pi-continual-harness 0.21.0 → 0.22.0

package/CHANGELOG.md

@@ -8,7 +8,32 @@ Releases are tag-driven (`vX.Y.Z`) and published by GitHub Actions via npm
 Trusted Publishing. This file begins at 0.7.0; earlier releases are recorded in
 the git tags (`git tag -l`) and the [GitHub release history](https://github.com/pungggi/pi-continual-harness/releases).
 
-## [Unreleased]
+## [0.22.0] — 2026-11-11
+
+### Added
+
+- **Classifier auto-discovery** (pi ≥ 1.5.1 interplay): when `classifier.model`
+  is unset, the classifier seam now falls back to a single known-good hosted
+  classifier — `github-copilot/microsoft-decision-1`, shipped by pi 1.5.1+
+  with GitHub Copilot auth — so classifier-confirmed dedupe and the auto-refine
+  gate work with zero setup for Copilot users. Discovery deliberately never
+  picks a local llama.cpp entry (the catalog can list one while no server
+  runs). An explicit `classifier.model` always wins; `"classifier": { "auto":
+  false }` restores explicit-only. `/harness status` and the setup/reminder
+  copy now treat the gate as the only required key when discovery is on.
+- **Determinism contract for the injected block** (pinned by a test): for an
+  unchanged store + config, `renderHarnessBlock` is byte-identical. This is
+  what lets pi 1.5.1+'s Anthropic message threads `continue` a conversation —
+  the thread-prefix hash covers the system prompt, so unchanged harness state
+  rides the server-side thread and the warmed prompt cache, while any harness
+  change mismatches the hash and pi correctly falls back to a full resend.
+  Idle prompt-cache warming (pi 1.5.1) similarly keeps the injected system
+  prefix warm between runs — no package change needed, both verified against
+  the pi-ai 1.5.2 source.
+- Dev dependencies bumped to `@earendil-works/pi-*` `^1.5.2` (runtime peers
+  stay `>=1.2.0`; nothing 1.5-only is imported at runtime).
+
+## [0.21.0] — 2026-11-07
 
 ### Added
 

package/README.md

@@ -195,7 +195,7 @@ Optional config at `~/.pi/agent/harness.json` (missing → defaults; malformed
   "remindRefine": { "enabled": true, "everyTurns": 25, "detailEvery": 4 },
   "autoRefine": { "enabled": false, "everyTurns": 100, "commit": false, "gate": false, "refineNoOpSkip": 3 },
   "outcomeImportance": { "enabled": false, "bump": 0.03 },
-  "classifier": { "model": "typesafe/jev-latest" },
+  "classifier": { "model": "github-copilot/microsoft-decision-1", "auto": true },
   "virtualBinding": "virtual"
 }
 ```
@@ -218,7 +218,7 @@ schemas (the same pattern pi ≥ 1.2 ships for its own `settings.json` /
 | `remindRefine` | on, every 25 turns | The notify-only `/refine` activation nudge; every 4th one is a detailed educational guide. |
 | `autoRefine` | off, every 100 turns | Opt-in autonomous `/refine` on a cadence; `gate` adds the classifier question, `refineNoOpSkip` the free history gate. |
 | `outcomeImportance` | off, bump 0.03 | Opt-in promotion of items you cite by `[h_xxxx]` tag. Promotion only — never deletes. |
-| `classifier` | unset | Classifier model (pi ≥ 0.99) enabling classifier-confirmed dedupe + the auto-refine gate. |
+| `classifier` | `model` unset, `auto: true` | Classifier enabling classifier-confirmed dedupe + the auto-refine gate. With `auto` (default), an unset `model` auto-resolves the hosted `github-copilot/microsoft-decision-1` on pi ≥ 1.5.1 with Copilot auth; set `model` to pin any classifier, or `auto: false` to require it. |
 | `virtualBinding` | `"virtual"` | Per-model isolation key for router (virtual) models; `"physical"` keys by the routed model. |
 
 The package's autonomous paths (`autoImport`, `autoRefine`, `outcomeImportance`)
@@ -542,8 +542,12 @@ migration; developed and tested against pi 1.5). Implemented:
   config key; the store is never changed by selection. Rendered on **every
   provider request** (`context_with_system`), so a mid-run `harness_mutate`
   shows up on the very next request.
