
• No automatic home-directory scan.
• Every history or folder source requires explicit consent and a preview.
• Stable and staging data, credentials, ports, volumes, logs, and services are fail-closed and independently addressable.
• Skills teach workflows; MCP and CLI provide live tools and controlled actions.
• Distribution is pinned to a release/tag and verified by checksum rather than executing instructions from a floating branch.
• Installation, repair, upgrade, rollback, and uninstall are idempotent and observable.
• The Electron application is a control plane; a background service owns runtime availability.

Ordered Sub-Issues

• feat: isolate stable and staging local Answer Engine channels #41 - isolate stable and staging local Answer Engine channels
• feat: add a cross-platform one-prompt installer and capability preflight #42 - add a cross-platform one-prompt installer and capability preflight
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 - install and verify Answer Engine skills, plugins, MCP, and CLI wiring
• feat: guide and reconcile the first agent-history import #44 - guide and reconcile the first agent-history import
• feat: add permissioned preview-first local folder ingestion #45 - add permissioned preview-first local folder ingestion
• feat: propose and apply reversible AI-assisted memory organization #46 - propose and apply reversible AI-assisted memory organization
• feat: prove first memory with a grounded cross-chat recall tutorial #47 - prove first memory with a grounded cross-chat recall tutorial
• feat: ship a lightweight desktop launcher and local release controls #48 - ship a lightweight desktop launcher and local release controls
• fix: adopt legacy stable installs and open the real desktop web runtime #68 - adopt legacy stable installs and open the real desktop web runtime
• feat: ship immutable installer downloads and dependency-aware bootstrap #71 - ship immutable installer downloads and dependency-aware bootstrap
• fix: keep local client credentials out of process arguments #74 - keep local client credentials out of process arguments
• docs: make OSS licensing and community policies unambiguous #76 - make OSS licensing and community policies unambiguous
• security: decide public Git history and contributor identity policy #77 - decide public Git history and contributor identity policy
• fix: remove unsupported GHCR visibility mutation from release workflow #78 - remove unsupported GHCR visibility mutation from release workflow
• security: harden GitHub settings for public contributions #79 - harden GitHub settings for public contributions
• fix: patch dependency advisories found by public-release audit #86 - patch dependency advisories found by public-release audit
• fix: avoid full-history archive churn on incremental transcript updates #88 - avoid full-history archive churn on incremental transcript updates
• security: rebuild the public runtime from a patched base and dependency graph #89 - rebuild the public runtime from a patched base and dependency graph
• fix: support OpenAI-compatible local model servers without LM Studio management APIs #93 - support OpenAI-compatible local model servers without LM Studio management APIs
• release: make the OSS repository and runtime image public #80 - make the OSS repository and runtime image public
• release: publish and anonymously verify v1.1.2 installer assets #81 - publish and anonymously verify v1.1.2 installer assets
• test: validate v1.1.2 on clean Apple Silicon macOS #82 - validate v1.1.2 on clean Apple Silicon macOS
• test: validate v1.1.2 on clean Windows 11 with WSL2 #83 - validate v1.1.2 on clean Windows 11 with WSL2
• test: prove clean-machine onboarding and stable-staging isolation #49 - prove clean-machine onboarding and stable-staging isolation

End-to-End User Journey

1. Copy the canonical install prompt from GitHub or the website and paste it into a supported agent.
2. Answer a short welcome interview: operating system, hardware, stable/staging intent, and agents used.
3. Review preflight results and approve installation.
4. Launch Answer Engine from a desktop icon and confirm healthy local services.
5. Review discovered agent histories and explicitly approve the first import.
6. Optionally select folders, review a dry-run inventory, and approve local ingestion.
7. Review proposed libraries and tags before applying organization changes.
8. Install and verify supported skills/plugins plus MCP or CLI wiring for selected clients.
9. Tell one client a harmless fact to remember, open a fresh chat, and recall it through a real Answer Engine tool with source evidence.
10. Develop against staging without changing stable memory.

Acceptance Criteria

• A canonical, version-controlled prompt drives an agent-led install without requiring the user to understand Docker, MCP, or model configuration.
• Clean installation passes on the supported Apple Silicon and Windows baselines.
• Stable and staging channels cannot share data, credentials, ports, volumes, or sync jobs.
• A desktop icon opens a lightweight control surface and the stable runtime survives closing that surface.
• Selected supported clients receive the appropriate global or packaged skills plus verified MCP/CLI wiring.
• Client capability detection explains any ChatGPT web or Cowork localhost limitation and offers only supported alternatives.
• First history import is previewed, consented, resumable, and reconciled to a final inventory.
• Folder ingestion is opt-in, preview-first, source-backed, removable, and integrity-verifiable.
• Organization suggestions are previewed before mutation and can be audited and undone.
• A cross-chat remember/recall tutorial succeeds through real tools and grounded source inspection.
• Clean-machine acceptance, desktop/mobile visual checks, and pnpm verify pass.
• No enterprise-only roles, RBAC, teams, billing, or permission-management code enters OSS.
• No npm package is published as part of this epic.

Related Foundations

• fix: connect local UI without manual API-key setup #4 — installer-managed local UI access without a manually created API key.
• epic: Open-source the complete single-user Answer Engine application #6 — complete single-user OSS application.
• test: prove complete local product parity on real imported history #12 — clean installer-managed product acceptance over real imported history.
• the-answerai/answer-engine-enterprise#964 — enterprise composition over the OSS core.

Sequencing

• Execute the ordered sub-issues serially because installer, runtime, client configuration, and desktop packaging surfaces overlap.
• Keep Graphify excluded.
• Only the first unblocked child should carry the ready label.
• Final acceptance must run from clean environments and explicitly prove stable/staging isolation.

[bradtaylorsf]

Alpha Loop preflight blocker

The required command was run exactly:

node ../alpha-loop/dist/cli.js run --epic 40 --dry-run --validate --batch --batch-size 1

It started at #41 with tests/review/learning/verification enabled and batch size 1, but it reported #42–#49 as skipped because they do not yet carry ready. It therefore found only #41 and would finalize the session after that child. The real run was not started.

Root cause: Alpha Loop 2.4.0 builds an epic queue once at session start and filters every unchecked child through the configured ready label. It has no repository-configurable mode that keeps exactly one dependency-ready child labeled ready, processes it, advances readiness, and then refreshes the remaining ordered queue in the same epic session. A second configuration mismatch is max_issues: 6 for nine unchecked children.

The tempting workarounds are intentionally rejected: labeling all children ready violates #40's single-ready-child policy and dependency gates; repeatedly invoking a partial session violates the required dry-run condition that no ordered child be reported skipped; changing the configured ready label would bypass the repository's automation posture.

Lesson Learned

Epic dry-run validation must distinguish ordered checklist membership from immediate execution eligibility. A regression test belongs in Alpha Loop: with an epic whose first child alone is ready and later children depend serially on predecessors, dry-run should show the full ordered queue without skip warnings, and the real runner should refresh readiness only after each successful merge/checklist/learning gate. That implementation is outside the allowed answer-engine-oss write scope, so no repository product branch or test can honestly fix this runner defect here.

Current safety state: no #40 branch, worktree, PR, or runner exists; the OSS worktree is clean; the live stable Answer Engine health endpoint still returns 200 and was not stopped or reconfigured.

Keep Graphify excluded. No npm publication or direct master push occurred.

[bradtaylorsf]

Preparatory configuration fix plan

The external ready-queue refresh blocker remains, but the independent repository mismatch can be fixed safely here:

1. Raise .alpha-loop.yaml max_issues from 6 to 9 so the configured cap can represent every ordered child in epic: deliver zero-to-first-memory desktop onboarding and isolated local channels #40.
2. Update the repository-posture regression assertion to pin the nine-child cap.
3. Run the focused repository posture test and pnpm verify.
4. Commit on chore/GH-40-expand-epic-cap, open a PR to master, and merge only after checks pass.
5. Re-run sync and the exact epic dry run after merge. The real epic run remains prohibited unless the separate no-skip queue defect is resolved.

This change does not touch stable services, data, credentials, ports, volumes, or sync jobs. Keep Graphify excluded; do not publish npm.

[bradtaylorsf]

Preflight repair result

PR #50 is merged into master. It raised the repository's max_issues cap from 6 to 9 and updated the exact posture assertion. Local focused verification passed 10/10 tests; full pnpm verify passed; GitHub verify and container checks passed.

After the merge, node ../alpha-loop/dist/cli.js sync introduced no changes. The exact required dry run was repeated and now reports Max Issues: 9, Batch Size: 1, and all test/review/learning/verification gates enabled.

The remaining blocker reproduced unchanged: #42–#49 are each reported as skipped because only the dependency-ready #41 carries ready; the runner finds one issue and would finalize after #41. The real run remains stopped under the explicit no-skipped-child gate.

