dsh-env-manager-plugin 0.1.0 → 0.1.1

package/docs/01-architecture/detailed-design.md

@@ -1,961 +1,961 @@
-# Detailed Design (V1)
-
-> Status: normative implementation design for V1. It refines
-> [`design.md`](design.md) into module-level contracts and depends on the verified
-> runtime facts in [`api-verification.md`](api-verification.md).
-> Record shapes are defined in [`data-model.md`](data-model.md); do not restate or
-> extend them here without updating that document.
-
-## 1. Scope
-
-This document specifies:
-
-1. the plugin configuration and its effective defaults;
-2. every Host module's public interface and internal algorithm;
-3. the exact `defineTool` contract for all ten tools;
-4. the plugin-owned HTTP API used by the Client panel;
-5. error mapping, limits, concurrency and test obligations.
-
-UI screens and copy are in [`ui-design.md`](ui-design.md); build and packaging are in
-[`design.md`](design.md) §15 and [`implementation-plan.md`](implementation-plan.md).
-
-## 2. Conventions
-
-- Host: TypeScript, ESM, `strict`.
-- Runtime validation: Zod. Tool parameter validation: DSH's schema DSL (F3).
-- Files: YAML 1.2 via the `yaml` package; no anchors/aliases are emitted.
-- Time: ISO-8601 UTC with milliseconds, produced by `new Date().toISOString()`.
-- Paths: POSIX separators inside resolved values; OS-native paths only at the
-  filesystem boundary.
-- All expected failures are `EnvManagerError extends HarnessError` (F5) with a code
-  from `data-model.md` §11.
-
-## 3. Configuration
-
-```ts
-// src/config.ts
-import z from '@deepseek-ai/schemastery'
-
-export const Config = z.object({
-  dataDir: z.string().description('Registry directory; relative paths resolve against the profile directory.'),
-  maxRecords: z.natural().default(2000),
-  registerSkill: z.boolean().default(true),
-  writePolicy: z.object({
-    allowUiProductionWrites: z.boolean().default(false),
-    requireApprovalForDelete: z.boolean().default(true),
-    requireApprovalForProductionSave: z.boolean().default(true),
-  }).default({}),
-  defaults: z.object({
-    environmentId: z.string().default(''),
-  }).default({}),
-})
-```
-
-| Field | Default | Meaning |
-|---|---|---|
-| `dataDir` | profile data dir | Registry root; see §4.1 |
-| `maxRecords` | 2000 | hard cap per record kind; exceeding it fails a write |
-| `registerSkill` | true | register the profile-wide `dsh-env-manager` discovery Skill |
-| `writePolicy.allowUiProductionWrites` | false | allow the HTTP API to write production records |
-| `writePolicy.requireApprovalForDelete` | true | `env_delete` asks the human first |
-| `writePolicy.requireApprovalForProductionSave` | true | `env_save` on a production environment asks first |
-| `defaults.environmentId` | empty | fallback when neither an argument nor the session default exists |
-
-Rules:
-
-- The config schema is schemastery (F1/F18), not Zod. Do not add a second config path.
-- Config is read-only at runtime; there is no write API. The Client shows effective
-  values only.
-- A blank `dataDir` means "use the resolution order in §4.1", not "current directory".
-
-## 4. Storage
-
-### 4.1 Data directory resolution
-
-```text
-1. config.dataDir, if set:
-     absolute  -> use as-is
-     relative  -> path.resolve(profileDir, config.dataDir)
-2. profileDir from ctx.get('profileContext')?.dir
-     -> <profileDir>/data/dsh-env-manager
-3. process.env.DSH_PROFILE_DIR
-     -> <DSH_PROFILE_DIR>/data/dsh-env-manager
-4. process.cwd()/data/dsh-env-manager
-```
-
-`profileDir` may be unavailable in an embedded host; step 3 and 4 keep the plugin
-usable. The resolved absolute path is memoized for the process lifetime, logged once,
-and reported by `env_get_context` and the UI overview. It is never derived from model
-input.
-
-### 4.2 Layout
-
-```text
-<dataDir>/
-├── environments/<environmentId>.yaml
-├── services/<serviceId>.yaml
-├── deployments/<deploymentId>.yaml
-├── rules/<ruleProfileId>.yaml
-├── operations/<bindingId>.yaml
-├── sessions/<sessionId>.json
-└── .backups/<kind>/<id>/<timestamp>.yaml      (rolling, max 5)
-```
-
-- The file name is always `<id>.yaml`; the record `id` must equal the file stem.
-- `sessions/` and `.backups/` are never scanned as registry records.
-- Directories are created lazily by the first write; a missing root on read yields an
-  empty registry, not an error.
-
-### 4.3 Id and path safety
-
-```ts
-const ID_PATTERN = /^[a-z0-9][a-z0-9-]{0,62}$/
-```
-
-`resolveRecordPath(kind, id)` rejects a non-matching id **before** touching the
-filesystem, then verifies `path.resolve(dir, id + '.yaml')` starts with
-`path.resolve(dir) + path.sep`. Session ids are validated as `[A-Za-z0-9_-]{1,128}`
-and hashed only if they fail that pattern.
-
-### 4.4 Atomic write protocol
-
-```ts
-async function writeRecord(kind, id, record) {
-  const target = resolveRecordPath(kind, id)
-  await mkdir(dirname(target), { recursive: true })
-  const tmp = join(dirname(target), `.${id}.${process.pid}.${randomUUID()}.tmp`)
-  const text = stringifyYaml(record)
-  const handle = await open(tmp, 'w')
-  try {
-    await handle.writeFile(text, 'utf8')
-    await handle.sync()
-  } finally {
-    await handle.close()
-  }
-  await rotateBackup(kind, id, target)   // copy previous file, prune to 5
-  await rename(tmp, target)              // atomic replace on POSIX and Windows
-  return { file: target, bytes: Buffer.byteLength(text) }
-}
-```
-
-Guarantees:
-
-- A reader never sees a partially written record.
-- A crash before `rename` leaves an ignorable `.tmp` file; startup cleanup removes
-  `*.tmp` older than one hour.
-- Backup failure does not block the write; it is reported as a warning.
-- `writeRecord` is serialized per target path by an in-process promise chain (§4.8).
-
-### 4.5 Read protocol
-
-```ts
-async function readRecord(kind, id): Promise<ReadResult> {
-  const target = resolveRecordPath(kind, id)
-  let text
-  try { text = await readFile(target, 'utf8') }
-  catch (e) { if (e.code === 'ENOENT') return { found: false }; throw e }
-  const parsed = parseYamlDocument(text)          // throws RECORD_INVALID
-  return { found: true, raw: parsed, text }
-}
-```
-
-Invalid YAML, a non-object document, an id/filename mismatch, or a schema failure
-produces a `Diagnostic` and the record is treated as absent for that request. The file
-is never rewritten or deleted.
-
-Unknown keys are preserved on read and reported as an `UNKNOWN_FIELD` warning, but
-`env_save` rejects them (`RECORD_INVALID`) so a typo cannot enter the store. Storage
-therefore parses twice: `.passthrough()` on read, strict on save.
-
-### 4.6 Delete
-
-```text
-1. Resolve the target and confirm it exists.
-2. Ask the registry whether any record references it (§6.3).
-3. If referenced -> RECORD_IN_USE listing referrers.
-4. Otherwise rotate a backup, then unlink the file.
-```
-
-Deleting a session file is a plain unlink and needs no reference check.
-
-### 4.7 Session files
-
-```json
-{
-  "schemaVersion": 1,
-  "sessionId": "01J...",
-  "defaultEnvironmentId": "test-a",
-  "updatedAt": "2026-11-01T03:00:00.000Z"
-}
-```
-
-Written with the same temp+rename protocol. A corrupt or unreadable file is reported
-as `persisted: false` and treated as "no default"; it is not deleted.
-
-### 4.8 Concurrency
-
-- One serialized chain per file path (`Map<string, Promise<void>>`), released when the
-  chain drains.
-- Registry loads are deduplicated: concurrent callers share one in-flight load.
-- No cross-process lock in V1. Two DSH processes writing the same file can lose one
-  update; the backup makes the loss recoverable. Documented as a limitation.
-
-### 4.9 Limits
-
-| Limit | Value | Failure |
-|---|---|---|
-| records per kind | `maxRecords` (2000) | `RECORD_LIMIT_EXCEEDED` |
-| record YAML bytes | 1 MiB | `RECORD_TOO_LARGE` |
-| session file bytes | 64 KiB | `SESSION_CONTEXT_UNPERSISTED` |
-
-## 5. Domain
-
-```ts
-// src/domain/types.ts
-export type RecordKind = 'environment' | 'service' | 'deployment' | 'ruleProfile' | 'operationBinding'
-export type AnyRecord = Environment | Service | Deployment | RuleProfile | OperationBinding
-```
-
-`src/domain/schemas.ts` exports one Zod schema per record plus:
-
-```ts
-export function parseRecord(kind: RecordKind, value: unknown, origin: string):
-  { ok: true; record: AnyRecord } | { ok: false; diagnostic: Diagnostic }
-```
-
-Validation responsibilities:
-
-- `RecordMeta`: id pattern, `schemaVersion === 1`, ISO timestamps, `updatedAt >= createdAt`.
-- Cross-field rules from `data-model.md`: at least one `deployModes`, host role ids
-  exist, `location.kind === runtimeMode`, unique host id / ip / cluster id /
-  endpoint id, unique account id per host, `loginRecords[].hostRef`/`accountId`
-  resolve within the environment, host `clusterRef` resolves and is present for
-  `k8s-all-in-one`/`k8s-worker`, `access` values are unique, `paths` keys and variable
-  names match `[A-Za-z_][A-Za-z0-9_]*`.
