@vincemakes/kiso-runtime 0.50.1 → 0.51.0

package/dist/agent.js

@@ -15,13 +15,18 @@ import { resolveContinuationScope } from "./provider/manifest.js";
 import { assessProfileDrift, buildProfile, profilePath, readProfile, writeProfile } from "./profile.js";
 import { EventLog, ToolRegistry } from "@vincemakes/kiso-core";
 import { AgentSession } from "./session.js";
+import { snapshotCheckpointSections } from "./summarize.js";
 /** @deprecated the canonical name is `Agent` (root export, 1.1.0); this alias is removed in the next major. */
 export class AgentRuntime {
     #definition;
+    /** 0.51.0: the context policy sessions receive — `checkpointSections` validated and copied at construction (a host that mutates its array afterwards changes nothing); every other field stays the host's own reference, as before. */
+    #contextPolicy;
     #registry;
     #adapterPromise;
     constructor(definition) {
         this.#definition = definition;
+        const sections = snapshotCheckpointSections(definition.contextPolicy?.checkpointSections);
+        this.#contextPolicy = sections === undefined || definition.contextPolicy === undefined ? definition.contextPolicy : { ...definition.contextPolicy, checkpointSections: sections };
         this.#registry = new ToolRegistry();
         for (const tool of definition.tools)
             this.#registry.register(tool);
@@ -178,7 +183,7 @@ export class AgentRuntime {
             ...(this.#definition.maxTokens !== undefined ? { maxTokens: this.#definition.maxTokens } : {}),
             ...(this.#definition.temperature !== undefined ? { temperature: this.#definition.temperature } : {}),
             ...(this.#definition.microcompact !== undefined ? { microcompact: this.#definition.microcompact } : {}),
-            ...(this.#definition.contextPolicy !== undefined ? { contextPolicy: this.#definition.contextPolicy } : {}),
+            ...(this.#contextPolicy !== undefined ? { contextPolicy: this.#contextPolicy } : {}),
             ...(this.#definition.maxRetries !== undefined ? { maxRetries: this.#definition.maxRetries } : {}),
             ...(this.#definition.streamIdleMs !== undefined ? { streamIdleMs: this.#definition.streamIdleMs } : {}),
             ...(this.#definition.toolRules !== undefined ? { toolRules: this.#definition.toolRules } : {}),

package/dist/index.d.ts

@@ -24,6 +24,8 @@ export { AgentSession, AgentSession as Session } from "./session.js";
 export type { SideQueryOptions } from "./session.js";
 export { PoisonedSessionError, ResumeBlockedError } from "./session.js";
 export type { ApprovalRequest, CompactInfo, ContextPolicy, SessionConfig, SummarizeResult } from "./session.js";
+export { DEFAULT_CHECKPOINT_SECTIONS } from "./summarize.js";
+export type { CheckpointSection } from "./summarize.js";
 export { Run, RunClosedError } from "./run.js";
 export { SessionStore, StaleWriterError, StoreCorruptionError } from "./store.js";
 export type { Event, SessionMeta, StoreRecord } from "./store.js";

package/dist/index.js

@@ -23,6 +23,8 @@ export { AgentRuntime, AgentRuntime as Agent, createAgent } from "./agent.js";
 // session
 export { AgentSession, AgentSession as Session } from "./session.js";
 export { PoisonedSessionError, ResumeBlockedError } from "./session.js";
+// 0.51.0: host-defined checkpoint sections (the type and the coding outline as data; the prompt machinery stays behind ./internal)
+export { DEFAULT_CHECKPOINT_SECTIONS } from "./summarize.js";
 // run
 export { Run, RunClosedError } from "./run.js";
 // store

package/dist/provider/metadata.d.ts

@@ -76,6 +76,15 @@ export interface ReasoningCapabilities {
         readonly levels: readonly Exclude<ReasoningEffort, "default">[];
         readonly default: Exclude<ReasoningEffort, "default"> | null;
         readonly wire: string;
+        /** a route that SPELLS a level differently on the wire (Command Code
+         *  takes `off` for no reasoning, and refuses `none`): the level a
+         *  person picks stays the native name, the wire carries the route's
+         *  word. Absent = every level goes as named. */
+        readonly values?: Readonly<Partial<Record<Exclude<ReasoningEffort, "default">, string>>>;
+        /** the levels came from the USER's profile, not from a dated source:
+         *  a claim of theirs, shown as such (the profile's `contextWindow` is
+         *  the same kind of claim). Absent on every registry row. */
+        readonly declared?: true;
     } | null;
     /** invalid combinations the provider documents (e.g. thinking-disabled
      *  at certain efforts). */
@@ -120,12 +129,20 @@ export interface ModelMetadataEntry {
     readonly capabilitiesSource?: string | null;
     readonly pricing: ModelPricing | null;
 }
+/** Every row of the table, for the gate that holds the whole table to its
+ *  contract (shape, dates, sources, unique keys). Internal; read-only. */
+export declare function modelMetadataRows(): readonly ModelMetadataEntry[];
 /**
  * Look a model up. `endpoint` narrows: an entry WITH an endpoint only
  * matches when the caller's endpoint origin equals it; an entry without
  * one matches any endpoint. Unknown model → null, never a default.
  */
 export declare function lookupModelMetadata(model: string, endpoint?: string): ModelMetadataEntry | null;
+export declare function declareEffort(model: string, endpoint: string | undefined, effort: Omit<NonNullable<ReasoningCapabilities["effort"]>, "declared" | "wire"> & {
+    readonly wire?: string;
+}): void;
+/** For tests and a config reload: forget every declaration. */
+export declare function clearDeclaredEfforts(): void;
 /** CW-1: the aliases, for the gate that holds each one to a dated source
  *  and a registered target. */
 export declare function modelAliases(): Readonly<Record<string, {

