@kindgi/cli 0.1.4-rc.5 → 0.1.5-rc.0

package/dist/sdk-skills/kindgi-authoring-providers/SKILL.md

@@ -5,8 +5,9 @@ description: >
   can actually call a real model. Covers four paths — hosted via
   Anthropic native adapter, Gemini on Vertex AI (Google Application
   Default Credentials, no API key), hosted via the OpenAI-compat adapter
-  (works with OpenAI + Groq + Together + Fireworks + OpenRouter +
-  Ollama + vLLM + any other OpenAI-compatible endpoint), and local
+  (works with OpenAI, Groq, self-hosted vLLM and Ollama, and any other
+  OpenAI-compatible endpoint, a hosted gateway such as OpenRouter
+  included), and local
   via the in-process ONNX adapter — plus the credential flow (in
   `kindgi dev` the key lives in the project's env files — `.env`, then
   `.env.local` — added by hand or with `kindgi secrets set`'s no-echo
@@ -22,9 +23,9 @@ description: >
   kindgi-getting-started.
 type: core
 library: "@kindgi/sdk"
-version: "0.9.5"
-sdk_version: "0.1.4-rc.5"
-pack_languages: [node, python]
+version: "0.9.11"
+sdk_version: "0.1.5-rc.0"
+pack_languages: [node, python, java, scala]
 sources:
   - packages/adapters/model-anthropic/src/provider.ts
   - packages/adapters/model-gemini/src/provider.ts
@@ -40,7 +41,8 @@ sources:
 > (`@kindgi/cli`), not a global command. Run it through the project's
 > package manager — `pnpm exec kindgi …`, `npx --no kindgi …` (npm),
 > `yarn kindgi …` or `bun run kindgi …`. A Python pack (`[tool.kindgi]` in
-> `pyproject.toml`) has no Node project: run the `kindgi` on `PATH`.
+> `pyproject.toml`) has no Node project: run the `kindgi` on `PATH`. A Java
+> or Scala pack (`kindgi.config.json`) runs the CLI it pins: `./kindgiw …`.
 > Commands below are written `kindgi …` for brevity.
 
 An **agent** is a versioned declaration; it needs a **provider** to run.
@@ -75,8 +77,9 @@ Three moving parts:
 1. **API key on disk** — in `kindgi dev` (environment `local`) the dotenv
    secret binding reads the project's own env files: `.env`, then
    `.env.local` on top (change the list with `dev.envFiles` in
-   `kindgi.config.ts`, or `envFiles` under `[tool.kindgi.dev]` in a Python
-   pack's `pyproject.toml`). A key already in the app's `.env` just works.
+   `kindgi.config.ts`, `envFiles` under `[tool.kindgi.dev]` in a Python
+   pack's `pyproject.toml`, or `dev.envFiles` in a Java or Scala pack's
+   `kindgi.config.json`). A key already in the app's `.env` just works.
    `kindgi secrets set` (interactive, no-echo) writes `.env.local`. For
    non-sensitive values (log levels, region names, feature flags),
    `kindgi env set NAME VALUE --env=local` writes the same file with a
@@ -106,7 +109,11 @@ yours.
 ## Path A — Hosted, native Anthropic
 
 Best fidelity to Anthropic's API (prompt caching, latest models, tool
-use, structured output). Requires an `ANTHROPIC_API_KEY`.
+use). Requires an `ANTHROPIC_API_KEY`.
+
+A model's `structured-output` feature is a routing label: the model can
+follow a JSON schema natively, but Kindgi's typed outputs use instructions,
+then parse, check against the schema and repair, on every provider.
 