-- **Classifier features (pi ≥ 0.99, opt-in)**: `"classifier": { "model": … }`
-  enables (1) classifier-confirmed dedupe — candidate merge pairs confirmed by
+- **Classifier features (pi ≥ 0.99, opt-in)**: a classifier resolves via
+  `"classifier": { "model": … }`, or — when `model` is unset — via
+  auto-discovery (default on, `"classifier": { "auto": false }` opts out)
+  of the hosted `github-copilot/microsoft-decision-1` that pi ≥ 1.5.1 ships
+  with GitHub Copilot auth. A resolved classifier enables (1)
+  classifier-confirmed dedupe — candidate merge pairs confirmed by
   one batched yes/no call (errors fall back to the rule-based plan) — and (2)
   the auto-refine gate `"autoRefine": { "gate": true }`, which skips a cadenced
   refine when the trajectory holds no durable correction.

package/dist/auto-refine.js

@@ -131,7 +131,7 @@ export async function handleAutoRefineTurnEnd(pi, event, ctx) {
     }
     // Opt-in classifier gate: cheap yes/no before spending the refine.
     if (config.autoRefine?.gate) {
-        const classify = buildClassify(ctx, config.classifier?.model);
+        const classify = buildClassify(ctx, config.classifier);
         if (classify) {
             const evidence = gatherEvidence(ctx, DEFAULT_LOOKBACK_TURNS);
             const res = await classify(buildGateRequest(evidence));

package/dist/classify.d.ts

@@ -65,14 +65,36 @@ interface ClassifierModelLike {
  * it then degrades to its no-classifier behavior.
  */
 export declare function resolveClassifierModel(registry: RegistryLike | undefined, configured?: string): ClassifierModelLike | undefined;
+/**
+ * The auto-discovered default classifier: a hosted bool/score classifier that
+ * pi ≥ 1.5.1 ships with GitHub Copilot authentication (billed to the user's
+ * Copilot plan; no local server, no extra API key). Keeping this a single
+ * known-good id — not "first classifier in the catalog" — makes discovery
+ * deterministic and reviewable; widen it only deliberately.
+ */
+export declare const DEFAULT_CLASSIFIER = "github-copilot/microsoft-decision-1";
+/**
+ * Auto-discovery fallback for buildClassify: the hosted default above when the
+ * registry lists it, otherwise undefined (degrade to no-classifier behavior).
+ */
+export declare function discoverClassifierModel(registry: RegistryLike | undefined): ClassifierModelLike | undefined;
 /** Build the ClassifyFn adapter over a resolved classifier model. */
 export declare function makeClassifyFn(registry: RegistryLike, model: ClassifierModelLike, signal?: AbortSignal): ClassifyFn;
+/** Structural slice of the normalized `classifier` config (config.ts). */
+export interface ClassifierConfigSlice {
+    model?: string | undefined;
+    /** Default true (opt-out): when `model` is unset, fall back to the hosted
+     *  default classifier when the registry lists one. */
+    auto?: boolean;
+}
 /**
  * Resolve the ClassifyFn for a session from config + ctx, or undefined.
  * The single construction point shared by refine (ProposeInput.classify) and
- * the auto-refine gate.
+ * the auto-refine gate. An explicit `classifier.model` always wins; with
+ * `auto` (default) an unset model falls back to discovery (see
+ * discoverClassifierModel).
  */
-export declare function buildClassify(ctx: ExtensionContext, configuredModel?: string): ClassifyFn | undefined;
+export declare function buildClassify(ctx: ExtensionContext, classifier?: ClassifierConfigSlice): ClassifyFn | undefined;
 /**
  * The auto-refine gate question: did the recent trajectory contain a durable,
  * reusable correction worth a /refine pass? Pure builder — unit-testable.

package/dist/classify.js

@@ -10,6 +10,15 @@
 //  2. dedupe confirmation (proposer.ts): a precision filter on candidate merge
 //     pairs — see classifyConfirmPairs().
 //
+// Model resolution has two tiers (buildClassify): an explicitly configured
+// `classifier.model` always wins; when it is unset, auto-discovery (default on,
+// `classifier.auto: false` opts out) falls back to ONE known hosted classifier —
+// github-copilot/microsoft-decision-1, shipped by pi ≥ 1.5.1 with GitHub
+// Copilot auth — so the gated paths work with zero setup for Copilot users.
+// Discovery deliberately never picks a local llama.cpp entry: the catalog can
+// list one while no server is running, and an unreachable classifier is worse
+// than none (the gates would soft-fail every time).
+//
 // The pi-ai classifier shapes (ClassifierContext/ClassifierQuestion/…) are kept
 // OUT of the public proposer interface: ClassifyFn below is a minimal structural
 // contract so proposers stay testable without pi runtime types.