Lesson Learned

Repository capacity and runner queue semantics are independent gates. The capacity mismatch was safely repaired in this repository, but changing capacity cannot make Alpha Loop refresh an ordered epic after readiness advances. The upstream regression described in the prior blocker comment is still required.

Current safety state: clean master, no #40 runner/branch/worktree, #41 remains the only ready child, and stable health remains HTTP 200. No npm publication or direct master push occurred; Graphify remains excluded.

[bradtaylorsf]

Queue unblocked by operator override

The operator explicitly requested that all ordered children be labeled ready, overriding the earlier single-ready-child policy for this run. Issues #41–#49 now all carry ready.

After node ../alpha-loop/dist/cli.js sync reported assets in sync and introduced no changes, the exact dry run passed:

node ../alpha-loop/dist/cli.js run --epic 40 --dry-run --validate --batch --batch-size 1

Evidence: 9 issues found, 9 batches of one, ordered #41 through #49, no skipped children, validation passed, max_issues: 9, and tests/review/learnings/verification all enabled.

The prior queue blocker is resolved by this explicit label override. The real run may proceed under the existing stop conditions, stable-memory protections, Graphify exclusion, no npm publication, and no direct master push.

[bradtaylorsf]

Epic Verification

Overall: PARTIAL

The merged code substantially implements epic #40's in-scope onboarding, isolation, import, ingestion, recall, and desktop-control capabilities: 43 of 48 evaluated criteria are met, and the integrated merged session passes pnpm verify with 116 server tests, 184 CLI tests, 63 MCP tests, 119 installer tests, 40 web tests, and 12 desktop tests. Five criteria remain partial because operational evidence is incomplete: simultaneous real Compose-channel startup was not demonstrated; the canonical installer references an unpublished npm package and nonexistent v1.1.0 Git tag; clean Apple Silicon and Windows installations were not executed; Codex and Claude Code verification used mocked command output rather than recorded live client calls; and the tutorial's same-client/cross-client paths were exercised through tests and fixtures rather than live configured clients.

Sub-issue	PR	Status
#41 feat: isolate stable and staging local Answer Engine channels	PR	partial (5/6)
#42 feat: add a cross-platform one-prompt installer and capability preflight	PR	partial (5/7)
#43 feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring	PR	partial (6/7)
#44 feat: guide and reconcile the first agent-history import	PR	pass (7/7)
#45 feat: add permissioned preview-first local folder ingestion	PR	pass (7/7)
#46 feat: propose and apply reversible AI-assisted memory organization	—	not yet merged
#47 feat: prove first memory with a grounded cross-chat recall tutorial	PR	partial (6/7)
#48 feat: ship a lightweight desktop launcher and local release controls	PR	pass (7/7)
#49 test: prove clean-machine onboarding and stable-staging isolation	—	not yet merged

Per-criterion findings (48)

• feat: isolate stable and staging local Answer Engine channels #41 • met — Stable and staging use distinct paths, Compose projects, ports, secrets, volumes, logs, and background services. — Runtime profiles define disjoint homes, data/log/archive and credential paths, Compose projects, databases, five reserved ports, named volumes, and launchd/systemd identifiers. Isolation validation rejects overlaps and matching credential fingerprints.
• feat: isolate stable and staging local Answer Engine channels #41 • met — Staging cannot discover or sync real history unless the user explicitly enables and confirms it. — Staging defaults both sync flags to false, and CLI policy requires both persisted history_sync.enabled=true and --confirm-staging-history-sync before one-shot, daemon, or service history access.
• feat: isolate stable and staging local Answer Engine channels #41 • met — A deliberate staging mutation leaves the stable database and archives unchanged. — Regression tests perform staging archive and cursor mutations and verify stable database/archive fixture bytes remain unchanged.
• feat: isolate stable and staging local Answer Engine channels #41 • partial — Concurrent startup works and health/status output identifies the active channel. — Health, CLI, web, and lifecycle status are channel-aware and reject mismatched identities, and start/stop routing is isolated. The recorded evidence does not demonstrate two real Compose stacks successfully starting and remaining healthy concurrently; lifecycle startup coverage uses mocked Docker and health adapters.
• feat: isolate stable and staging local Answer Engine channels #41 • met — Automated regression tests reject every known collision class. — Tests cover home/data/log/archive overlaps, symlinks, credential paths and fingerprints, Compose projects, ports, database names, volumes, service identifiers, configuration drift, and changed Compose ownership.
• feat: isolate stable and staging local Answer Engine channels #41 • met — pnpm verify passes. — The authoritative integrated rerun against the merged epic session completed successfully.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • partial — A non-technical user can paste one prompt and complete setup through a short guided interview. — The synchronized plain-language prompt and consented interview are implemented, but the prompt invokes @answer-engine/create@1.1.0 and a v1.1.0 raw GitHub URL. The npm package currently returns 404 and no v1.1.0 Git tag exists, so a user cannot complete the advertised copy-paste path.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — Preflight returns actionable pass, warning, or unsupported results without mutating the machine. — Preflight produces ordered human/JSON results with remediation, distinguishes warnings from blockers, detects supported and unsupported hosts, and has zero-write coverage for absent, partial, and existing installations.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • partial — Clean Apple Silicon and Windows baseline installations reach a healthy UI without manual API-key creation. — Capability tests model both baselines and installer code automatically captures the local key, but no clean Apple Silicon or Windows installation was recorded. The PR explicitly deferred this live path, and the unpublished installer/tag currently prevents it.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — Repeated installation preserves data and reports no unintended changes. — Tests verify legacy adoption without data/archive changes, preservation of credentials and encryption material, recovery from a partial final stage, and byte-stable completed retries that become explicit no-ops.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — Tampered or mismatched release artifacts are rejected. — The Zod-validated manifest enforces version/tag and official immutable prompt agreement, exact checksum coverage, versioned or digest-pinned image rules, and SHA-256 rejection of modified bundled artifacts.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — The canonical prompt and surfaced copies cannot drift. — A repository regression compares the bracketed prompt bytes across INSTALL_AGENT.md, README.md, and packages/create/README.md.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — pnpm verify passes. — Although the original PR metadata records a failed hosted check, later integrated code and the authoritative rerun pass the complete pnpm verify gate.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — The welcome flow allows selecting all applicable clients. — The installer has a maintained client capability matrix, interactive multi-selection, --clients parsing, detection of applicable installed surfaces, Cowork mode selection, and a compatible --agents alias.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — Each selected supported client receives the appropriate skills/plugin and tool configuration. — Code plans and applies Codex and Claude plugin installations, installer-owned runtime MCP launchers, ChatGPT Desktop shared plugin wiring, and JSON MCP/CLI handoffs for supported desktop/adaptor clients.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • partial — Codex and Claude Code complete a real Answer Engine query through their configured integration. — Production verification launches codex and claude and requires correlated raw recall tool events containing the expected marker and content ID. Recorded verification only supplies mocked command output, however, and the acceptance record states that no live client smoke test was performed.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — ChatGPT web and Cowork limitations are detected and explained; no false localhost claim is made. — Hosted ChatGPT and remote Cowork are classified as unavailable for localhost wiring, while WSL2 also refuses unusable Linux-home configuration for Windows-host desktop applications.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — Configuration changes are idempotent, backed up, redacted, and reversible. — Integration tests cover private ledgers and backups, byte-stable retries, secret redaction, preservation of unrelated settings, drift handling, owned-entry removal, restoration, and idempotent removal.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — Skill activation and representative workflow evals pass. — Persisted evidence records 6/6 activation decisions and 14/14 independently graded packaged-skill assertions, with repository tests validating fixture hashes, coverage, and passing grades.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — pnpm verify passes. — The authoritative integrated rerun passes all installer, repository, package, and build gates.
• feat: guide and reconcile the first agent-history import #44 • met — Nothing is imported before explicit approval. — Discovery is metadata-only; state transitions reject progress or import before approval, and apply/resume revalidate the approved owner-only bundle manifest before reading transcript bodies.
• feat: guide and reconcile the first agent-history import #44 • met — The user can select any subset of discovered sources. — The API and UI accept any available subset of Claude Code, Codex, and Cowork sources while preventing selection of unavailable sources.
• feat: guide and reconcile the first agent-history import #44 • met — Interrupted imports resume without duplication or archive corruption. — Durable session items, sync cursors, deterministic content-addressed archives, atomic publication, cancellation checks, unique fingerprints, and --resume coverage provide convergent retry behavior.
• feat: guide and reconcile the first agent-history import #44 • met — Final counts reconcile and every imported history row has summary/raw-manifest integrity. — Completion reconciles imported, duplicate, failed, and skipped outcomes against discovery, while database constraints and service verification require imported content IDs and matching archive manifests with non-empty summaries.
• feat: guide and reconcile the first agent-history import #44 • met — Failures identify the source and recovery action without exposing secrets or private content in logs. — Errors are reduced to source labels, safe codes, and recovery actions; tests cover inaccessible and changed sources without returning transcript bodies, raw filesystem errors, credentials, or secrets.
• feat: guide and reconcile the first agent-history import #44 • met — Desktop and 375px visual verification pass. — The approved browser workflow passed interaction, responsive width, focus, reduced-motion, contrast, accessibility, and browser-error checks at 1440x900 and 375x812. Screenshot RPC failure did not invalidate the completed semantic visual checks.
• feat: guide and reconcile the first agent-history import #44 • met — pnpm verify passes. — The authoritative integrated rerun passes, including first-import server, CLI, UI, migration-order, and build coverage.
• feat: add permissioned preview-first local folder ingestion #45 • met — No filesystem content is ingested before folder-level approval. — Full-byte reads, archives, and imports require durable approval bound to an exact channel-local manifest; legacy direct local_dir synchronization fails closed.
• feat: add permissioned preview-first local folder ingestion #45 • met — Dry-run and applied inventories reconcile or explain differences. — Every preview item receives an applied or explained outcome, while refresh persists added, changed, unchanged, missing, and excluded classifications.
• feat: add permissioned preview-first local folder ingestion #45 • met — Unsupported, oversized, hidden, binary, permission-denied, and symlink cases are safely reported. — Bounded no-follow discovery and regression tests cover all listed classes plus missing paths, aggregate limits, traversal rejection, and apply-time identity changes.
• feat: add permissioned preview-first local folder ingestion #45 • met — Sampled archived content matches recorded SHA-256 values. — Archive tests read sampled stored bytes and compare them with the recorded SHA-256, with pre/post streaming identity validation and atomic archive publication.
• feat: add permissioned preview-first local folder ingestion #45 • met — Removing a source has explicit keep/delete behavior and a complete audit record. — Removal requires an explicit keep or delete choice, limits deletion to mapped content, reconciles archive handling, and records preparation, completion, counts, and failures.
• feat: add permissioned preview-first local folder ingestion #45 • met — Desktop and 375px visual verification pass. — Real browser verification passed consent gating, progress, responsive layout, keyboard focus, reduced motion, and accessibility at desktop and 375px, with screenshots recorded.
• feat: add permissioned preview-first local folder ingestion #45 • met — pnpm verify passes. — The authoritative integrated rerun passes folder-ingestion CLI, API, UI, integrity, and build coverage.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — The tutorial verifies a tool invocation rather than accepting an unsourced model answer. — Completion requires ordered tenant-scoped import, recall-result content ID, and lineage audit events bound to validated transport and client identity; typed answer text cannot complete the challenge.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — The fresh chat contains no copied secret or answer-bearing context. — The fresh-chat prompt contains only an opaque marker and tool instructions, excluding the generated fact, answer, credentials, and answer-bearing context.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • partial — Same-client and cross-client paths work wherever capabilities permit. — Capability-aware same-client and client-identifiable cross-client paths are implemented and tested, including wrong-client refusal. Recorded execution used server tests and an in-memory browser fixture rather than live configured agent clients, so operational cross-chat interoperability is not fully demonstrated.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — Unsupported cross-client combinations are identified before the demo starts. — Challenge creation rejects hosted, remote Cowork, WSL-to-Windows-host, and identity-ambiguous shared-plugin combinations with client-specific guidance.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — Failure diagnostics distinguish runtime, wiring, permission, indexing, and retrieval causes. — The state machine exposes distinct runtime, wiring, local access, indexing, retrieval, and source-evidence diagnostics with targeted recovery actions; local permission failures are intentionally named access to preserve the OSS boundary.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — Desktop and 375px visual verification pass. — The browser workflow passed planned-to-verified interaction, responsiveness, focus, reduced motion, accessibility, and error checks at 1440x900 and 375x812.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — pnpm verify passes. — The authoritative integrated rerun passes tutorial server, CLI, MCP, installer, web, boundary, and build coverage.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Clicking an installed icon reaches a healthy UI or an actionable repair screen. — A packaged macOS application was built and launched successfully; renderer/controller logic selects healthy controls or an actionable repair state, and the generated application contains the embedded icon.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Stable continues running after all Electron windows close. — Window close only hides the UI, quitting the launcher never invokes runtime stop, and regression tests separate window, tray, and background-runtime lifetimes.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Tray/menu status matches actual service health and selected channel. — Tray labels derive from polled lifecycle status for the validated selected channel, refresh periodically, and have tests for healthy, unhealthy, stable, and staging identities.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Staging is visually unmistakable on desktop and at 375px. — Electron verification and committed screenshots show explicit staging identity at 1440x900 and 375x812 with no overflow, accessibility violations, focus defects, or reduced-motion defects.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Update and rollback preserve stable data and reject unverified artifacts. — The desktop app delegates update and rollback to installer ownership, manifest, immutable-image, checksum, and release-state guards and exposes no purge or uninstall path. Data-preserving and artifact-rejection regressions pass.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Electron security regression tests and agent-browser/Electron visual verification pass. — Tests pass for context isolation, renderer sandboxing, disabled Node integration, strict CSP, sender and input validation, narrow preload IPC, navigation restrictions, and fixture isolation; Electron visual verification passed at both required widths.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — pnpm verify passes. — The authoritative integrated rerun passes all 12 desktop tests, desktop lint/type-check/build, and every repository package gate.