-- `structural` shape only. Referential integrity is the registry's job (§6.3), because
-  it needs the whole snapshot.
-
-`parseRecord` never throws for bad data; it returns a diagnostic with `path` pointing
-at the failing field.
-
-Tool input is parsed by `parseRecordInput`, which is identical except `createdAt` and
-`updatedAt` are optional. The app then calls `finalizeRecord(record, previous?)` to
-stamp them before `parseRecord` and before writing.
-
-## 6. Registry
-
-### 6.1 Snapshot
-
-```ts
-interface RegistrySnapshot {
-  environments: Map<string, Environment>
-  services: Map<string, Service>
-  deployments: Map<string, Deployment>
-  ruleProfiles: Map<string, RuleProfile>
-  operationBindings: Map<string, OperationBinding>
-  byAlias: Map<string, string>          // "service:<alias>" -> id
-  diagnostics: Diagnostic[]
-  loadedAt: string
-}
-```
-
-`byAlias` keys are `${kind}:${alias}` and resolve aliases for `env_get`/`env_find`.
-Duplicate aliases are a warning; the first record wins, sorted by file name for
-determinism.
-
-### 6.2 Load algorithm
-
-```text
-for each kind directory (environments, services, deployments, rules, operations):
-  entries = readdir(dir, { withFileTypes: true })
-  files   = entries where isFile() and name endsWith('.yaml') and not name startsWith('.')
-  sort files by name
-  for each file:
-    read + parse
-    parseRecord -> on failure: diagnostic (file, code, path, message); skip
-    if map already has record.id: diagnostic DUPLICATE_ID; skip later file
-    else insert
-refresh byAlias and run integrity checks
-```
-
-A missing kind directory is normal (empty). An unreadable directory is an error
-diagnostic, and the snapshot keeps the other kinds.
-
-### 6.3 Referential integrity
-
-| Reference | Rule | Code |
-|---|---|---|
-| `deployment.environmentId` | environment exists | `ENV_NOT_FOUND` |
-| `deployment.serviceId` | service exists | `SERVICE_NOT_FOUND` |
-| `deployment.location.hostRef` | host exists in that environment | `HOST_NOT_FOUND` |
-| `deployment.location.clusterRef` | cluster exists in that environment | `CLUSTER_NOT_FOUND` |
-| environment `hosts[].clusterRef` | cluster exists; required for K8s node kinds | `RECORD_INVALID` |
-| environment `endpoints[].hostRef` / `clusterRef` | target exists in that environment | `RECORD_INVALID` |
-| `service.ruleProfileRefs[]` | rule profile exists | `RULE_PROFILE_NOT_FOUND` |
-| operation binding `match.environmentId` | environment exists | `ENV_NOT_FOUND` |
-| operation binding `match.serviceId` | service exists | `SERVICE_NOT_FOUND` |
-| operation binding `match.deploymentId` | deployment exists | `DEPLOYMENT_NOT_FOUND` |
-| any `credentialRef` | non-empty string matching the agreed ref pattern | `RECORD_INVALID` |
-
-Schema validation rejects an environment with zero logical roles and a host with zero
-roles before integrity checks run; both fail as `RECORD_INVALID`. A host role id that is
-not declared in `Environment.roles` is reported as `ROLE_NOT_FOUND` by `env_validate`;
-the panel prevents selecting an undeclared role in the node form.
-
-Integrity failures are **diagnostics**, not load failures: the registry still serves
-the other records, and `env_validate` reports the broken references. A resolve or
-route that needs a missing record fails with the specific code.
-
-### 6.4 Reload policy
-
-V1 reloads the registry at the start of every application request. Rationale: users
-edit YAML by hand, and a stale cache is worse than a few milliseconds. Concurrent
-requests share one in-flight load. A snapshot is replaced only after a successful
-load; a load that throws keeps the previous snapshot and reports an error diagnostic.
-
-Performance target: 500 records load in under 100 ms on a local SSD. If that is
-missed, add an mtime-keyed cache in a follow-up, not in V1.
-
-## 7. Resolver
-
-### 7.1 Template parsing
-
-```ts
-type TemplatePart =
-  | { kind: 'literal'; text: string }
-  | { kind: 'ref'; path: string }
-
-function parseTemplate(text: string): TemplatePart[]
-```
-
-Scanner rules:
-
-1. Find `${`. Text before it is a literal.
-2. Find the next `}`. A missing `}` is `RULE_TEMPLATE_INVALID`.
-3. The inner text must match `^[A-Za-z_][A-Za-z0-9_]*(\.[A-Za-z_][A-Za-z0-9_]*)*$`;
-   otherwise `RULE_TEMPLATE_INVALID`.
-4. Anything that is not `${…}` is literal text, including a lone `$`.
-5. No nesting, escaping, calls, operators, or env access.
-6. Template length is capped at 2048 characters.
-
-### 7.2 Namespaces
-
-| Reference | Resolves to |
-|---|---|
-| `env.id`, `env.name`, `env.category` | environment record |
-| `service.id`, `service.name`, `service.category` | service record |
-| `service.artifact.name` | service artifact name, when declared |
-| `deployment.id`, `deployment.name`, `deployment.runtimeMode` | deployment record |
-| `deployment.location.<field>` | deployment location, runtime-specific |
-| `<variable>` | layer variable of that name |
-| `variables.<variable>` | same, explicit form |
-| `paths.<pathName>` | another resolved path |
-
-`service.artifact.name` resolves only when `artifact` exists. A reference to a
-namespace that is unavailable for this context is `RULE_UNDEFINED_VARIABLE`, not an
-empty string.
-
-### 7.3 Layers
-
-```ts
-interface Layer {
-  source: RuleSource
-  variables?: Record<string, string | number | boolean>
-  paths?: Record<string, string>
-  logSources?: Partial<Record<RuntimeMode, LogSource>>
-  classLocations?: Partial<Record<RuntimeMode, ClassLocation>>
-}
-```
-
-Built from `design.md` §6.3:
-
-1. built-in global defaults (empty in V1);
-2. each `service.ruleProfileRefs` entry, in order;
-3. `environment.ruleOverrides`;
-4. `deployment.variables` (variables only), then `deployment.ruleOverrides`.
-
-Flatten to a field map:
-
-```text
-variables.<name>
-paths.<name>
-logSources.<runtime>.<property>
-classLocations.<runtime>.<property>
-```
-
-### 7.4 Merge
-
-```ts
-interface MergedField {
-  key: string
-  raw: string | number | boolean
-  definedBy: RuleSource
-  shadowed: RuleSource[]
-}
-```
-
-Iterate layers in ascending order. For each field, replace the winner and append the
-previous winner to `shadowed`. This is the only merge rule; there is no deep merge and
-no array concatenation.
-
-### 7.5 Evaluation
-
-```text
-1. Build the reference graph: node = merged field key; edge = each ref used by a
-   string field. Constants (env/service/deployment) are not nodes.
-2. DFS with an active stack. A back edge is RULE_CYCLE with the full cycle path.
-3. A ref with no node and no constant is RULE_UNDEFINED_VARIABLE
-   (message names the field and the ref).
-4. Resolve nodes in post-order; substitute into an ordered list of parts.
-5. For every resolved `paths.*` value, split on `/` and reject `..`
-   (RULE_PATH_TRAVERSAL). Absolute/relative shape is otherwise preserved.
-6. Convert non-string numbers/booleans to their literal text when interpolated.
-7. Select `logSources[runtimeMode]` and `classLocations[runtimeMode]` after their
-   strings resolve.
-```
-
-Cycle detection is iterative (explicit stack), not recursive, so a pathological graph
-cannot overflow the JS stack.
-
-### 7.6 Provenance
-
-For every requested field:
-
-```ts
-interface ProvenanceEntry {
-  field: string
-  value: string | number | boolean
-  definedBy: RuleSource
-  inputs: Array<{ field: string; source: RuleSource }>
-  shadowed: RuleSource[]
-}
-```
-
-`inputs` records the source of the winning definition of each referenced field, so a
-value that depends on an environment override names the environment layer. `shadowed`
-comes from the merge.
-
-`paths` and `variables` produce one entry per name. `logSource` and `classLocation` are
-structured: each resolved property produces its own entry
-(`logSources.kubernetes.namespace`), returned as a list so a mixed-layer object is still
-explainable.
-
-### 7.7 Public API
-
-```ts
-interface ResolveInput {
-  environment: Environment
-  service: Service
-  deployment?: Deployment
-  ruleProfiles: Map<string, RuleProfile>
-}
-
-function resolveContext(input: ResolveInput): ResolveResult
-```
-
-`ResolveResult` is `{ ok: true; context: ServiceContext } | { ok: false; diagnostics: Diagnostic[] }`.
-The function is pure; it never reads files, time, or config.
-
-## 8. Router
-
-### 8.1 Eligibility
-
-A binding is eligible when:
-
-- `enabled === true`;
-- `operation` equals the requested id exactly;
-- every declared `match` field matches the context:
-
-| Field | Matches |
-|---|---|
-| `environmentId` | context environment id |
-| `serviceId` | context service id |
-| `deploymentId` | context deployment id (requires a deployment) |
-| `deploymentMode` | context runtime mode |
-| `serviceCategory` | service category |
-
-An empty `match` matches everything.
-
-A binding whose `match` is not one of the five documented shapes is `RECORD_INVALID` at
-load (data-model §7); the router never sees such a binding.
-
-### 8.2 Level and selection
-
-Rank: `deployment` 1, `environment-service` 2, `service` 3, `category-mode` 4,
-`global` 5. Level is derived from which match fields are present, with
-`deploymentId` winning over `environmentId`+`serviceId`, which wins over `serviceId`,
-then `serviceCategory`+`deploymentMode`, then empty.