@@ -35,6 +44,21 @@ export function resolveClassifierModel(registry, configured) {
         return cm.id === configured || `${cm.provider}/${cm.id}` === configured;
     }) ?? undefined);
 }
+/**
+ * The auto-discovered default classifier: a hosted bool/score classifier that
+ * pi ≥ 1.5.1 ships with GitHub Copilot authentication (billed to the user's
+ * Copilot plan; no local server, no extra API key). Keeping this a single
+ * known-good id — not "first classifier in the catalog" — makes discovery
+ * deterministic and reviewable; widen it only deliberately.
+ */
+export const DEFAULT_CLASSIFIER = "github-copilot/microsoft-decision-1";
+/**
+ * Auto-discovery fallback for buildClassify: the hosted default above when the
+ * registry lists it, otherwise undefined (degrade to no-classifier behavior).
+ */
+export function discoverClassifierModel(registry) {
+    return resolveClassifierModel(registry, DEFAULT_CLASSIFIER);
+}
 /** Build the ClassifyFn adapter over a resolved classifier model. */
 export function makeClassifyFn(registry, model, signal) {
     const label = `${model.provider}/${model.id}`;
@@ -102,12 +126,17 @@ export function makeClassifyFn(registry, model, signal) {
 /**
  * Resolve the ClassifyFn for a session from config + ctx, or undefined.
  * The single construction point shared by refine (ProposeInput.classify) and
- * the auto-refine gate.
+ * the auto-refine gate. An explicit `classifier.model` always wins; with
+ * `auto` (default) an unset model falls back to discovery (see
+ * discoverClassifierModel).
  */
-export function buildClassify(ctx, configuredModel) {
+export function buildClassify(ctx, classifier) {
     const registry = ctx.modelRegistry;
-    const model = resolveClassifierModel(registry, configuredModel);
-    if (!registry || !model)
+    if (!registry)
+        return undefined;
+    const model = resolveClassifierModel(registry, classifier?.model) ??
+        (classifier?.auto !== false ? discoverClassifierModel(registry) : undefined);
+    if (!model)
         return undefined;
     return makeClassifyFn(registry, model, ctx.signal);
 }

package/dist/config.d.ts

@@ -72,12 +72,18 @@ export interface HarnessConfig {
          *  DEFAULT_REFINE_NOOP_SKIP (3); 0 disables the history gate. */
         refineNoOpSkip?: number;
     };
-    /** Classifier model for the cheap gated paths (pi ≥ 0.99): "provider/id" or
+    /** Classifier for the cheap gated paths (pi ≥ 0.99): "provider/id" or
      *  a bare id of a classifier catalog entry (e.g. "typesafe/jev-latest" or a
-     *  local llama.cpp classifier). When unset or unresolvable, every classifier
-     *  feature degrades to its no-classifier behavior. */
+     *  local llama.cpp classifier). When `model` is unset or unresolvable,
+     *  `auto` (default true) falls back to a hosted zero-setup classifier when
+     *  the registry lists one — pi ≥ 1.5.1 ships
+     *  "github-copilot/microsoft-decision-1" with GitHub Copilot auth. When
+     *  nothing resolves, every classifier feature degrades to its no-classifier
+     *  behavior. */
     classifier?: {
         model?: string | undefined;
+        /** Default true (opt-out). Set false to require an explicit `model`. */
+        auto?: boolean;
     };
     /** Virtual-model binding policy (pi ≥ 0.99). Default "virtual". */
     virtualBinding?: VirtualBinding;

package/dist/config.js

@@ -50,7 +50,7 @@ export const DEFAULT_CONFIG = {
         omitStaleDrafts: true,
         refineNoOpSkip: DEFAULT_REFINE_NOOP_SKIP,
     },
-    classifier: { model: undefined },
+    classifier: { model: undefined, auto: true },
     virtualBinding: "virtual",
     proposer: "steering",
     outcomeImportance: { enabled: false, bump: DEFAULT_REF_BUMP },
@@ -113,6 +113,7 @@ function mergeConfig(over) {
             model: typeof over.classifier?.model === "string" && over.classifier.model
                 ? over.classifier.model
                 : undefined,
+            auto: over.classifier?.auto !== false,
         },
         virtualBinding: over.virtualBinding === "physical" ? "physical" : "virtual",
         proposer: over.proposer ?? "steering",

package/dist/index.d.ts

@@ -13,7 +13,7 @@ export { preview, renderListReport, selectListItems, type ListOptions, type List
 export { DEFAULT_INJECTION, estimateTokens, normalizeInjection, selectForInjection, type InjectionConfig, type NormalizedInjection, type SelectionResult, } from "./select.js";
 export { branchEntries, isVirtualModel, physicalKeyFromBranch, resolveBindingKey, VIRTUAL_MODEL_API, } from "./vmodel.js";
 export { buildDetailGuide, buildNudge, reminderState, type ReminderState, } from "./remind.js";
-export { buildGateRequest, gateDecision, type ClassifyBoolAnswer, type ClassifyBoolQuestion, type ClassifyBoolResult, type ClassifyFn, } from "./classify.js";
+export { DEFAULT_CLASSIFIER, buildGateRequest, discoverClassifierModel, gateDecision, type ClassifyBoolAnswer, type ClassifyBoolQuestion, type ClassifyBoolResult, type ClassifyFn, type ClassifierConfigSlice, } from "./classify.js";
 export { AUTO_REFINE_ENTRY, buildOmissionDrafts, buildRefineDraft, turnEndedBadly, type OmissionDraft, type ProjectedLike, } from "./auto-refine.js";
 export { messageText } from "./refine.js";
 export { deepMergeConfig, readConfigRaw, saveConfig, type SaveConfigResult, } from "./config.js";

package/dist/index.js

@@ -51,8 +51,9 @@ export { branchEntries, isVirtualModel, physicalKeyFromBranch, resolveBindingKey
 // richer /harness status) share the same copy source.