---

Verified by alpha-loop

[bradtaylorsf]

Epic Verification

Overall: PARTIAL

Most implementation criteria for merged issues #41–#48 are supported by code, automated tests, migrations, and recorded verification. The gate is not a full pass because the pinned installer artifacts were not published or exercised on clean Apple Silicon and Windows baselines, real Codex/Claude Code integration execution was not recorded, several UI issues lack the required screenshot evidence, same-client and cross-client tutorial paths were verified only through fixtures, and the desktop launcher’s real web-app action targets the unused development port rather than the API-served UI.

Sub-issue	PR	Status
#41 feat: isolate stable and staging local Answer Engine channels	PR	pass (6/6)
#42 feat: add a cross-platform one-prompt installer and capability preflight	PR	partial (5/7)
#43 feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring	PR	partial (6/7)
#44 feat: guide and reconcile the first agent-history import	PR	partial (6/7)
#45 feat: add permissioned preview-first local folder ingestion	PR	pass (7/7)
#46 feat: propose and apply reversible AI-assisted memory organization	PR	partial (6/7)
#47 feat: prove first memory with a grounded cross-chat recall tutorial	PR	partial (5/7)
#48 feat: ship a lightweight desktop launcher and local release controls	PR	partial (6/7)
#68 fix: adopt legacy stable installs and open the real desktop web runtime	—	not yet merged
#49 test: prove clean-machine onboarding and stable-staging isolation	—	not yet merged

Per-criterion findings (55)