package/dist/provider/metadata.js

@@ -18,6 +18,7 @@
  * Lives under runtime/internal — the curated root surface (44 names)
  * does not move; core is untouched.
  */
+import { OPENROUTER_ROWS } from "./openrouter-rows.js";
 /**
  * REG-1 — the OFF-PEAK rates, read from the vendor's pricing table on
  * 2026-09-17. The table carries TWO sets: off-peak and peak, the peak
@@ -63,12 +64,16 @@ const ANTHROPIC_PRICING_SOURCE = "https://platform.claude.com/docs/en/about-clau
 const ANTHROPIC_MODELS_SOURCE = "https://platform.claude.com/docs/en/models/overview";
 const ANTHROPIC_EFFORT_SOURCE = "https://platform.claude.com/docs/en/build-with-claude/effort";
 const ANTHROPIC_ASOF = "2026-09-07";
-const anthropicPricing = (inputPerM, outputPerM, cacheReadPerM, cacheWritePerM) => ({
+/** The 5.5 line (Opus 5.5, Sonnet 5.5, Haiku 5.5): the models overview, the
+ *  effort page, the thinking table and the pricing page, all read 2026-10-17. */
+const ANTHROPIC_ASOF_55 = "2026-10-17";
+const ANTHROPIC_THINKING_SOURCE = "https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting";
+const anthropicPricing = (inputPerM, outputPerM, cacheReadPerM, cacheWritePerM, asOf = ANTHROPIC_ASOF) => ({
     inputPerM,
     outputPerM,
     cacheReadPerM,
     cacheWritePerM,
-    asOf: ANTHROPIC_ASOF,
+    asOf,
     source: ANTHROPIC_PRICING_SOURCE,
 });
 /** The 5-line's effort control: five native levels, default high, wire
@@ -77,24 +82,24 @@ const anthropicPricing = (inputPerM, outputPerM, cacheReadPerM, cacheWritePerM)
  *  gives the modes per model. */
 const FIVE_LEVELS = { levels: ["low", "medium", "high", "xhigh", "max"], default: "high", wire: "output_config.effort" };
 function anthropicLine() {
-    const row = (model, capabilities, pricing) => ({
+    const row = (model, capabilities, pricing, asOf = ANTHROPIC_ASOF) => ({
         model,
         providerId: "anthropic",
         capabilities,
-        capabilitiesAsOf: ANTHROPIC_ASOF,
+        capabilitiesAsOf: asOf,
         capabilitiesSource: ANTHROPIC_MODELS_SOURCE,
         pricing,
     });
-    const fiveLine = (thinkingModes, forbidden) => ({
+    const fiveLine = (thinkingModes, forbidden, opts = {}) => ({
         contextWindow: 1_000_000,
         maxOutputTokens: 128_000,
         promptCaching: "explicit",
         reasoning: {
             emitsThinkingStream: true,
             thinking: { modes: thinkingModes, default: "adaptive" },
-            effort: FIVE_LEVELS,
+            effort: opts.effortDefault === undefined ? FIVE_LEVELS : { ...FIVE_LEVELS, default: opts.effortDefault },
             ...(forbidden !== undefined ? { forbidden } : {}),
-            asOf: ANTHROPIC_ASOF,
+            asOf: opts.asOf ?? ANTHROPIC_ASOF,
             source: ANTHROPIC_EFFORT_SOURCE,
         },
         inputModalities: ["text", "image"],
@@ -121,6 +126,30 @@ function anthropicLine() {
         // Haiku 4.5: the pinned id and its alias (the registry matches exact ids).
         row("claude-haiku-4-5-20251001", haiku, haikuPricing),
         row("claude-haiku-4-5", haiku, haikuPricing),
+        // The 5.5 line (read 2026-10-17; every model id is its own pinned
+        // snapshot, so there is no dated alias). All three: 1M window, 128K
+        // output, five native effort levels, wire output_config.effort.
+        //
+        // Opus 5.5: adaptive ALWAYS on ("disabled" is a 400 at every level);
+        // the default effort is MEDIUM, one below Opus 5's. Cache hits cost 5%
+        // of input.
+        row("claude-opus-5-5", fiveLine(["adaptive"], undefined, { effortDefault: "medium", asOf: ANTHROPIC_ASOF_55 }), anthropicPricing(4, 20, 0.2, 5, ANTHROPIC_ASOF_55), ANTHROPIC_ASOF_55),
+        // Sonnet 5.5: adaptive, on by default. "disabled" is a 400 at every
+        // level; the vendor's own off switch is the new thinking type
+        // `between_tools` (accepted only here, only at effort high or below),
+        // which the registry's thinking vocabulary does not have, so it is
+        // not offered. Default effort high. Cache hits cost 5% of input.
+        row("claude-sonnet-5-5", fiveLine(["adaptive"], undefined, { asOf: ANTHROPIC_ASOF_55 }), anthropicPricing(2, 10, 0.1, 2.5, ANTHROPIC_ASOF_55), ANTHROPIC_ASOF_55),
+        // Haiku 5.5: adaptive, on by default; "disabled" accepted at effort
+        // high or below, a 400 with xhigh or max (the forbidden pair). Default
+        // effort medium. PRICE: the vendor prices by PROMPT LENGTH. The row
+        // carries the rates for prompts up to 100,000 tokens; a prompt over
+        // that pays $0.50 in / $2.50 out / $0.05 cache read / $0.625 cache
+        // write, five times the row. ModelPricing has one set of rates, so a
+        // bill computed from this row for a long prompt is a fifth of the
+        // truth: the one-rate-against-two-rates shape the DeepSeek peak
+        // rates already have.
+        row("claude-haiku-5-5", fiveLine(["adaptive", "disabled"], [{ thinking: "disabled", effort: "xhigh" }, { thinking: "disabled", effort: "max" }], { effortDefault: "medium", asOf: ANTHROPIC_ASOF_55 }), anthropicPricing(0.1, 0.5, 0.01, 0.125, ANTHROPIC_ASOF_55), ANTHROPIC_ASOF_55),
     ];
 }
 const OPENAI_MODELS_ASOF = "2026-09-08";
@@ -141,7 +170,7 @@ const openaiRow = (model, effortDefault, rate, page, opts = {}) => ({
         }, inputModalities: null },
     capabilitiesAsOf: opts.asOf ?? OPENAI_MODELS_ASOF,
     capabilitiesSource: page,
-    pricing: { ...rate, cacheWritePerM: 0, asOf: opts.asOf ?? OPENAI_MODELS_ASOF, source: page },
+    pricing: { ...rate, cacheWritePerM: rate.cacheWritePerM ?? 0, asOf: opts.asOf ?? OPENAI_MODELS_ASOF, source: page },
 });
 /** The subscription backend's row for the same id: the presets' four
  *  levels (no `none`), the presets' context window, no price, and a cache
@@ -171,12 +200,24 @@ const ULTRA_LADDER = ["low", "medium", "high", "xhigh", "max", "ultra"];
  *  The owner's `sol` profile moved to `gpt-6.1-sol` on 2026-10-08, and with