 **Step 1 — set the key:**
 ```sh
@@ -123,9 +130,19 @@ credential on argv.
 
 **Step 2 — register it, from the preset:**
 ```sh
-kindgi providers register --preset=anthropic                          # Opus 5.5, Sonnet 5.5, Haiku 4.5
-kindgi providers register --preset=anthropic --models=claude-haiku-4-5  # just one
+kindgi providers register --preset=anthropic                          # Opus 5.5, Sonnet 5.5 (default), Haiku 5.5, Haiku 4.5
+kindgi providers register --preset=anthropic --models=claude-sonnet-5-5  # just one
 ```
+Before pinning a Claude model, check its status on Anthropic's model
+deprecations page (https://platform.claude.com/docs/en/about-claude/model-deprecations): a turn routed to a retired model fails. Prefer the
+preset's default. Each
+preset names a default model (`metadata.defaultModel`, marked `(default)`
+when it registers), which an agent with no preference gets. A preset
+registered before 0.1.4 has none: unregister it and register it again.
+The Claude 5.5 and GPT-6 models take no `temperature` (`"sampling": false`:
+the call goes without it, with a `sampling-unsupported` warning), and a
+model's `thinking` says how it thinks; thinking counts against
+`maxOutputTokens` and bills as output.
 The preset carries the models, context windows, output limits and current
 prices (`kindgi providers presets` lists the presets and when their prices
 were checked); `--max-output-tokens=<n>` sets another output limit. In a pack it refuses until the key is in the pack's env files —
@@ -139,8 +156,8 @@ needs under `kindgi dev`:
 ```ts
 // in kindgi.config.ts
 providers: [
-  { preset: 'anthropic', models: ['claude-haiku-4-5'] },   // key ANTHROPIC_API_KEY, from the env files
-  { preset: 'gemini', project: 'acme-gcp', models: ['gemini-2.5-flash'] },
+  { preset: 'anthropic', models: ['claude-sonnet-5-5'] },  // key ANTHROPIC_API_KEY, from the env files
+  { preset: 'gemini', project: 'acme-gcp', models: ['gemini-3.8-flash'] },
   { spec: { /* the provider.json body below */ } },
 ],
 ```
@@ -148,14 +165,17 @@ providers: [
 # in pyproject.toml: one table per provider, same keys
 [[tool.kindgi.providers]]
 preset = "anthropic"
-models = ["claude-haiku-4-5"]
+models = ["claude-sonnet-5-5"]
 ```
+In a Java or Scala pack's `kindgi.config.json`, the same keys:
+`"providers": [{"preset": "anthropic", "models": ["claude-sonnet-5-5"]}]`.
 - A preset entry takes `models`, `project`, `secret` (the key's name, in place
-  of the preset's) and `maxOutputTokens`, spelled the same in `pyproject.toml`;
+  of the preset's) and `maxOutputTokens`, spelled the same in `pyproject.toml`
+  and `kindgi.config.json`;
   a `spec` entry is a `--spec` body. A
   key is always a secret's name (`secret_ref`); a credential in
   `adapter_config` is refused.
-- Each boot prints `Providers from kindgi.config.ts:` with one line each:
+- Each boot prints `Providers from kindgi.config.ts:` (the pack's config file) with one line each:
   `registered`, `unchanged`, `registered again (changed in kindgi.config.ts)`,
   `unregistered (no longer in kindgi.config.ts)`, or ⚠ `not registered: <KEY>
   is not in .env, .env.local` (set the key, then restart: the config isn't
@@ -167,7 +187,7 @@ models = ["claude-haiku-4-5"]
   providers with `kindgi providers register`.
 