• feat: isolate stable and staging local Answer Engine channels #41 • met — Stable and staging use distinct paths, Compose projects, ports, secrets, volumes, logs, and background services. — Runtime profiles define distinct homes, projects, databases, ports, credentials, volumes, logs, archives, and launchd/systemd identifiers, with collision validation and regression coverage.
• feat: isolate stable and staging local Answer Engine channels #41 • met — Staging cannot discover or sync real history unless the user explicitly enables and confirms it. — Staging sync defaults off and the CLI requires both persisted opt-in and explicit command confirmation.
• feat: isolate stable and staging local Answer Engine channels #41 • met — A deliberate staging mutation leaves the stable database and archives unchanged. — Regression tests mutate staging data, archives, and cursors while asserting stable fixtures remain byte-identical; recorded verification also reports no stable resource use.
• feat: isolate stable and staging local Answer Engine channels #41 • met — Concurrent startup works and health/status output identifies the active channel. — Channel-specific projects and ports permit concurrent operation, and health, lifecycle status, CLI, and web contracts validate explicit channel identity.
• feat: isolate stable and staging local Answer Engine channels #41 • met — Automated regression tests reject every known collision class. — Tests cover path and symlink overlap, ports, projects, databases, volumes, service identifiers, credential paths and fingerprints, and configuration/ownership drift.
• feat: isolate stable and staging local Answer Engine channels #41 • met — pnpm verify passes. — The issue records a passing implementation gate and independent post-review verification; the merged aggregate later passed the same repository gate.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • partial — A non-technical user can paste one prompt and complete setup through a short guided interview. — The synchronized prompt and interview are implemented, but they reference unpublished v1.1.0 npm/GHCR release artifacts, so recorded evidence does not show that the prompt can currently complete a real installation.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — Preflight returns actionable pass, warning, or unsupported results without mutating the machine. — The Zod-validated preflight reports ordered severity, remediation, and blocking state; tests and a built-command check confirm a nonexistent home remains absent.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • missing — Clean Apple Silicon and Windows baseline installations reach a healthy UI without manual API-key creation. — Only simulated baseline preflight tests exist. The PR explicitly states that the immutable package and image references were not published and that clean-machine live acceptance could not run.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — Repeated installation preserves data and reports no unintended changes. — Scaffold writes are byte-compared and atomic, durable completion state prevents false no-ops, and retry/idempotency tests cover healthy, partial, repair, upgrade, and rollback paths.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — Tampered or mismatched release artifacts are rejected. — Release validation pins the official tag and prompt URL, enforces version and exact checksum coverage, and has tampered-template and mismatch regressions.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — The canonical prompt and surfaced copies cannot drift. — Repository tests compare the marked prompt blocks byte-for-byte across INSTALL_AGENT.md, README.md, and the installer package README.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — pnpm verify passes. — Local verification recorded all suites and builds passing, and subsequent merged aggregate verification passed despite the original PR metadata showing a failed CI check.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — The welcome flow allows selecting all applicable clients. — The installer supports interactive multi-selection and validated client IDs, with --clients and the compatibility --agents alias.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — Each selected supported client receives the appropriate skills/plugin and tool configuration. — The capability matrix and integration planner install Codex and Claude plugins, MCP launchers, or JSON/CLI handoffs according to client and host capabilities.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • unclear — Codex and Claude Code complete a real Answer Engine query through their configured integration. — The shipped verifier launches client commands and requires correlated recall tool events, but the acceptance record says no global configuration or live Codex/Claude smoke test was performed and no isolated API was running.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — ChatGPT web and Cowork limitations are detected and explained; no false localhost claim is made. — Hosted ChatGPT and remote Cowork are explicitly classified as unable to use localhost, and WSL2 host/guest desktop limitations prevent unusable Linux-home wiring.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — Configuration changes are idempotent, backed up, redacted, and reversible. — Integration changes use a private ownership ledger, hashes, backups, owned-entry merging, redacted output, byte-stable retries, and reversible removal.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — Skill activation and representative workflow evals pass. — Recorded evidence includes passing plugin validators, all activation cases, and 14 of 14 independently graded packaged-skill assertions with retained evidence.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — pnpm verify passes. — The acceptance record and subsequent aggregate verification report all relevant tests, checks, and builds passing.
• feat: guide and reconcile the first agent-history import #44 • met — Nothing is imported before explicit approval. — Discovery is metadata-only, progress/import operations require durable approval, and approved bundle fingerprints are revalidated before transcript bodies are read.
• feat: guide and reconcile the first agent-history import #44 • met — The user can select any subset of discovered sources. — The API and UI accept arbitrary subsets of available Claude Code, Codex, and Cowork sources while preventing selection of unavailable sources.
• feat: guide and reconcile the first agent-history import #44 • met — Interrupted imports resume without duplication or archive corruption. — Persisted items, cursors, deterministic archives, atomic staging, cancellation checks, and owner-only manifests support safe resume and reject changed bundles.
• feat: guide and reconcile the first agent-history import #44 • met — Final counts reconcile and every imported history row has summary/raw-manifest integrity. — Completion reconciles every discovered item to imported, duplicate, failed, or skipped outcomes and validates content IDs, non-empty summaries, and matching archive manifests.
• feat: guide and reconcile the first agent-history import #44 • met — Failures identify the source and recovery action without exposing secrets or private content in logs. — Errors are reduced to safe source labels, codes, and recovery actions; tests cover inaccessible and changed sources without transcript bodies, raw errors, or credentials.
• feat: guide and reconcile the first agent-history import #44 • partial — Desktop and 375px visual verification pass. — Responsive interaction, focus, reduced motion, contrast, accessibility, and overflow checks passed at both widths, but the required screenshot/recording RPC repeatedly stalled and produced no visual artifact.
• feat: guide and reconcile the first agent-history import #44 • met — pnpm verify passes. — Recorded verification reports the full gate passing, including server, CLI, MCP, installer, web, and build checks.
• feat: add permissioned preview-first local folder ingestion #45 • met — No filesystem content is ingested before folder-level approval. — Only bounded metadata classification occurs before durable approval; full-byte reads, archives, imports, and legacy direct local-directory sync fail closed.
• feat: add permissioned preview-first local folder ingestion #45 • met — Dry-run and applied inventories reconcile or explain differences. — Every preview item receives an applied or explained result, and refresh persists added, changed, unchanged, missing, and excluded categories.
• feat: add permissioned preview-first local folder ingestion #45 • met — Unsupported, oversized, hidden, binary, permission-denied, and symlink cases are safely reported. — Traversal is bounded and no-follow; tests cover each listed classification plus missing paths and aggregate-size limits.
• feat: add permissioned preview-first local folder ingestion #45 • met — Sampled archived content matches recorded SHA-256 values. — Archive tests read sampled stored bytes and compare them with recorded SHA-256 values.
• feat: add permissioned preview-first local folder ingestion #45 • met — Removing a source has explicit keep/delete behavior and a complete audit record. — Removal requires an explicit retention choice and records preparation, completion, affected counts, archive handling, and failures.
• feat: add permissioned preview-first local folder ingestion #45 • met — Desktop and 375px visual verification pass. — Recorded browser verification covers both widths, consent gating, screenshots, focus, reduced motion, accessibility, and absence of horizontal overflow.
• feat: add permissioned preview-first local folder ingestion #45 • met — pnpm verify passes. — Both local and required GitHub verification passed after the archive-safety integration.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — Proposal generation does not mutate user data. — Proposal creation writes workflow evidence only and does not alter imported content, tags, libraries, memberships, or archives.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — Every suggestion links to supporting source examples and can be individually accepted or rejected. — Suggestions carry confidence, rationale, and one to three validated source examples; apply requires a decision for every suggestion.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — Approved changes use existing tenant-scoped APIs and reconcile to the preview. — Apply is parameterized, tenant-scoped, transactional, dependency-aware, and rechecks the source and taxonomy snapshot before mutation.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — Undo restores the pre-organization state without deleting imported content. — Undo removes only plan-owned memberships and deactivates only plan-created taxonomy; imported content and archives are not deleted or rewritten.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — Rerunning does not create duplicate tags, libraries, or memberships. — Content-addressed suggestions, unique fingerprints, slug reuse, conflict-safe membership writes, and apply/undo/reapply tests demonstrate idempotence.
• feat: propose and apply reversible AI-assisted memory organization #46 • partial — Desktop and 375px visual verification pass. — Interaction, layout, focus, reduced motion, contrast, and accessibility checks passed at both widths, but screenshot RPCs stalled and no PNG evidence was produced.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — pnpm verify passes. — The final repository gate passed with all organization, package, and build checks.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — The tutorial verifies a tool invocation rather than accepting an unsourced model answer. — Completion requires ordered import, recall-result ID, and lineage audit events bound to tenant, content, transport, and validated client identity; typed answer text is ignored.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — The fresh chat contains no copied secret or answer-bearing context. — The fresh-chat prompt contains only an opaque marker and instructions to invoke recall and source inspection, not the generated fact or secrets.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • unclear — Same-client and cross-client paths work wherever capabilities permit. — Capability checks and client-identity regressions cover both paths, but recorded execution used an API-shaped browser fixture rather than genuine fresh chats in installed clients.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — Unsupported cross-client combinations are identified before the demo starts. — Hosted, remote Cowork, WSL host/guest, and identity-ambiguous shared-plugin combinations are rejected before challenge creation with specific guidance.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — Failure diagnostics distinguish runtime, wiring, permission, indexing, and retrieval causes. — Distinct runtime, wiring, access-key, indexing, retrieval, and evidence states are persisted and tested; access is the OSS-safe name for permission failures.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • partial — Desktop and 375px visual verification pass. — Semantic interaction, responsiveness, focus, reduced motion, accessibility, and overflow checks passed, but the screenshot RPC stalled and produced no PNG.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — pnpm verify passes. — The final repository gate passed across server, CLI, MCP, installer, web, and builds.
• feat: ship a lightweight desktop launcher and local release controls #48 • partial — Clicking an installed icon reaches a healthy UI or an actionable repair screen. — A packaged macOS app launched successfully in fixture mode, but real mode opens the web app on reserved port 3200 while production serves it from the API origin, and legacy installs can encounter an ownership error that Repair cannot resolve.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Stable continues running after all Electron windows close. — Window close only hides the window, quitting does not invoke runtime stop, and runtime-lifetime independence is covered by regression tests.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Tray/menu status matches actual service health and selected channel. — The real controller delegates status to channel-aware installer lifecycle checks, and tray presentation is refreshed from that status every 15 seconds.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Staging is visually unmistakable on desktop and at 375px. — Committed desktop and mobile screenshots plus measured Electron checks show a permanent staging banner, distinct styling, and no overflow.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Update and rollback preserve stable data and reject unverified artifacts. — Electron delegates these actions to installer lifecycle and release-manifest guards and exposes no uninstall or purge path; integrity regressions remain passing.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Electron security regression tests and agent-browser/Electron visual verification pass. — Tests cover sandboxing, context isolation, disabled Node integration, strict CSP, sender validation, navigation restrictions, and narrow IPC; Electron checks and committed screenshots cover both viewports.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — pnpm verify passes. — The aggregate gate passed all server, CLI, MCP, installer, web, desktop, lint, type-check, and build checks.