  *  no row kiso sent it no effort: the remembered `high` never reached the
  *  wire and no reasoning summary came back. The id answered on the
- *  subscription backend that night (the 0.47.3 live sessions). No
- *  first-party row: no model page was read for it. */
+ *  subscription backend that night (the 0.47.3 live sessions). The
+ *  first-party row was added 2026-10-17 (OPENAI_MODELS_ASOF_4). */
 const OPENAI_MODELS_ASOF_3 = "2026-10-09";
 const CHATGPT_PRESETS_SOURCE_3 = "https://github.com/openai/codex/blob/d63a9b8344cfe58bc78bbe319b560378fc8756ef/codex-rs/models-manager/models.json";
+/** The first-party pages for the GPT-6.1 / GPT-6 sol and luna rows, read
+ *  2026-10-17 (the model page is the source for the levels, the window, the
+ *  output cap and the price alike, as for every first-party row). */
+const OPENAI_MODELS_ASOF_4 = "2026-10-17";
 /** The v1 table. Nulls outnumber numbers ON PURPOSE: only values with a
  *  named source enter; everything else waits for one. */
+/** Command Code: what the gateway ACCEPTS (probe, 2026-10-17) intersected with
+ *  each model's native levels; `none` is spelled `off` on this route. */
+const COMMAND_CODE_EFFORT_SOURCE = "probe 2026-10-17: POST https://api.commandcode.ai/provider/v1/chat/completions with a sentinel reasoning_effort names the legal set; native levels per model: the first-party deepseek-flash row and https://openrouter.ai/api/v1/models";
+const COMMAND_CODE_EFFORT = {
+    "deepseek/deepseek-v4.1-flash": { levels: ["none", "low", "high", "max"], default: "high", wire: "reasoning_effort", values: { none: "off" } },
+    "z-ai/glm-5.3-flash": { levels: ["low", "high", "max"], default: null, wire: "reasoning_effort" },
+    "xai/grok-4.7": { levels: ["low", "medium", "high", "xhigh"], default: null, wire: "reasoning_effort" },
+};
 const ENTRIES = [
     {
         model: "deepseek-chat",
@@ -418,43 +459,98 @@ const ENTRIES = [
     chatgptRow("gpt-6-astra", { levels: ULTRA_LADDER, default: "low", source: CHATGPT_PRESETS_SOURCE_2, asOf: OPENAI_MODELS_ASOF_2 }),
     chatgptRow("gpt-5.6-sol", { levels: ULTRA_LADDER, default: "low", source: CHATGPT_PRESETS_SOURCE_2, asOf: OPENAI_MODELS_ASOF_2 }),
     chatgptRow("gpt-6.1-sol", { levels: ULTRA_LADDER, default: "low", source: CHATGPT_PRESETS_SOURCE_3, asOf: OPENAI_MODELS_ASOF_3 }),
-    // The compat table's SECOND row (roadmap R3.2 §1a step 1): GLM 5.3 Flash
-    // through OpenRouter, read from OpenRouter's models API on 2026-09-09. The
-    // id is OpenRouter's, so the row is keyed to that origin; the
-    // continuation scope stamps such a profile `custom` at openrouter.ai
-    // (manifest.ts's KNOWN_ORIGINS is deliberately not widened — the scope
-    // identity is a durable fact and this row needs only the lookup). The
-    // window is the TOP provider's (1,048,576), not the listing's 1,310,720:
-    // what a request actually gets. No default effort — the page states
-    // none. `reasoning_effort` is in the model's supported parameters; the
-    // stream's think arrives as `reasoning` (GLM-F1, the compat adapter).
-    {
-        model: "z-ai/glm-5.3-flash",
-        endpoint: "https://openrouter.ai",
+    // 2026-10-17: the first-party rows the 0.48.0 note said were missing
+    // ("no model page was read for it"). gpt-6.1-sol: five levels, and the
+    // page says `none` and `minimal` are not supported; default medium.
+    // gpt-6-sol and gpt-6-luna: `none` through `max`, default medium.
+    // PRICE: each page also lists a LONG-PROMPT tier: a prompt over 272,000
+    // input tokens pays 2x the input and cache rates and 1.5x the output
+    // rate for the whole request. The rows carry the standard rates; a bill
+    // computed from them for a long prompt is low (the one-rate-against-
+    // two-rates shape the DeepSeek peak rates already have). Cache writes
+    // are priced on these pages and carried (sol $2.50, luna $0.125).
+    openaiRow("gpt-6.1-sol", "medium", { inputPerM: 2, outputPerM: 10, cacheReadPerM: 0.1, cacheWritePerM: 2.5 }, "https://developers.openai.com/api/docs/models/gpt-6.1-sol", { levels: ["low", "medium", "high", "xhigh", "max"], asOf: OPENAI_MODELS_ASOF_4 }),
+    openaiRow("gpt-6-sol", "medium", { inputPerM: 2, outputPerM: 10, cacheReadPerM: 0.2, cacheWritePerM: 2.5 }, "https://developers.openai.com/api/docs/models/gpt-6-sol", { levels: ["none", "low", "medium", "high", "xhigh", "max"], asOf: OPENAI_MODELS_ASOF_4 }),
+    openaiRow("gpt-6-luna", "medium", { inputPerM: 0.1, outputPerM: 0.5, cacheReadPerM: 0.01, cacheWritePerM: 0.125 }, "https://developers.openai.com/api/docs/models/gpt-6-luna", { levels: ["none", "low", "medium", "high", "xhigh", "max"], asOf: OPENAI_MODELS_ASOF_4 }),
+    // The OpenRouter route (origin https://openrouter.ai): generated from
+    // OpenRouter's own model list by scripts/registry-or-rows.mjs and kept in
+    // openrouter-rows.ts, dated 2026-10-17. It replaces the hand-written
+    // z-ai/glm-5.3-flash row of 2026-09-09 (roadmap R3.2 §1a step 1), which
+    // the list has since moved past: the efforts are now max / high / low.
+    ...OPENROUTER_ROWS,
+    // The Command Code gateway (origin https://api.commandcode.ai), read by a
+    // PROBE on 2026-10-17: a chat completion with a nonsense reasoning_effort
+    // and max_tokens 1 is refused before generation and the 400 names the
+    // legal set, `off | low | medium | high | xhigh | max`. That enum belongs
+    // to the gateway, not to a model: the same list came back for every model
+    // probed. So a row here offers the INTERSECTION of that set with the
+    // model's own native levels. `none` goes on the wire as `off`: the gateway
+    // refuses `none` (400, probed the same day) and accepts `off` (200). Each
+    // model's native levels: deepseek from the first-party deepseek-flash row
+    // (live-probed 2026-09-14); glm and grok from OpenRouter's list, which is
+    // a router's statement and is named as such in `source`. Mimo has no