 **Step 2 (by hand) — write `provider.json`** at the pack root. One connection,
-three models — matches how the Anthropic SDK actually works (the API
+two models — matches how the Anthropic SDK actually works (the API
 key is per-vendor; the model is per-call):
 ```json
 {
@@ -194,16 +214,6 @@ key is per-vendor; the model is per-call):
           "completionUsdPer1kTokens": 0.01
         },
         "description": "Balanced performance/cost."
-      },
-      {
-        "name": "claude-haiku-4-5",
-        "contextWindow": 200000,
-        "features": ["tool-use"],
-        "cost": {
-          "promptUsdPer1kTokens": 0.001,
-          "completionUsdPer1kTokens": 0.005
-        },
-        "description": "Fastest and cheapest — routing, classification, simple calls."
       }
     ],
     "description": "Anthropic Claude via native adapter."
@@ -261,9 +271,9 @@ Works with **any** OpenAI-compatible endpoint. Same adapter, different
 | Groq | `https://api.groq.com/openai/v1` |
 | Together | `https://api.together.xyz/v1` |
 | Fireworks | `https://api.fireworks.ai/inference/v1` |
-| OpenRouter | `https://openrouter.ai/api/v1` |
 | DeepSeek | `https://api.deepseek.com/v1` |
 | LiteLLM proxy | `http://localhost:4000/v1` |
+| OpenRouter (a hosted gateway) | `https://openrouter.ai/api/v1` |
 
 The connection carries the `baseURL` (in `adapter_config`);
 each endpoint is a separate provider row because each has its own API
@@ -485,24 +495,16 @@ outside Google Cloud: put a service-account key (its JSON) in a secret
     "region": "global",
     "models": [
       {
-        "name": "gemini-2.5-pro",
+        "name": "gemini-3.8-flash",
         "contextWindow": 1048576,
-        "features": ["tool-use"],
+        "features": ["tool-use", "structured-output", "long-context"],
         "maxOutputTokens": 65536,
-        "cost": {
-          "promptUsdPer1kTokens": 0.00125,
-          "completionUsdPer1kTokens": 0.01,
-          "longContext": {
-            "thresholdTokens": 200000,
-            "promptUsdPer1kTokens": 0.0025,
-            "completionUsdPer1kTokens": 0.015
-          }
-        }
+        "cost": { "promptUsdPer1kTokens": 0.00075, "completionUsdPer1kTokens": 0.00375 }
       },
       {
-        "name": "gemini-2.5-flash",
+        "name": "gemini-3.5-flash-lite",
         "contextWindow": 1048576,
-        "features": ["tool-use"],
+        "features": ["tool-use", "structured-output", "long-context"],
         "maxOutputTokens": 65536,
         "cost": { "promptUsdPer1kTokens": 0.0003, "completionUsdPer1kTokens": 0.0025 }
       }
@@ -517,11 +519,17 @@ outside Google Cloud: put a service-account key (its JSON) in a secret
 - `metadata.region` is the Vertex location: `global`, or a region such as
   `us-central1` or `northamerica-northeast1` when data must stay in one
   place. `unspecified` means `global`. Different locations are different
-  provider rows.
+  provider rows. Check that the location serves the model: Gemini 3.8 Flash
+  isn't served from `us-central1`.
+- Don't register `gemini-2.5-pro` or `gemini-2.5-flash`: Vertex AI retires
+  both on 2026-10-20.
 - Rates are per 1K tokens, from Google's published pricing; check them
-  before relying on budgets. Thinking tokens bill as output.
-  `longContext` switches the whole call to the higher rates past the
-  threshold; `cachedPromptMultiplier` (default 0.25) prices cached
+  before relying on budgets. Thinking tokens bill as output, and Gemini 3.8
+  Flash thinks by default. `gemini-3.8-flash`'s rates above are Google's
+  launch price, through 2026-12-31 ($0.0015 / $0.0075 from 2027-01-01).
+  `longContext` (`{ thresholdTokens, promptUsdPer1kTokens,
+  completionUsdPer1kTokens }` in a model's `cost`) switches the whole call
+  to the higher rates past the threshold; `cachedPromptMultiplier` (default 0.25) prices cached
   prompt tokens.
 
 **Step 3 — register and check:**
@@ -533,7 +541,8 @@ Or skip step 2: `kindgi providers register --preset=gemini --project=<your-gcp-p
 registers both models above.