---

Verified by alpha-loop

[bradtaylorsf]

Epic Verification

Overall: PARTIAL

The merged session branch implements most of epic #40’s scoped functionality, and the authoritative integrated pnpm verify run passed with 552 tests passing and one environment-gated test skipped. Of 62 evaluated criteria, 54 are met, six are partial, and two are missing. The principal blocker is distribution: the canonical v1.1.0 prompt URL returns 404, the Git tag is absent, the npm installer is unpublished, and the runtime image is unavailable, so the promised one-prompt and clean-baseline installations cannot work. Additional partials reflect fixture/contract evidence where criteria call for actual concurrent runtimes or real Codex/Claude client execution, plus mutable-tag runtime verification.

Sub-issue	PR	Status
#41 feat: isolate stable and staging local Answer Engine channels	PR	partial (4/6)
#42 feat: add a cross-platform one-prompt installer and capability preflight	PR	partial (4/7)
#43 feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring	PR	partial (6/7)
#44 feat: guide and reconcile the first agent-history import	PR	pass (7/7)
#45 feat: add permissioned preview-first local folder ingestion	PR	pass (7/7)
#46 feat: propose and apply reversible AI-assisted memory organization	PR	pass (7/7)
#47 feat: prove first memory with a grounded cross-chat recall tutorial	PR	partial (6/7)
#48 feat: ship a lightweight desktop launcher and local release controls	PR	partial (6/7)
#68 fix: adopt legacy stable installs and open the real desktop web runtime	PR	pass (7/7)
#49 test: prove clean-machine onboarding and stable-staging isolation	—	not yet merged

Per-criterion findings (62)