+    // stated efforts anywhere, so it has no row: unknown stays unknown. No
+    // default, window, output cap or price is claimed here: the window falls
+    // to the model's identity row, and the gateway's plan has no per-token
+    // price.
+    ...["deepseek/deepseek-v4.1-flash", "z-ai/glm-5.3-flash", "xai/grok-4.7"].map((model) => ({
+        model,
+        endpoint: "https://api.commandcode.ai",
         capabilities: {
-            contextWindow: 1_048_576,
-            maxOutputTokens: 131_072,
-            promptCaching: "automatic",
+            contextWindow: null,
+            maxOutputTokens: null,
+            promptCaching: null,
             reasoning: {
-                emitsThinkingStream: true,
+                emitsThinkingStream: null,
                 thinking: null,
-                effort: { levels: ["low", "medium", "high"], default: null, wire: "reasoning_effort" },
-                asOf: "2026-09-09",
-                source: "https://openrouter.ai/docs/use-cases/reasoning-tokens",
+                effort: COMMAND_CODE_EFFORT[model],
+                asOf: "2026-10-17",
+                source: COMMAND_CODE_EFFORT_SOURCE,
             },
             inputModalities: null,
         },
-        capabilitiesAsOf: "2026-09-09",
-        capabilitiesSource: "https://openrouter.ai/api/v1/models",
-        pricing: { inputPerM: 0.075, outputPerM: 0.25, cacheReadPerM: 0.015, cacheWritePerM: 0, asOf: "2026-09-09", source: "https://openrouter.ai/api/v1/models" },
-    },
+        capabilitiesAsOf: "2026-10-17",
+        capabilitiesSource: COMMAND_CODE_EFFORT_SOURCE,
+        pricing: null,
+    })),
 ];
+/** Every row of the table, for the gate that holds the whole table to its
+ *  contract (shape, dates, sources, unique keys). Internal; read-only. */
+export function modelMetadataRows() {
+    return ENTRIES;
+}
 /**
  * Look a model up. `endpoint` narrows: an entry WITH an endpoint only
  * matches when the caller's endpoint origin equals it; an entry without
  * one matches any endpoint. Unknown model → null, never a default.
  */
 export function lookupModelMetadata(model, endpoint) {
+    const base = lookupRegistryRow(model, endpoint);
+    const declared = DECLARED_EFFORT.get(declarationKey(model, endpoint));
+    if (declared === undefined)
+        return base;
+    // a profile's declaration wins over the registry's effort for that
+    // (model, endpoint); the rest of the row (window, price, thinking modes)
+    // stays the registry's, and a model the registry does not know becomes a
+    // row that claims nothing but the declared levels
+    const reasoning = {
+        emitsThinkingStream: base?.capabilities.reasoning?.emitsThinkingStream ?? null,
+        thinking: base?.capabilities.reasoning?.thinking ?? null,
+        effort: declared,
+        ...(base?.capabilities.reasoning?.forbidden !== undefined ? { forbidden: base.capabilities.reasoning.forbidden } : {}),
+        asOf: null,
+        source: "declared in the profile",
+    };
+    if (base !== null)
+        return { ...base, capabilities: { ...base.capabilities, reasoning } };
+    return {
+        model,
+        ...(endpoint !== undefined ? { endpoint: originOf(endpoint) } : {}),
+        capabilities: { contextWindow: null, maxOutputTokens: null, promptCaching: null, reasoning, inputModalities: null },
+        pricing: null,
+    };
+}
+function lookupRegistryRow(model, endpoint) {
     for (const entry of ENTRIES) {
         if (entry.model !== model)
             continue;
@@ -464,6 +560,18 @@ export function lookupModelMetadata(model, endpoint) {
     }
     return null;
 }
+/** A profile's own effort levels (the CLI reads them from the user's config
+ *  at startup). Keyed by (model, endpoint origin); a later declaration for
+ *  the same key replaces the earlier. Process-wide, like the table. */
+const DECLARED_EFFORT = new Map();
+const declarationKey = (model, endpoint) => `${model}@${endpoint === undefined ? "" : originOf(endpoint)}`;
+export function declareEffort(model, endpoint, effort) {
+    DECLARED_EFFORT.set(declarationKey(model, endpoint), { ...effort, wire: effort.wire ?? "reasoning_effort", declared: true });
+}
+/** For tests and a config reload: forget every declaration. */
+export function clearDeclaredEfforts() {
+    DECLARED_EFFORT.clear();
+}
 /**
  * CW-1 — names one set of weights answers to across routes, each a dated
  * vendor statement. The key is the FOLDED name a route serves (vendor
@@ -473,6 +581,11 @@ const MODEL_ALIASES = {
     // The pricing table: `deepseek-flash` IS DeepSeek-V4.1-Flash. Gateways
     // serve it by the version name (`deepseek-v4.1-flash`).
     "deepseek-v4.1-flash": { is: "deepseek-flash", asOf: "2026-09-17", source: "https://api-docs.deepseek.com/quick_start/pricing/" },
+    // OpenRouter spells the 5.5 line with a dot; the vendor's ids carry dashes
+    // (the models overview, read 2026-10-17). Same weights.
+    "claude-opus-5.5": { is: "claude-opus-5-5", asOf: "2026-10-17", source: "https://platform.claude.com/docs/en/models/overview" },
+    "claude-sonnet-5.5": { is: "claude-sonnet-5-5", asOf: "2026-10-17", source: "https://platform.claude.com/docs/en/models/overview" },
+    "claude-haiku-5.5": { is: "claude-haiku-5-5", asOf: "2026-10-17", source: "https://platform.claude.com/docs/en/models/overview" },
 };
 /** CW-1: the aliases, for the gate that holds each one to a dated source
  *  and a registered target. */
@@ -565,7 +678,7 @@ export function resolveReasoning(model, setting, endpoint) {
             const levels = r.effort === null ? "none known" : r.effort.levels.join("/");
             return { ok: false, reason: `${model} does not support effort "${setting.effort}" (native: ${levels})` };
         }
-        wire.effort = setting.effort;
+        wire.effort = r.effort.values?.[setting.effort] ?? setting.effort;
     }
     for (const f of r.forbidden ?? []) {
         if (f.thinking === setting.thinking && f.effort === setting.effort) {

