dsh-env-manager-plugin 0.0.0-stage → 0.1.0

package/docs/01-architecture/tool-contract.md

@@ -0,0 +1,438 @@
+# Tool Contract (V1)
+
+> Status: normative model-facing contract. Registration mechanics are in
+> [`api-verification.md`](api-verification.md) F2–F5; implementation in
+> [`detailed-design.md`](detailed-design.md) §11.
+> An implementation whose tool schema differs from this document is a bug in one of
+> them; update both in the same change.
+
+## 1. Conventions
+
+### 1.1 Registration
+
+Every tool is a `defineTool` definition registered on `ctx.tools` during `apply`:
+
+```ts
+ctx.tools.register(defineTool({
+  name, description, parameters,
+  output: { schema, render },
+  async execute(args, exec) { /* returns the declared JSON value */ },
+}))
+```
+
+- The tool name is stable and matches `^env_[a-z_]+$`.
+- `parameters` uses DSH's value-schema DSL, not raw JSON Schema. Property-level
+  `required: true` marks required arguments. `additionalProperties` must be explicit.
+- `output.schema` is the enforced JSON Schema subset: `type`, `oneOf`, `properties`,
+  `required`, `additionalProperties`, `items`, `enum`, `const` plus annotations.
+- `render(args, value)` returns `ContentBlock[]`. V1 emits exactly one text block: a
+  one-line summary plus the ids a human needs. The canonical JSON value is what the
+  model receives.
+- Descriptions start with the decision rule ("Use when …"), not with a restatement of
+  the name.
+
+### 1.2 Errors
+
+Success returns the declared value. Every expected failure throws:
+
+```ts
+class EnvManagerError extends HarnessError {
+  constructor(code: ErrorCode, message: string, details?: JsonValue)
+}
+```
+
+`HarnessError` is imported from `@deepseek-ai/dsh-llm` and constructed as
+`super(message, code)`. DSH records `{ name, code }` beside the message, so the model
+and the UI can branch on `code` without parsing text.
+
+Rules:
+
+- No tool returns `{ ok: false }`.
+- No tool throws a bare string.
+- Messages never contain a credential value, a file body, or a stack trace.
+- Unexpected exceptions are wrapped as `INTERNAL_ERROR` with a redacted message.
+- A cancelled call surfaces DSH's own abort error; do not translate it.
+
+Codes are defined in [`data-model.md`](data-model.md) §11.
+
+### 1.3 Approval
+
+| Tool | Approval |
+|---|---|
+| `env_delete` | always, when `confirm: true` and `writePolicy.requireApprovalForDelete` |
+| `env_save` | when the record is a production environment, or a deployment whose environment is production, and `writePolicy.requireApprovalForProductionSave` |
+| everything else | no |
+
+Approval uses `ctx.approval.request` with the executing agent, the tool name, the call
+id, an English `reason`, and a localized `displayReason`. Only `'allowed-once'` grants;
+`'rejected'`, `'cancelled'` and `'unavailable'` all become `POLICY_DENIED`.
+
+A session-scoped tool invoked without `exec.agent` fails with
+`SESSION_CONTEXT_UNAVAILABLE`; it never guesses a session.
+
+### 1.4 Session default
+
+`env_resolve`, `env_resolve_operation` and `env_set_context` operate on the calling
+agent's session. When `environmentId` is omitted, the order is:
+
+1. the explicit argument;
+2. the session default (`sessions/<sessionId>.json`);
+3. `config.defaults.environmentId`;
+4. `INVALID_ARGUMENT` naming both ways to supply an environment.
+
+### 1.5 Shared value shapes
+
+```ts
+interface Diagnostic {
+  severity: 'info' | 'warning' | 'error'
+  code: string
+  message: string
+  file?: string
+  path?: string
+}
+
+interface RecordSummary {
+  kind: RecordKind
+  id: string
+  name: string
+  summary?: string        // kind-specific one-liner
+}
+
+interface WarningList { warnings: Diagnostic[] }
+```
+
+`RecordKind = 'environment' | 'service' | 'deployment' | 'ruleProfile' | 'operationBinding'`.
