10. Quality Requirements
This section makes the quality goals QG-01–QG-05 testable using accepted ADRs. The scenarios specify observable responses and acceptance criteria; they do not report a completed product or accept new architecture.
10.1 Quality tree and verification priorities
Section titled “10.1 Quality tree and verification priorities”| Goal | Quality branches | Scenarios |
|---|---|---|
| QG-01 — Authority | Current permissions, exact identity, artifact integrity, confidentiality and enforceable restrictions | Q02, Q07–Q11, Q16, Q17 |
| QG-02 — Failure and recovery | Atomic publication, bounded stopping, preserved uncertainty, resource bounds and reliable audit | Q03, Q04, Q10, Q13, Q18 |
| QG-03 — Extensibility | Independent extension releases, replaceable providers, versioned contracts and explicitly supported platform mappings | Q05, Q15, Q17, Q19 |
| QG-04 — Local simplicity | Static help, ordinary CLI behavior, noninteractive use and optional organization integrations | Q01, Q12, Q15 |
| QG-05 — Responsiveness | Minimal invocation latency, bounded queues and separately measured preparation/dependency costs | Q06, Q13, Q14 |
P1 prioritizes verification of authority, data integrity, ownership and truthful outcomes: a violation can admit an unauthorized action, lose data or report false success. P2 prioritizes usability, replaceability and performance within those boundaries. These are verification priorities derived from failure consequences, not a new delivery order or permission to omit P2 requirements.
A zero-violation criterion is an invariant checked over the declared test cases and fault schedules; a passing run is not proof for every workload. A selected bound comes from the cited ADR and its effective owner profile; tighter applicable limits win. A timeout or quota is an admission/stopping rule, not a latency or capacity promise. Where a product-level number has not been agreed, the scenario states what to measure and leaves its acceptance target open.
10.2 Quality scenarios
Section titled “10.2 Quality scenarios”
Q01 — Base help without a runtime
Section titled “Q01 — Base help without a runtime”P2 · QG-04. ADR-0002, ADR-0006.
| Element | Scenario |
|---|---|
| Source | CLI user or static completion client |
| Stimulus | Request base help or static command information when the extension runtime is absent |
| Artifact | Verified descriptor, visibility check and core help renderer |
| Environment | Admitted local metadata; runtime uninstalled; allowed and denied visibility; supported locales |
| Response | Render permitted static information without executing extension code, preparing a runtime or initiating sign-in. Keep contextual enrichment a separate invocation. |
| Measure | Zero extension starts, preparation attempts or implicit authentication flows; zero protected metadata disclosed on failed visibility. Apply the descriptor validation limits. |
| Verification and evidence | Instrument launch/preparation/authentication boundaries and compare declared routes and language fallback with output. ADR-0002 has reference help checks; the full descriptor/tooling path remains unverified. |
Q02 — Provider failure, expiry or revocation
Section titled “Q02 — Provider failure, expiry or revocation”P1 · QG-01. ADR-0005, ADR-0011, ADR-0012.
| Element | Scenario |
|---|---|
| Source | Authority provider, owner policy or authentication coordinator |
| Stimulus | Denial, outage, expiry, revocation or a late allow races with queued action admission |
| Artifact | Bound decision, authentication session and final admission gate |
| Environment | Every selected validity mode; concurrent cores; delayed audit; clock discontinuity or lost synchronization |
| Response | Recheck current fences, exact context and remaining validity at actual dispatch. An outage permits reuse only where the previously selected bounded-reuse contract still allows it. |
| Measure | Zero dispatches after an invalidation ordered before admission, outside original scope or using expired/mismatched authority; zero permissive mode changes. Respect decision, dispatch-delay and coordination bounds. |
| Verification and evidence | Schedule races around the admission transaction; inject stale replies, missing state and clock changes. Existing broker stubs cover only limited failures; ADR-0011/0012 coordination and validity checks remain unrun. |
Q03 — Update during a call or publication crash
Section titled “Q03 — Update during a call or publication crash”P1 · QG-02. ADR-0007, ADR-0009.
