Skip to content

Evidence index ​

Public engineering evidence: architecture decisions, verification records, implementation notes, assessments, and compatibility reports. This index is for contributors, maintainers, and reviewers. It is not user guidance; start at the docs home or the contributing hub for the user and contributor paths.

Every verification record dated 2026-11-11 or later must be linked here. npm run docs:check enforces that rule so a current record cannot disappear from the public evidence set.

Architecture decisions ​

The ADR index lists every accepted decision, its status, and its tickets.

ADRDecision
0001Native-skill boundary: native execution is authoritative
0002Server-authoritative durable patch feed
0003One Ensemble service per host realm, with lease-safe recovery
0004Layered dispatch budgets, continuation groups, and one durable ledger
0005Work status taxonomy and delivery/lease separation
0006Three-pane workspace shell
0007Apply agent behavior changes to the existing native session
0008Task-level history and sticky rollup
0009Conversation identity and per-conversation native sessions
0010Native turn ingestion is service-owned and identity-based
0011$Name routing grammar and server-owned budgets
0012Terminal turns, root-only synthesis, and explicit continuation

Verification records ​

  • Ticket #301 documentation acceptance: the five-minute human quickstart walkthrough, the heading-anchor and terminal-hero fixes, the full 185-route desktop/mobile/light/dark scan, zero axe violations, and the remaining external Pages/ruleset actions
  • Ticket #301 final documentation review corrections: rendered-site anchors from the VitePress renderer and built HTML crawl, stable-0.1.0 operations reconciliation, non-destructive synthetic media capture, dynamic privacy markers, fast/full docs CI split, Markdown lint, exact evidence links, terminology, and truthful README badges
  • Documentation quality gates and quickstart audit: the extended docs contract (inventory, frontmatter, links/anchors, fences, shell, canonical claims, reference, media, site build), the packed-install quickstart TUI E2E and 29.4 s timing, the docs/rationale PR guard, and the unrun human-usability limit
  • Documentation visuals and accessibility: the synthetic terminal replay and screenshots, rendered Mermaid diagrams with accessible titles/descriptions, manifest and alt/caption checks, keyboard/heading/contrast/reduced-motion measurements, and the remaining unrun browser and install limits
  • Public release readiness — parent #268: the released 0.1.0-rc.2 state, every child ticket result including #277 macOS verification and #281–#287, the final-tip install/fake/soak evidence, green required CI at a7f1eb3, the full macOS checklist including 1.7, the public-registry smoke on 2.2.7/2.3.0, npm latest/next behavior, and the remaining announcement decision
  • Live TUI readiness after opening Ensemble: statusline-gated live_open_ensemble, a header-only guard test, the deterministic command-line retype ordering, and passing 2.3.0/2.2.7 fake suites
  • Four-hour enforced soak — 0.1.0-rc.2 final tip: PASS on both supported hosts, OpenCode 2.2.7 and 2.3.0, with 240 samples over 14,439 s each, 0 failures, settled queue/native state, and growth within both gates
  • Release workflow, branch protection, and publish provenance: the applied main ruleset with the real check contexts, the tag/dry-run workflow, the token boundary and provenance, the CycloneDX SBOM, the changelog-based release, the sign-off checklist, and the local dry-run evidence
  • OpenCode 2.3.0 compatibility: the host-range expansion, the 2.3.0 and 2.2.7 fake E2E runs, the passing host probe, the high-contrast focus-tone harness fix, the size matrix, and the loopback install proof
  • Operations runbook, support intake, and templates: the support channel and response targets, diagnostics and rollback runbook, support matrix, issue/PR forms with redaction rules, and the docs-check support assertions
  • MIT license and npm package identity: the MIT license and scoped package metadata, the runtime-only contents allowlist (119 modules), the pack-content guard, and the passing verify/docs/pack dry run