• feat: isolate stable and staging local Answer Engine channels #41 • met — Stable and staging use distinct paths, Compose projects, ports, secrets, volumes, logs, and background services. — Runtime profiles define disjoint homes, projects, ports, databases, credential files, volumes, archives, logs, and launchd/systemd identifiers; collision validation covers these resources.
• feat: isolate stable and staging local Answer Engine channels #41 • met — Staging cannot discover or sync real history unless the user explicitly enables and confirms it. — Staging defaults history sync off, rejects client wiring, and CLI policy requires both persisted opt-in and --confirm-staging-history-sync.
• feat: isolate stable and staging local Answer Engine channels #41 • partial — A deliberate staging mutation leaves the stable database and archives unchanged. — A regression mutates isolated staging data/archive fixtures and verifies stable bytes are unchanged, but recorded verification does not exercise a real staging database/archive mutation alongside the stable runtime.
• feat: isolate stable and staging local Answer Engine channels #41 • partial — Concurrent startup works and health/status output identifies the active channel. — Disjoint resources and channel-aware health/status behavior are implemented and tested, including rejection of wrong-channel health responses, but no recorded run starts both real Compose channels concurrently.
• feat: isolate stable and staging local Answer Engine channels #41 • met — Automated regression tests reject every known collision class. — Tests reject path and symlink overlap, projects, ports, databases, volumes, service identifiers, credential paths/fingerprints, unresolved identities, and intra-channel collisions.
• feat: isolate stable and staging local Answer Engine channels #41 • met — pnpm verify passes. — The authoritative integrated run passed all boundary, lint, type-check, test, and build stages.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • missing — A non-technical user can paste one prompt and complete setup through a short guided interview. — The interview and prompt exist in code, but the canonical tagged prompt currently returns HTTP 404 and @answer-engine/create@1.1.0 returns npm E404, so the copy-pasted prompt cannot start or complete setup.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — Preflight returns actionable pass, warning, or unsupported results without mutating the machine. — Read-only preflight reports ordered pass/warning/unsupported checks with remediation; tests cover supported macOS/WSL2, constrained hardware, unsupported platforms, occupied ports, and unchanged homes.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • missing — Clean Apple Silicon and Windows baseline installations reach a healthy UI without manual API-key creation. — Only simulated host/preflight coverage exists. The merged review explicitly deferred clean-machine execution, the Windows package was not run, and the required prompt, npm package, tag, and runtime image are unavailable.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — Repeated installation preserves data and reports no unintended changes. — Durable completion state, byte-stable scaffold writes, partial-run resume, healthy no-op behavior, and data-preserving lifecycle retries are covered by tests.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • partial — Tampered or mismatched release artifacts are rejected. — Bundled templates have exact SHA-256 coverage and version/tag/host mismatches are rejected, but the Answer Engine runtime is allowlisted by a mutable version tag rather than a digest, and the outer prompt/npm artifact cannot currently be retrieved and verified.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — The canonical prompt and surfaced copies cannot drift. — A repository test compares the prompt block byte-for-byte across INSTALL_AGENT.md, the root README, and the installer README.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — pnpm verify passes. — Although PR feat: batch implementation (#42) #57’s recorded GitHub verify check failed, the current integrated session branch passes the complete pnpm verify gate.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — The welcome flow allows selecting all applicable clients. — The maintained capability matrix and interactive/--clients selection cover Codex, ChatGPT surfaces, Claude surfaces, Cowork modes, and adapter-based clients.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — Each selected supported client receives the appropriate skills/plugin and tool configuration. — Installer plans provide Codex and Claude plugins, runtime-backed MCP launchers, desktop/adapter MCP configuration, CLI handoff, and no writes for unsupported surfaces.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • partial — Codex and Claude Code complete a real Answer Engine query through their configured integration. — Production code invokes the real codex and claude executables and requires correlated recall tool events, but recorded verification used injected command output and explicitly did not perform a live Codex/Claude smoke test.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — ChatGPT web and Cowork limitations are detected and explained; no false localhost claim is made. — Hosted ChatGPT and remote Cowork are classified as unable to reach localhost, and WSL2 host/guest desktop boundaries are reported without writing unusable configuration.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — Configuration changes are idempotent, backed up, redacted, and reversible. — The private integration ledger tracks owned changes and backups, preserves unrelated configuration, redacts credentials, supports byte-stable retries, and restores/removes owned entries.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — Skill activation and representative workflow evals pass. — Recorded independent evidence passes all six activation cases and all 14 packaged-skill assertions, with hashed evidence retained in the repository.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — pnpm verify passes. — The current integrated pnpm verify run passed, including all 129 installer tests.
• feat: guide and reconcile the first agent-history import #44 • met — Nothing is imported before explicit approval. — Discovery is metadata-only; execution requires durable source approval and rejects changed bundle fingerprints before reading transcript bodies.
• feat: guide and reconcile the first agent-history import #44 • met — The user can select any subset of discovered sources. — The API/UI allow any available subset while unavailable sources remain visible but unselectable.
• feat: guide and reconcile the first agent-history import #44 • met — Interrupted imports resume without duplication or archive corruption. — Session items, cursors, owner-only manifests, deterministic archives, atomic replacement, cancellation checks, and resume commands provide idempotent recovery.
• feat: guide and reconcile the first agent-history import #44 • met — Final counts reconcile and every imported history row has summary/raw-manifest integrity. — The lifecycle reconciles all discovered outcomes, and imported results require content IDs, non-empty summaries, and matching archive manifests; database integration coverage passed.
• feat: guide and reconcile the first agent-history import #44 • met — Failures identify the source and recovery action without exposing secrets or private content in logs. — Errors expose safe source labels, codes, and recovery actions while suppressing transcript contents, credentials, and raw filesystem errors.
• feat: guide and reconcile the first agent-history import #44 • met — Desktop and 375px visual verification pass. — Recorded pnpm browser:ui interaction, responsive-width, focus, reduced-motion, error, and accessibility checks passed at 1440×900 and 375×812; screenshot RPC failure was separately documented.
• feat: guide and reconcile the first agent-history import #44 • met — pnpm verify passes. — The current integrated verification passes all affected server, CLI, web, and build gates.
• feat: add permissioned preview-first local folder ingestion #45 • met — No filesystem content is ingested before folder-level approval. — Full-byte reads, archives, and imports require durable approval bound to the exact channel-local inventory; direct local_dir sync fails closed.
• feat: add permissioned preview-first local folder ingestion #45 • met — Dry-run and applied inventories reconcile or explain differences. — Every preview row receives an applied or explained outcome, and refresh records added, changed, unchanged, missing, and excluded files.
• feat: add permissioned preview-first local folder ingestion #45 • met — Unsupported, oversized, hidden, binary, permission-denied, and symlink cases are safely reported. — Bounded no-follow discovery and tests cover each listed case, missing files, aggregate limits, traversal, and apply-time identity changes.
• feat: add permissioned preview-first local folder ingestion #45 • met — Sampled archived content matches recorded SHA-256 values. — Archive tests read sampled stored bytes and compare them with recorded SHA-256 values.
• feat: add permissioned preview-first local folder ingestion #45 • met — Removing a source has explicit keep/delete behavior and a complete audit record. — Removal requires an explicit retention mode and records preparation, completion, affected counts, archive handling, and failures.
• feat: add permissioned preview-first local folder ingestion #45 • met — Desktop and 375px visual verification pass. — Production UI checks passed at both widths with consent gating, screenshots, no overflow, visible focus, reduced motion, and zero Axe violations.
• feat: add permissioned preview-first local folder ingestion #45 • met — pnpm verify passes. — The authoritative integrated verification passed the folder workflow’s server, CLI, UI, and build coverage.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — Proposal generation does not mutate user data. — Proposal generation writes workflow evidence only and leaves imported content, taxonomy, libraries, memberships, and archives unchanged.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — Every suggestion links to supporting source examples and can be individually accepted or rejected. — Strict proposal contracts require supporting sampled content IDs, rationale, confidence, and a complete individual accept/reject decision set.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — Approved changes use existing tenant-scoped APIs and reconcile to the preview. — Transactional, parameterized, tenant-scoped apply rechecks the source/taxonomy fingerprint and refuses stale or incomplete proposals.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — Undo restores the pre-organization state without deleting imported content. — Undo removes only plan-owned memberships and deactivates plan-created taxonomy with stale-write guards; imported content is not rewritten or deleted.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — Rerunning does not create duplicate tags, libraries, or memberships. — Content-addressed suggestions, unique fingerprints/slugs, conflict-safe membership insertion, and apply–undo–reapply tests prevent duplication.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — Desktop and 375px visual verification pass. — Recorded production UI checks completed apply, undo, and reapply at both widths with no overflow, visible focus, reduced motion, and zero Axe violations.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — pnpm verify passes. — The integrated full verification passed organization service, route, CLI, UI, and build coverage.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — The tutorial verifies a tool invocation rather than accepting an unsourced model answer. — Completion requires ordered audited remember, recall-result-ID, and lineage events bound to tenant, content, transport, and client identity; typed answers are ignored.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — The fresh chat contains no copied secret or answer-bearing context. — The fresh-chat prompt contains only an opaque marker and tool instructions, not the generated fact, answer, secret, or prior-chat context.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • partial — Same-client and cross-client paths work wherever capabilities permit. — The state machine, identity checks, and Codex-to-Claude UI/API fixture path are implemented and tested, but recorded verification did not execute a fresh chat through actual installed agent clients.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — Unsupported cross-client combinations are identified before the demo starts. — Capability preflight rejects hosted, remote Cowork, WSL host/guest, and identity-ambiguous shared-plugin combinations before challenge creation.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — Failure diagnostics distinguish runtime, wiring, permission, indexing, and retrieval causes. — Distinct runtime, wiring, access, indexing, retrieval, and source-evidence diagnostic states are implemented with specific recovery guidance.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — Desktop and 375px visual verification pass. — The production React workflow passed interaction, responsive-width, focus, reduced-motion, accessibility, and browser-error checks at both required widths.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — pnpm verify passes. — The authoritative integrated run passed tutorial server, CLI, MCP, installer, web, and build coverage.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Clicking an installed icon reaches a healthy UI or an actionable repair screen. — A packaged macOS app launched successfully, and later native fix: adopt legacy stable installs and open the real desktop web runtime #68 evidence showed the real controller reporting stable healthy and opening the HTTP 200 API-served web UI.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Stable continues running after all Electron windows close. — Window close hides the window while retaining the tray process, quit contains no runtime-stop call, and lifecycle-independence regression tests pass.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Tray/menu status matches actual service health and selected channel. — Tray labels derive from refreshed selected-channel status; native real-controller evidence showed stable healthy and staging independently not installed.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Staging is visually unmistakable on desktop and at 375px. — Committed desktop/mobile screenshots and native browser checks show permanent staging identity with no overflow at 1440×900 and 375×812.
• feat: ship a lightweight desktop launcher and local release controls #48 • partial — Update and rollback preserve stable data and reject unverified artifacts. — Desktop delegates to data-preserving guarded lifecycle controls and rejects arbitrary unpinned image inputs, but the bundled Answer Engine image itself is accepted by mutable version tag rather than an immutable digest/checksum.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Electron security regression tests and agent-browser/Electron visual verification pass. — Security tests cover sandboxing, context isolation, disabled Node integration, strict CSP, sender validation, IPC, and navigation; fix: adopt legacy stable installs and open the real desktop web runtime #68 adds host-launched native Electron checks at both widths.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — pnpm verify passes. — The integrated run passed all desktop lint, type-check, 18 tests, and build stages along with the repository-wide gate.
• fix: adopt legacy stable installs and open the real desktop web runtime #68 • met — Fixture mode is visibly identified and cannot report that a no-op external action succeeded. — Demo mode is visible in renderer/tray state, reports no installed or healthy runtime, and disabled URL/log actions return opened: false with truthful messages.
• fix: adopt legacy stable installs and open the real desktop web runtime #68 • met — A healthy installed desktop launcher opens the packaged web UI on the runtime’s actual origin. — Production URLs preserve the API-served 5050/5150 origin, and native real-controller evidence opened the stable web app with an independent HTTP 200 check.
• fix: adopt legacy stable installs and open the real desktop web runtime #68 • met — A valid legacy stable home can be safely adopted from the launcher without editing database, blob, archive, or volume bytes. — Stable-only launcher adoption delegates to validated installer logic that writes channel metadata/ownership only; fixture database and archive bytes are asserted unchanged.
• fix: adopt legacy stable installs and open the real desktop web runtime #68 • met — Invalid, symlinked, wrong-project, or channel-conflicting legacy homes still fail closed with actionable errors. — Regression coverage rejects full and partial symlinks, wrong projects/channels, malformed files, unsafe Compose topology, and conflicting channel resources before ownership is written.
• fix: adopt legacy stable installs and open the real desktop web runtime #68 • met — Legacy adoption is not blocked solely by unrelated reserved development ports that it does not start or bind. — The metadata-only adoption branch runs before Docker, model, owned-port, and reserved-port readiness checks; tests confirm those probes are not called.
• fix: adopt legacy stable installs and open the real desktop web runtime #68 • met — Regression coverage exists for fixture truthfulness, production URL selection, legacy adoption, and ownership refusal. — Installer and desktop suites cover all four requested areas, including refusal to run lifecycle actions without a valid ownership marker.
• fix: adopt legacy stable installs and open the real desktop web runtime #68 • met — Desktop browser checks at desktop and 375px plus pnpm verify pass. — Host-launched native Electron checks passed at 1440×900 and 375×812 with focus, reduced motion, no overflow/errors, and zero Axe violations; the authoritative integrated pnpm verify run also passed.

---

Verified by alpha-loop

[bradtaylorsf]

Epic Verification

Overall: PARTIAL

The merged implementation substantially satisfies epic #40’s in-scope children: channel isolation, consent-first imports, folder ingestion, reversible organization, grounded recall auditing, desktop controls, legacy adoption, and immutable bootstrap safeguards are implemented with broad regression coverage. The integrated head passed pnpm verify with 576 tests passing and one environment-gated database test skipped, and pnpm audit --prod --audit-level high found no known vulnerabilities. The gate remains partial because clean-machine one-prompt installation has only simulated/candidate evidence, live Codex and Claude Code integration queries were not recorded, the cross-client tutorial was exercised through fixtures rather than actual supported clients, and issues #44, #46, and #47 lack completed screenshot-based visual evidence after the approved browser RPC stalled. Issue #49 was not merged and is excluded.