package/dist/provider/openrouter-rows.d.ts

@@ -0,0 +1,7 @@
+/**
+ * GENERATED by scripts/registry-or-rows.mjs from https://openrouter.ai/api/v1/models on 2026-10-17.
+ * Do not edit by hand: run the script and commit the diff. The rules for what
+ * a row says (and does not) are in the script's header.
+ */
+import type { ModelMetadataEntry } from "./metadata.js";
+export declare const OPENROUTER_ROWS: readonly ModelMetadataEntry[];

package/dist/provider/openrouter-rows.js

@@ -0,0 +1,14 @@
+export const OPENROUTER_ROWS = [
+    { "model": "anthropic/claude-opus-5.5", "endpoint": "https://openrouter.ai", "capabilities": { "contextWindow": 1000000, "maxOutputTokens": 128000, "promptCaching": null, "reasoning": { "emitsThinkingStream": null, "thinking": null, "effort": { "levels": ["low", "medium", "high", "xhigh", "max"], "default": null, "wire": "reasoning_effort" }, "asOf": "2026-10-17", "source": "https://openrouter.ai/api/v1/models" }, "inputModalities": ["text", "image", "file"] }, "capabilitiesAsOf": "2026-10-17", "capabilitiesSource": "https://openrouter.ai/api/v1/models", "pricing": null },
+    { "model": "anthropic/claude-sonnet-5.5", "endpoint": "https://openrouter.ai", "capabilities": { "contextWindow": 1000000, "maxOutputTokens": 128000, "promptCaching": null, "reasoning": { "emitsThinkingStream": null, "thinking": null, "effort": { "levels": ["low", "medium", "high", "xhigh", "max"], "default": null, "wire": "reasoning_effort" }, "asOf": "2026-10-17", "source": "https://openrouter.ai/api/v1/models" }, "inputModalities": ["text", "image", "file"] }, "capabilitiesAsOf": "2026-10-17", "capabilitiesSource": "https://openrouter.ai/api/v1/models", "pricing": null },
+    { "model": "anthropic/claude-haiku-5.5", "endpoint": "https://openrouter.ai", "capabilities": { "contextWindow": 1000000, "maxOutputTokens": 128000, "promptCaching": null, "reasoning": { "emitsThinkingStream": null, "thinking": null, "effort": { "levels": ["none", "low", "medium", "high", "xhigh", "max"], "default": null, "wire": "reasoning_effort" }, "asOf": "2026-10-17", "source": "https://openrouter.ai/api/v1/models" }, "inputModalities": ["text", "image", "file"] }, "capabilitiesAsOf": "2026-10-17", "capabilitiesSource": "https://openrouter.ai/api/v1/models", "pricing": null },
+    { "model": "openai/gpt-6.1-sol", "endpoint": "https://openrouter.ai", "capabilities": { "contextWindow": 1050000, "maxOutputTokens": 128000, "promptCaching": null, "reasoning": { "emitsThinkingStream": null, "thinking": null, "effort": { "levels": ["low", "medium", "high", "xhigh", "max"], "default": null, "wire": "reasoning_effort" }, "asOf": "2026-10-17", "source": "https://openrouter.ai/api/v1/models" }, "inputModalities": ["file", "image", "text"] }, "capabilitiesAsOf": "2026-10-17", "capabilitiesSource": "https://openrouter.ai/api/v1/models", "pricing": null },
+    { "model": "openai/gpt-6-sol", "endpoint": "https://openrouter.ai", "capabilities": { "contextWindow": 1050000, "maxOutputTokens": 128000, "promptCaching": null, "reasoning": { "emitsThinkingStream": null, "thinking": null, "effort": { "levels": ["none", "low", "medium", "high", "xhigh", "max"], "default": null, "wire": "reasoning_effort" }, "asOf": "2026-10-17", "source": "https://openrouter.ai/api/v1/models" }, "inputModalities": ["file", "image", "text"] }, "capabilitiesAsOf": "2026-10-17", "capabilitiesSource": "https://openrouter.ai/api/v1/models", "pricing": null },
+    { "model": "openai/gpt-6-luna", "endpoint": "https://openrouter.ai", "capabilities": { "contextWindow": 1050000, "maxOutputTokens": 128000, "promptCaching": null, "reasoning": { "emitsThinkingStream": null, "thinking": null, "effort": { "levels": ["none", "low", "medium", "high", "xhigh", "max"], "default": null, "wire": "reasoning_effort" }, "asOf": "2026-10-17", "source": "https://openrouter.ai/api/v1/models" }, "inputModalities": ["file", "image", "text"] }, "capabilitiesAsOf": "2026-10-17", "capabilitiesSource": "https://openrouter.ai/api/v1/models", "pricing": null },
+    { "model": "deepseek/deepseek-v4.1-flash", "endpoint": "https://openrouter.ai", "capabilities": { "contextWindow": 1048576, "maxOutputTokens": null, "promptCaching": null, "reasoning": { "emitsThinkingStream": null, "thinking": null, "effort": { "levels": ["none", "low", "high", "max"], "default": null, "wire": "reasoning_effort" }, "asOf": "2026-10-17", "source": "https://openrouter.ai/api/v1/models" }, "inputModalities": ["text", "image"] }, "capabilitiesAsOf": "2026-10-17", "capabilitiesSource": "https://openrouter.ai/api/v1/models", "pricing": null },
+    { "model": "x-ai/grok-4.7", "endpoint": "https://openrouter.ai", "capabilities": { "contextWindow": 500000, "maxOutputTokens": null, "promptCaching": null, "reasoning": { "emitsThinkingStream": null, "thinking": null, "effort": { "levels": ["low", "medium", "high", "xhigh"], "default": null, "wire": "reasoning_effort" }, "asOf": "2026-10-17", "source": "https://openrouter.ai/api/v1/models" }, "inputModalities": ["text", "image", "file"] }, "capabilitiesAsOf": "2026-10-17", "capabilitiesSource": "https://openrouter.ai/api/v1/models", "pricing": null },
+    { "model": "z-ai/glm-5.3-flash", "endpoint": "https://openrouter.ai", "capabilities": { "contextWindow": 1048576, "maxOutputTokens": null, "promptCaching": "automatic", "reasoning": { "emitsThinkingStream": true, "thinking": null, "effort": { "levels": ["low", "high", "max"], "default": null, "wire": "reasoning_effort" }, "asOf": "2026-10-17", "source": "https://openrouter.ai/api/v1/models" }, "inputModalities": ["text", "image", "video"] }, "capabilitiesAsOf": "2026-10-17", "capabilitiesSource": "https://openrouter.ai/api/v1/models", "pricing": null },
+    { "model": "moonshotai/kimi-k3", "endpoint": "https://openrouter.ai", "capabilities": { "contextWindow": 1048576, "maxOutputTokens": null, "promptCaching": null, "reasoning": { "emitsThinkingStream": null, "thinking": null, "effort": { "levels": ["none", "low", "high", "max"], "default": null, "wire": "reasoning_effort" }, "asOf": "2026-10-17", "source": "https://openrouter.ai/api/v1/models" }, "inputModalities": ["text", "image", "video"] }, "capabilitiesAsOf": "2026-10-17", "capabilitiesSource": "https://openrouter.ai/api/v1/models", "pricing": null },
+    { "model": "xiaomi/mimo-v2.6-flash", "endpoint": "https://openrouter.ai", "capabilities": { "contextWindow": 1050000, "maxOutputTokens": 131072, "promptCaching": null, "reasoning": null, "inputModalities": ["text", "image", "video", "audio"] }, "capabilitiesAsOf": "2026-10-17", "capabilitiesSource": "https://openrouter.ai/api/v1/models", "pricing": null },
+    { "model": "google/gemini-4-argon", "endpoint": "https://openrouter.ai", "capabilities": { "contextWindow": 1048576, "maxOutputTokens": 65536, "promptCaching": null, "reasoning": { "emitsThinkingStream": null, "thinking": null, "effort": { "levels": ["low", "medium", "high"], "default": null, "wire": "reasoning_effort" }, "asOf": "2026-10-17", "source": "https://openrouter.ai/api/v1/models" }, "inputModalities": ["audio", "file", "image", "text", "video"] }, "capabilitiesAsOf": "2026-10-17", "capabilitiesSource": "https://openrouter.ai/api/v1/models", "pricing": null },
+];