  • Clean-install release candidate: the loopback-registry opencode plugin add proof, the installed server tools and package source, and the exact TUI source-package load failure deferred to #279
  • Built package for the installed TUI plugin: the Solid-compiled dist build, runtime dependency resolution in the plugin cache, the passing 0.1.0-rc.2 TUI route and /ensemble command, and the runtime-only pack guard
  • macOS/Herdr verification via SSH: the macOS 0.1.0-rc.2 install and six server tools, the installed TUI route in Herdr/Ghostty, the full checklist including 1.7 pane-bounded Shift+drag with explicit copy, the macOS fake E2E, and the Git-spec/Herdr-prefix findings
  • Security and privacy audit: threat model and permission review, the handoff symlink containment fix, npm audit baseline and accepted risks, Dependabot configuration, and the passing verify/docs gates
  • Backup/restore and upgrade/rollback: the versioned backup document, the service RPC and operator CLI, the automatic rollback slot, the injected-failure recovery rehearsal, restored-state doctor compatibility, and the upgrade/rollback guides
  • Scale/soak and local diagnostics: the fake soak harness, the tested operating envelope, the redacted support bundle, the no-telemetry policy, and the settled native-session cap fix; the final-tip 4-hour result is in the soak record
  • Production-readiness partials full E2E: the 17-suite matrix on OpenCode 2.2.7, the 2.2.6/2.2.7 compatibility runs, the Node 22/24/26 unit matrix, the latency smoke and reference profiles, and zero retries across the fake suites
  • CI stabilization after the first PR run: the linked E2E record, the settle-aware skill-validation probe, the Ensemble-scoped reduced-motion still frame, and the passing verify/docs/fake matrix
  • Production-readiness partials review fixes: the runtime-seam removal, pure matrix decision, header-scoped live assertion, compatibility artifacts, real card geometry, and the passing verify/latency/fake matrix
  • Full deterministic E2E in CI: cost-split PR jobs, nightly latency and opt-in real-model jobs, failure artifacts, the OpenCode 2.2.7 pin, the job-graph checker, and the passing deterministic matrix
  • Documentation freshness: canonical contract/host/Node constants, the docs index, ADR 0011 at contract v7, and the docs:check CI gate
  • Data retention, export, delete, and roster disclosure: the data.export/data.delete RPC actions and CLI, the documented retention and export format, post-delete workspace health, log deletion, and the v7 AGENTS.md title/role disclosure guidance
  • Live harness flake policy: timing-only retry, flake budget, failure artifacts, and the issue/owner quarantine rule
  • Live harness flake policy verification: readiness-based wizard alignment, classified timing retry, artifact preservation, and the passing deterministic matrix
  • Latency budget and SLO: split rail/patch-card/first-token metrics, reference and smoke profiles, the deterministic patch-card fixture, and the deliberate regression gate
  • Latency budget verification: baseline samples, reference-host gate, regression evidence, and the verify/latency results
  • Node support policy and CI matrix: npm run verify on Node 22.23.4, 24.22.1, and 26.12.1, the >=22.23.4 engine floor, the load-time version gate, and the separate Node matrix workflow
  • Accessibility audit: contrast matrix for dark/light/custom themes, non-color Work status, keyboard-only reachability, reduced-motion/NO_COLOR/ASCII fallbacks, and the ASCII glyph fix
  • Terminal accessibility checklist: manual screen-reader, high-contrast, font-scaling, reduced-motion, ASCII, keyboard-only, and host-limit checks
  • OpenCode peer compatibility matrix: supported 2.2.6/2.2.7 hosts, peer range, runtime warning, and local verify/probe evidence
  • Vera regression and cutover verification: unset-store default-on restart, one leader turn with explicit delegation plus multi-mention routing without duplicates, no completion bounce, exact generated roster table, one-card-per-row size matrix, rollback/restore durability, and the measured AGENTS block delta
  • Native-first default on and rollback: persisted default, explicit off/shadow, tool-surface rollback, global commands, restart requirement, and the passing live matrix
