ytdb-slate 0.12.0 → 0.13.0

package/docs/configuration.md

@@ -71,7 +71,7 @@ The project must be trusted before Slate reads this file. Start a new session af
 | `doctrineExtraPath` | string | — | Markdown whose **content** is appended to the orchestrator doctrine (project-specific workflow additions). |
 | `reviewPerspectivesPath` | string | — | Review charters, each declaring its own finding-ID prefix. The doctrine references this **path**; the orchestrator reads the file alongside the shipped review rules. |
 | `router.models` | object | shipped six-model pool | Ordinary membership and definitions. `include` replaces the starting membership, including with an empty list. `add` accepts complete new definitions. `replace` changes selected fields. `exclude` applies last. Within each model definition, lists and provider maps replace shipped fields. The home and project configuration files merge first. |
-| `router.compressor.models` | array of `{ model, effort }` | `[{"model":"claude-sonnet-5","effort":"medium"}]` | Independent ordered compressor list. An explicit empty list blocks work. |
+| `router.compressor.models` | array of `{ model, effort }` | `[{"model":"claude-sonnet-5.5","effort":"medium"}]` | Independent ordered compressor list. An explicit empty list blocks work. |
 
 ## Extended example
 
@@ -91,7 +91,7 @@ The `docs/agents/...` values below are placeholders. Point them at Markdown file
   "reviewPerspectivesPath": "docs/agents/review-perspectives.md",
   "router": {
     "models": { "include": ["luna-6", "claude-opus-5.5"] },
-    "compressor": { "models": [{ "model": "claude-sonnet-5", "effort": "medium" }] }
+    "compressor": { "models": [{ "model": "claude-sonnet-5.5", "effort": "medium" }] }
   }
 }
 ```

package/docs/context-budget.md

@@ -170,9 +170,9 @@ Slate reports each refused prompt once. The orchestrator must use one state-save
 worker at a time. It must wait for the result and verify success. It must report
 incomplete preparation rather than claim that the research log was saved.
 
-Remote steer or follow-up, compaction-buffered input, and input accepted before
-the pause can bypass the hook. These are accepted limits of the pinned Pi input
-surface.
+The remote procedure call (RPC) commands `steer` and `follow_up` pass through the Pi input hook.
+Pi also applies the hook when it sends input buffered during compaction.
+Input accepted before the pause can still reach the agent.
 
 ## What the always-loaded doctrine costs the budget
 
@@ -196,28 +196,28 @@ characters. Tool names use 64 characters. Descriptions use 140 characters.
 
 | fixture | paths | portable characters | lines |
 | --- | ---: | ---: | ---: |
-| shipped logical section | 0 | 3,892 | 11 |
-| trusted shipped-default doctrine | 5 | 8,764 | 88 |
+| shipped logical section | 0 | 3,596 | 11 |
+| trusted shipped-default doctrine | 5 | 8,468 | 88 |
 | untrusted doctrine without a home configuration file | 4 | 2,775 | 47 |
-| draft pull requests, shipped policy | 6 | 8,839 | 89 |
-| capped workers, shipped policy | 5 | 10,111 | 98 |
-| draft plus capped workers | 6 | 10,186 | 99 |
-| deferred issues plus capped workers | 5 | 10,185 | 99 |
-| routing recommendations plus capped workers | 5 | 10,208 | 99 |
-| draft plus deferred issues and capped workers | 6 | 10,260 | 100 |
-| draft plus routing recommendations and capped workers | 6 | 10,279 | 100 |
-| deferred issues plus routing recommendations and capped workers | 5 | 10,282 | 100 |
-| canonical maximal baseline with all workflow options | 6 | 10,353 | 101 |
-| maximal baseline with an open change | 6 | 10,551 | 102 |
-| maximal baseline with source and legacy root log | 6 | 10,781 | 104 |
-| dogfood `.pi/slate.json` and its two extension units | 6 | 9,478 | 100 |
+| draft pull requests, shipped policy | 6 | 8,543 | 89 |
+| capped workers, shipped policy | 5 | 9,815 | 98 |
+| draft plus capped workers | 6 | 9,890 | 99 |
+| deferred issues plus capped workers | 5 | 9,889 | 99 |
+| routing recommendations plus capped workers | 5 | 9,912 | 99 |
+| draft plus deferred issues and capped workers | 6 | 9,964 | 100 |
+| draft plus routing recommendations and capped workers | 6 | 9,983 | 100 |
+| deferred issues plus routing recommendations and capped workers | 5 | 9,986 | 100 |
+| canonical maximal baseline with all workflow options | 6 | 10,057 | 101 |
+| maximal baseline with an open change | 6 | 10,255 | 102 |
+| maximal baseline with source and legacy root log | 6 | 10,485 | 104 |
+| dogfood `.pi/slate.json` and its two extension units | 6 | 9,182 | 100 |
 | runtime boundary plus draft and capped workers, source and legacy log | 6 | 26,122 | 102 |
 | runtime boundary plus draft and deferred issues, source and legacy log | 6 | 26,196 | 103 |
 | runtime boundary plus routing recommendations, source and legacy log | 5 | 26,144 | 102 |
 | runtime boundary plus deferred issues and routing recommendations, source and legacy log | 5 | 26,218 | 103 |
 | runtime boundary plus draft and routing recommendations, source and legacy log | 6 | 26,215 | 103 |
 | same boundary composition plus deferred issues | 6 | 26,289 | 104 |
-| valid shipped policy plus 88 capped worker tools, source and legacy log | 6 | 28,455 | 187 |
+| valid shipped policy plus 88 capped worker tools, source and legacy log | 6 | 28,159 | 187 |
 
 An untrusted session with an empty home `slate.json` receives the same doctrine
 as the trusted shipped-default fixture above. It includes routing and writing
@@ -228,7 +228,7 @@ project preferences can.