package/dist/session.d.ts

@@ -31,6 +31,7 @@ import { EventLog, type AbortSignalLike, type Adapter, type Event, type KisoExte
 import { type OverflowMeasure } from "./overflow-belt.js";
 import { type ContinuationScope } from "@vincemakes/kiso-core";
 import { type ReasoningSetting } from "./provider/metadata.js";
+import type { CheckpointSection } from "./summarize.js";
 import { type SessionStore, type StoreRecord } from "./store.js";
 import { Run } from "./run.js";
 import { type TaskDeliveryOptions } from "./tasks/delivery.js";
@@ -432,6 +433,17 @@ export declare class AgentSession {
  * /compact affordance point.
  */
 export interface ContextPolicy {
+    /**
+     * 0.51.0 — the sections of every checkpoint this session writes (the run-start
+     * auto-summary, the in-run tiers, the manual compact). Static data: a list of
+     * `{ heading, instruction }`, rendered in place of the coding outline; EVERY
+     * listed heading must appear as its own `## <heading>` line or the checkpoint
+     * is rejected (the usual route: serialized retry, then the breaker). Absent:
+     * the coding outline of 0.50.0, byte for byte. Validated and copied when the
+     * agent or session is constructed; extend the default with
+     * `[...DEFAULT_CHECKPOINT_SECTIONS, { heading, instruction }]`.
+     */
+    readonly checkpointSections?: readonly CheckpointSection[];
     /**
      * A — the auto-summary: at run start, when the projected context
      * crosses the trigger AND enough uncovered rounds exist, the

package/dist/session.js

@@ -35,7 +35,7 @@ import { denialResult } from "@vincemakes/kiso-core";
 import { buildProfile, readProfile, writeProfile, writeSummary } from "./profile.js";
 import { summarizeEvents } from "./session-summary.js";
 import { lookupModelMetadata, resolveReasoning } from "./provider/metadata.js";
-import { DROP_PLACEHOLDER, estimateSummarySavings, KEEP_RECENT_ROUNDS, KEEP_TOKENS_DEFAULT, lastSummaryPoint, MAX_SUMMARY_FAILURES, policyTriggerFromWindow, serializeCovered, MANUAL_SUMMARY_BUDGET, SUMMARY_MAX_OUTPUT, summarizeConversation, SummaryBudgetExhausted, summaryBoundarySeq, } from "./summarize.js";
+import { DROP_PLACEHOLDER, estimateSummarySavings, KEEP_RECENT_ROUNDS, KEEP_TOKENS_DEFAULT, lastSummaryPoint, MAX_SUMMARY_FAILURES, policyTriggerFromWindow, serializeCovered, MANUAL_SUMMARY_BUDGET, SUMMARY_MAX_OUTPUT, resolveCheckpoint, snapshotCheckpointSections, summarizeConversation, SummaryBudgetExhausted, summaryBoundarySeq, } from "./summarize.js";
 import { canonicalizeUsageForModel } from "./usage/canonical.js";
 import { contextAnchor } from "./context-anchor.js";
 import { appendFileSync, mkdirSync } from "node:fs";
@@ -102,6 +102,8 @@ export class AgentSession {
     // constructor, setAdapter, and setModelBinding are the only writers).
     #adapter;
     #config;
+    /** 0.51.0: the checkpoint every summary call is built from — resolved ONCE at construction from a snapshot of `contextPolicy.checkpointSections`; the raw host array is never read again. */
+    #checkpoint;
     // PH-1a (finding PH-F8, P0): the LIVE model binding. #config froze the
     // startup model/provider, so a /model switch replaced the adapter while
     // every later run kept sending the OLD model id and canonicalizing
@@ -193,6 +195,10 @@ export class AgentSession {
         }
         const composedHooks = composeHooks(config.hooks, config.extensions ?? []);
         this.#config = composedHooks === undefined ? config : { ...config, hooks: composedHooks };
+        // 0.51.0: validate and snapshot the host's checkpoint sections here, at
+        // construction (a direct `new AgentSession` included); an invalid list throws.
+        const sections = snapshotCheckpointSections(config.contextPolicy?.checkpointSections);
+        this.#checkpoint = sections === undefined ? undefined : resolveCheckpoint(sections);
         this.#model = config.model;
         this.#provider = config.provider;
         this.#baseUrl = config.baseUrl;
@@ -415,6 +421,7 @@ export class AgentSession {
                 await this.#config.hooks?.onRetry?.(info, {});
             },
             ...(args.onProgress !== undefined ? { onProgress: args.onProgress } : {}),
+            ...(this.#checkpoint !== undefined ? { checkpoint: this.#checkpoint } : {}),
         };
         try {
             // The in-band call carries the run's whole context, so the overflow