  • Contract v6 and team roster: native-first rules, generated roster table, content-aware reconciliation, refresh triggers, and the measured prompt delta
  • Contract v7 terminal-turn workflow: root-only synthesis, terminal-turn references, ensemble_dispatch continuation, regenerated repository block, and the passing Vera regression
  • Native-first full E2E: all 12 unit/live suites passing on OpenCode 2.2.7, including the real TUI acceptance
  • Native ingestion state and AGENTS v7 full E2E: the 13-suite matrix on OpenCode 2.2.7 covering native ingestion, the v7 terminal-turn contract, Vera, accessibility sizes, and the real TUI
  • One native ingestion state value: one NativeSessionState per session, event order, steering, retry, interruption, adoption, and restart
  • Terminal turns and root-only synthesis E2E: root synthesis, intermediate relay, terminal-turn references, dispatch continuation, and the passing full live matrix
  • Native probe version set and stale live harness: 2.2.6/2.2.7 contract set, Native mirror settings row, and Work interaction before HISTORY
  • Native-first review fixes: privacy toggle, recovered attachment chips, adoption race guard, kill-switch tool path, shadow delta, parallel ingestion, delegated prompt context, and standards cleanup
  • Native-first cutover verification: passing live matrix for direct/workspace turns, mention routing, attachments, adoption, retry/interruption, restart/reconnect, Work history, rollback, shadow comparison, latency, and tokens
  • Native shadow mode: invisible shadow ingestion, content-free comparison rows, no visible writes or routes, and the passing 65/65 + 16/16 live matrix
  • Native prompt attachment size: host data payload derives a bounded positive chip size; live attachment check passes
  • Native-first child-ticket evidence index: where each #210–#230 verification lives
  • Retry and interruption for native turns: one durable turn and Work card across retry, interruption, and settlement
  • Native reconciliation and support surface: reconciliation triggers, context reads, and support-surface limits
  • Legacy direct-message transcript contract: one conversation-channel store for direct messages and the retired per-binding branch
  • Compact synthesis context: root synthesis without the bounded group transcript
  • Multi-ticket and token verification: multi-ticket roots, command-cap refusal, prompt/schema token counts, and the Vera regression
  • Native host event contract: measured OpenCode 2.2.6 order, identities, retry, interruption, background subagent, and move behavior
  • Composer attachments: path and clipboard-image paste, bounded store, prompt files, transcript chips, cleanup, and terminal limits
  • Direct native turn ingestion: direct durable messages, exact settlement, Work card, restart recovery, and the nativeFirst rollback
  • Workspace native turn ingestion: workspace group messages, implicit addressee, subagent children, isolation, bounded stream checkpoints, and the #215 review fixes
  • Immediate prompt after adoption: steered multi-turn executions settle each participating turn with its own exact row
  • Tool retirement and contract shrink: native vs legacy tool surfaces, per-turn context injection, measured prompt/schema reduction, and the one-release rollback
  • Real live workspace-wizard timeout: pane-selection harness sequence, viewport-independent real assertions, and a passing npm run test:live
  • macOS/Herdr manual checklist: repeatable, manual, and unexecuted until a run is recorded
  • v6 verification: nine-scene motion parity, Work-card size matrix, load probe, reduced motion, and NO_COLOR
  • v5.9 verification: live-matrix guardrails, CI, reviewer standards, unused checks, and tracker hygiene
  • v5.8 verification: System Agent init and deliberate skill invocation
  • v5.7 verification: persistent modes, native key ownership, and interaction polish
  • v5.6 verification: modal interactions, composer editing, mouse copy, and statusline

Earlier verification records ​

These older records predate the current evidence index. They are retained unchanged and linked here so the published site keeps every record reachable from the contributing hub.

Implementation and assessments ​

Compatibility evidence ​

Contributor context ​

Released under the MIT License.