 Then pin it from an agent with `preferredProvider: 'gemini'` (and a model
 with `preferredModel`; in Python, `preferred_provider="gemini"` and
-`preferred_model=…`), or let the router pick by capability.
+`preferred_model=…`; in Java, `.set("preferredProvider", "gemini")`), or let
+the router pick by capability.
 
 ## How the router picks between multiple providers + models
 
@@ -548,13 +557,16 @@ tenant policy), then sorts survivors in this order:
      - Only `preferredModel` set → promote any provider exposing that model.
      - Only `preferredProvider` set → promote every model of that provider.
    `defineAgent` takes both (`preferredProvider`, `preferredModel`), and
-   so does a Python `Agent` (`preferred_provider=`, `preferred_model=`).
+   so does a Python `Agent` (`preferred_provider=`, `preferred_model=`) and
+   a Java `Agent.define(…)` (`set("preferredProvider", …)`).
 2. **`capability.prefer[]` weights.** If the agent's capability
    declares `prefer: [{feature: 'thinking', weight: 3}, ...]`, tuples
    with matching model features (or provider attributes) get higher
    scores. Sorted by summed score, descending.
-3. **Deterministic lexical tiebreak.** When scores tie, tuples sort by
-   `(providerId, modelName)` alphabetically — replay-safe and stable.
+3. **Deterministic tiebreak.** When scores tie, tuples sort by provider
+   id, then the provider's `defaultModel` before its other models, then
+   model name — replay-safe and stable. A provider without a
+   `defaultModel` falls back to its first model by name.
 
 **Practical rule:** preferences are soft — they rank, they don't
 exclude. To guarantee which model runs, make it a hard requirement in
@@ -645,7 +657,8 @@ defineAgent({
    be the FULL npm package name of the adapter — `"@kindgi/adapter-model-anthropic"`,
    NOT `"anthropic"`. Adapters are registered with the runtime under
    their full package names, and a short name matches none of them, so
-   the registration fails. The model adapters are
+   registering is refused: the runtime has no adapter by that name
+   (`✗ /adapter_id: …`). The model adapters are
    `@kindgi/adapter-model-anthropic`, `@kindgi/adapter-model-gemini`,
    `@kindgi/adapter-model-openai-compat` and
    `@kindgi/adapter-model-in-process`; `kindgi providers presets` shows
@@ -682,8 +695,9 @@ defineAgent({
    Then re-register.
 
 6. **Key not found by the runtime.** `kindgi dev` reads the env files
-   at the PACK ROOT (the directory with `kindgi.config.ts`, or a Python
-   pack's `pyproject.toml` with `[tool.kindgi]`) — `.env` and
+   at the PACK ROOT (the directory with `kindgi.config.ts`, a Python
+   pack's `pyproject.toml` with `[tool.kindgi]`, or a Java or Scala pack's
+   `kindgi.config.json`) — `.env` and
    `.env.local`, or whatever `dev.envFiles` lists; the boot log prints
    which files it found. A `KINDGI_`-prefixed name is Kindgi runtime
    config and never resolves as a secret. Outside `kindgi dev`, the
@@ -712,6 +726,21 @@ defineAgent({
    that names nothing registered, and the tenant's policy. `kindgi
    providers list` shows what is registered.
 
+10. **A setting the adapter can't use.** Registering checks the spec
+    against its adapter (no network call, no key read) and refuses what
+    it can't use: `422 provider-config-invalid`, nothing stored, one line
+    per problem with its JSON-pointer path:
+    ```text
+    Error [invalid-request]: Provider "ollama" doesn't fit adapter @kindgi/adapter-model-openai-compat: adapter_config.api must be one of responses, chat-completions.
+      ✗ /adapter_config/api: adapter_config.api must be one of responses, chat-completions.
+    ```
+    Fix each `✗` line's setting and register again. A key the adapter
+    needs is checked too (`✗ /secret_ref: …`); whether the key works, or
+    the endpoint answers, isn't (the first turn finds out). A
+    registration stored before 0.1.5 wasn't checked: `kindgi doctor`
+    names its problems (`GET /v1/providers/<id>/check`); unregister it
+    and register it again.
+
 ## Verifying end-to-end
 