| Element | Scenario |
|---|---|
| Source | Concurrent owner updates, storage failure or core restart |
| Stimulus | Publish configuration/commands while calls hold references; interrupt a write or lose its acknowledgement |
| Artifact | Active profile, immutable releases, commit journal and protective references |
| Environment | Local storage with its declared durability mapping; competing writers, readers and collection |
| Response | Expose one committed configuration/command pair; keep active calls pinned. Reconcile an uncertain transaction without replay, rollback to a stale policy or premature collection. |
| Measure | Zero mixed revisions, stale overwrites or deletion of referenced content. Each recovery reports a verified commit, confirmed non-commitment or explicit uncertainty/corruption; never unsupported success. |
| Verification and evidence | Fault each write/flush/publication boundary, race capture against collection and exhaust storage quotas. ADR-0007/0009 storage and recovery checks remain unrun. |
Q04 — Cancellation during output, audit or cleanup
Section titled “Q04 — Cancellation during output, audit or cleanup”P1 · QG-02. ADR-0002, ADR-0003, ADR-0004.
| Element | Scenario |
|---|---|
| Source | User, owning scope, invocation deadline or failed execution |
| Stimulus | Stop races with completion, blocked output, required audit, descendants or final commitment |
| Artifact | Lifecycle owner, invocation outcome and remaining finalization budget |
| Environment | Managed and process-only calls; preparation through finalization; orderly shutdown and abrupt owner loss |
| Response | Preserve the first accepted cancellation cause, close admission and use one non-renewable shutdown budget. Retain uncertain effects/references. A dead owner cannot fabricate a committed result. |
| Measure | At most one committed outcome; zero new protected admissions after finalization; no restarted cleanup timer. Stopping, draining and audit share ADR-0003/0004 bounds. stdin EOF alone does not cancel. |
| Verification and evidence | Record monotonic phase/admission/commit times; inject all first-cause races and owner loss. Earlier reference checks are partial; the later lifecycle clarifications and ADR-0004 checks remain unrun. |
Q05 — Replace a provider or embed the core
Section titled “Q05 — Replace a provider or embed the core”P2 · QG-03. ADR-0005, ADR-0008, ADR-0010.
| Element | Scenario |
|---|---|
| Source | CLI owner or trusted integration maintainer |
| Stimulus | Replace a provider with equivalent advertised facilities or compose another host |
| Artifact | Trusted registrations, provider contracts, invocation semantics and ownership |
| Environment | Explicit configuration migration; concurrent old/new profiles; owned and borrowed provider instances |
| Response | Validate required contracts before publication; preserve extension declarations and canonical meanings. Active invocations retain their bindings; unsupported guarantees fail explicitly. |
| Measure | Zero semantic differences in the common conformance fixtures, silent rebinding or disposal of borrowed instances. An equivalent provider change requires no extension-specific vendor client or declaration rewrite. |
| Verification and evidence | Run the same inputs, failures and cancellation schedules against both adapters/hosts; record changed integration surfaces and capability gaps. No development-time target is agreed; ADR-0008/0010 implementation checks remain unrun. |
Q06 — Minimal warmed process invocation
Section titled “Q06 — Minimal warmed process invocation”P2 · QG-05. ADR-0001, ADR-0003.
| Element | Scenario |
|---|---|
| Source | Caller of the minimal process reference workload |
| Stimulus | Start a fresh process, execute its minimal operation, then close; exchange small/large reference IPC payloads |
| Artifact | Process/adapter initialization, complete lifecycle and established IPC |
| Environment | Comparable workstation, exact tested adapter and warmed caches; no runtime installation, remote broker or real command work |
| Response | Complete the selected lifecycle without omitting validation, authorization or required cleanup to improve timing. |
| Measure | Meet ADR-0001’s unchanged p95 regression limits for readiness, complete invocation and both IPC payloads. Report median/p95, maxima, errors and sample counts separately for each adapter. |
| Verification and evidence | Repeat the recorded sampling method and retain raw timings/environment metadata. ADR-0003’s Python readiness and complete-invocation p95 values exceed the targets; acceptance did not relax them. |
Q07 — Tool changes after listing
Section titled “Q07 — Tool changes after listing”P1 · QG-01. ADR-0013, ADR-0007.