 The canonical maximal baseline uses all six shipped logical definitions, all
 three workflow options, and the fixed capped worker roster. Its largest variant
 has an open change, one direct source folder, and a legacy root log. Its
-five-percent requirement is `ceil(10,781 × 1.05) = 11,321`.
+five-percent requirement is `ceil(10,485 × 1.05) = 11,010`.
 The 26,300 whole-doctrine ceiling exceeds it. The table pins all eight
 combinations of the three workflow options with the capped worker roster. The
 dogfood row reads the actual project configuration. A change to `.pi/slate.json` requires a fresh render.
@@ -245,8 +245,8 @@ the first value beyond the ceiling.
 
 The over-cap counterfactual uses a valid shipped logical policy and one supported
 worker-extension unit with 88 capped tools. It also includes an open change,
-its source, and the legacy root log. Its 28,455 portable characters exceed
-the whole-doctrine ceiling by 2,155. The check therefore detects whole-doctrine
+its source, and the legacy root log. Its 28,159 portable characters exceed
+the whole-doctrine ceiling by 1,859. The check therefore detects whole-doctrine
 growth without using a retired model-row shape or an assumed per-tool increment.
 Worker-extension doctrine has no runtime size cap, so the fixture reaches the
 whole-doctrine guard rather than failing logical-policy validation first.
@@ -301,9 +301,9 @@ the compact JSON form after pi converts the reminder to a provider request user
 message with one text part. It excludes the volatile timestamp.
 
 The 319-byte figure is one compact persisted `custom_message` record from pi
-0.83.0, including its identifiers, timestamp, JSON framing, and final line-feed
-byte. These figures contain no installed path, so the portable and rendered
-counts are identical.
+1.0.0. It includes two eight-character identifiers, the timestamp, JSON framing
+and the final line-feed byte. These figures contain no installed path, so the
+portable and rendered counts are identical.
 
 Before automatic compaction, transcript storage grows in proportion to the number
 of turns. Cumulative request appearances grow with the square of the number of

package/docs/model-routing.md

@@ -69,11 +69,11 @@ forecasts. Membership changes do not rescale them.
 | logical model | ordinary effort | capability | cost | preferred provider | exact initial permission | guidance | cautions |
 | --- | --- | ---: | ---: | --- | --- | --- | --- |
 | `luna-6` | `max` | 45 | 5 | `openai` | `openai/gpt-6-luna` | Auxiliary tasks only, such as file location, check-result collection, or routine text management such as modifying research logs. Never primary research, implementation, design, or review. Consumer-contract work only when auxiliary. | scope and complex-text cautions below |
-| `claude-sonnet-5` | `high` | 40 | 90 | `anthropic` | `anthropic/claude-sonnet-5` | none | scope caution below |
-| `sol-6` | `high` | 58 | 20 | `openai` | `openai/gpt-6-sol` | Default thread choice. | blocked-action caution below |
+| `claude-sonnet-5.5` | `high` | 45 | 54 | `anthropic` | `anthropic/claude-sonnet-5-5` | none | scope caution below |
+| `sol-6.1` | `high` | 85 | 20 | `openai` | `openai/gpt-6.1-sol` | Default thread choice. | blocked-action caution below |
 | `gemini-3.8-flash` | `medium` | 55 | 30 | `google-vertex` | `google-vertex/gemini-3.8-flash` | Default thread choice. Generally prefer over Luna when available. A more specific active guideline overrides this general Flash preference. | reviewer and evidence caution below |
-| `claude-opus-5.5` | `high` | 90 | 65 | `anthropic` | `anthropic/claude-opus-5-5` | Concurrency, data-loss, and performance work. Prefer over Astra for code reviews of focus areas that require high-level design, except non-local logic. Do not select Opus 5.5 as the default implementer. If a lower-capability model repeatedly fails at implementation, first ask Opus 5.5 to investigate and provide detailed repair instructions. Let the implementer try those instructions. Use Opus 5.5 as the implementer only if that guided attempt also fails. Treat that use as an exception. Select another suitable model for later implementation work. Existing approval requirements and repair limits still apply. | scope caution below |