-
-```text
-eligible = filter(bindings, eligible)
-if eligible is empty: BINDING_NOT_FOUND
-top = min(rank)
-winners = eligible where rank == top
-if winners.length > 1: BINDING_CONFLICT (list every candidate)
-else selected = winners[0]
-```
-
-Lower levels are ignored once a winner exists, even if they also match.
-
-### 8.3 Skill availability
-
-```ts
-interface SkillDirectory {
-  snapshot(opts): Promise<{ skills: SkillSummary[]; complete: boolean }>
-}
-```
-
-Adapter over `ctx.skills` (F20):
-
-| Condition | Result |
-|---|---|
-| `ctx.skills` absent | `available: 'unknown'` + `SKILL_REGISTRY_UNAVAILABLE` warning, `ok: true` |
-| `complete === false` | same |
-| name not found | `ok: false`, `SKILL_UNAVAILABLE` |
-| `expectedSource` set and `source` differs | `ok: false`, `SKILL_UNAVAILABLE` |
-| `expectedProvider` set and `provider` differs | `ok: false`, `SKILL_UNAVAILABLE` |
-| otherwise | `available: true` |
-
-Digest pinning is not supported: DSH's catalog exposes no digest. The record field is
-removed in this design revision; do not reintroduce it without a provider that can
-verify it.
-
-### 8.4 Policy
-
-```text
-write = operation in WRITE_OPERATIONS
-requiresApproval = policy.requiresApproval ?? write
-if context.environmentCategory === 'production' and policy.allowProduction !== true:
-    POLICY_DENIED
-else: return policy
-```
-
-`requiresApproval` and `allowProduction` are returned to the caller; V1 does not
-execute the Skill. A future executor must enforce them, not just display them.
-
-### 8.5 Parameters
-
-- Strings are templates resolved through the same engine, with the context namespaces
-  and `paths.*`.
-- Numbers and booleans are stringified.
-- A null, array, or object value is `PARAMETER_RESOLUTION_FAILED`.
-- At most 50 parameters; keys must match `[A-Za-z_][A-Za-z0-9_]*`.
-
-### 8.6 Public API
-
-```ts
-function resolveOperation(input: {
-  operation: string
-  context: ServiceContext
-  bindings: OperationBinding[]
-  skills: SkillDirectory | null
-}): OperationResolution
-```
-
-Pure except for `skills.snapshot()`, which is an injected async capability. Tests use
-a fake directory.
-
-## 9. Session store
-
-```ts
-interface SessionContext {
-  sessionId: string
-  defaultEnvironmentId: string | null
-  persisted: boolean
-}
-```
-
-- `get(sessionId)`: read `sessions/<id>.json`; missing → `null` default; corrupt →
-  `persisted: false`.
-- `set(sessionId, environmentId)`: validate the environment exists, then write the
-  file; a write failure returns `persisted: false` and `SESSION_CONTEXT_UNPERSISTED`.
-- `set(..., null)` writes `defaultEnvironmentId: null` (an explicit clear).
-- Two sessions never share a file.
-- A process-local `Map<sessionId, string | null>` caches the last value. `get` prefers
-  the file and falls back to the cache when the file is missing or corrupt; a failed
-  write still updates the cache so the current process honors the choice. The cache is
-  not durable and is empty after a restart.
-
-## 10. Application service
-
-`src/app/app.ts` is the single place where storage, registry, resolver, router and
-session are composed. Tools and HTTP both call it; neither contains business logic.
-
-```ts
-interface App {
-  meta(): Promise<MetaResult>
-  list(kind: RecordKind): Promise<ListResult>
-  get(kind: RecordKind, id: string): Promise<GetResult>
-  find(query: string): Promise<FindResult>
-  save(kind: RecordKind, record: unknown, options?: SaveOptions): Promise<SaveResult>
-  saveYaml(kind: RecordKind, yaml: string, options?: SaveOptions): Promise<SaveResult>
-  remove(kind: RecordKind, id: string, actor: Actor, options?: RemoveOptions): Promise<RemoveResult>
-  validate(): Promise<ValidationReport>
-  resolve(input: ResolveRequest): Promise<ResolveResult>
-  route(input: RouteRequest): Promise<OperationResolution>
-  setCredential(ref: string, value: string): Promise<{ ref: string; configured: true }>
-  unsetCredential(ref: string): Promise<{ ref: string; removed: true }>
-  getContext(sessionId: string): Promise<SessionContext>
-  setContext(sessionId: string, environmentId: string | null): Promise<SessionContext>
-}
-
-interface SaveOptions { dryRun?: boolean }
-interface RemoveOptions { dryRun?: boolean }
-```
-
-Rules:
-
-- `save` validates with Zod, sets `createdAt` (create only) and `updatedAt`, checks the
-  record cap, writes, then reloads the registry. Timestamps supplied by the caller are
-  ignored; the stored record always carries both.
-- A schema failure throws `RECORD_INVALID` with
-  `details.issues: [{ path, message }]`, where `path` is a record-relative field path
-  (`hosts[1].ip`). The UI renders these as inline diagnostics.
-- `save` returns `created: boolean` by comparing existence before the write.
-- `save` re-runs integrity checks after the write and returns warnings for broken
-  references; it does not roll back, because the record is individually valid.
-- `remove` refuses when referenced (`RECORD_IN_USE`) and supports a dry-run flag used
-  by the UI's confirmation dialog.
-- `resolve` and `route` take explicit ids; the **tool layer** supplies the session
-  default when the model omits an id.
-- `Actor` is `'agent' | 'ui'`; the UI actor cannot write production unless
-  `writePolicy.allowUiProductionWrites` is true.
-- `resolve` fills `context.credentials` through an injected
-  `CredentialDirectory { describe(ref); set(ref, value); unset(ref) }`. The Host
-  adapter calls `ctx.credentials.describe`; tests use a fake. A missing ref yields
-  `{ ref, configured: false }` plus `CREDENTIAL_MISSING`; no value is ever requested.
-- `setCredential` accepts only `env-manager/<environmentId>/<accountId>` refs and a
-  non-empty value, forwards them to `ctx.credentials.set`, and returns only
-  `{ ref, configured: true }`. It never writes the value to a record or log.
-  `unsetCredential` uses the same namespace check and calls `ctx.credentials.unset`.
-  A missing service yields `CREDENTIAL_UNAVAILABLE`; provider failures yield
-  `CREDENTIAL_WRITE_FAILED` without exposing the provider message to the UI.
-
-## 11. Tools
-
-### 11.1 Shared shape
-
-```ts
-function toolOutput<T extends JsonSchemaNode>(schema: T, render: Renderer): ToolOutputDefinition
-```
-
-- Every tool returns a plain object matching its declared schema; DSH validates it.
-- Every failure throws `EnvManagerError`; no tool returns `{ ok: false }`.
-- `render` emits one text block: a one-line summary plus the essential ids. The
-  canonical JSON value is what the model receives.
-
-`EnvManagerError`:
-
-```ts
-class EnvManagerError extends HarnessError {
-  constructor(code: ErrorCode, message: string, details?: JsonValue) {
-    super(message, code)
-    this.details = details
-  }
-}
-```
-
-### 11.2 Tool table
-
-| Tool | Parameters | Returns | Session | Approval |
-|---|---|---|---|---|
-| `env_list` | `kind?`, `query?` | `{ items, warnings }` | no | no |
-| `env_get` | `kind`, `id` | `{ record, yaml }` | no | no |
-| `env_find` | `query` | `{ hits }` | no | no |
-| `env_resolve` | `environmentId?`, `serviceId`, `deploymentId?` | `{ context, warnings }` | default env | no |
-| `env_resolve_operation` | `operation`, `environmentId?`, `serviceId?`, `deploymentId?` | `{ resolution }` | yes | no |
-| `env_get_context` | — | `{ sessionId, defaultEnvironmentId, persisted }` | yes | no |
-| `env_set_context` | `environmentId \| null` | same | yes | no |
-| `env_save` | `kind`, `record`, `dryRun?` | `{ id, kind, created, dryRun, warnings }` | no | production env |
-| `env_delete` | `kind`, `id`, `confirm` | `{ id, kind, deleted, blockers }` | no | yes |
-| `env_validate` | — | `{ ok, counts, diagnostics }` | no | no |
-
-### 11.3 Parameter definitions
-
-```ts
-const kindSchema = {
-  type: 'string',
-  enum: ['environment', 'service', 'deployment', 'ruleProfile', 'operationBinding'],
-  description: 'Record kind.',
-} as const
-```
-
-`env_list`:
-
-```ts
-parameters: {
-  kind: { ...kindSchema, required: false },
-  query: { type: 'string', required: false, description: 'Case-insensitive match over id, name and alias.' },
-}
-```
-
-`env_get`: `kind` and `id` required, `id` description names the id pattern.
-
-`env_find`: `query` required, min length 2.
-
-`env_resolve`: `serviceId` required; `environmentId` and `deploymentId` optional.
-Resolving an environment/service pair with more than one deployment and no
-`deploymentId` returns `DEPLOYMENT_AMBIGUOUS` listing candidates; zero deployments
-returns `DEPLOYMENT_NOT_FOUND`.
-
-`env_resolve_operation`: `operation` required, `enum` of the 13 standard ids plus a
-free-form custom id — model-visible enum lists only the standard ids; a custom id is
-rejected by the schema unless it is declared in a binding. Because the DSL cannot
-express "either enum or binding-declared", the parameter is a plain string with the
-standard ids in the description and is validated in `execute`
-(`OPERATION_UNKNOWN` when neither standard nor bound).
-
-`env_set_context`: `environmentId` is `oneOf` string / `null`.
-
-`env_save`: `record` is `{ type: 'json', required: true, description: 'Complete record object.' }`.