| Element | Scenario |
|---|---|
| Source | MCP client and concurrent profile/policy publisher |
| Stimulus | Call a listed tool after its profile, release or visibility binding changes |
| Artifact | Derived MCP projection, exact tool name, cursor and invocation capture |
| Environment | Concurrent listing/calling/publication; cached names; different callers and surviving active invocations |
| Response | Reject obsolete or invisible names safely, require fresh discovery and never route by a similar display name. Keep already admitted calls pinned with current action checks. |
| Measure | Zero replacement dispatches or cross-caller metadata leaks; zero extension starts/preparations during listing; one profile source for CLI and MCP. Invalidated cursors do not merge revisions. |
| Verification and evidence | Race list pages and exact-name capture against publication and visibility changes; inspect dispatch identity, not just labels. ADR-0013 projection and adapter checks remain unrun. |
Q08 — Forged authority or concurrent delegated spending
Section titled “Q08 — Forged authority or concurrent delegated spending”P1 · QG-01. ADR-0014, ADR-0012.
| Element | Scenario |
|---|---|
| Source | Agent caller, delegated child or parallel core process |
| Stimulus | Forge a task/actor binding, broaden delegation, race the last allowance or reconnect after expenditure |
| Artifact | Trusted issuance, task lineage, current-rights intersection and shared admission ledger |
| Environment | Independently bound issuer; multiple processes/children; restart, expiry, revocation and retained reservations |
| Response | Bind authority at final admission; consume counted attempts and reserve enforceable maximum costs atomically across ancestors. Apply issuance limits to automatic roots; expiry blocks new admission without becoming a new process timeout. |
| Measure | Zero scope expansion, overspend or balance reset. Spent plus reserved cost stays within every applicable ceiling; uncertain effects retain reservations. Unknown maximum cost cannot satisfy a required hard ceiling. |
| Verification and evidence | Race independent processes, duplicate references and root issuance; restart at each ledger boundary. Check selected lineage, lifetime and storage bounds. No issuer/coordinator implementation is confirmed; independent authorized roots need an explicit aggregate policy for an aggregate bound. |
Q09 — Required confinement is unavailable
Section titled “Q09 — Required confinement is unavailable”P1 · QG-01. ADR-0015, ADR-0008.
| Element | Scenario |
|---|---|
| Source | Command requirements, owner profile or changing platform state |
| Stimulus | Require restrictions the selected executor cannot establish, or replace a prepared resource before launch |
| Artifact | Exact restriction plan, runtime/executable binding and supervised execution scope |
| Environment | Each claimed OS/runtime mapping; descendants; mutable borrowed environments; separately authorized preparation |
| Response | Verify all required restrictions before extension startup or reject. Ordinary explicit host-trusted execution remains possible with its actual OS authority. |
| Measure | Zero starts under missing mandatory restrictions, weaker fallback or unsupported claims. Probe every advertised restriction, including descendants and direct access; uncertain termination retains references. |
| Verification and evidence | Use platform-specific attempts to violate filesystem, network, inheritance and resource restrictions under the exact plan. ADR-0015 selects no proven adapter; process supervision and MCP exposure are not confinement evidence. |
Q10 — Effect sent but response lost
Section titled “Q10 — Effect sent but response lost”P1 · QG-01, QG-02. ADR-0016, ADR-0014.