package/dist/summarize.d.ts

@@ -159,8 +159,57 @@ export declare const DSML_MARKERS: readonly ["<tool_call", "<tool_use", "<invoke
  *  guards the trust-critical tail). */
 export declare const REQUIRED_SECTIONS: readonly ["## Current work", "## Next steps"];
 /** null = pass; an error string = reject. Empty text is the existing
- *  no-text rule's domain, reported here too (defense in depth). */
-export declare function validateSummary(text: string): string | null;
+ *  no-text rule's domain, reported here too (defense in depth).
+ *  `headings` (a host's `checkpointSections`, canonical form): when given,
+ *  EVERY listed heading must appear as its own line (`## <heading>`,
+ *  trimmed) instead of the two default tail sections matched by substring —
+ *  `## Story` is a substring of `## Storyboard`, so the custom path cannot
+ *  match that way. */
+export declare function validateSummary(text: string, headings?: readonly string[]): string | null;
+/**
+ * 0.51.0 — host-defined checkpoint sections (plan-checkpoint-sections rev 2.1;
+ * ADR-0055 Amendment). Kiso owns the mechanics of a checkpoint — when it is
+ * written, its boundary, its validation, its retries; the host owns what the
+ * checkpoint records. The list is static data: a callback could read time,
+ * state or spend, and two requests from one configuration would differ.
+ */
+export interface CheckpointSection {
+    /** One line, rendered as `## <heading>`. */
+    readonly heading: string;
+    /** What the section must hold; rendered under the heading. */
+    readonly instruction: string;
+}
+/** The seven sections of the coding outline, as data. Deep-frozen: a host
+ *  extends it with `[...DEFAULT_CHECKPOINT_SECTIONS, { heading, instruction }]`
+ *  and cannot rewrite the shared default. */
+export declare const DEFAULT_CHECKPOINT_SECTIONS: readonly CheckpointSection[];
+/** The kiso-owned rule the CUSTOM path adds: strict every-section validation
+ *  must not push the model to invent content for a section with nothing in it.
+ *  The default path does not carry it — its bytes stay 0.50.0's. */
+export declare const CHECKPOINT_NONE_RECORDED_RULE = "Every listed section must appear. If the covered conversation holds nothing for a section, write that none is recorded; never invent content to fill it.";
+/** The custom-path checkpoint prompt for `sections` (canonical form). */
+export declare function renderCheckpointPrompt(sections: readonly CheckpointSection[]): string;
+/** The rendered sections block's cap, UTF-8 bytes. */
+export declare const MAX_CHECKPOINT_SECTIONS_BYTES = 8192;
+/**
+ * Validate and snapshot a host's list: a deep copy, trimmed, frozen at both
+ * levels, or `undefined` for an absent list. Throws
+ * `Invalid contextPolicy.checkpointSections: …` — fail fast, never truncate.
+ * The snapshot is taken at construction (the runtime's, and a direct
+ * session's); the raw host array is never read again.
+ */
+export declare function snapshotCheckpointSections(input: readonly CheckpointSection[] | undefined): readonly CheckpointSection[] | undefined;
+/** What one summary call is built from: the system prompt of the serialized
+ *  form, the instruction of the in-band form, and the headings validation
+ *  requires (`null` = the default two-section substring rule). */
+export interface ResolvedCheckpoint {
+    readonly prompt: string;
+    readonly inBand: string;
+    readonly headings: readonly string[] | null;
+}
+/** Absent sections return the EXISTING constants (the path stays byte-identical
+ *  to 0.50.0), not a rebuild of them. `sections` must already be a snapshot. */
+export declare function resolveCheckpoint(sections?: readonly CheckpointSection[]): ResolvedCheckpoint;
 export interface SummarizeConversationOptions {
     readonly adapter: Adapter;
     readonly model: string;
@@ -196,6 +245,9 @@ export interface SummarizeConversationOptions {
      *  the final usage arrive. Observation only: a throw is swallowed, and
      *  the request is identical with or without it. */
     readonly onProgress?: (progress: SummaryProgress) => void;
+    /** 0.51.0: the checkpoint the call is built from (`resolveCheckpoint`).
+     *  Absent, the coding outline of 0.50.0, byte for byte. */
+    readonly checkpoint?: ResolvedCheckpoint;
 }
 /**
  * 0.40.0 — the summary call's progress against its output budget.