-| `gpt-6-astra` | `medium` | 86 | 60 | `openai` | `openai/gpt-6-astra` | Prefer when available for design reviews of all focus areas that require high-level design. Prefer for code reviews of non-local logic defects. Security and performance work. Do not select Astra as the default implementer. Use Astra for review only when assigned to a specific focus area. Use Astra for research when appropriate. | none |
+| `claude-opus-5.5` | `high` | 90 | 65 | `anthropic` | `anthropic/claude-opus-5-5` | Reviewer for each proved DESIGN-TRIGGERING focus area, in adversarial design review and in implementation area review. Do not select Opus 5.5 as the default implementer. If a lower-capability model repeatedly fails at implementation, first ask Opus 5.5 to investigate and provide detailed repair instructions. Let the implementer try those instructions. Use Opus 5.5 as the implementer only if that guided attempt also fails. Treat that use as an exception. Select another suitable model for later implementation work. Existing approval requirements and repair limits still apply. | scope caution below |
+| `gpt-6-astra` | `medium` | 86 | 60 | `openai` | `openai/gpt-6-astra` | Do not select Astra as the default implementer. Prefer Sol 6.1 over Astra. | none |
 
 Sonnet and Opus use this caution: `May exceed explicit scope or infer permission
 from earlier requests. Check changes against stated exclusions and approval
@@ -89,14 +89,15 @@ behavior.` Fable is not active because it has no complete
 approved definition.
 
 The independent shipped compressor list contains one entry. It is
-`claude-sonnet-5` at `medium`. Ordinary Sonnet remains at `high`.
+`claude-sonnet-5.5` at `medium`. Ordinary Sonnet remains at `high`.
 
-The ratings are project judgments informed by DeepSWE v1.1 by DataCurve and
-reviewed supporting evidence. The live source is
+The Sol 6.1 and Sonnet 5.5 capability and cost ratings are project judgments
+set on 2026-09-30 without new benchmark evidence. Other ratings are project
+judgments informed by DeepSWE v1.1 by DataCurve and reviewed supporting evidence.
+The source for those other ratings is
 <https://deepswe.datacurve.ai/artifacts/v1.1/leaderboard-live.json>. The project
-retrieved it on 2026-09-11. The 2026-09-18 rating decision reused that evidence
-without a refresh. The live source has no confirmed licence bridge for every
-newer result. Slate publishes no raw task, trajectory, leaderboard row, copied
+retrieved it on 2026-09-11. Those ratings use that evidence without a later
+refresh. The live source has no confirmed licence bridge for every newer result. Slate publishes no raw task, trajectory, leaderboard row, copied
 publisher prose, or copied table structure. This attribution is not a new
 licence grant.
 
@@ -121,13 +122,13 @@ new session.
 {
   "router": {
     "models": {
-      "include": ["luna-6", "claude-sonnet-5", "gpt-6-astra"],
+      "include": ["luna-6", "claude-sonnet-5.5", "gpt-6-astra"],
       "add": [],
       "replace": [],
       "exclude": []
     },
     "compressor": {
-      "models": [{ "model": "claude-sonnet-5", "effort": "medium" }]
+      "models": [{ "model": "claude-sonnet-5.5", "effort": "medium" }]
     }
   }
 }
@@ -198,12 +199,12 @@ Complete `replace` example:
     "models": {
       "replace": [
         {
-          "model": "sol-6",
+          "model": "sol-6.1",
           "effort": "max",
-          "capabilityRating": 58,
+          "capabilityRating": 85,
           "costRating": 20,
           "preferredProvider": "gateway",
-          "providers": { "gateway": "openai/gpt-6-sol" },
+          "providers": { "gateway": "openai/gpt-6.1-sol" },
           "guidelines": ["repository-wide implementation"],
           "cautions": []
         }
@@ -226,9 +227,11 @@ block logical work before dispatch. Slate reports each critical error. Legacy
 `router.allowUnmeasuredEffort`, and `router.showWarnings` values are named and
 ignored. Slate performs no automatic migration.
 
-The shipped names `gpt-5.6-luna`, `gpt-5.6-sol`, `gpt-5.6-terra`, and
-`claude-opus-5` are withdrawn. Use `luna-6`, `sol-6`, and
-`claude-opus-5.5` in `include`, `replace`, `exclude`, and compressor entries.
+The shipped names `gpt-5.6-luna`, `gpt-5.6-sol`, `gpt-5.6-terra`,
+`claude-opus-5`, `sol-6`, and `claude-sonnet-5` are withdrawn. Use `luna-6`
+for `gpt-5.6-luna`. Use `sol-6.1` for `gpt-5.6-sol` and `sol-6`. Use
+`claude-opus-5.5` for `claude-opus-5` and `claude-sonnet-5.5` for
+`claude-sonnet-5` in `include`, `replace`, `exclude`, and compressor entries.