+
+## 2. Read tools
+
+### 2.1 `env_list`
+
+Use when the caller needs an inventory or wants to discover ids before a specific call.
+
+```ts
+parameters: {
+  kind: {
+    type: 'string',
+    required: false,
+    enum: ['environment', 'service', 'deployment', 'ruleProfile', 'operationBinding'],
+    description: 'Restrict to one record kind.',
+  },
+  query: {
+    type: 'string',
+    required: false,
+    description: 'Case-insensitive match over id, name and alias.',
+  },
+}
+```
+
+Returns:
+
+```json
+{ "items": [ { "kind": "environment", "id": "test-a", "name": "Test A", "summary": "test · 2 hosts" } ],
+  "warnings": [] }
+```
+
+- Sorted by kind, then id.
+- An empty registry is `{ "items": [], "warnings": [] }`, not an error.
+- Registry diagnostics are summarized in `warnings` (at most 20, with a count of the
+  rest).
+
+Errors: `INVALID_ARGUMENT` for an unknown `kind`.
+
+Render: `"5 environments, 2 services (2 warnings)."`
+
+### 2.2 `env_get`
+
+Use when the caller already has an id and needs the full record.
+
+```ts
+parameters: {
+  kind: { type: 'string', required: true, enum: [/* five kinds */] },
+  id: { type: 'string', required: true, description: 'Record id, e.g. test-a.' },
+}
+```
+
+Returns:
+
+```json
+{ "record": { "id": "test-a", "schemaVersion": 1, "...": "..." },
+  "yaml": "id: test-a\n..." }
+```
+
+- `record` is the parsed document; `yaml` is the on-disk text (helpful for edits).
+- Unknown fields in the file are preserved in both.
+
+Errors: `ENV_NOT_FOUND`, `SERVICE_NOT_FOUND`, `DEPLOYMENT_NOT_FOUND`,
+`RULE_PROFILE_NOT_FOUND`, `BINDING_NOT_FOUND` depending on `kind`; `INVALID_ARGUMENT`
+for a malformed id (fails before any I/O).
+
+Render: `"environment test-a (Test A), updated 2026-11-01."`
+
+### 2.3 `env_find`
+
+Use when the caller has a name or fragment and wants candidate ids. Prefer this over
+guessing an id.
+
+```ts
+parameters: {
+  query: {
+    type: 'string',
+    required: true,
+    description: 'Fragment of an id, name, alias or category (min 2 characters).',
+  },
+}
+```
+
+Returns:
+
+```json
+{ "hits": [ { "kind": "service", "id": "aicoservice", "name": "Aico Service", "field": "alias" } ] }
+```
+
+- Matching is case-insensitive over `id`, `name`, `aliases`, `category`.
+- Ranked: exact id, prefix id, name, alias, category. At most 25 hits.
+- No hit is not an error.
+
+Errors: `INVALID_ARGUMENT` when `query` is shorter than 2 characters.
+
+Render: `"3 matches for \"aico\"."`
+
+### 2.4 `env_resolve`
+
+Use when the caller needs concrete paths, log sources, class locations or variables
+for a service in an environment. This is the authoritative answer; do not reconstruct
+paths from templates.
+
+```ts
+parameters: {
+  environmentId: { type: 'string', required: false, description: 'Defaults to the session environment.' },
+  serviceId: { type: 'string', required: true },
+  deploymentId: {
+    type: 'string',
+    required: false,
+    description: 'Required when the service has more than one binding in the environment; a single binding is selected automatically.',
+  },
+}
+```
+
+Returns the Service Context from `data-model.md` §8.2, including `provenance` and
+`warnings`.