-Zod still validates it; the JSON parameter only avoids duplicating every record schema
-in the model-facing tool. `dryRun: true` validates and returns the same shape without
-writing; a schema failure still throws `RECORD_INVALID`.
-
-`env_delete`: `confirm` is a required boolean. `false` performs a dry run: it returns
-`deleted: false` plus `blockers` and writes nothing.
-
-### 11.4 Output schemas
-
-Small results are precise; open records use `additionalProperties: true`.
-
-```ts
-const diagnosticSchema = {
-  type: 'object', additionalProperties: false,
-  properties: {
-    severity: { type: 'string', enum: ['info', 'warning', 'error'], required: true },
-    code: { type: 'string', required: true },
-    message: { type: 'string', required: true },
-    file: { type: 'string' },
-    path: { type: 'string' },
-  },
-}
-const recordSchema = { type: 'object', additionalProperties: true }
-```
-
-`env_list` output:
-
-```ts
-{ type: 'object', additionalProperties: false, properties: {
-  items: { type: 'array', required: true, items: { type: 'object', additionalProperties: false, properties: {
-    kind: { type: 'string', required: true },
-    id: { type: 'string', required: true },
-    name: { type: 'string', required: true },
-    summary: { type: 'string' },
-  } } },
-  warnings: { type: 'array', required: true, items: diagnosticSchema },
-} }
-```
-
-`env_get` output: `{ record: recordSchema, yaml: { type: 'string', required: true } }`.
-
-`env_resolve` output: `{ context: recordSchema, warnings: arrayOf(diagnosticSchema) }`.
-
-`env_validate` output: counts per kind plus `diagnostics`.
-
-### 11.5 Execution flow
-
-```text
-execute(args, exec):
-  1. assert exec.signal not aborted
-  2. resolve target ids (session default when allowed)
-  3. call app.<operation>
-  4. on EnvManagerError: rethrow
-  5. on unexpected error: wrap as INTERNAL_ERROR with a redacted message
-  6. return the declared value
-```
-
-Approval helper:
-
-```ts
-async function requireApproval(exec, toolName: string, reason: string): Promise<void> {
-  if (!exec.agent) throw new EnvManagerError('SESSION_CONTEXT_UNAVAILABLE', '...')
-  const outcome = await ctx.approval.request({
-    agent: exec.agent,
-    toolName,
-    callId: exec.callId,
-    reason,
-    displayReason: { en: reason, zh: reasonZh },
-    signal: exec.signal,
-  })
-  if (outcome !== 'allowed-once') throw new EnvManagerError('POLICY_DENIED', '...')
-}
-```
-
-The `displayReason` is localized; the `reason` is the stable English fallback.
-
-## 12. HTTP transport
-
-### 12.1 Route
-
-```ts
-ctx.webServer.register({ kind: 'prefix', path: '/env-manager/api', handler })
-```
-
-- `POST` only; other methods → 405 with `Allow: POST`.
-- `Content-Type: application/json; charset=utf-8` required.
-- Body limit 2 MiB; exceeding it → 413.
-- Request header `x-env-manager-request: 1` required, otherwise 403.
-- If an `Origin` header is present it must equal `ctx.webServer.getOrigin()`,
-  otherwise 403. No CORS headers are ever sent.
-- The handler never reads cookies or credentials. A credential value may appear only in
-  the `credential` request body, is forwarded directly to the credential service and is
-  never logged or echoed.
-
-### 12.2 Envelope
-
-Request:
-
-```json
-{ "op": "record.save", "payload": { "kind": "environment", "yaml": "id: test-a\n..." } }
-```
-
-Success `200`:
-
-```json
-{ "ok": true, "data": { ... } }
-```
-
-Failure:
-
-```json
-{ "ok": false, "error": { "code": "RECORD_INVALID", "message": "...", "details": { ... } } }
-```
-
-Status map: `INVALID_ARGUMENT` 400, `POLICY_DENIED`/`FORBIDDEN` 403, `*_NOT_FOUND` 404,
-`RECORD_IN_USE`/`BINDING_CONFLICT`/`DEPLOYMENT_AMBIGUOUS` 409, `RECORD_INVALID`/
-`RULE_*`/`PARAMETER_*` 422, `CREDENTIAL_UNAVAILABLE` 503, anything else 500 (including
-`CREDENTIAL_WRITE_FAILED`). The body always carries the code; the status is for tooling.
-
-### 12.3 Operations
-
-| op | payload | data |
-|---|---|---|
-| `meta` | — | dataDir, version, counts, warnings |
-| `list` | `{ kind }` | summaries |
-| `get` | `{ kind, id }` | `{ record, yaml }` |
-| `find` | `{ query }` | hits |
-| `save` | `{ kind, record?, yaml?, dryRun? }` | save result |
-| `remove` | `{ kind, id, confirm?, dryRun? }` | remove result |
-| `validate` | — | validation report |
-| `resolve` | `{ environmentId, serviceId, deploymentId? }` | context |
-| `route` | `{ operation, environmentId, serviceId, deploymentId? }` | resolution |
-| `credential` | `{ ref, value }` | `{ ref, configured }` |
-| `credentialUnset` | `{ ref }` | `{ ref, removed }` |
-
-The UI cannot read or write session context; that is agent-only in V1 (§9).
-
-### 12.4 Actor policy
-
-`record.save` and `record.remove` run with actor `'ui'`:
-
-- production environment save is denied unless `allowUiProductionWrites`;
-- `record.remove` requires `confirm: true` unless `dryRun: true`;
-- every write is logged with actor, op, kind, id and outcome (no record content).
-
-`credential` and `credentialUnset` are UI-only, restricted to the `env-manager/`
-namespace and never write to the registry. They log only the ref and outcome.
-
-## 13. Client module
-
-Structural contract:
-
-```js
-return {
-  inject: ['slots', 'locale'],
-  apply(ctx) {
-    ctx.effect(() => ctx.locale.register('envManager', { en, zh }))
-    ctx.slots.inject('sidebar.panellist', () => ctx.slots.register(
-      { name: 'sidebar.panellist', id: 'env-manager', order: 40 }, Icon))
-    ctx.slots.inject('main', () => ctx.slots.register(
-      { name: 'main', key: 'env-manager' }, Panel))
-  },
-}
-```
-
-- All data access goes through `createApi()` in `client.js`, which posts to
-  `/env-manager/api` with the required header and maps failures to a UI error object.
-- The panel keeps state locally (`useState`/`useEffect`); no global store.
-- Rendering and interactions are specified in [`ui-design.md`](ui-design.md).
-
-## 14. Memory bridge
-
-```ts
-interface MemoryBridge {
-  readonly available: boolean
-  emit(record: OperationRecord): Promise<void>
-  observe(scope: MemoryScope): Promise<readonly Observation[]>
-  propose(candidate: CandidateChange): Promise<string>   // returns proposal id
-}
-```
-
-V1 ships `NullMemoryBridge`: `available: false`, `emit` resolves without writing,
-`observe` returns `[]`, `propose` throws `MEMORY_BRIDGE_UNAVAILABLE`. It is selected
-when no memory provider is registered; a future adapter plugs in without changing the
-tools because nothing in V1 calls `emit` (no execution).
-
-## 15. Cross-cutting
-
-### 15.1 Logging
-
-`src/logger.ts` wraps `ctx.get('logger')` when it is callable, else `console`.
-Log levels: `info` for lifecycle (dataDir, loaded counts), `warn` for skipped records,
-`error` for failed loads. Records and credential refs are logged as ids only; never a
-whole record, never a value.
-
-### 15.2 Redaction
-
-`redactValue(text)` removes anything that looks like a URL userinfo, a bearer token,
-or a private key block from messages before they reach the model or the UI. Tool
-results never include `credentialRef` values, only configured state.
-
-### 15.3 Cancellation
-
-`execute` receives `exec.signal`; long loops (registry load, resolve) check it between
-files and between fields. A cancelled call surfaces DSH's own abort error unchanged.
-
-### 15.4 Performance budgets
-
-| Operation | Budget |
-|---|---|
-| registry load (500 records) | 100 ms |
-| resolve (50 fields) | 5 ms |
-| route with skill snapshot | 50 ms |
-| HTTP round trip (local) | 150 ms |
-
-## 16. Test obligations
-
-| Module | Tests | From AGENTS §13 |
-|---|---|---|
-| `storage` | atomic write, missing file, invalid YAML, schema mismatch, backups, traversal rejection | required |
-| `registry` | load/lookup, duplicate id, broken references, alias collision | required |
-| `resolver` | precedence, field merge, provenance, undefined/cycle/traversal, runtime selection | required |
-| `router` | each level, conflict, missing skill, source/provider mismatch, policy, parameters | required |
-| `session` | set/get/clear, isolation, corrupt fallback | required |
-| `tools` | output matches schema, error codes, no credential values, approval denial | required |
-| `app` | save→get round trip, delete in use, validate report | new |
-| `http` | auth header, origin, method, size cap, status map, production denial | new |
-| `client` | manual checklist in `ui-design.md` §9 | manual |
-
-Tests live in `tests/<module>/*.test.ts`, fixtures in `tests/fixtures/<scenario>/`.
-
-## 17. Traceability
-
-| Design section | Implements | Acceptance |
-|---|---|---|
-| §4 storage | ADR-0002 | AC-01, AC-02 |
-| §7 resolver | ADR-0005 | AC-03, AC-04 |
-| §8 router | ADR-0003 | AC-05…AC-07 |
-| §9 session | ADR-0007 | AC-08 |
-| §12 HTTP | ADR-0009 | UI manual QA (`ui-design.md` §12) |
-| §11 tools | tool-contract.md | AC-02…AC-08 |
-
-## 18. Deliberate omissions
-
-- No cache beyond one in-flight load.
-- No cross-process locking.
-- No Skill execution.
-- No session-event projection.