| Element | Scenario |
|---|---|
| Source | Protected service, transport failure or crash during effect commitment |
| Stimulus | Lose a response, repeat a permit, or substitute payload/resource/preconditions after approval |
| Artifact | Exact effect binding, one-use permit, durable dispatch intent and cost reservation |
| Environment | Parallel callers; crashes before/after durable admission and dispatch; service-specific idempotency/consistency |
| Response | Admit only the approved binding; retain unknown effects and costs. Reconcile using authorized evidence without command replay or rewriting the original outcome. |
| Measure | At most one commitment per permit; zero dispatch on a changed binding or unconfirmed admission. No automatic replay, false no-effect proof or release of unknown cost. Local records do not promise universal exactly-once service execution. |
| Verification and evidence | Fault every commitment boundary and vary lost replies, key expiry, payload mismatch and object replacement. Count actual service effects separately from invocation outcomes. Ledger/adapters and reconciliation checks remain unrun. |
Q11 — Altered or invalid publication
Section titled “Q11 — Altered or invalid publication”P1 · QG-01. ADR-0006, ADR-0007, ADR-0009.
| Element | Scenario |
|---|---|
| Source | Publisher, catalog/artifact source or malformed package |
| Stimulus | Change bytes behind a reference; supply an unapproved digest, invalid descriptor, conflicting route or unsafe archive |
| Artifact | Owner trust binding, verified metadata, extraction staging and candidate command tree |
| Environment | Local files and approved source adapters; redirected/moved references; compressed oversized payloads |
| Response | Independently validate authority, exact bytes, paths, schemas and mounted routes before activation. Publisher checks or KCL compilation cannot replace client validation. |
| Measure | Zero execution/activation of rejected content, writes outside owned staging or partial active-tree changes. Enforce both declared and observed size limits and the bounded validation deadline. |
| Verification and evidence | Mutate digests, aliases, targets and archive paths; test boundary sizes and independently supplied owner trust. A signing service is not required by the baseline. ADR-0006/0007/0009 checks remain unrun. |
Q12 — Terminal, pipelines and headless execution
Section titled “Q12 — Terminal, pipelines and headless execution”P1 · QG-04, QG-02. ADR-0002, ADR-0010, ADR-0013.
| Element | Scenario |
|---|---|
| Source | User, pipeline consumer or noninteractive CLI/MCP caller |
| Stimulus | Redirect binary streams, close a consumer, request input without an approved terminal, resize or cancel an active view |
| Artifact | Command streams, terminal binding, dialog/view ownership and protocol stdout |
| Environment | stdout/stderr independently redirected; physical terminals and headless hosts; competing interactions |
| Response | Preserve redirected bytes, render terminal text safely and route keyboard input only through the core. Reject missing required interaction; never invent an answer or open a hidden terminal. |
| Measure | Redirected command data retains byte fidelity, without injected UI/control framing or implicit use of pipeline input as answers. Account separately for owner-selected diagnostics on stderr; MCP stdout contains only its protocol. One foreground input owner; restoration and consequential input follow declared delivery/cleanup bounds. |
| Verification and evidence | Compare binary fixtures; exercise focus, keyboard cancellation, resize, stale events and terminal restoration. ADR-0002’s model is headless; real terminal and assistive-technology support need their own evidence. |
Q13 — Saturated queues and concurrent calls
Section titled “Q13 — Saturated queues and concurrent calls”P1 · QG-02, QG-05. ADR-0003, ADR-0010, ADR-0013.
| Element | Scenario |
|---|---|
| Source | Noisy extension, slow output consumer or concurrent callers |
| Stimulus | Fill pending requests, output/log queues, observers or invocation capacity while cancellation arrives |
| Artifact | Bounded parsers/queues, lifecycle capacity, host handles and transport writers |
| Environment | At the selected limits and one unit beyond; simultaneous read/write; stalled partial frames |
| Response | Reject excess admission, count permitted log drops and preserve control progress. Do not silently drop consequential input; close a failed transport and finalize its owned work. |
| Measure | No configured queue/concurrency bound exceeded; fixed partial-read/write timers never reset with progress. Cancellation/finalization obey existing bounds. Report peak RAM/CPU/handles; quotas do not establish a whole-process memory or throughput target. |
| Verification and evidence | Saturate each bound separately and in combination; record queue peaks, rejected work and cleanup obligations. ADR-0003 has limited channel/load evidence; host/MCP capacity remains unverified. |
Q14 — Cold preparation and slow dependencies
Section titled “Q14 — Cold preparation and slow dependencies”P2 · QG-05, QG-02. ADR-0005, ADR-0007, ADR-0008.