 ```sh

package/dist/sdk-skills/kindgi-authoring-tools/SKILL.md

@@ -13,7 +13,7 @@ description: >
 type: core
 library: "@kindgi/sdk"
 version: "0.4.4"
-sdk_version: "0.1.4-rc.5"
+sdk_version: "0.1.5-rc.0"
 pack_languages: [node]
 sources:
   - packages/tools/src/types.ts
@@ -132,6 +132,32 @@ const defined = defineTool({
 
 The runtime resolves every declared secret on every call, for the call's tenant, in its env (`KINDGI_ENV`; in `kindgi dev`, `local`: the pack's `.env` and `.env.local`). It checks each value against its schema, and fails the call, naming the secret, when one is missing or doesn't match. Every declared secret is required, so in a runtime call `ctx.secrets` holds them all; it's optional in the type because a unit test builds its own context and passes `secrets: { CITATOR_KEY: '…' }`.
 
+A value that differs per tenant, org or project but isn't secret (a base URL, a region, an account id) is an **env value**: declared in `needsSpec.env`, read from `ctx.env`:
+
+```ts
+const defined = defineTool({
+  // …id, description, version, input, output, effects…
+  needsSpec: {
+    env: {
+      ORDERS_BASE_URL: { type: 'string', pattern: '^https://' },
+      ORDERS_REGION: { type: 'string', enum: ['eu', 'us'], default: 'eu' },
+    },
+  },
+  handler: async ({ orderId }, ctx) => ({
+    url: `${ctx.env?.ORDERS_BASE_URL}/${ctx.env?.ORDERS_REGION}/orders/${orderId}`,
+  }),
+});
+```
+
+- **Which value a call gets:** its project's, else its org's, else the tenant's, in the runtime's env; a schema `default` makes a name optional. The values a call used are recorded with it, so a retry or a resume sees the same ones.
+- **Setting them:** `kindgi env set ORDERS_REGION us --scope=project:<project-id> --env=local` (or `--scope=tenant`, for every project). Changing a value that's already set takes `--force`.
+- **A declared value nobody set** stops the call before the tool runs. For a tool `acme-orders.needs-account` that declares `ACME_ACCOUNT_ID`, the message reads:
+  ```text
+  precondition-failed: Tool "acme-orders.needs-account" was not run: env-value-missing: tool "acme-orders.needs-account" needs env value "ACME_ACCOUNT_ID" in env "local", and none is set for project e889c1f5-eae7-45dc-8669-5bd029a5d85c, its org, or the tenant. Set it: kindgi env set ACME_ACCOUNT_ID <value> --scope=project:e889c1f5-eae7-45dc-8669-5bd029a5d85c --env=local (or --scope=tenant, for every project)
+  ```
+- **Not secret:** env values are recorded with each run that uses them and shown in its journal. A credential is a secret (`needsSpec.secrets`), never an env value.
+- **In a unit test:** pass `env: { … }` in the context `invokeTool` gets.
+
 Everything else comes from the process environment: `process.env.CITATOR_URL`. The pack service runs with the pack's env files in `kindgi dev`, and with the container's environment in an image. Declare the names your code reads in `kindgi.config.ts`, `env: { required: ['CITATOR_URL'], optional: [...] }`: a deployment injects exactly those, a pack service missing a required one isn't ready and says which, and `kindgi dev` warns about it. Values per environment go in `environments.<name>.env`, secrets only as references.
 
 ## Declarative HTTP spec

package/dist/sdk-skills/kindgi-framework-feedback/SKILL.md

@@ -14,9 +14,9 @@ description: >
   diagnostic output into durable input for framework improvement.
 type: core
 library: "@kindgi/sdk"
-version: "0.4.0"
-sdk_version: "0.1.4-rc.5"
-pack_languages: [node, python]
+version: "0.4.2"
+sdk_version: "0.1.5-rc.0"
+pack_languages: [node, python, java, scala]
 ---
 
 # Capturing framework feedback
@@ -25,7 +25,8 @@ pack_languages: [node, python]
 > (`@kindgi/cli`), not a global command. Run it through the project's
 > package manager — `pnpm exec kindgi …`, `npx --no kindgi …` (npm),
 > `yarn kindgi …` or `bun run kindgi …`. A Python pack (`[tool.kindgi]` in
-> `pyproject.toml`) has no Node project: run the `kindgi` on `PATH`.
+> `pyproject.toml`) has no Node project: run the `kindgi` on `PATH`. A Java
+> or Scala pack (`kindgi.config.json`) runs the CLI it pins: `./kindgiw …`.
 > Commands below are written `kindgi …` for brevity.
 