 Terra has no replacement. Slate provides no automatic alias. A withdrawn name
 in one of those reference positions produces that position's explicit unknown-model
 error and blocks routing. A project can still define a withdrawn spelling as a
@@ -331,11 +334,11 @@ for an unknown or ambiguous route. Main recovery stays available during handoff
 pause. An unresolved stop keeps worker dispatch paused while brief writing waits
 for the user choice.
 
-Pi 0.85.1 model and effort switches are session-only unless the caller explicitly
+Pi 1.0.0 model and effort switches are session-only unless the caller explicitly
 requests persistence. Slate retains shared saved-default ownership and its
-compatibility restoration guard. The guard still protects explicit persistence
-and older behavior. A restoration failure is visible and can leave the changed
-default in place.
+compatibility restoration guard. The guard covers explicit persistence and
+hosts that persist switches. A restoration failure is visible and can leave the
+changed default in place.
 
 ## Package reliance boundary
 

package/extension/handoff.ts

@@ -556,18 +556,18 @@ export function registerSlateHandoff(
 				let adopted = false;
 				try {
 				// Keep the whole adoption block inside the compatibility restore guard.
-				// Pi 0.85.1 extension setters are session-only, so this is normally a
-				// zero-write no-op. The guard still covers older or explicitly persistent
-				// host behavior, including a thinking-level-only write.
+				// Pi 1.0.0 extension setters are session-only, so this is normally a
+				// zero-write no-op. The guard covers hosts that persist switches,
+				// including a thinking-level-only write.
 				await withGlobalModelDefaultRestored(
 					pi,
 					ctx,
 					getConfig(),
 					{ provider, id },
 					async () => {
-						// True once a Pi setter has been called. Pi 0.85.1 does not persist
-						// this call, but the compatibility guard must conservatively cover hosts
-						// where the same call can write a default.
+						// True once a Pi setter has been called. Pi 1.0.0 does not persist
+						// this call. The compatibility guard also covers hosts where the call
+						// can write a default.
 						let calledSetter = false;
 						try {
 							// Equality guard: avoid a recorded session switch when the fresh

package/extension/logical-model-definitions.ts

@@ -1,12 +1,16 @@
 export const LOGICAL_MODEL_EFFORTS = ["off", "minimal", "low", "medium", "high", "xhigh", "max"] as const;
 export type LogicalModelEffort = (typeof LOGICAL_MODEL_EFFORTS)[number];
 
-export interface LogicalFieldSource {
+export type LogicalFieldSource = {
 	publisher: "DeepSWE/DataCurve";
 	retrieved: "2026-09-11";
 	sourceUrl: "https://deepswe.datacurve.ai/artifacts/v1.1/leaderboard-live.json";
 	basis: "Project judgment informed by DeepSWE v1.1 and reviewed supporting evidence";
-}
+} | {
+	publisher: "Slate project";
+	decided: "2026-09-30";
+	basis: "Project judgment set on 2026-09-30 without new benchmark evidence";
+};
 
 export interface LogicalModelSource {
 	capabilityRating?: Readonly<LogicalFieldSource>;
@@ -66,8 +70,11 @@ function fieldSource(): LogicalFieldSource {
 	};
 }
 
-function source(): LogicalModelSource {
-	return { capabilityRating: fieldSource(), costRating: fieldSource(), guidelines: fieldSource() };
+function source(projectRatings = false): LogicalModelSource {
+	const ratingSource: LogicalFieldSource = projectRatings
+		? { publisher: "Slate project", decided: "2026-09-30", basis: "Project judgment set on 2026-09-30 without new benchmark evidence" }
+		: fieldSource();
+	return { capabilityRating: ratingSource, costRating: ratingSource, guidelines: fieldSource() };
 }
 
 const SHIPPED_DEFINITIONS: LogicalModelDefinition[] = [
@@ -78,14 +85,14 @@ const SHIPPED_DEFINITIONS: LogicalModelDefinition[] = [
 		cautions: [LUNA_CAUTION, "Do not use Luna to handle complex texts."], source: source(),
 	},
 	{
-		model: "claude-sonnet-5", capabilityRating: 40, effort: "high", costRating: 90,
-		preferredProvider: "anthropic", providers: providers("anthropic", "claude-sonnet-5"),
-		guidelines: [], cautions: [SONNET_OPUS_CAUTION], source: source(),
+		model: "claude-sonnet-5.5", capabilityRating: 45, effort: "high", costRating: 54,
+		preferredProvider: "anthropic", providers: providers("anthropic", "claude-sonnet-5-5"),
+		guidelines: [], cautions: [SONNET_OPUS_CAUTION], source: source(true),
 	},
 	{
-		model: "sol-6", capabilityRating: 58, effort: "high", costRating: 20,
-		preferredProvider: "openai", providers: providers("openai", "gpt-6-sol"),