| Element | Scenario |
|---|---|
| Source | First permitted caller, runtime/source provider or external authority |
| Stimulus | Execute with a missing/cold environment, delayed transfer or metadata response; cancel one shared waiter |
| Artifact | Authorized preparation plan, environment readiness/reference and dependency calls |
| Environment | Cached versus uncached artifacts; local versus remote sources; concurrent callers sharing one exact key |
| Response | Prepare only the approved plan; publish verified readiness once; keep other authorized waiters’ ownership. Stop on fixed deadlines or exhausted quotas without evicting referenced/uncertain data. |
| Measure | Stay within each selected phase/writer/cache bound; no duplicate writer publication or extension start on incomplete state. Measure resolve, transfer, verify, acquire, authority and launch times separately; no agreed end-to-end cold-start SLO exists. |
| Verification and evidence | Vary cache state, service delay, transfer inactivity, cancellation and writer crashes; record bytes, reservations and wait times. Runtime-provider and storage checks remain unrun; website use of mise supplies no product evidence. |
Q15 — Local operation and organization integration
Section titled “Q15 — Local operation and organization integration”P2 · QG-03, QG-04. ADR-0001, ADR-0005, ADR-0009, ADR-0011.
| Element | Scenario |
|---|---|
| Source | Small-team owner or organization integrator |
| Stimulus | Use an approved local catalog, then explicitly add required identity/permission or platform adapters |
| Artifact | Catalog, owner profile, extension declarations and platform/provider bindings |
| Environment | Local artifacts with no required remote service; separately configured corporate/CI profiles; exact OS/runtime/store tuple |
| Response | Keep the local baseline useful without mandatory login, identity/permission service, database server or client KCL. Add integrations through registered contracts; unsupported mappings and unavailable required services fail explicitly. |
| Measure | Zero unrequested infrastructure/authentication startup in the local fixture; unchanged extension semantics for equivalent integrations. Every claimed platform facility passes its mapping checks; no global offline or all-platform guarantee follows. |
| Verification and evidence | Run the same command fixtures with local policy and declared organization adapters, recording actual endpoint/store requirements. Early simulated identity and Windows/WSL checks are not proof of corporate or general platform support. |
Q16 — Secret handling, renewal and logout
Section titled “Q16 — Secret handling, renewal and logout”P1 · QG-01. ADR-0011, ADR-0012.
| Element | Scenario |
|---|---|
| Source | Authentication provider, secret store or concurrent account-management caller |
| Stimulus | Lock storage, lose a refresh response, change account or log out while an old writer returns |
| Artifact | Protected credential references, session epochs, refresh family and request-bound credential use |
| Environment | Persistent/explicit volatile modes; multiple cores; noninteractive CI with a registered service identity; approved targets and audiences |
| Response | Preserve durable invalidation and uncertainty, block replay of a possibly consumed refresh token and keep general credentials behind approved scoped adapters. Limited credential issuance has its own authorization contract. Never fall back to plaintext or another account. |
| Measure | No stale-session resurrection or unauthorized credential use. General identity/service/registry credentials never escape to extensions; separately authorized limited credentials retain their recipient/scope/expiry contract. Neither appears in diagnostics. One refresh exchange per admitted family attempt; confirmed logout fences later admissions. |
| Verification and evidence | Use identifiable test secrets; inspect defined output/storage channels; race refresh, logout and restart at publication boundaries. Do not infer isolation from arbitrary host-trusted code. Credential/coordinator checks remain unrun. |
Q17 — Contract or session mismatch
Section titled “Q17 — Contract or session mismatch”P1 · QG-01, QG-03. ADR-0003, ADR-0006, ADR-0013.