-- No digest pinning.
-- No plugin-side bundler.
-- No database.
+# Detailed Design (V1)
+
+> Status: normative implementation design for V1. It refines
+> [`design.md`](design.md) into module-level contracts and depends on the verified
+> runtime facts in [`api-verification.md`](api-verification.md).
+> Record shapes are defined in [`data-model.md`](data-model.md); do not restate or
+> extend them here without updating that document.
+
+## 1. Scope
+
+This document specifies:
+
+1. the plugin configuration and its effective defaults;
+2. every Host module's public interface and internal algorithm;
+3. the exact `defineTool` contract for all ten tools;
+4. the plugin-owned HTTP API used by the Client panel;
+5. error mapping, limits, concurrency and test obligations.
+
+UI screens and copy are in [`ui-design.md`](ui-design.md); build and packaging are in
+[`design.md`](design.md) §15 and [`implementation-plan.md`](implementation-plan.md).
+
+## 2. Conventions
+
+- Host: TypeScript, ESM, `strict`.
+- Runtime validation: Zod. Tool parameter validation: DSH's schema DSL (F3).
+- Files: YAML 1.2 via the `yaml` package; no anchors/aliases are emitted.
+- Time: ISO-8601 UTC with milliseconds, produced by `new Date().toISOString()`.
+- Paths: POSIX separators inside resolved values; OS-native paths only at the
+  filesystem boundary.
+- All expected failures are `EnvManagerError extends HarnessError` (F5) with a code
+  from `data-model.md` §11.
+
+## 3. Configuration
+
+```ts
+// src/config.ts
+import z from '@deepseek-ai/schemastery'
+
+export const Config = z.object({
+  dataDir: z.string().description('Registry directory; relative paths resolve against the profile directory.'),
+  maxRecords: z.natural().default(2000),
+  registerSkill: z.boolean().default(true),
+  writePolicy: z.object({
+    allowUiProductionWrites: z.boolean().default(false),
+    requireApprovalForDelete: z.boolean().default(true),
+    requireApprovalForProductionSave: z.boolean().default(true),
+  }).default({}),
+  defaults: z.object({
+    environmentId: z.string().default(''),
+  }).default({}),
+})
+```
+
+| Field | Default | Meaning |
+|---|---|---|
+| `dataDir` | profile data dir | Registry root; see §4.1 |
+| `maxRecords` | 2000 | hard cap per record kind; exceeding it fails a write |
+| `registerSkill` | true | register the profile-wide `dsh-env-manager` discovery Skill |
+| `writePolicy.allowUiProductionWrites` | false | allow the HTTP API to write production records |
+| `writePolicy.requireApprovalForDelete` | true | `env_delete` asks the human first |
+| `writePolicy.requireApprovalForProductionSave` | true | `env_save` on a production environment asks first |
+| `defaults.environmentId` | empty | fallback when neither an argument nor the session default exists |
+
+Rules:
+
+- The config schema is schemastery (F1/F18), not Zod. Do not add a second config path.
+- Config is read-only at runtime; there is no write API. The Client shows effective
+  values only.
+- A blank `dataDir` means "use the resolution order in §4.1", not "current directory".
+
+## 4. Storage
+
+### 4.1 Data directory resolution
+
+```text
+1. config.dataDir, if set:
+     absolute  -> use as-is
+     relative  -> path.resolve(profileDir, config.dataDir)
+2. profileDir from ctx.get('profileContext')?.dir
+     -> <profileDir>/data/dsh-env-manager
+3. process.env.DSH_PROFILE_DIR
+     -> <DSH_PROFILE_DIR>/data/dsh-env-manager
+4. process.cwd()/data/dsh-env-manager
+```
+
+`profileDir` may be unavailable in an embedded host; step 3 and 4 keep the plugin
+usable. The resolved absolute path is memoized for the process lifetime, logged once,
+and reported by `env_get_context` and the UI overview. It is never derived from model
+input.
+
+### 4.2 Layout
+
+```text
+<dataDir>/
+├── environments/<environmentId>.yaml
+├── services/<serviceId>.yaml
+├── deployments/<deploymentId>.yaml
+├── rules/<ruleProfileId>.yaml
+├── operations/<bindingId>.yaml
+├── sessions/<sessionId>.json
+└── .backups/<kind>/<id>/<timestamp>.yaml      (rolling, max 5)
+```
+
+- The file name is always `<id>.yaml`; the record `id` must equal the file stem.
+- `sessions/` and `.backups/` are never scanned as registry records.
+- Directories are created lazily by the first write; a missing root on read yields an
+  empty registry, not an error.
+
+### 4.3 Id and path safety
+
+```ts
+const ID_PATTERN = /^[a-z0-9][a-z0-9-]{0,62}$/
+```
+
+`resolveRecordPath(kind, id)` rejects a non-matching id **before** touching the
+filesystem, then verifies `path.resolve(dir, id + '.yaml')` starts with
+`path.resolve(dir) + path.sep`. Session ids are validated as `[A-Za-z0-9_-]{1,128}`
+and hashed only if they fail that pattern.
+
+### 4.4 Atomic write protocol
+
+```ts
+async function writeRecord(kind, id, record) {
+  const target = resolveRecordPath(kind, id)
+  await mkdir(dirname(target), { recursive: true })
+  const tmp = join(dirname(target), `.${id}.${process.pid}.${randomUUID()}.tmp`)
+  const text = stringifyYaml(record)
+  const handle = await open(tmp, 'w')
+  try {
+    await handle.writeFile(text, 'utf8')
+    await handle.sync()
+  } finally {
+    await handle.close()
+  }
+  await rotateBackup(kind, id, target)   // copy previous file, prune to 5
+  await rename(tmp, target)              // atomic replace on POSIX and Windows
+  return { file: target, bytes: Buffer.byteLength(text) }
+}
+```
+
+Guarantees:
+
+- A reader never sees a partially written record.
+- A crash before `rename` leaves an ignorable `.tmp` file; startup cleanup removes
+  `*.tmp` older than one hour.
+- Backup failure does not block the write; it is reported as a warning.
+- `writeRecord` is serialized per target path by an in-process promise chain (§4.8).
+
+### 4.5 Read protocol
+
+```ts
+async function readRecord(kind, id): Promise<ReadResult> {
+  const target = resolveRecordPath(kind, id)
+  let text
+  try { text = await readFile(target, 'utf8') }
+  catch (e) { if (e.code === 'ENOENT') return { found: false }; throw e }
+  const parsed = parseYamlDocument(text)          // throws RECORD_INVALID
+  return { found: true, raw: parsed, text }
+}
+```
+
+Invalid YAML, a non-object document, an id/filename mismatch, or a schema failure
+produces a `Diagnostic` and the record is treated as absent for that request. The file
+is never rewritten or deleted.
+
+Unknown keys are preserved on read and reported as an `UNKNOWN_FIELD` warning, but
+`env_save` rejects them (`RECORD_INVALID`) so a typo cannot enter the store. Storage
+therefore parses twice: `.passthrough()` on read, strict on save.
+
+### 4.6 Delete
+
+```text
+1. Resolve the target and confirm it exists.
+2. Ask the registry whether any record references it (§6.3).
+3. If referenced -> RECORD_IN_USE listing referrers.
+4. Otherwise rotate a backup, then unlink the file.
+```
+
+Deleting a session file is a plain unlink and needs no reference check.
+
+### 4.7 Session files
+
+```json
+{
+  "schemaVersion": 1,
+  "sessionId": "01J...",
+  "defaultEnvironmentId": "test-a",
+  "updatedAt": "2026-11-01T03:00:00.000Z"
+}
+```
+
+Written with the same temp+rename protocol. A corrupt or unreadable file is reported
+as `persisted: false` and treated as "no default"; it is not deleted.
+
+### 4.8 Concurrency
+
+- One serialized chain per file path (`Map<string, Promise<void>>`), released when the
+  chain drains.
+- Registry loads are deduplicated: concurrent callers share one in-flight load.
+- No cross-process lock in V1. Two DSH processes writing the same file can lose one
+  update; the backup makes the loss recoverable. Documented as a limitation.
+
+### 4.9 Limits
+
+| Limit | Value | Failure |
+|---|---|---|
+| records per kind | `maxRecords` (2000) | `RECORD_LIMIT_EXCEEDED` |
+| record YAML bytes | 1 MiB | `RECORD_TOO_LARGE` |
+| session file bytes | 64 KiB | `SESSION_CONTEXT_UNPERSISTED` |
+
+## 5. Domain
+
+```ts
+// src/domain/types.ts
+export type RecordKind = 'environment' | 'service' | 'deployment' | 'ruleProfile' | 'operationBinding'
+export type AnyRecord = Environment | Service | Deployment | RuleProfile | OperationBinding
+```
+
+`src/domain/schemas.ts` exports one Zod schema per record plus:
+
+```ts
+export function parseRecord(kind: RecordKind, value: unknown, origin: string):
+  { ok: true; record: AnyRecord } | { ok: false; diagnostic: Diagnostic }
+```
+
+Validation responsibilities:
+
+- `RecordMeta`: id pattern, `schemaVersion === 1`, ISO timestamps, `updatedAt >= createdAt`.
+- Cross-field rules from `data-model.md`: at least one `deployModes`, host role ids
+  exist, `location.kind === runtimeMode`, unique host id / ip / cluster id /
+  endpoint id, unique account id per host, `loginRecords[].hostRef`/`accountId`
+  resolve within the environment, host `clusterRef` resolves and is present for
+  `k8s-all-in-one`/`k8s-worker`, `access` values are unique, `paths` keys and variable
+  names match `[A-Za-z_][A-Za-z0-9_]*`.
+- `structural` shape only. Referential integrity is the registry's job (§6.3), because
+  it needs the whole snapshot.