-		guidelines: ["default thread choice"], cautions: [SOL_CAUTION], source: source(),
+		model: "sol-6.1", capabilityRating: 85, effort: "high", costRating: 20,
+		preferredProvider: "openai", providers: providers("openai", "gpt-6.1-sol"),
+		guidelines: ["default thread choice"], cautions: [SOL_CAUTION], source: source(true),
 	},
 	{
 		model: "gemini-3.8-flash", capabilityRating: 55, effort: "medium", costRating: 30,
@@ -96,12 +103,12 @@ const SHIPPED_DEFINITIONS: LogicalModelDefinition[] = [
 	{
 		model: "claude-opus-5.5", capabilityRating: 90, effort: "high", costRating: 65,
 		preferredProvider: "anthropic", providers: providers("anthropic", "claude-opus-5-5"),
-		guidelines: ["concurrency work", "data-loss work", "performance work", "prefer over Astra for code reviews of focus areas that require high-level design, except non-local logic", "Do not select Opus 5.5 as the default implementer. If a lower-capability model repeatedly fails at implementation, first ask Opus 5.5 to investigate and provide detailed repair instructions. Let the implementer try those instructions. Use Opus 5.5 as the implementer only if that guided attempt also fails. Treat that use as an exception. Select another suitable model for later implementation work. Existing approval requirements and repair limits still apply."], cautions: [SONNET_OPUS_CAUTION], source: source(),
+		guidelines: ["Reviewer for each proved DESIGN-TRIGGERING focus area, in adversarial design review and in implementation area review.", "Do not select Opus 5.5 as the default implementer. If a lower-capability model repeatedly fails at implementation, first ask Opus 5.5 to investigate and provide detailed repair instructions. Let the implementer try those instructions. Use Opus 5.5 as the implementer only if that guided attempt also fails. Treat that use as an exception. Select another suitable model for later implementation work. Existing approval requirements and repair limits still apply."], cautions: [SONNET_OPUS_CAUTION], source: source(),
 	},
 	{
 		model: "gpt-6-astra", capabilityRating: 86, effort: "medium", costRating: 60,
 		preferredProvider: "openai", providers: providers("openai", "gpt-6-astra"),
-		guidelines: ["prefer when available for design reviews of all focus areas that require high-level design", "prefer for code reviews of non-local logic defects", "security work", "performance work", "Do not select Astra as the default implementer. Use Astra for review only when assigned to a specific focus area. Use Astra for research when appropriate."],
+		guidelines: ["Do not select Astra as the default implementer.", "Prefer Sol 6.1 over Astra."],
 		cautions: [], source: source(),
 	},
 ];
@@ -109,11 +116,11 @@ const SHIPPED_DEFINITIONS: LogicalModelDefinition[] = [
 export const SHIPPED_LOGICAL_MODELS: readonly LogicalModelDefinition[] = deepFreeze(SHIPPED_DEFINITIONS);
 
 export const SHIPPED_COMPRESSOR_MODELS: readonly LogicalCompressorEntry[] = deepFreeze([
-	{ model: "claude-sonnet-5", effort: "medium" },
+	{ model: "claude-sonnet-5.5", effort: "medium" },
 ]);
 
 export const LOGICAL_MODEL_SOURCE_NOTES = deepFreeze({
-	ratings: "Fixed project judgments informed by DeepSWE v1.1 by DataCurve and reviewed supporting evidence",
+	ratings: "Fixed project judgments. Sol 6.1 and Sonnet 5.5 ratings were set on 2026-09-30 without new benchmark evidence. Other ratings are informed by DeepSWE v1.1 by DataCurve and reviewed supporting evidence",
 	guidance: "Project-authored advisory guidance informed by DeepSWE v1.1 focus analysis and approved project inference",
 	source: "https://deepswe.datacurve.ai/artifacts/v1.1/leaderboard-live.json",
 	retrieved: "2026-09-11",

package/extension/mode.ts

@@ -598,14 +598,13 @@ export function registerSlateMode(
 		if (store.orchestratorMode && choice.enabled) displaySummary(ctx);
 	};
 
-	// pi runs a REGISTERED extension command before it emits the input event: in
-	// the pinned pi 0.83.0, AgentSession.prompt() calls
-	// _tryExecuteExtensionCommand(text) first and returns when a command claims
-	// the text, and it emits the input event only for text that no command
-	// claimed. /slate resume and /slate handoff therefore never reach this
-	// handler, so it needs no command exemption. Text that only LOOKS like a
-	// command — a different letter case, or a sendUserMessage() call, which skips
-	// command handling — is ordinary user input and is refused like any other.
+	// In pinned pi 1.0.0, AgentSession.prompt() handles registered extension
+	// commands before the input event. It returns when a command claims the text.
+	// /slate resume and /slate handoff therefore never reach this handler.