| Element | Scenario |
|---|---|
| Source | Extension adapter or CLI/MCP client |
| Stimulus | Supply an unsupported version/feature, wrong session, malformed frame or a value with ambiguous representation |
| Artifact | Static capability checks, private session, typed input/result schema and outer protocol adapter |
| Environment | Every advertised binding/runtime tuple; split frames, concurrent launches and old/new compatible versions |
| Response | Reject known incompatibility before startup; initialization mismatch permits no invocation dispatch. Parse supported values once without silent narrowing, downgrade or channel substitution. |
| Measure | Zero pre-confirmation handler calls, cross-session authority substitution or unsupported result acceptance. Equivalent CLI/MCP typed fixtures preserve values, defaults and explicitly supplied IDs; private IPC and outer MCP remain separate. |
| Verification and evidence | Exercise invalid encodings, numeric edges, duplicate IDs/keys, inherited-handle isolation and result schemas. ADR-0003 covers specific Go/Python Windows adapters; descriptor and MCP refinements remain unverified. |
Q18 — Diagnostics and mandatory audit failure
Section titled “Q18 — Diagnostics and mandatory audit failure”P1 · QG-02, QG-01. ADR-0002, ADR-0003, ADR-0016.
| Element | Scenario |
|---|---|
| Source | Extension log producer, diagnostic destination or mandatory audit service |
| Stimulus | Forge log metadata, include a known secret, block delivery or fail audit before/after an action |
| Artifact | Core-origin correlation, diagnostic routing, audit obligation and final outcome |
| Environment | Concurrent invocations; local/remote sinks; saturation, cancellation and an already admitted external effect |
| Response | Keep publisher text separate from trusted metadata; redact known secrets and bound ordinary delivery. Required pre-action audit gates dispatch; failed post-action audit prevents false success and preserves uncertainty. |
| Measure | Zero metadata spoofing or diagnostic bytes in command/protocol stdout. No protected dispatch before required pre-audit confirmation; no ordinary success with an unmet mandatory obligation. Diagnostic loss is counted and cannot extend shutdown. |
| Verification and evidence | Inject sink failures and inspect trusted correlation, typed reason and outcome rather than trusting log prose. Reference fixtures cover limited audit failures; durable audit/retention and effect-service guarantees remain unverified. |
Q19 — Independent extension evolution
Section titled “Q19 — Independent extension evolution”P2 · QG-03. ADR-0001, ADR-0006, ADR-0007, ADR-0010, ADR-0013.
| Element | Scenario |
|---|---|
| Source | Team maintaining an ordinary command extension |
| Stimulus | Publish a compatible release adding/updating a command or subcommand; alternatively require a feature absent from an older core |
| Artifact | Extension release, mounted command/alias tree, common input contract and optional MCP projection |
| Environment | Existing branded CLI build and trusted registrations; running calls pinned to the previous release; supported and unsupported contract versions |
| Response | Admit and activate a compatible release without rebuilding the CLI or changing its registered integrations. Preserve canonical identity and access meaning; validate explicit MCP eligibility/exposure. Reject collisions or unsupported required features without downgrade or partial publication. |
| Measure | Host build and common input meanings remain unchanged for an ordinary compatible update; zero ambiguous routes, aliases that alter authority or silently retargeted running calls. Both CLI and any selected MCP projection derive from the same committed profile. |
| Verification and evidence | Install two compatible releases against one host build, retain an active old call, then test collisions and a required unsupported feature. Compare routes, typed inputs, grants and release bindings; record actual change surfaces/effort without promising developer hours. Integrated publication and MCP checks remain unrun. |
10.3 Reproducible verification envelope
Section titled “10.3 Reproducible verification envelope”A result is meaningful only with its workload and boundary recorded. The following are measurement classes, not additional supported capacities or product SLAs.