 You just spent time diagnosing a Kindgi-framework issue. That diagnostic

package/dist/sdk-skills/kindgi-getting-started/SKILL.md

@@ -15,7 +15,7 @@ description: >
 type: core
 library: "@kindgi/sdk"
 version: "0.3.8"
-sdk_version: "0.1.4-rc.5"
+sdk_version: "0.1.5-rc.0"
 pack_languages: [node]
 ---
 

package/dist/sdk-skills/kindgi-java-authoring-agents/SKILL.md

@@ -0,0 +1,220 @@
+---
+name: kindgi-java-authoring-agents
+description: >
+  Covers writing agents for a Kindgi pack in Java (`com.kindgi:kindgi-pack`):
+  `Agent.define(id)` as a `public static final` field, wiring tools (`Tool`
+  objects or an id with a version range) and guardrails, capabilities and
+  model choice (preferredProvider / preferredModel), conversation policy,
+  turn budgets, prompt parameters, a typed answer from a record, and
+  tool-error retries, all as data. Load this whenever you are authoring or
+  editing code in a Java pack's agents packages (a pack whose
+  `kindgi.config.json` says `"language": "java"`), defining an agent, or
+  when the user asks to add, change or refactor one. Java tools are covered
+  by kindgi-java-authoring-tools, Java guardrails by
+  kindgi-java-authoring-guardrails, connecting a real model by
+  kindgi-authoring-providers.
+type: core
+library: "kindgi-pack (Java)"
+version: "0.1.0"
+sdk_version: "0.1.5-rc.0"
+pack_languages: [java]
+sources:
+  - sdks/java/kindgi-pack/README.md
+  - sdks/java/kindgi-pack/src/main/java/com/kindgi/pack/Agent.java
+  - packages/specs/schemas/agent.schema.json
+  - packages/specs/schemas/pack-index.schema.json
+---
+
+# Authoring Kindgi agents in Java
+
+> **Running `kindgi`:** the pack pins its CLI (`"cli"` in
+> `kindgi.config.json`), and `./kindgiw` runs that version, so every
+> `kindgi <command>` below runs as `./kindgiw <command>`. Maven runs as
+> `./mvnw`.
+>
+> Java support is in preview: tested and supported, but the API may still
+> change in 0.1.6 without the usual deprecation period.
+
+An **agent** is a versioned, model-driven orchestrator. It's made of:
+- instructions (a prompt template);
+- the tools it may call;
+- the capabilities its model needs;
+- guardrails that gate its answer;
+- optionally, a conversation policy.
+
+In a Java pack it is **data**: a `public static final Agent` field of a class
+in an `agents` package. The model runs in the Kindgi runtime, not in your
+JVM. Your Java code runs only inside the agent's tools and guardrail checks.
+
+## Ask before building
+
+"Add an agent" is a conversation opener, not a ticket. Before writing a
+file, ask:
+
+- **What should the agent do?** The purpose drives everything else.
+- **Which tools does it need?** New ones, or existing ones?
+- **Multi-turn or one-shot?** History changes the shape.
+- **Any rules it must respect?** Those become guardrails.
+
+The pack's sample agent proves the runtime works end to end. It is not the
+shape to imitate unless the user asks for that.
+
+## An agent
+
+```java
+// src/main/java/acme/agents/BriefWriter.java
+package acme.agents;
+
+import acme.guardrails.ResponseNotEmpty;
+import acme.tools.Echo;
+import com.kindgi.pack.Agent;
+import java.util.List;
+import java.util.Map;
+
+/** acme.brief-writer: drafts a brief's argument from the case facts. */
+public final class BriefWriter {
+  /** The typed answer: the final message must be JSON of this shape. */
+  public record Brief(String argument, List<String> citations) {}
+
+  public static final Agent AGENT = Agent.define("acme.brief-writer")
+      .version("0.1.0")
+      .name("Brief Writer")
+      .description("Drafts appellate briefs from a case file; cites precedents.")
+      .instructions("You are drafting a brief in {{ jurisdiction }}. The user gives the case facts; you write "