Sub-issue	PR	Status
#41 feat: isolate stable and staging local Answer Engine channels	PR	pass (6/6)
#42 feat: add a cross-platform one-prompt installer and capability preflight	PR	partial (5/7)
#43 feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring	PR	partial (6/7)
#44 feat: guide and reconcile the first agent-history import	PR	partial (6/7)
#45 feat: add permissioned preview-first local folder ingestion	PR	pass (7/7)
#46 feat: propose and apply reversible AI-assisted memory organization	PR	partial (6/7)
#47 feat: prove first memory with a grounded cross-chat recall tutorial	PR	partial (5/7)
#48 feat: ship a lightweight desktop launcher and local release controls	PR	pass (7/7)
#68 fix: adopt legacy stable installs and open the real desktop web runtime	PR	pass (7/7)
#71 feat: ship immutable installer downloads and dependency-aware bootstrap	PR	pass (10/10)
#49 test: prove clean-machine onboarding and stable-staging isolation	—	not yet merged

Per-criterion findings (72)

• feat: isolate stable and staging local Answer Engine channels #41 • met — Stable and staging use distinct paths, Compose projects, ports, secrets, volumes, logs, and background services. — Runtime profiles define disjoint homes, Compose projects, ports, database names, credentials, volumes, logs, archives, and launchd/systemd identifiers, with collision validation and regression tests.
• feat: isolate stable and staging local Answer Engine channels #41 • met — Staging cannot discover or sync real history unless the user explicitly enables and confirms it. — Staging sync defaults off and requires both persisted opt-in and explicit command confirmation; CLI and service-generation tests cover both gates.
• feat: isolate stable and staging local Answer Engine channels #41 • met — A deliberate staging mutation leaves the stable database and archives unchanged. — CLI and installer regressions mutate staging archives/cursors/data while asserting stable database and archive fixture bytes remain unchanged.
• feat: isolate stable and staging local Answer Engine channels #41 • met — Concurrent startup works and health/status output identifies the active channel. — Channel-aware lifecycle and health contracts support independent ports/projects and propagate validated channel identity through status, API health, CLI, web, and desktop surfaces.
• feat: isolate stable and staging local Answer Engine channels #41 • met — Automated regression tests reject every known collision class. — Tests reject overlapping or unresolved homes, symlinks, ports, projects, databases, volumes, services, credential paths/fingerprints, and drifted runtime ownership.
• feat: isolate stable and staging local Answer Engine channels #41 • met — pnpm verify passes. — The integrated branch passed the current pnpm verify gate.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • partial — A non-technical user can paste one prompt and complete setup through a short guided interview. — The synchronized prompt, interview, consent plan, and installer stages exist, but no recorded end-to-end non-technical clean-machine run completed from the canonical release prompt; that proof remains deferred to test: prove clean-machine onboarding and stable-staging isolation #49.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — Preflight returns actionable pass, warning, or unsupported results without mutating the machine. — Typed preflight reports capability severity and blocking state, with tests covering supported, constrained, unsupported, and zero-write paths.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • partial — Clean Apple Silicon and Windows baseline installations reach a healthy UI without manual API-key creation. — Platform fixtures, automatic local-key capture, health checks, and UI readiness logic are covered, but recorded evidence explicitly defers real clean-machine macOS and Windows installation to test: prove clean-machine onboarding and stable-staging isolation #49.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — Repeated installation preserves data and reports no unintended changes. — Durable completion checkpoints, byte-stable writes, interrupted-stage recovery, and healthy-install no-op tests preserve data and credentials.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — Tampered or mismatched release artifacts are rejected. — Release validation checks version, tag, official immutable URLs, exact artifact coverage, checksums, and manifest consistency before mutation.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — The canonical prompt and surfaced copies cannot drift. — A repository test compares the marked prompt blocks byte-for-byte across INSTALL_AGENT.md, README.md, and the installer README.
• feat: add a cross-platform one-prompt installer and capability preflight #42 • met — pnpm verify passes. — The integrated branch passed the current pnpm verify gate.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — The welcome flow allows selecting all applicable clients. — The installer exposes typed interactive and --clients selection for maintained client surfaces, with --agents retained as an alias.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — Each selected supported client receives the appropriate skills/plugin and tool configuration. — Client-specific Codex and Claude plugins, MCP launchers, JSON adapters, CLI handoff, and guided desktop paths are implemented and checksum-covered.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • partial — Codex and Claude Code complete a real Answer Engine query through their configured integration. — Production code launches each real non-interactive client and requires correlated recall tool-event evidence, but verification used mocked command execution and explicitly did not perform a live global-client smoke test against an isolated API.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — ChatGPT web and Cowork limitations are detected and explained; no false localhost claim is made. — The capability matrix rejects hosted ChatGPT and remote Cowork localhost wiring and also handles WSL-to-Windows desktop boundaries explicitly.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — Configuration changes are idempotent, backed up, redacted, and reversible. — Private ledgers, backups, byte-stable retries, owned-entry restoration/removal, secret redaction, symlink refusal, and drift detection are covered.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — Skill activation and representative workflow evals pass. — Current hash-bound evidence records 6/6 activation decisions and 14/14 packaged-skill assertions passing, enforced by repository tests.
• feat: install and verify Answer Engine skills, plugins, MCP, and CLI wiring #43 • met — pnpm verify passes. — The integrated branch passed the current pnpm verify gate.
• feat: guide and reconcile the first agent-history import #44 • met — Nothing is imported before explicit approval. — Discovery is metadata-only; transcript reads and imports require approved sources plus revalidation of the owner-only complete-bundle manifest.
• feat: guide and reconcile the first agent-history import #44 • met — The user can select any subset of discovered sources. — The API, CLI, and UI support arbitrary available-source subsets while disabling unavailable sources.
• feat: guide and reconcile the first agent-history import #44 • met — Interrupted imports resume without duplication or archive corruption. — Durable items/cursors, deterministic hash-verified archives, cancellation, retry, and resume paths are covered by CLI and database tests.
• feat: guide and reconcile the first agent-history import #44 • met — Final counts reconcile and every imported history row has summary/raw-manifest integrity. — Completion reconciles imported, duplicate, failed, and skipped outcomes; imported rows require non-empty summaries and matching archive manifests.
• feat: guide and reconcile the first agent-history import #44 • met — Failures identify the source and recovery action without exposing secrets or private content in logs. — Errors use safe source labels, codes, and recovery actions while suppressing transcript content, credentials, and raw filesystem details.
• feat: guide and reconcile the first agent-history import #44 • partial — Desktop and 375px visual verification pass. — Real-browser interaction, responsive dimensions, focus, reduced motion, contrast, accessibility, and error checks passed at both widths, but the approved screenshot/recording RPC repeatedly stalled and produced no visual artifact.
• feat: guide and reconcile the first agent-history import #44 • met — pnpm verify passes. — The integrated branch passed the current pnpm verify gate.
• feat: add permissioned preview-first local folder ingestion #45 • met — No filesystem content is ingested before folder-level approval. — Full-byte reads, archives, and imports require durable approval bound to an exact channel-local manifest; direct local-directory sync fails closed.
• feat: add permissioned preview-first local folder ingestion #45 • met — Dry-run and applied inventories reconcile or explain differences. — Every preview row receives an applied or explained result, and refresh records added, changed, unchanged, missing, and excluded files.
• feat: add permissioned preview-first local folder ingestion #45 • met — Unsupported, oversized, hidden, binary, permission-denied, and symlink cases are safely reported. — Bounded no-follow discovery classifies each listed case, aggregate-limit failures, and missing paths without unsafe traversal.
• feat: add permissioned preview-first local folder ingestion #45 • met — Sampled archived content matches recorded SHA-256 values. — Archive tests compare sampled stored bytes with recorded SHA-256 values; later holistic hardening also validates manifest identity, containment, and no-follow payload reads.
• feat: add permissioned preview-first local folder ingestion #45 • met — Removing a source has explicit keep/delete behavior and a complete audit record. — Removal requires an explicit retention choice and records preparation, completion, counts, archive handling, and failures.
• feat: add permissioned preview-first local folder ingestion #45 • met — Desktop and 375px visual verification pass. — Browser verification passed at 1440x900 and 375x812 with consent gating, screenshots, no overflow, focus, reduced motion, and zero Axe violations.
• feat: add permissioned preview-first local folder ingestion #45 • met — pnpm verify passes. — The integrated branch passed the current pnpm verify gate.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — Proposal generation does not mutate user data. — Proposal creation writes workflow/audit evidence only and does not mutate imported content, taxonomy, libraries, memberships, or archives.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — Every suggestion links to supporting source examples and can be individually accepted or rejected. — Suggestions require supporting content IDs/examples and complete per-suggestion accept/reject decisions before apply.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — Approved changes use existing tenant-scoped APIs and reconcile to the preview. — Apply is parameterized, tenant-scoped, transactional, dependency-aware, and rejects changed source/taxonomy snapshots.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — Undo restores the pre-organization state without deleting imported content. — Undo removes only plan-owned memberships/taxonomy state, preserves content, and now retains PostgreSQL microsecond precision for stale-write guards.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — Rerunning does not create duplicate tags, libraries, or memberships. — Apply, undo, reapply, and second-undo tests reuse taxonomy IDs and prevent duplicate memberships.