| Class | Inputs to record and vary | Required observations |
|---|---|---|
| Minimal warmed process | ADR-0001/0003 reference workload, adapter/runtime versions, payload sizes, warmups and samples | Readiness, complete lifecycle and established IPC distributions; retain the original timer boundaries for comparison |
| Cold local path | Core startup and OS-cache state; runtime present/absent; installed/uncached artifacts; descriptor/command counts and sizes | Separate host, metadata/trust, preparation and execution times; record actual transfers, launches and filesystem work |
| Local metadata growth | Catalog entries, commands, aliases, locales, input sizes/depth and page counts from a small fixture through selected bounds | Validation/help/list latency, memory, cancellation response and exact rejection at each exceeded bound |
| Dependent services | Adapter/endpoint type, selected validity mode, injected latency, failure, synchronization and credential/store state | Core time versus dependency time, admission decisions, timeout/error counts and absence of unauthorized fallback/replay |
| Concurrency and saturation | Active calls/cores/roots, shared preparation keys, queue occupancy, slow consumers and payload mix | Admission/rejection counts, queue/resource peaks, common-budget conservation, fairness observations and stopping behavior |
| Failure and recovery | Exact fault point before/after write, flush, admission, dispatch or acknowledgement; filesystem/platform mapping | Authoritative state after recovery, retained references/reservations, real effect count and truthful uncertainty |
Record the hardware, OS/filesystem, terminal mode, adapter/build/configuration revisions, owner limits, cache definition, background load, clock, percentile method and raw samples. Report failed/timed-out attempts separately; do not discard slow successful samples. For new workload classes, record median/p95/p99 where sample size supports them, maxima, throughput, CPU, peak memory, disk/staging usage and handle counts as relevant. These measurements describe the tested envelope; only the applicable ADR’s agreed threshold determines a pass.
Verify each selected limit at its boundary and immediately beyond it, including interactions where the smaller message, invocation, task or owner bound wins. Fault schedules must cover concurrent processes and restart where the contract requires durable coordination. Test results identify fixtures, observed outcome and evidence artifacts; a diagram render or documentation check is not product conformance.
10.4 Evidence and unresolved targets
Section titled “10.4 Evidence and unresolved targets”| Recorded evidence | What it supports and what it does not |
|---|---|
| ADR-0001, 2026-09-20 | Historical minimal warmed Go/Node measurements and process feasibility. Common checks passed on the recorded Windows/Ubuntu environments; Windows terminal/tree checks retained 3 failures concerning Node line-input resize. Temporary prototypes/raw timing data are no longer in this repository. |
| ADR-0002, 2026-09-20 | 137 reference checks on Windows/Node cover selected semantics, real child streams and a headless interaction model. Providers, audit and UI events are controlled fixtures; this is not a physical terminal renderer or production SDK. |
| ADR-0003, 2026-09-21 | 126 checks cover the recorded Windows Go/Python inherited-channel reference. Python p95 readiness 166.002 ms and full invocation 185.434 ms exceed the unchanged comparison limits; established IPC meets its comparison limits. The later lifecycle clarifications and CLI/MCP refinements are not covered by that run. |
| ADR-0004–0016 | Acceptance establishes the contracts. Their stated implementation checks remain outstanding; older fixtures do not establish persistent coordination, production providers, MCP, confinement or effect reconciliation. |
The following acceptance targets still require a concrete deployment/workload decision and measurements; this page does not assign them new values:
| Open target | Required specification before making a claim |
|---|---|
| Real-command responsiveness | End-to-end cold/warm and local/dependent-service latency/error targets, with phase boundaries and representative commands; reference IPC targets do not supply them |
| Supported operating envelope | Sustained throughput, catalog/profile/task scale and total CPU/RAM/disk/handle limits for exact platform/adapter combinations; existing quotas remain protective bounds |
| Accessibility and platform support | Physical-terminal/keyboard/resize and assistive-technology test matrix, supported OS/runtime/storage combinations and any claimed accessibility standard |
| Corporate operation | Required endpoints, proxy/private-CA and credential-store capabilities, disconnected validity policy, distribution/revocation delays and external-service guarantees for the chosen deployment |
| Operational continuity | Availability and recovery-time/data-loss objectives, durable audit retention/access and storage maintenance beyond the accepted recovery contracts; no service-wide SLA or universal RTO/RPO is currently established |
See risks and technical debt for the consequences of these evidence gaps. Exact field contracts, quotas, deadlines and error codes remain authoritative in their defining ADRs.