+	// Command matching is exact and case-sensitive. sendUserMessage() skips
+	// command handling by default. Setting expandPromptTemplates to true enables
+	// command handling. A command-looking string that reaches this handler is
+	// ordinary user input and receives the same pause rule.
 	pi.on("input", async (event, ctx) => {
 		if (store.orchestratorMode && store.paused) {
 			reportPausedInput(ctx);

package/extension/model-default.ts

@@ -2,12 +2,11 @@
  * Global model-default restore: put the user's GLOBAL pi defaults back after a
  * slate-initiated model switch.
  *
- * Pi 0.85.1 keeps ordinary extension model and thinking-level switches inside
- * the current session. It changes global defaults only when a lower-level caller
+ * Pi 1.0.0 keeps ordinary extension model and thinking-level switches inside
+ * the current session. AgentSession changes global defaults only when a caller
  * explicitly requests persistence. Slate requests no persistence. This guard
- * remains for compatibility with explicit persistence fixtures and older host
- * behavior. If a switch does write global defaults, the guard restores exactly
- * the residue that the switch produced.
+ * covers explicit persistence fixtures and hosts that persist switches. If a
+ * switch writes global defaults, the guard restores exactly its matching residue.
  *
  * TWO INDEPENDENT DECISIONS, NEVER ONE PER SWITCH AND NEVER ONE PER KEY:
  *  - the defaultProvider + defaultModel PAIR, decided and written as ONE UNIT;
@@ -15,23 +14,20 @@
  * Each is judged from two fresh disk reads taken immediately before and after
  * the switch: restore only what CHANGED, and only when what it changed to is
  * exactly what slate's own switch should have produced. Anything unchanged is
- * never touched (which also makes this a zero-write no-op in EVERY state on a
- * future pi that stops persisting switches — upstream #5263); anything holding
- * a different value belongs to somebody else and is left alone.
+ * never touched, so session-only switches make this a zero-write no-op.
+ * A different value belongs to another caller and stays unchanged.
  *
- * Why the pair is ONE unit while the thinking level is separate: pi's ONLY
- * writer of those two keys is setDefaultModelAndProvider, which writes them
- * together, so any combination other than a pair that actually existed is a
- * state pi cannot produce and no consumer expects (a provider that does not own
- * the model, or a model with no provider at all). Deciding the halves
- * separately would manufacture exactly that: pre-switch provider + a third
- * party's model. So the pair is restored only when EVERY half that changed
- * matches the switch's expectation, and then the pre-switch pair is written
- * back EXACTLY, both halves, absence included — a state that demonstrably
- * existed a moment ago. The thinking level is not part of that pair and stays
- * independent, because the model setter's cascade can write it even when
- * provider and model already name the target, and handoff adoption can write it
- * ALONE when its equality guard skips the model setter.
+ * Restore provider and model as one pair. AgentSession persists that pair with
+ * setDefaultModelAndProvider. SettingsManager also exposes separate setters,
+ * so other callers can change either key. Restoring halves separately could
+ * combine Slate's saved provider with another caller's model. Restore the pair
+ * only when every changed half matches the switch expectation. Restore both
+ * saved halves, including absence.
+ *
+ * Restore the thinking level independently. AgentSession.setModel does not
+ * persist the thinking level by itself. An explicitly persistent thinking
+ * setter or a persistence fixture can write it separately. Handoff can call
+ * only the thinking setter when its model equality guard skips the model switch.
  *
  * Invariants baked in here, each one a defect this shape had to reverse:
  *  - NO STATE OUTLIVES A SINGLE SWITCH. Everything consulted is read during the
@@ -67,7 +63,7 @@
  *
  * Residue, stated honestly: a restore that exhausts its retry budget is
  * ABANDONED with a loud console warning and never repaired — the switched value
- * then persists globally exactly as it does today.
+ * remains in the global file when the switch requested persistence.
  */
 
 import { existsSync, statSync } from "node:fs";
@@ -90,8 +86,8 @@ type ThinkingLevel = ReturnType<ExtensionAPI["getThinkingLevel"]>;
  * busy-waits SYNCHRONOUSLY, so an attempt's cost is set by contention, not by
  * slate — a count-based bound would leave the stall unbounded.
  *
- * WORST-CASE STALL, derived from this file as written (pi 0.82.1: a contended
- * acquisition spins 10 × 20 ms ≈ 180 ms before it succeeds or throws):
+ * WORST-CASE STALL, derived from this file as written (pi 1.0.0: a contended
+ * acquisition makes ten attempts with nine 20 ms waits, about 180 ms):
  * five lock acquisitions per attempt — post-switch read (construction IS a
  * read), write-manager construction, the pair write, the separate
  * thinking-level write, and the verifying read — so a fully contended attempt
@@ -124,7 +120,7 @@ const REPORT_MAX_CHARS = 500;
 /** Explicit end-marker for a truncated report line. */
 const REPORT_TRUNCATION_MARK = " […truncated]";
 
-/** The three global keys pi's model setter can write. */
+/** The three global defaults covered by the restore guard. */
 interface GlobalDefaults {
 	provider: string | undefined;
 	model: string | undefined;
@@ -322,9 +318,8 @@ function readGlobalDefaults(cwd: string): ReadResult {
  * module header for why a per-half decision is unsafe). Returns the exact
  * pre-switch pair to write back, or undefined to leave BOTH keys alone.