• feat: propose and apply reversible AI-assisted memory organization #46 • partial — Desktop and 375px visual verification pass. — Interaction, responsive width, focus, reduced motion, browser-error, and accessibility checks passed at both widths, but screenshot RPCs stalled and no PNG was produced.
• feat: propose and apply reversible AI-assisted memory organization #46 • met — pnpm verify passes. — The integrated branch passed the current pnpm verify gate.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — The tutorial verifies a tool invocation rather than accepting an unsourced model answer. — Completion requires ordered, audited remember, recall-result-ID, and lineage events for the exact tenant, content ID, transport, and client identity.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — The fresh chat contains no copied secret or answer-bearing context. — The fresh-chat prompt contains only an opaque marker and tool instructions, never the generated fact, answer, key, or copied context.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • partial — Same-client and cross-client paths work wherever capabilities permit. — Supported paths, per-client identity enforcement, and wrong-client refusal are implemented and fixture-tested, but recorded verification did not execute an actual supported-client same/cross-client chat pair.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — Unsupported cross-client combinations are identified before the demo starts. — Hosted, remote Cowork, WSL/desktop-host, and ambiguous shared-plugin combinations fail before challenge creation with specific guidance.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — Failure diagnostics distinguish runtime, wiring, permission, indexing, and retrieval causes. — Distinct runtime, wiring, access, indexing, retrieval, and source-evidence diagnostic states are implemented with recovery guidance.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • partial — Desktop and 375px visual verification pass. — The browser workflow passed semantic interaction, layout, focus, reduced-motion, accessibility, and error checks at both widths, but its screenshot RPC stalled and produced no PNG.
• feat: prove first memory with a grounded cross-chat recall tutorial #47 • met — pnpm verify passes. — The integrated branch passed the current pnpm verify gate.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Clicking an installed icon reaches a healthy UI or an actionable repair screen. — The packaged macOS application launched successfully, and later native fix: adopt legacy stable installs and open the real desktop web runtime #68 evidence showed the real controller reporting healthy stable status and opening the HTTP 200 UI.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Stable continues running after all Electron windows close. — Window close hides the window while tray/runtime lifetimes remain independent; quitting Electron never invokes runtime stop.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Tray/menu status matches actual service health and selected channel. — Tray labels derive from refreshed selected-channel status, with tests for stable health, staging identity, and fixture truthfulness.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Staging is visually unmistakable on desktop and at 375px. — Committed desktop/mobile evidence and browser checks show explicit staging treatment without overflow.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Update and rollback preserve stable data and reject unverified artifacts. — Desktop delegates to installer lifecycle guards; immutable manifest/image validation, ownership checks, and data-preserving update/rollback regressions pass.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — Electron security regression tests and agent-browser/Electron visual verification pass. — Tests cover context isolation, sandboxing, disabled Node integration, CSP, sender checks, navigation, IPC, and runtime independence; committed fixture and later native Electron visual evidence covers both widths.
• feat: ship a lightweight desktop launcher and local release controls #48 • met — pnpm verify passes. — The integrated branch passed the current pnpm verify gate, including desktop lint, type-check, tests, and build.
• fix: adopt legacy stable installs and open the real desktop web runtime #68 • met — Fixture mode is visibly identified and cannot report that a no-op external action succeeded. — Fixture status and tray are labeled Demo mode, report no installed/healthy runtime, and return explicit opened: false results for disabled URL/log actions.
• fix: adopt legacy stable installs and open the real desktop web runtime #68 • met — A healthy installed desktop launcher opens the packaged web UI on the runtime’s actual origin. — Production URL resolution retains the API-served 5050/5150 origin; native real-controller evidence opened stable port 5050 and independently received HTTP 200.
• fix: adopt legacy stable installs and open the real desktop web runtime #68 • met — A valid legacy stable home can be safely adopted from the launcher without editing database, blob, archive, or volume bytes. — Launcher adoption delegates to a stable-only metadata adopter that writes channel ownership metadata without starting services; unchanged data/archive fixtures are asserted.
• fix: adopt legacy stable installs and open the real desktop web runtime #68 • met — Invalid, symlinked, wrong-project, or channel-conflicting legacy homes still fail closed with actionable errors. — Adoption validates regular files, home/config symlinks, channel/project identity, Compose topology, and isolation before writing ownership.
• fix: adopt legacy stable installs and open the real desktop web runtime #68 • met — Legacy adoption is not blocked solely by unrelated reserved development ports that it does not start or bind. — Metadata-only adoption runs before Docker, model, and reserved-port readiness; regression coverage verifies unrelated occupied ports do not block it.
• fix: adopt legacy stable installs and open the real desktop web runtime #68 • met — Regression coverage exists for fixture truthfulness, production URL selection, legacy adoption, and ownership refusal. — Desktop and installer suites cover all named cases, including continued lifecycle refusal when an ownership marker is absent.
• fix: adopt legacy stable installs and open the real desktop web runtime #68 • met — Desktop browser checks at desktop and 375px plus pnpm verify pass. — Committed native Electron evidence covers 1440x900, 375x812, and the real stable runtime; the integrated pnpm verify gate passed.
• feat: ship immutable installer downloads and dependency-aware bootstrap #71 • met — The canonical install prompt no longer requires an npm registry publication; Bash and PowerShell paths download an immutable versioned installer asset and verify SHA-256 before executing it. — The synchronized prompt uses exact-tag GitHub Release bootstrap paths with pinned hashes, checksum-file agreement, and pre-execution verification; npm is not used.
• feat: ship immutable installer downloads and dependency-aware bootstrap #71 • met — A release workflow reproducibly builds installer assets, checksum metadata, and provenance from an exact tag/commit without an npm publish step. — The manually gated workflow validates exact inputs, checks out the supplied commit, produces deterministic installer/CLI archives, SHA256SUMS, manifest, and provenance, and contains no npm publication step.
• feat: ship immutable installer downloads and dependency-aware bootstrap #71 • met — Install, update, and rollback reject mutable runtime tags and persist a verified sha256: runtime digest. — Scaffold and lifecycle paths require digest references before mutation; managed lifecycle execution is bound to persisted verified digest state, including pending transition recovery.
• feat: ship immutable installer downloads and dependency-aware bootstrap #71 • met — Read-only preflight reports supported Apple Silicon macOS and Windows 11/WSL2 dependency state, reuses compatible installed dependencies, and distinguishes required, optional, unsupported, and privileged prerequisites. — Typed dependency inventory and deterministic platform fixtures cover Node, Docker, Compose, WSL2, hardware/GPU, and model runtime policies without mutation.
• feat: ship immutable installer downloads and dependency-aware bootstrap #71 • met — Supported missing dependencies are installed only after the user approves the displayed source, version, and command; downloads are verified and readiness is rerun. — Only the official user-scoped Node 22.16.0 archive is automatable; source, version, checksum, destination, and command are displayed, explicit approval is required, and resumed readiness is tested.
• feat: ship immutable installer downloads and dependency-aware bootstrap #71 • met — Cancellation and unsupported/privileged dependency paths make no system or Answer Engine runtime changes and provide actionable instructions. — Consent-refusal and unsupported paths assert unchanged homes and provide platform-specific manual guidance for Docker Desktop, WSL2, drivers, privileged components, and model runtimes.
• feat: ship immutable installer downloads and dependency-aware bootstrap #71 • met — Session worktree preparation is re-entrant when the exact worktree-owned Compose project is already running, while foreign or stable port listeners remain fail-closed. — Repository posture tests reuse only matching worktree-owned Compose identity and reject foreign listeners, stable ports/projects/homes, and ownership mismatches.
• feat: ship immutable installer downloads and dependency-aware bootstrap #71 • met — Automated tests cover Apple Silicon and Windows/WSL command generation, installed/missing dependency states, consent refusal, reruns, tampered checksums, mismatched artifacts, and mutable runtime references. — Release, bootstrap, preflight, lifecycle, scaffold, and repository suites cover every enumerated case, including cross-manifest mismatch and production-verifier invocation.
• feat: ship immutable installer downloads and dependency-aware bootstrap #71 • met — Documentation describes the release download, dependency policy, supported baselines, and manual fallback paths. — INSTALL_AGENT.md, package documentation, installer-release documentation, and integration safety references describe the exact downloads, supported platforms, dependency policy, and manual remediations.
• feat: ship immutable installer downloads and dependency-aware bootstrap #71 • met — pnpm verify and pnpm audit --prod --audit-level high pass. — This authoritative pass reran both commands successfully: 576 tests passed, one environment-gated database test was skipped, all builds passed, and no known production vulnerabilities were reported.

---

Verified by alpha-loop