+
+Errors: `ENV_NOT_FOUND`, `SERVICE_NOT_FOUND`, `DEPLOYMENT_NOT_FOUND`,
+`DEPLOYMENT_AMBIGUOUS` (lists candidates), `RULE_*`, `PARAMETER_RESOLUTION_FAILED`.
+
+Render: `"Resolved aicoservice in test-a (kubernetes): 7 paths, 2 variables, 1 warning."`
+
+### 2.5 `env_resolve_operation`
+
+Use before acting: ask which Skill handles an operation for this context and what
+parameters it expects. This tool **does not execute** the Skill.
+
+```ts
+parameters: {
+  operation: {
+    type: 'string',
+    required: true,
+    description: 'deploy | redeploy | restart | rollback | collect-logs | health-check | … (standard id or a bound custom id).',
+  },
+  environmentId: { type: 'string', required: false },
+  serviceId: { type: 'string', required: false },
+  deploymentId: { type: 'string', required: false },
+}
+```
+
+- At least one of `serviceId` / `deploymentId` is required; `environmentId` falls back
+  to the session default.
+
+Returns `OperationResolution` from `data-model.md` §8.3:
+
+```json
+{
+  "ok": true,
+  "operation": "redeploy",
+  "requested": { "environmentId": "test-a", "serviceId": "aicoservice", "deploymentId": "deploy-001" },
+  "matchedBindingId": "binding-001",
+  "matchLevel": "environment-service",
+  "skill": { "name": "redeploy-k8s-service", "available": true, "source": "project-dsh", "provider": "dsh-skills" },
+  "parameters": { "namespace": "dv-sop", "workload": "aicoservice", "cluster": "k8s-01" },
+  "policy": { "requiresApproval": true, "allowProduction": false },
+  "diagnostics": []
+}
+```
+
+Errors: `ENV_NOT_FOUND`, `SERVICE_NOT_FOUND`, `DEPLOYMENT_NOT_FOUND`,
+`DEPLOYMENT_AMBIGUOUS`, `BINDING_NOT_FOUND`, `BINDING_CONFLICT` (lists every candidate),
+`SKILL_UNAVAILABLE`, `POLICY_DENIED`, `OPERATION_UNKNOWN`, `PARAMETER_RESOLUTION_FAILED`.
+
+Incomplete Skill catalog: `skill.available: "unknown"` plus a
+`SKILL_REGISTRY_UNAVAILABLE` warning; the call still succeeds.
+
+Render: `"redeploy aicoservice/test-a → redeploy-k8s-service (approval required)."`
+
+### 2.6 `env_get_context`
+
+Use to check what environment the current session defaults to.
+
+```ts
+parameters: {}
+```
+
+Returns:
+
+```json
+{ "sessionId": "01J…", "defaultEnvironmentId": "test-a", "persisted": true }
+```
+
+Errors: `SESSION_CONTEXT_UNAVAILABLE` when there is no owning agent.
+
+Render: `"Session default: test-a."`
+
+### 2.7 `env_set_context`
+
+Use when the user says "work on test-a" and later calls should omit `environmentId`.
+
+```ts
+parameters: {
+  environmentId: {
+    oneOf: [{ type: 'string' }, { type: 'null' }],
+    required: true,
+    description: 'Environment id, or null to clear the session default.',
+  },
+}
+```
+
+Returns the same shape as `env_get_context`. Writing an unknown environment fails with
+`ENV_NOT_FOUND`.
+
+A file write failure returns `persisted: false` plus a `SESSION_CONTEXT_UNPERSISTED`
+warning; it does not throw, because the process-local cache still honors the choice for
+this process.
+
+Render: `"Session default set to test-a."` / `"Session default cleared."`
+
+## 3. Management tools
+
+### 3.1 `env_save`
+
+Use to create or replace a record. Prefer `env_get` first, edit the YAML, then save the
+whole record; do not send a partial patch.