  *
- * Do not re-split this into two decisions: pi's only writer of these keys
- * writes them together, so a pair mixing a pre-switch half with a post-switch
- * half is a state pi can never produce.
+ * Keep the restore decision atomic for the provider and model pair. Restoring
+ * only one half could combine the saved value with another caller's value.
  */
 function planPairRestore(
 	pre: GlobalDefaults,
@@ -518,16 +513,19 @@ function switchMayHavePersisted<T>(predicate: ((result: T) => boolean) | undefin
 }
 
 /**
- * Wrap ONE slate-initiated model switch so the user's global model defaults are
- * left as slate found them. `performSwitch` must contain every setter call of
- * that switch — the post-switch read has to follow the LAST write, since
- * adoption can persist the thinking-level key alone.
+ * Wrap one slate-initiated model switch to preserve the user's global defaults.
+ * `performSwitch` must contain every setter call for that switch.
+ * The post-switch read must follow the last write.
+ * Explicit persistence or a host that persists switches can write the thinking
+ * default alone during handoff adoption. Ordinary pi 1.0.0 switches write no
+ * global defaults.
  *
- * The restore runs even when `performSwitch` THROWS (pi's model setter persists
- * the pair before the cascade and the model-select emission that can throw, so
- * a throw can still leave residue); the original rejection is then re-thrown
- * unchanged. Apart from that, the wrapper never throws and never alters what
- * `performSwitch` returns: a restore failure is reported, not propagated.
+ * The restore runs even when `performSwitch` throws.
+ * With persist:true, pi's model setter writes the pair before updating thinking
+ * and emitting model_select. That emission can throw after the write.
+ * The wrapper rethrows the original rejection unchanged.
+ * Otherwise, the wrapper never throws or changes what `performSwitch` returns.
+ * It reports a restore failure without propagating it.
  *
  * `mayHavePersisted` is an optional escape hatch for the opposite case: return
  * false to state that the callback provably called NO pi setter, so nothing can

package/extension/worker.ts

@@ -78,7 +78,7 @@ export function isJudgementThreadType(type: unknown): type is (typeof JUDGEMENT_
 	return (JUDGEMENT_THREAD_TYPES as readonly unknown[]).includes(type);
 }
 
-// pi-ai 0.85.1 declares this value in
+// pi-ai 1.0.0 declares this value in
 // @earendil-works/pi-ai/dist/api/openai-prompt-cache.js. pi's extension loader
 // aliases the package root to the compat.js FILE, so an extension cannot safely
 // import that deep subpath. The node test pins this copy to pi-ai's declaration.
@@ -162,7 +162,7 @@ function installPromptCacheKey(session: WorkerSession, promptCacheKey?: string):
 		}
 
 		try {
-			// pi-ai 0.85.1's api/openai-responses.js buildParams always creates
+			// pi-ai 1.0.0 api/openai-responses.js buildParams always creates
 			// prompt_cache_key and assigns undefined when resolved cacheRetention is
 			// "none". Own-property presence therefore distinguishes that deliberate
 			// opt-out from a payload shape that never considered this field.
@@ -295,15 +295,11 @@ export function createWorkerRequestContract(): WorkerRequestContract {
  *  · the wait happens BEFORE the SDK converts the conversation into a provider
  *    payload, so a waiting request holds no converted copy (PF5).
  *
- * THE AUTHENTICATION CONTRACT IS UNCHANGED by this wrapper.
- * `AgentSession._getSummarizationRequestAuth` (agent-session.js:195) takes its
- * STRICT branch only while `agent.streamFunction === streamSimple`. A session
- * built by `createAgentSession` never satisfies that test: the SDK installs its
- * own stream function on the Agent (dist/core/sdk.js, `streamFn:` in the agent
- * runtime options), so the lenient branch is already the branch every slate
- * worker session used before this wrapper existed. A measured probe on the
- * pinned pi 0.85.1 confirmed `streamFunction === streamSimple` is false for a
- * fresh worker session, and test/request-throttle-worker.test.ts pins it.
+ * The wrapper preserves the software development kit (SDK) authentication branch.
+ * AgentSession._getSummarizationRequestAuth takes its strict branch only when
+ * agent.streamFunction === streamSimple. In pinned pi 1.0.0, createAgentSession
+ * installs its own stream function, so fresh worker sessions take the lenient
+ * branch. test/request-throttle-worker.test.ts checks that identity.