+
+`parseRecord` never throws for bad data; it returns a diagnostic with `path` pointing
+at the failing field.
+
+Tool input is parsed by `parseRecordInput`, which is identical except `createdAt` and
+`updatedAt` are optional. The app then calls `finalizeRecord(record, previous?)` to
+stamp them before `parseRecord` and before writing.
+
+## 6. Registry
+
+### 6.1 Snapshot
+
+```ts
+interface RegistrySnapshot {
+  environments: Map<string, Environment>
+  services: Map<string, Service>
+  deployments: Map<string, Deployment>
+  ruleProfiles: Map<string, RuleProfile>
+  operationBindings: Map<string, OperationBinding>
+  byAlias: Map<string, string>          // "service:<alias>" -> id
+  diagnostics: Diagnostic[]
+  loadedAt: string
+}
+```
+
+`byAlias` keys are `${kind}:${alias}` and resolve aliases for `env_get`/`env_find`.
+Duplicate aliases are a warning; the first record wins, sorted by file name for
+determinism.
+
+### 6.2 Load algorithm
+
+```text
+for each kind directory (environments, services, deployments, rules, operations):
+  entries = readdir(dir, { withFileTypes: true })
+  files   = entries where isFile() and name endsWith('.yaml') and not name startsWith('.')
+  sort files by name
+  for each file:
+    read + parse
+    parseRecord -> on failure: diagnostic (file, code, path, message); skip
+    if map already has record.id: diagnostic DUPLICATE_ID; skip later file
+    else insert
+refresh byAlias and run integrity checks
+```
+
+A missing kind directory is normal (empty). An unreadable directory is an error
+diagnostic, and the snapshot keeps the other kinds.
+
+### 6.3 Referential integrity
+
+| Reference | Rule | Code |
+|---|---|---|
+| `deployment.environmentId` | environment exists | `ENV_NOT_FOUND` |
+| `deployment.serviceId` | service exists | `SERVICE_NOT_FOUND` |
+| `deployment.location.hostRef` | host exists in that environment | `HOST_NOT_FOUND` |
+| `deployment.location.clusterRef` | cluster exists in that environment | `CLUSTER_NOT_FOUND` |
+| environment `hosts[].clusterRef` | cluster exists; required for K8s node kinds | `RECORD_INVALID` |
+| environment `endpoints[].hostRef` / `clusterRef` | target exists in that environment | `RECORD_INVALID` |
+| `service.ruleProfileRefs[]` | rule profile exists | `RULE_PROFILE_NOT_FOUND` |
+| operation binding `match.environmentId` | environment exists | `ENV_NOT_FOUND` |
+| operation binding `match.serviceId` | service exists | `SERVICE_NOT_FOUND` |
+| operation binding `match.deploymentId` | deployment exists | `DEPLOYMENT_NOT_FOUND` |
+| any `credentialRef` | non-empty string matching the agreed ref pattern | `RECORD_INVALID` |
+
+Schema validation rejects an environment with zero logical roles and a host with zero
+roles before integrity checks run; both fail as `RECORD_INVALID`. A host role id that is
+not declared in `Environment.roles` is reported as `ROLE_NOT_FOUND` by `env_validate`;
+the panel prevents selecting an undeclared role in the node form.
+
+Integrity failures are **diagnostics**, not load failures: the registry still serves
+the other records, and `env_validate` reports the broken references. A resolve or
+route that needs a missing record fails with the specific code.
+
+### 6.4 Reload policy
+
+V1 reloads the registry at the start of every application request. Rationale: users
+edit YAML by hand, and a stale cache is worse than a few milliseconds. Concurrent
+requests share one in-flight load. A snapshot is replaced only after a successful
+load; a load that throws keeps the previous snapshot and reports an error diagnostic.
+
+Performance target: 500 records load in under 100 ms on a local SSD. If that is
+missed, add an mtime-keyed cache in a follow-up, not in V1.
+
+## 7. Resolver
+
+### 7.1 Template parsing
+
+```ts
+type TemplatePart =
+  | { kind: 'literal'; text: string }
+  | { kind: 'ref'; path: string }
+
+function parseTemplate(text: string): TemplatePart[]
+```
+
+Scanner rules:
+
+1. Find `${`. Text before it is a literal.
+2. Find the next `}`. A missing `}` is `RULE_TEMPLATE_INVALID`.
+3. The inner text must match `^[A-Za-z_][A-Za-z0-9_]*(\.[A-Za-z_][A-Za-z0-9_]*)*$`;
+   otherwise `RULE_TEMPLATE_INVALID`.
+4. Anything that is not `${…}` is literal text, including a lone `$`.
+5. No nesting, escaping, calls, operators, or env access.
+6. Template length is capped at 2048 characters.
+
+### 7.2 Namespaces
+
+| Reference | Resolves to |
+|---|---|
+| `env.id`, `env.name`, `env.category` | environment record |
+| `service.id`, `service.name`, `service.category` | service record |
+| `service.artifact.name` | service artifact name, when declared |
+| `deployment.id`, `deployment.name`, `deployment.runtimeMode` | deployment record |
+| `deployment.location.<field>` | deployment location, runtime-specific |
+| `<variable>` | layer variable of that name |
+| `variables.<variable>` | same, explicit form |
+| `paths.<pathName>` | another resolved path |
+
+`service.artifact.name` resolves only when `artifact` exists. A reference to a
+namespace that is unavailable for this context is `RULE_UNDEFINED_VARIABLE`, not an
+empty string.
+
+### 7.3 Layers
+
+```ts
+interface Layer {
+  source: RuleSource
+  variables?: Record<string, string | number | boolean>
+  paths?: Record<string, string>
+  logSources?: Partial<Record<RuntimeMode, LogSource>>
+  classLocations?: Partial<Record<RuntimeMode, ClassLocation>>
+}
+```
+
+Built from `design.md` §6.3:
+
+1. built-in global defaults (empty in V1);
+2. each `service.ruleProfileRefs` entry, in order;
+3. `environment.ruleOverrides`;
+4. `deployment.variables` (variables only), then `deployment.ruleOverrides`.
+
+Flatten to a field map:
+
+```text
+variables.<name>
+paths.<name>
+logSources.<runtime>.<property>
+classLocations.<runtime>.<property>
+```
+
+### 7.4 Merge
+
+```ts
+interface MergedField {
+  key: string
+  raw: string | number | boolean
+  definedBy: RuleSource
+  shadowed: RuleSource[]
+}
+```
+
+Iterate layers in ascending order. For each field, replace the winner and append the
+previous winner to `shadowed`. This is the only merge rule; there is no deep merge and
+no array concatenation.
+
+### 7.5 Evaluation
+
+```text
+1. Build the reference graph: node = merged field key; edge = each ref used by a
+   string field. Constants (env/service/deployment) are not nodes.
+2. DFS with an active stack. A back edge is RULE_CYCLE with the full cycle path.
+3. A ref with no node and no constant is RULE_UNDEFINED_VARIABLE
+   (message names the field and the ref).
+4. Resolve nodes in post-order; substitute into an ordered list of parts.
+5. For every resolved `paths.*` value, split on `/` and reject `..`
+   (RULE_PATH_TRAVERSAL). Absolute/relative shape is otherwise preserved.
+6. Convert non-string numbers/booleans to their literal text when interpolated.
+7. Select `logSources[runtimeMode]` and `classLocations[runtimeMode]` after their
+   strings resolve.
+```
+
+Cycle detection is iterative (explicit stack), not recursive, so a pathological graph
+cannot overflow the JS stack.
+
+### 7.6 Provenance
+
+For every requested field:
+
+```ts
+interface ProvenanceEntry {
+  field: string
+  value: string | number | boolean
+  definedBy: RuleSource
+  inputs: Array<{ field: string; source: RuleSource }>
+  shadowed: RuleSource[]
+}
+```
+
+`inputs` records the source of the winning definition of each referenced field, so a
+value that depends on an environment override names the environment layer. `shadowed`
+comes from the merge.
+
+`paths` and `variables` produce one entry per name. `logSource` and `classLocation` are
+structured: each resolved property produces its own entry
+(`logSources.kubernetes.namespace`), returned as a list so a mixed-layer object is still
+explainable.
+
+### 7.7 Public API
+
+```ts
+interface ResolveInput {
+  environment: Environment
+  service: Service
+  deployment?: Deployment
+  ruleProfiles: Map<string, RuleProfile>
+}
+
+function resolveContext(input: ResolveInput): ResolveResult
+```
+
+`ResolveResult` is `{ ok: true; context: ServiceContext } | { ok: false; diagnostics: Diagnostic[] }`.
+The function is pure; it never reads files, time, or config.
+
+## 8. Router
+
+### 8.1 Eligibility
+
+A binding is eligible when:
+
+- `enabled === true`;
+- `operation` equals the requested id exactly;
+- every declared `match` field matches the context:
+
+| Field | Matches |
+|---|---|
+| `environmentId` | context environment id |
+| `serviceId` | context service id |
+| `deploymentId` | context deployment id (requires a deployment) |
+| `deploymentMode` | context runtime mode |
+| `serviceCategory` | service category |
+
+An empty `match` matches everything.
+
+A binding whose `match` is not one of the five documented shapes is `RECORD_INVALID` at
+load (data-model §7); the router never sees such a binding.
+
+### 8.2 Level and selection
+
+Rank: `deployment` 1, `environment-service` 2, `service` 3, `category-mode` 4,
+`global` 5. Level is derived from which match fields are present, with
+`deploymentId` winning over `environmentId`+`serviceId`, which wins over `serviceId`,
+then `serviceCategory`+`deploymentMode`, then empty.