package/dist/summarize.js

@@ -243,8 +243,13 @@ export const DSML_MARKERS = ["<tool_call", "<tool_use", "<invoke", "tool_calls",
  *  guards the trust-critical tail). */
 export const REQUIRED_SECTIONS = ["## Current work", "## Next steps"];
 /** null = pass; an error string = reject. Empty text is the existing
- *  no-text rule's domain, reported here too (defense in depth). */
-export function validateSummary(text) {
+ *  no-text rule's domain, reported here too (defense in depth).
+ *  `headings` (a host's `checkpointSections`, canonical form): when given,
+ *  EVERY listed heading must appear as its own line (`## <heading>`,
+ *  trimmed) instead of the two default tail sections matched by substring —
+ *  `## Story` is a substring of `## Storyboard`, so the custom path cannot
+ *  match that way. */
+export function validateSummary(text, headings) {
     const trimmed = text.trim();
     if (trimmed === "")
         return "the summary is empty";
@@ -253,12 +258,117 @@ export function validateSummary(text) {
         if (lower.includes(marker))
             return `the summary carries a tool-call marker (${marker}) — reject`;
     }
+    if (headings !== undefined) {
+        const lines = new Set(trimmed.split(/\r?\n/).map((l) => l.trim()));
+        for (const h of headings) {
+            if (!lines.has(`## ${h}`))
+                return `the summary is missing the required section ## ${h} — a truncated or incomplete checkpoint`;
+        }
+        return null;
+    }
     for (const section of REQUIRED_SECTIONS) {
         if (!trimmed.includes(section))
             return `the summary is missing the required section ${section} — a truncated or incomplete checkpoint`;
     }
     return null;
 }
+/** The seven sections of the coding outline, as data. Deep-frozen: a host
+ *  extends it with `[...DEFAULT_CHECKPOINT_SECTIONS, { heading, instruction }]`
+ *  and cannot rewrite the shared default. */
+export const DEFAULT_CHECKPOINT_SECTIONS = Object.freeze([
+    { heading: "Goal", instruction: "The user's goal and the acceptance criterion, in one or two sentences." },
+    { heading: "Constraints", instruction: "The constraints, requirements, and rulings the work must honor." },
+    {
+        heading: "User requests",
+        instruction: `Every user request of the task — those an earlier checkpoint lists and
+those in the covered range — enumerated one by one, each with what it
+asked for and what was done about it.`,
+    },
+    {
+        heading: "Files and changes",
+        instruction: `Every file touched — exact paths, what changed, and why. Include the
+precise code-level changes later turns may need to continue.`,
+    },
+    {
+        heading: "Errors and fixes",
+        instruction: `Every error encountered and its resolution; commands run and their
+outcomes.`,
+    },
+    {
+        heading: "Current work",
+        instruction: `The current state of the work — what is done, what is not. Quote the
+current task's criterion VERBATIM if one exists.`,
+    },
+    { heading: "Next steps", instruction: "The concrete next steps, in order." },
+].map((x) => Object.freeze(x)));
+/** The kiso-owned rule the CUSTOM path adds: strict every-section validation
+ *  must not push the model to invent content for a section with nothing in it.
+ *  The default path does not carry it — its bytes stay 0.50.0's. */
+export const CHECKPOINT_NONE_RECORDED_RULE = "Every listed section must appear. If the covered conversation holds nothing for a section, write that none is recorded; never invent content to fill it.";
+const SECTIONS_LEAD = "Produce the checkpoint with exactly these sections, in this order:\n\n";
+// The frame is cut from the constant itself — the host's sections replace only
+// the block between these two marks, never a copy of the frame.
+const FRAME_HEAD = SUMMARY_PROMPT.slice(0, SUMMARY_PROMPT.indexOf(SECTIONS_LEAD) + SECTIONS_LEAD.length);
+const FRAME_TAIL = SUMMARY_PROMPT.slice(SUMMARY_PROMPT.indexOf("Preserve concrete identifiers VERBATIM"));
+const renderSectionBlocks = (sections) => sections.map((x) => `## ${x.heading}\n${x.instruction}`).join("\n\n");
+/** The custom-path checkpoint prompt for `sections` (canonical form). */
+export function renderCheckpointPrompt(sections) {
+    return `${FRAME_HEAD}${renderSectionBlocks(sections)}\n\n${CHECKPOINT_NONE_RECORDED_RULE}\n\n${FRAME_TAIL}`;
+}
+const MAX_SECTIONS = 16;
+const MAX_HEADING_CHARS = 80;
+/** The rendered sections block's cap, UTF-8 bytes. */
+export const MAX_CHECKPOINT_SECTIONS_BYTES = 8192;
+/**
+ * Validate and snapshot a host's list: a deep copy, trimmed, frozen at both
+ * levels, or `undefined` for an absent list. Throws
+ * `Invalid contextPolicy.checkpointSections: …` — fail fast, never truncate.
+ * The snapshot is taken at construction (the runtime's, and a direct
+ * session's); the raw host array is never read again.
+ */
+export function snapshotCheckpointSections(input) {
+    if (input === undefined)
+        return undefined;
+    const fail = (why) => {
+        throw new Error(`Invalid contextPolicy.checkpointSections: ${why}`);
+    };
+    if (!Array.isArray(input))
+        return fail("it must be an array");
+    if (input.length < 1 || input.length > MAX_SECTIONS)
+        return fail(`it must hold 1 to ${MAX_SECTIONS} sections, got ${input.length}`);
+    const seen = new Set();
+    const out = input.map((x, i) => {
+        const heading = typeof x?.heading === "string" ? x.heading.trim() : "";
+        const instruction = typeof x?.instruction === "string" ? x.instruction.trim() : "";
+        if (heading === "")
+            return fail(`section ${i + 1} has an empty heading`);
+        if (/[\r\n]/.test(heading))
+            return fail(`section ${i + 1}: the heading must be one line`);
+        if (heading.length > MAX_HEADING_CHARS)
+            return fail(`section ${i + 1}: the heading is ${heading.length} characters, the cap is ${MAX_HEADING_CHARS}`);
+        if (heading.startsWith("#"))
+            return fail(`section ${i + 1}: the heading must not start with "#" (kiso adds "## ")`);
+        const key = heading.toLowerCase();
+        if (seen.has(key))
+            return fail(`section ${i + 1}: the heading "${heading}" repeats an earlier one`);
+        seen.add(key);
+        if (instruction === "")
+            return fail(`section ${i + 1} ("${heading}") has an empty instruction`);
+        return Object.freeze({ heading, instruction });
+    });
+    const bytes = Buffer.byteLength(renderSectionBlocks(out), "utf8");
+    if (bytes > MAX_CHECKPOINT_SECTIONS_BYTES)
+        fail(`the rendered sections are ${bytes} bytes, the cap is ${MAX_CHECKPOINT_SECTIONS_BYTES}`);
+    return Object.freeze(out);
+}
+/** Absent sections return the EXISTING constants (the path stays byte-identical
+ *  to 0.50.0), not a rebuild of them. `sections` must already be a snapshot. */
+export function resolveCheckpoint(sections) {
+    if (sections === undefined)
+        return { prompt: SUMMARY_PROMPT, inBand: SUMMARY_IN_BAND, headings: null };
+    const prompt = renderCheckpointPrompt(sections);
+    return { prompt, inBand: `${SUMMARY_IN_BAND.slice(0, SUMMARY_IN_BAND.length - SUMMARY_PROMPT.length)}${prompt}`, headings: sections.map((x) => x.heading) };
+}
 /** A thrown value the adapter CLASSIFIED — `{ code, retryable, message }`
  *  (ADR-0005 rule 1). The core's own guard is private, and this asks a
  *  narrower question than that one: is this a failure someone already
@@ -385,12 +495,13 @@ async function summaryAttempt(options) {
     const budget = options.maxOutputTokens ?? null;
     report({ produced: 0, budget, reasoningUnseen: false, reported: false });
     const inBand = options.inBand;
+    const cp = options.checkpoint;
     for await (const ev of adapter.stream({
         model,
         ...(inBand === undefined
-            ? { messages, systemPrompt: SUMMARY_PROMPT }
+            ? { messages, systemPrompt: cp?.prompt ?? SUMMARY_PROMPT }
             : {
-                messages: [...messages, { role: "user", content: inBand.focus === undefined ? SUMMARY_IN_BAND : `${SUMMARY_IN_BAND}\n\nFocus the summary on: ${inBand.focus}` }],
+                messages: [...messages, { role: "user", content: inBand.focus === undefined ? (cp?.inBand ?? SUMMARY_IN_BAND) : `${cp?.inBand ?? SUMMARY_IN_BAND}\n\nFocus the summary on: ${inBand.focus}` }],
                 ...(inBand.systemPrompt !== undefined ? { systemPrompt: inBand.systemPrompt } : {}),
                 tools: inBand.tools,
             }),
@@ -453,7 +564,7 @@ async function summaryAttempt(options) {
     }
     // E6 (b): a non-checkpoint summary is an honest failure — throw, the
     // caller reports it, nothing is persisted (the auto-T5-1 regression).
-    const invalid = validateSummary(trimmed);
+    const invalid = validateSummary(trimmed, cp?.headings ?? undefined);
     if (invalid !== null) {
         throw new Error(`the summary call produced an invalid summary: ${invalid}`);
     }

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@vincemakes/kiso-runtime",
-	"version": "0.50.1",
+	"version": "0.51.0",
 	"description": "kiso runtime \u2014 durable multi-turn agent sessions: AgentDefinition, AgentRuntime, AgentSession, Run, append-only JSONL store.",
 	"type": "module",
 	"license": "MIT",
@@ -25,12 +25,12 @@
 		"test": "vitest run"
 	},
 	"dependencies": {
-		"@vincemakes/kiso-core": "0.50.1"
+		"@vincemakes/kiso-core": "0.51.0"
 	},
 	"peerDependencies": {
-		"@vincemakes/kiso-provider-anthropic": "0.50.1",
-		"@vincemakes/kiso-provider-openai": "0.50.1",
-		"@vincemakes/kiso-provider-openai-responses": "0.50.1"
+		"@vincemakes/kiso-provider-anthropic": "0.51.0",
+		"@vincemakes/kiso-provider-openai": "0.51.0",
+		"@vincemakes/kiso-provider-openai-responses": "0.51.0"
 	},
 	"peerDependenciesMeta": {
 		"@vincemakes/kiso-provider-anthropic": {
@@ -44,7 +44,7 @@
 		}
 	},
 	"devDependencies": {
-		"@vincemakes/kiso-evals": "0.50.1",
+		"@vincemakes/kiso-evals": "0.51.0",
 		"@types/node": "^26.1.2",
 		"typescript": "^5.7.2",
 		"vitest": "^3.0.0"