  *
  * A THROWN admission error is NOT swallowed. A cancelled wait rejects, pi turns
  * the rejection into the same aborted or errored assistant outcome it produces
@@ -455,24 +451,24 @@ export async function openWorkerSession(opts: {
 	// user has NOT trusted. Carry the host session's actual trust decision into
 	// both the resource loader and the session.
 	//
-	// READ-ONLY view (AF8/AF9): model failover may call session.setModel on a
-	// live worker, and setModel persists the new model as the default via
-	// SettingsManager (setThinkingLevel likewise). A file-backed manager would
-	// write a worker's failover model into the USER'S global settings.json.
-	// Instead, snapshot the settings once here and serve them through a custom
-	// SettingsStorage whose withLock discards the callback's return value —
-	// the supported no-op write path (persistScopedSettings only hands the new
-	// JSON back as that return value; there is no error path). Reads, merge
-	// semantics, and trust gating are identical to a file-backed manager.
+	// READ-ONLY view (AF8/AF9): pi 1.0.0 keeps ordinary session.setModel and
+	// setThinkingLevel calls inside the session. Slate requests no persistence.
+	// This storage protects global settings from calls with persist:true and
+	// from hosts that persist switches. A file-backed manager would save those
+	// explicit writes as global defaults, including a worker's failover choice.
+	// Snapshot the settings once and serve them through a custom SettingsStorage.
+	// Its withLock discards the callback's return value, which drops writes.
+	// persistScopedSettings returns the new JSON through that callback.
+	// Reads, merge semantics and trust gating match a file-backed manager.
 	//
-	// CN4: the snapshot is taken via a throwaway file-backed SettingsManager,
-	// NOT a raw readFileSync — pi's settings writer holds a lockfile during
-	// its non-atomic writes (e.g. the orchestrator's own failover setModel
-	// persisting a new global default), so an unlocked read could tear. The
-	// throwaway does the locked, error-tolerant read; its per-scope snapshots
-	// are re-serialized for the storage below (fromStorage re-parses and
-	// re-migrates them — idempotent), and it never writes: no setter is ever
-	// called on it.
+	// CN4: a throwaway file-backed SettingsManager takes the snapshot under pi's
+	// settings lock. Pi holds that lock during its non-atomic settings writes.
+	// A call with persist:true or a direct SettingsManager setter can write
+	// global defaults. An unlocked read could capture an incomplete write.
+	// The throwaway manager reads under the lock and handles read errors.
+	// Its per-scope snapshots become JSON for the storage below.
+	// fromStorage parses and migrates that JSON again without changing it.
+	// The throwaway manager never writes because no caller invokes its setters.
 	const trusted = ctx.isProjectTrusted();
 	const snapshot = SettingsManager.create(ctx.cwd, agentDir, { projectTrusted: trusted });
 	// Pi's background warmer bypasses agent.streamFunction and therefore Slate's
@@ -597,13 +593,11 @@ export async function openWorkerSession(opts: {
 		? SessionManager.open(opts.sessionFile)
 		: SessionManager.create(ctx.cwd, ensureRuntimeDirectory(ctx.cwd, opts.runtimeFolder ?? defaultArtifactFolder(), "threads"));
 
-	// No modelRuntime passed: createAgentSession (pi >= 0.80.8) defaults to a
-	// ModelRuntime replacing the AuthStorage + ModelRegistry setup this code
-	// hand-built before those SDK options were removed. Credential/config
-	// sources are unchanged — global agentDir auth.json + models.json, never
-	// project-local — but the default is a superset: it also reads/writes
-	// agentDir/models-store.json and may run a throttled (~4h-cached)
-	// create-time network catalog refresh (disabled by PI_OFFLINE).
+	// In pi 1.0.0, createAgentSession creates a ModelRuntime when none is passed.
+	// It uses auth.json, models.json and models-store.json under agentDir.
+	// SDK creation enables network catalog refresh only with allowModelNetwork.
+	// This call supplies no runtime option, so creation uses no network catalog
+	// refresh. PI_OFFLINE also disables model-network access.
 	const { session } = await createAgentSession({
 		cwd: ctx.cwd,
 		agentDir,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ytdb-slate",
-  "version": "0.12.0",
+  "version": "0.13.0",
   "description": "Multi-agent orchestration extension for the pi coding agent: bounded worker sessions, structured episode summaries, and a risk-based review workflow.",
   "license": "Apache-2.0",
   "repository": {
@@ -36,9 +36,9 @@
     "typebox": "*"
   },
   "devDependencies": {
-    "@earendil-works/pi-ai": "0.87.1",
-    "@earendil-works/pi-coding-agent": "0.87.1",
-    "@earendil-works/pi-tui": "0.87.1",
+    "@earendil-works/pi-ai": "1.0.0",
+    "@earendil-works/pi-coding-agent": "1.0.0",
+    "@earendil-works/pi-tui": "1.0.0",
     "@types/node": "24.13.3",
     "typebox": "1.3.27",
     "typescript": "5.9.3"