+          + "a Section IV argument citing at least two precedents. Check every cite with the echo tool "
+          + "before using it. Never invent one.")
+      .capability(Map.of("needs", List.of(Map.of("feature", "tool-use"))))
+      .tool(Echo.TOOL)
+      .guardrail(ResponseNotEmpty.GUARDRAIL)
+      .set("parameters", List.of(Map.of("name", "jurisdiction", "type", "string", "required", true)))
+      .set("conversationPolicy", Map.of("historyLimit", 20))
+      .set("budget", Map.of("maxSteps", 8, "maxCostUsd", 0.5, "maxWallMs", 60_000))
+      .set("toolErrors", Map.of("maxRetries", 1, "retryOn", List.of("invalid-arguments", "unknown-tool")))
+      .output(Brief.class)
+      .build();
+
+  private BriefWriter() {}
+}
+```
+
+- Tools and guardrails are the pack's own fields, imported like any class:
+  `.tool(Echo.TOOL)`, `.guardrail(ResponseNotEmpty.GUARDRAIL)`.
+- **`build()` needs a version, a name and instructions.** Anything missing
+  is an error where the agent is defined, and the indexer reports it with
+  the file.
+- **Every other field is `set(field, value)`, keyed as on the wire:** the
+  field and its maps are camelCase (`conversationPolicy`, `maxSteps`,
+  `historyLimit`), as `agent.schema.json` names them. The indexer checks
+  each agent against the pack index's schema, and `kindgi dev` reports a
+  mistake with its file.
+- A class may hold several agents, each in its own `static final` field.
+
+## Field by field
+
+- **`id`:** `<pack-id>.<agent-name>`, kebab-case, dot-namespaced.
+- **`version`:** an exact semver. Conversations pin the version they started
+  on.
+- **`name`, `description`**, and `set("tags", List.of(…))`: for people and
+  listings. The model never sees the description.
+- **`instructions`:** a LiquidJS template. `{{ variable }}` comes from
+  `parameters` or the runtime's own variables (`today`, `now`, `agent.*`,
+  `conversation.*`). It's rendered strictly: an unknown variable fails the
+  turn. Write it as a brief for a capable colleague: what to do, which tools
+  to prefer, what to refuse, the quality bar. Name a tool by what it does
+  ("the verify-citation tool"), never by its dotted id. The model sees ids in
+  its provider's form (`acme__verify-citation` for Anthropic and
+  OpenAI-compatible models), and a dotted id in the instructions can make it
+  call a name it wasn't given. `instructions(Map.of("prompt", …, "version", …))`
+  references a registered prompt block instead.
+- **`capability(Map)`:** what the model must support, such as
+  `Map.of("needs", List.of(Map.of("feature", "tool-use")))`. Call it once per
+  capability. The turn routes its first capability to pick a provider and
+  model; with none declared, the turn fails.
+- **`tool(Tool)`:** pins that tool's version (its own, or the pack's).
+  **`tool(id, range)`** is a tool of another pack, with a semver **range**
+  (`"^1.0.0"`); the highest active matching version is picked at turn
+  start. An agent with no tools is chat-only.
+- **`guardrail(Guardrail)`** or **`guardrail(id)`:** evaluated once per turn
+  on the final answer, before it is stored. An id with no registered
+  guardrail fails the turn.
+- **`set("parameters", List.of(Map.of("name", …, "type", …, "required", …)))`:**
+  inputs the caller supplies per run. They fill `{{ … }}` in the
+  instructions.
+- **`set("preferredProvider", "anthropic")`, `set("preferredModel", "claude-haiku-4-5")`:**
+  soft hints. The router prefers them when they satisfy the capabilities.
+  To *require* a model, put it in the capability:
+  `Map.of("needs", List.of(Map.of("feature", "tool-use"), Map.of("models", Map.of("allow", List.of("claude-haiku-4-5")))))`.
+- **`set("conversationPolicy", …)`:** `historyLimit` caps the prior messages