+
+```text
+eligible = filter(bindings, eligible)
+if eligible is empty: BINDING_NOT_FOUND
+top = min(rank)
+winners = eligible where rank == top
+if winners.length > 1: BINDING_CONFLICT (list every candidate)
+else selected = winners[0]
+```
+
+Lower levels are ignored once a winner exists, even if they also match.
+
+### 8.3 Skill availability
+
+```ts
+interface SkillDirectory {
+  snapshot(opts): Promise<{ skills: SkillSummary[]; complete: boolean }>
+}
+```
+
+Adapter over `ctx.skills` (F20):
+
+| Condition | Result |
+|---|---|
+| `ctx.skills` absent | `available: 'unknown'` + `SKILL_REGISTRY_UNAVAILABLE` warning, `ok: true` |
+| `complete === false` | same |
+| name not found | `ok: false`, `SKILL_UNAVAILABLE` |
+| `expectedSource` set and `source` differs | `ok: false`, `SKILL_UNAVAILABLE` |
+| `expectedProvider` set and `provider` differs | `ok: false`, `SKILL_UNAVAILABLE` |
+| otherwise | `available: true` |
+
+Digest pinning is not supported: DSH's catalog exposes no digest. The record field is
+removed in this design revision; do not reintroduce it without a provider that can
+verify it.
+
+### 8.4 Policy
+
+```text
+write = operation in WRITE_OPERATIONS
+requiresApproval = policy.requiresApproval ?? write
+if context.environmentCategory === 'production' and policy.allowProduction !== true:
+    POLICY_DENIED
+else: return policy
+```
+
+`requiresApproval` and `allowProduction` are returned to the caller; V1 does not
+execute the Skill. A future executor must enforce them, not just display them.
+
+### 8.5 Parameters
+
+- Strings are templates resolved through the same engine, with the context namespaces
+  and `paths.*`.
+- Numbers and booleans are stringified.
+- A null, array, or object value is `PARAMETER_RESOLUTION_FAILED`.
+- At most 50 parameters; keys must match `[A-Za-z_][A-Za-z0-9_]*`.
+
+### 8.6 Public API
+
+```ts
+function resolveOperation(input: {
+  operation: string
+  context: ServiceContext
+  bindings: OperationBinding[]
+  skills: SkillDirectory | null
+}): OperationResolution
+```
+
+Pure except for `skills.snapshot()`, which is an injected async capability. Tests use
+a fake directory.
+
+## 9. Session store
+
+```ts
+interface SessionContext {
+  sessionId: string
+  defaultEnvironmentId: string | null
+  persisted: boolean
+}
+```
+
+- `get(sessionId)`: read `sessions/<id>.json`; missing → `null` default; corrupt →
+  `persisted: false`.
+- `set(sessionId, environmentId)`: validate the environment exists, then write the
+  file; a write failure returns `persisted: false` and `SESSION_CONTEXT_UNPERSISTED`.
+- `set(..., null)` writes `defaultEnvironmentId: null` (an explicit clear).
+- Two sessions never share a file.
+- A process-local `Map<sessionId, string | null>` caches the last value. `get` prefers
+  the file and falls back to the cache when the file is missing or corrupt; a failed
+  write still updates the cache so the current process honors the choice. The cache is
+  not durable and is empty after a restart.
+
+## 10. Application service
+
+`src/app/app.ts` is the single place where storage, registry, resolver, router and
+session are composed. Tools and HTTP both call it; neither contains business logic.
+
+```ts
+interface App {
+  meta(): Promise<MetaResult>
+  list(kind: RecordKind): Promise<ListResult>
+  get(kind: RecordKind, id: string): Promise<GetResult>
+  find(query: string): Promise<FindResult>
+  save(kind: RecordKind, record: unknown, options?: SaveOptions): Promise<SaveResult>
+  saveYaml(kind: RecordKind, yaml: string, options?: SaveOptions): Promise<SaveResult>
+  remove(kind: RecordKind, id: string, actor: Actor, options?: RemoveOptions): Promise<RemoveResult>
+  validate(): Promise<ValidationReport>
+  resolve(input: ResolveRequest): Promise<ResolveResult>
+  route(input: RouteRequest): Promise<OperationResolution>
+  setCredential(ref: string, value: string): Promise<{ ref: string; configured: true }>
+  unsetCredential(ref: string): Promise<{ ref: string; removed: true }>
+  getContext(sessionId: string): Promise<SessionContext>
+  setContext(sessionId: string, environmentId: string | null): Promise<SessionContext>
+}
+
+interface SaveOptions { dryRun?: boolean }
+interface RemoveOptions { dryRun?: boolean }
+```
+
+Rules:
+
+- `save` validates with Zod, sets `createdAt` (create only) and `updatedAt`, checks the
+  record cap, writes, then reloads the registry. Timestamps supplied by the caller are
+  ignored; the stored record always carries both.
+- A schema failure throws `RECORD_INVALID` with
+  `details.issues: [{ path, message }]`, where `path` is a record-relative field path
+  (`hosts[1].ip`). The UI renders these as inline diagnostics.
+- `save` returns `created: boolean` by comparing existence before the write.
+- `save` re-runs integrity checks after the write and returns warnings for broken
+  references; it does not roll back, because the record is individually valid.
+- `remove` refuses when referenced (`RECORD_IN_USE`) and supports a dry-run flag used
+  by the UI's confirmation dialog.
+- `resolve` and `route` take explicit ids; the **tool layer** supplies the session
+  default when the model omits an id.
+- `Actor` is `'agent' | 'ui'`; the UI actor cannot write production unless
+  `writePolicy.allowUiProductionWrites` is true.
+- `resolve` fills `context.credentials` through an injected
+  `CredentialDirectory { describe(ref); set(ref, value); unset(ref) }`. The Host
+  adapter calls `ctx.credentials.describe`; tests use a fake. A missing ref yields
+  `{ ref, configured: false }` plus `CREDENTIAL_MISSING`; no value is ever requested.
+- `setCredential` accepts only `env-manager/<environmentId>/<accountId>` refs and a
+  non-empty value, forwards them to `ctx.credentials.set`, and returns only
+  `{ ref, configured: true }`. It never writes the value to a record or log.
+  `unsetCredential` uses the same namespace check and calls `ctx.credentials.unset`.
+  A missing service yields `CREDENTIAL_UNAVAILABLE`; provider failures yield
+  `CREDENTIAL_WRITE_FAILED` without exposing the provider message to the UI.
+
+## 11. Tools
+
+### 11.1 Shared shape
+
+```ts
+function toolOutput<T extends JsonSchemaNode>(schema: T, render: Renderer): ToolOutputDefinition
+```
+
+- Every tool returns a plain object matching its declared schema; DSH validates it.
+- Every failure throws `EnvManagerError`; no tool returns `{ ok: false }`.
+- `render` emits one text block: a one-line summary plus the essential ids. The
+  canonical JSON value is what the model receives.
+
+`EnvManagerError`:
+
+```ts
+class EnvManagerError extends HarnessError {
+  constructor(code: ErrorCode, message: string, details?: JsonValue) {
+    super(message, code)
+    this.details = details
+  }
+}
+```
+
+### 11.2 Tool table
+
+| Tool | Parameters | Returns | Session | Approval |
+|---|---|---|---|---|
+| `env_list` | `kind?`, `query?` | `{ items, warnings }` | no | no |
+| `env_get` | `kind`, `id` | `{ record, yaml }` | no | no |
+| `env_find` | `query` | `{ hits }` | no | no |
+| `env_resolve` | `environmentId?`, `serviceId`, `deploymentId?` | `{ context, warnings }` | default env | no |
+| `env_resolve_operation` | `operation`, `environmentId?`, `serviceId?`, `deploymentId?` | `{ resolution }` | yes | no |
+| `env_get_context` | — | `{ sessionId, defaultEnvironmentId, persisted }` | yes | no |
+| `env_set_context` | `environmentId \| null` | same | yes | no |
+| `env_save` | `kind`, `record`, `dryRun?` | `{ id, kind, created, dryRun, warnings }` | no | production env |
+| `env_delete` | `kind`, `id`, `confirm` | `{ id, kind, deleted, blockers }` | no | yes |
+| `env_validate` | — | `{ ok, counts, diagnostics }` | no | no |
+
+### 11.3 Parameter definitions
+
+```ts
+const kindSchema = {
+  type: 'string',
+  enum: ['environment', 'service', 'deployment', 'ruleProfile', 'operationBinding'],
+  description: 'Record kind.',
+} as const
+```
+
+`env_list`:
+
+```ts
+parameters: {
+  kind: { ...kindSchema, required: false },
+  query: { type: 'string', required: false, description: 'Case-insensitive match over id, name and alias.' },
+}
+```
+
+`env_get`: `kind` and `id` required, `id` description names the id pattern.
+
+`env_find`: `query` required, min length 2.
+
+`env_resolve`: `serviceId` required; `environmentId` and `deploymentId` optional.
+Resolving an environment/service pair with more than one deployment and no
+`deploymentId` returns `DEPLOYMENT_AMBIGUOUS` listing candidates; zero deployments
+returns `DEPLOYMENT_NOT_FOUND`.
+
+`env_resolve_operation`: `operation` required, `enum` of the 13 standard ids plus a
+free-form custom id — model-visible enum lists only the standard ids; a custom id is
+rejected by the schema unless it is declared in a binding. Because the DSL cannot
+express "either enum or binding-declared", the parameter is a plain string with the
+standard ids in the description and is validated in `execute`
+(`OPERATION_UNKNOWN` when neither standard nor bound).
+
+`env_set_context`: `environmentId` is `oneOf` string / `null`.
+
+`env_save`: `record` is `{ type: 'json', required: true, description: 'Complete record object.' }`.
+Zod still validates it; the JSON parameter only avoids duplicating every record schema
+in the model-facing tool. `dryRun: true` validates and returns the same shape without
+writing; a schema failure still throws `RECORD_INVALID`.