+
+```ts
+parameters: {
+  kind: { type: 'string', required: true, enum: [/* five kinds */] },
+  record: {
+    type: 'json',
+    required: true,
+    description: 'Complete record; id and schemaVersion are required. Host timestamps are set by the plugin and ignored if supplied.',
+  },
+  dryRun: {
+    type: 'boolean',
+    required: false,
+    description: 'Validate and report without writing.',
+  },
+}
+```
+
+Returns:
+
+```json
+{ "id": "test-a", "kind": "environment", "created": false, "dryRun": false,
+  "warnings": [ { "severity": "warning", "code": "CREDENTIAL_MISSING", "message": "…" } ] }
+```
+
+Behavior:
+
+- Zod validation, id/filename equality, record cap, then atomic write, then reload.
+- `created` is `true` only when no file existed before.
+- Broken references after the write are returned as warnings; the write is not rolled
+  back because the record itself is valid.
+- Production records require approval unless the config disables it.
+
+Errors: `RECORD_INVALID` (with `details.issues`), `RECORD_LIMIT_EXCEEDED`,
+`RECORD_TOO_LARGE`, `STORAGE_UNWRITABLE`, `POLICY_DENIED`.
+
+Render: `"Saved environment test-a (updated, 1 warning)."`
+
+### 3.2 `env_delete`
+
+Use only when the user explicitly asks to remove a record.
+
+```ts
+parameters: {
+  kind: { type: 'string', required: true, enum: [/* five kinds */] },
+  id: { type: 'string', required: true },
+  confirm: {
+    type: 'boolean',
+    required: true,
+    description: 'false checks references only; true deletes after approval.',
+  },
+}
+```
+
+Returns:
+
+```json
+{ "id": "test-a", "kind": "environment", "deleted": true, "blockers": [] }
+```
+
+Behavior:
+
+- `confirm: false` never deletes; it returns `blockers` describing references.
+- `confirm: true` asks for approval, then refuses with `RECORD_IN_USE` when referenced.
+- A backup copy is written before the unlink.
+
+Errors: not found by kind, `RECORD_IN_USE`, `POLICY_DENIED`, `STORAGE_UNWRITABLE`.
+
+Render: `"Deleted environment test-a (backup kept)."`
+
+### 3.3 `env_validate`
+
+Use after hand-editing YAML or before trusting a resolve result.
+
+```ts
+parameters: {}
+```
+
+Returns:
+
+```json
+{
+  "ok": false,
+  "counts": {
+    "environment": 2, "service": 1, "deployment": 1,
+    "ruleProfile": 1, "operationBinding": 2
+  },
+  "diagnostics": [
+    { "severity": "error", "code": "RULE_PROFILE_NOT_FOUND",
+      "message": "service aicoservice references unknown rule profile java-commn",
+      "file": "services/aicoservice.yaml", "path": "ruleProfileRefs[0]" }
+  ]
+}
+```
+
+- Always succeeds at the tool level; `ok: false` is a report, not a failure.
+- Read-only: no write, no reload side effect beyond the normal request load.
+
+Render: `"Registry invalid: 2 errors, 1 warning."`
+
+## 4. Cross-tool rules
+
+1. **Ids are data, not paths.** A tool never accepts a filesystem path.
+2. **No credentials.** `credentialRef` status is reported as
+   `{ ref, configured, source }`; values are never resolved.
+3. **No execution.** No tool runs a Skill, a shell command, or a network request.
+4. **Determinism.** Same registry and inputs produce the same result; no random ids,
+   no wall-clock ordering.
+5. **Read-only means read-only.** `env_validate`, `env_get_context`, `env_list`,
+   `env_get`, `env_find`, `env_resolve`, `env_resolve_operation` never write a file.
+6. **Fail closed.** A conflict, missing Skill, broken reference or production denial is
+   an error, never a fallback.
+
+## 5. Non-goals for V1 tools
+
+- Executing an operation or invoking a Skill.
+- Writing operation records to memory.
+- Bulk import/export.
+- Editing plugin configuration.
+- Reading or writing session context for a session other than the caller's.
+- Any network or shell access.