+  loaded, and `hitl` configures approval gates. Absent, the turn loads the
+  full history with no gates. A tenant's `hitl` policy can tighten the gates
+  (a shorter timeout, a higher reviewer role, stricter per tool), never
+  loosen them.
+- **`set("budget", …)`:** per turn. `maxSteps` counts model calls (default
+  8); `maxCostUsd`; `maxWallMs` (default 120 000). Exceeding the steps or the
+  cost fails the turn (`budget-exceeded`); running out of wall time aborts
+  it. Leave room for real models: a turn with tool calls can take tens of
+  seconds.
+- **`output(Type.class)`:** a typed answer, with its schema derived from the
+  record. The final answer must be JSON matching it. A wrong one goes back
+  to the model with the problems (`maxRepairs`, default 1), and then the
+  turn fails (`output-schema-violation`). For `maxRepairs` or a schema of
+  your own, use `set("output", Map.of("schema", …, "maxRepairs", 2))`. The
+  parsed answer is the turn result's `output`. In a flow it is
+  `nodeOutputs.<step>.output.<field>`.
+- **`set("toolErrors", …)`:** `maxRetries` and `retryOn`. A failed tool call
+  goes back to the model as the call's result, so it can fix the call. The
+  default kinds are `invalid-arguments` and `unknown-tool` (nothing ran). Add
+  `tool-error` only when retrying the tool is safe. Each retry costs a step.
+
+## Which model answers
+
+Agents run on a registered model provider: the router picks one whose
+models satisfy the capabilities. `kindgi dev` gives a new pack `dev-echo`, a
+**fallback** that answers only while no other provider fits. It calls the
+first tool and replies "⚠ dev-echo isn't a real model: …", then "Tool
+responded: …". The turn carries the `fallback-provider` and
+`dev-echo-not-a-model` warnings. dev-echo can't fill in any other tool
+input, and can't give a typed answer: an agent with `output` fails with
+`output-schema-violation` until a real model is registered. To register one,
+see `kindgi-authoring-providers` (`./kindgiw providers register --preset=anthropic`,
+or `"providers": [{"preset": "anthropic"}]` in `kindgi.config.json`).
+
+## Iterating
+
+Save the file: `kindgi dev` recompiles, re-indexes, and the next run uses
+it, with no restart. Bump `version` when you break what callers rely on (a
+removed parameter, an incompatible output), not on every save.
+
+Run an agent from another terminal in the pack directory:
+
+```sh
+./kindgiw runs start --agent=acme.brief-writer --input='{"userMessage":"…","parameters":{"jurisdiction":"US"}}'
+```
+
+or from a Java app with the client (`kindgi-java-getting-started`).
+
+## Common mistakes
+
+1. **Building an agent without asking what it should do.** Copying the
+   sample's shape answers the wrong question.
+2. **No `capability(…)`.** The turn can't pick a model.
+3. **Java names inside the maps.** `set("budget", Map.of("max_steps", 4))`
+   isn't a budget: the keys are the wire's camelCase (`maxSteps`,
+   `historyLimit`, `maxRetries`).
+4. **An unregistered guardrail id.** Pass the `Guardrail` field, or make sure
+   the id is one the tenant has.
+5. **`{{ variable }}` not in `parameters`.** The turn fails when the
+   instructions render.
+6. **An agent that isn't a `static final` field.** Only static fields are
+   indexed.
+7. **A tight `maxWallMs` with a real model.** 15 s aborts real turns under
+   load; 60 s is a safer start.
+8. **Expecting dev-echo to give a typed answer.** It can't: register a real
+   model first.
+
+## When the framework itself is the problem
+
+If the bug is in Kindgi or kindgi-pack (the index dropping a field, a
+misleading error, the router picking the wrong model) and not in the pack's
+code, load `kindgi-framework-feedback` and file it with
+`./kindgiw feedback write`.