+
+`env_delete`: `confirm` is a required boolean. `false` performs a dry run: it returns
+`deleted: false` plus `blockers` and writes nothing.
+
+### 11.4 Output schemas
+
+Small results are precise; open records use `additionalProperties: true`.
+
+```ts
+const diagnosticSchema = {
+  type: 'object', additionalProperties: false,
+  properties: {
+    severity: { type: 'string', enum: ['info', 'warning', 'error'], required: true },
+    code: { type: 'string', required: true },
+    message: { type: 'string', required: true },
+    file: { type: 'string' },
+    path: { type: 'string' },
+  },
+}
+const recordSchema = { type: 'object', additionalProperties: true }
+```
+
+`env_list` output:
+
+```ts
+{ type: 'object', additionalProperties: false, properties: {
+  items: { type: 'array', required: true, items: { type: 'object', additionalProperties: false, properties: {
+    kind: { type: 'string', required: true },
+    id: { type: 'string', required: true },
+    name: { type: 'string', required: true },
+    summary: { type: 'string' },
+  } } },
+  warnings: { type: 'array', required: true, items: diagnosticSchema },
+} }
+```
+
+`env_get` output: `{ record: recordSchema, yaml: { type: 'string', required: true } }`.
+
+`env_resolve` output: `{ context: recordSchema, warnings: arrayOf(diagnosticSchema) }`.
+
+`env_validate` output: counts per kind plus `diagnostics`.
+
+### 11.5 Execution flow
+
+```text
+execute(args, exec):
+  1. assert exec.signal not aborted
+  2. resolve target ids (session default when allowed)
+  3. call app.<operation>
+  4. on EnvManagerError: rethrow
+  5. on unexpected error: wrap as INTERNAL_ERROR with a redacted message
+  6. return the declared value
+```
+
+Approval helper:
+
+```ts
+async function requireApproval(exec, toolName: string, reason: string): Promise<void> {
+  if (!exec.agent) throw new EnvManagerError('SESSION_CONTEXT_UNAVAILABLE', '...')
+  const outcome = await ctx.approval.request({
+    agent: exec.agent,
+    toolName,
+    callId: exec.callId,
+    reason,
+    displayReason: { en: reason, zh: reasonZh },
+    signal: exec.signal,
+  })
+  if (outcome !== 'allowed-once') throw new EnvManagerError('POLICY_DENIED', '...')
+}
+```
+
+The `displayReason` is localized; the `reason` is the stable English fallback.
+
+## 12. HTTP transport
+
+### 12.1 Route
+
+```ts
+ctx.webServer.register({ kind: 'prefix', path: '/env-manager/api', handler })
+```
+
+- `POST` only; other methods → 405 with `Allow: POST`.
+- `Content-Type: application/json; charset=utf-8` required.
+- Body limit 2 MiB; exceeding it → 413.
+- Request header `x-env-manager-request: 1` required, otherwise 403.
+- If an `Origin` header is present it must equal `ctx.webServer.getOrigin()`,
+  otherwise 403. No CORS headers are ever sent.
+- The handler never reads cookies or credentials. A credential value may appear only in
+  the `credential` request body, is forwarded directly to the credential service and is
+  never logged or echoed.
+
+### 12.2 Envelope
+
+Request:
+
+```json
+{ "op": "record.save", "payload": { "kind": "environment", "yaml": "id: test-a\n..." } }
+```
+
+Success `200`:
+
+```json
+{ "ok": true, "data": { ... } }
+```
+
+Failure:
+
+```json
+{ "ok": false, "error": { "code": "RECORD_INVALID", "message": "...", "details": { ... } } }
+```
+
+Status map: `INVALID_ARGUMENT` 400, `POLICY_DENIED`/`FORBIDDEN` 403, `*_NOT_FOUND` 404,
+`RECORD_IN_USE`/`BINDING_CONFLICT`/`DEPLOYMENT_AMBIGUOUS` 409, `RECORD_INVALID`/
+`RULE_*`/`PARAMETER_*` 422, `CREDENTIAL_UNAVAILABLE` 503, anything else 500 (including
+`CREDENTIAL_WRITE_FAILED`). The body always carries the code; the status is for tooling.
+
+### 12.3 Operations
+
+| op | payload | data |
+|---|---|---|
+| `meta` | — | dataDir, version, counts, warnings |
+| `list` | `{ kind }` | summaries |
+| `get` | `{ kind, id }` | `{ record, yaml }` |
+| `find` | `{ query }` | hits |
+| `save` | `{ kind, record?, yaml?, dryRun? }` | save result |
+| `remove` | `{ kind, id, confirm?, dryRun? }` | remove result |
+| `validate` | — | validation report |
+| `resolve` | `{ environmentId, serviceId, deploymentId? }` | context |
+| `route` | `{ operation, environmentId, serviceId, deploymentId? }` | resolution |
+| `credential` | `{ ref, value }` | `{ ref, configured }` |
+| `credentialUnset` | `{ ref }` | `{ ref, removed }` |
+
+The UI cannot read or write session context; that is agent-only in V1 (§9).
+
+### 12.4 Actor policy
+
+`record.save` and `record.remove` run with actor `'ui'`:
+
+- production environment save is denied unless `allowUiProductionWrites`;
+- `record.remove` requires `confirm: true` unless `dryRun: true`;
+- every write is logged with actor, op, kind, id and outcome (no record content).
+
+`credential` and `credentialUnset` are UI-only, restricted to the `env-manager/`
+namespace and never write to the registry. They log only the ref and outcome.
+
+## 13. Client module
+
+Structural contract:
+
+```js
+return {
+  inject: ['slots', 'locale'],
+  apply(ctx) {
+    ctx.effect(() => ctx.locale.register('envManager', { en, zh }))
+    ctx.slots.inject('sidebar.panellist', () => ctx.slots.register(
+      { name: 'sidebar.panellist', id: 'env-manager', order: 40 }, Icon))
+    ctx.slots.inject('main', () => ctx.slots.register(
+      { name: 'main', key: 'env-manager' }, Panel))
+  },
+}
+```
+
+- All data access goes through `createApi()` in `client.js`, which posts to
+  `/env-manager/api` with the required header and maps failures to a UI error object.
+- The panel keeps state locally (`useState`/`useEffect`); no global store.
+- Rendering and interactions are specified in [`ui-design.md`](ui-design.md).
+
+## 14. Memory bridge
+
+```ts
+interface MemoryBridge {
+  readonly available: boolean
+  emit(record: OperationRecord): Promise<void>
+  observe(scope: MemoryScope): Promise<readonly Observation[]>
+  propose(candidate: CandidateChange): Promise<string>   // returns proposal id
+}
+```
+
+V1 ships `NullMemoryBridge`: `available: false`, `emit` resolves without writing,
+`observe` returns `[]`, `propose` throws `MEMORY_BRIDGE_UNAVAILABLE`. It is selected
+when no memory provider is registered; a future adapter plugs in without changing the
+tools because nothing in V1 calls `emit` (no execution).
+
+## 15. Cross-cutting
+
+### 15.1 Logging
+
+`src/logger.ts` wraps `ctx.get('logger')` when it is callable, else `console`.
+Log levels: `info` for lifecycle (dataDir, loaded counts), `warn` for skipped records,
+`error` for failed loads. Records and credential refs are logged as ids only; never a
+whole record, never a value.
+
+### 15.2 Redaction
+
+`redactValue(text)` removes anything that looks like a URL userinfo, a bearer token,
+or a private key block from messages before they reach the model or the UI. Tool
+results never include `credentialRef` values, only configured state.
+
+### 15.3 Cancellation
+
+`execute` receives `exec.signal`; long loops (registry load, resolve) check it between
+files and between fields. A cancelled call surfaces DSH's own abort error unchanged.
+
+### 15.4 Performance budgets
+
+| Operation | Budget |
+|---|---|
+| registry load (500 records) | 100 ms |
+| resolve (50 fields) | 5 ms |
+| route with skill snapshot | 50 ms |
+| HTTP round trip (local) | 150 ms |
+
+## 16. Test obligations
+
+| Module | Tests | From AGENTS §13 |
+|---|---|---|
+| `storage` | atomic write, missing file, invalid YAML, schema mismatch, backups, traversal rejection | required |
+| `registry` | load/lookup, duplicate id, broken references, alias collision | required |
+| `resolver` | precedence, field merge, provenance, undefined/cycle/traversal, runtime selection | required |
+| `router` | each level, conflict, missing skill, source/provider mismatch, policy, parameters | required |
+| `session` | set/get/clear, isolation, corrupt fallback | required |
+| `tools` | output matches schema, error codes, no credential values, approval denial | required |
+| `app` | save→get round trip, delete in use, validate report | new |
+| `http` | auth header, origin, method, size cap, status map, production denial | new |
+| `client` | manual checklist in `ui-design.md` §9 | manual |
+
+Tests live in `tests/<module>/*.test.ts`, fixtures in `tests/fixtures/<scenario>/`.
+
+## 17. Traceability
+
+| Design section | Implements | Acceptance |
+|---|---|---|
+| §4 storage | ADR-0002 | AC-01, AC-02 |
+| §7 resolver | ADR-0005 | AC-03, AC-04 |
+| §8 router | ADR-0003 | AC-05…AC-07 |
+| §9 session | ADR-0007 | AC-08 |
+| §12 HTTP | ADR-0009 | UI manual QA (`ui-design.md` §12) |
+| §11 tools | tool-contract.md | AC-02…AC-08 |
+
+## 18. Deliberate omissions
+
+- No cache beyond one in-flight load.
+- No cross-process locking.
+- No Skill execution.
+- No session-event projection.
+- No digest pinning.
+- No plugin-side bundler.
+- No database.