 export { buildDetailGuide, buildNudge, reminderState, } from "./remind.js";
 // Classifier seam (pi ≥ 0.99, 0.12.0): the minimal structural ClassifyFn
-// contract proposers see, plus the gate builders (pure, testable).
-export { buildGateRequest, gateDecision, } from "./classify.js";
+// contract proposers see, plus the gate builders (pure, testable) and the
+// auto-discovery default (0.22.0: hosted fallback for unset classifier.model).
+export { DEFAULT_CLASSIFIER, buildGateRequest, discoverClassifierModel, gateDecision, } from "./classify.js";
 // Auto-refine boundary helpers (pi ≥ 0.87, 0.12.0–0.13.0): the steering draft,
 // the stale-draft context-edit omissions, and the shared aborted/error-turn
 // predicate (pi BoundaryState.outcome). Re-exported for tests and companion

package/dist/inject.d.ts

@@ -30,6 +30,14 @@ export declare function resetInjectionStats(): void;
  * defining the marker when any such item is surfaced; `priority` is honored
  * inside the selection (trusted items win budget contention).
  *
+ * DETERMINISM CONTRACT: for an unchanged store + config the returned block is
+ * byte-identical (stable item ids, no timestamps, no random ordering — pinned
+ * by test/model-binding.test.ts). This is what lets pi ≥ 1.5.1 Anthropic message
+ * threads `continue` a conversation: the thread-prefix hash covers the system
+ * message, so an unchanged harness rides the server-side thread (and the
+ * warmed prompt cache) while any harness change mismatches the hash and pi
+ * correctly falls back to a full resend. Never add volatile content here.
+ *
  * Note: every non-empty render folds into the session injection stats
  * (InjectionStats above; /harness status surfaces them). Tests that render
  * directly reset with resetInjectionStats().

package/dist/inject.js

@@ -51,6 +51,14 @@ export function resetInjectionStats() {
  * defining the marker when any such item is surfaced; `priority` is honored
  * inside the selection (trusted items win budget contention).
  *
+ * DETERMINISM CONTRACT: for an unchanged store + config the returned block is
+ * byte-identical (stable item ids, no timestamps, no random ordering — pinned
+ * by test/model-binding.test.ts). This is what lets pi ≥ 1.5.1 Anthropic message
+ * threads `continue` a conversation: the thread-prefix hash covers the system
+ * message, so an unchanged harness rides the server-side thread (and the
+ * warmed prompt cache) while any harness change mismatches the hash and pi
+ * correctly falls back to a full resend. Never add volatile content here.
+ *
  * Note: every non-empty render folds into the session injection stats
  * (InjectionStats above; /harness status surfaces them). Tests that render
  * directly reset with resetInjectionStats().

package/dist/refine.js

@@ -103,7 +103,7 @@ async function buildProposeInput(ctx, evidence, lookback, options) {
         complete: buildComplete(ctx),
         // Inject the classifier adapter when `classifier.model` is configured and
         // resolvable (pi ≥ 0.99) — the dedupe proposer uses it to confirm merges.
-        classify: buildClassify(ctx, config.classifier?.model),
+        classify: buildClassify(ctx, config.classifier),
         ownerKey,
         lessons,
         evidence,

package/dist/remind.js

@@ -151,9 +151,9 @@ export const DETAIL_GUIDES = [
         '  "remindRefine": { "enabled": true, "everyTurns": 25, "detailEvery": 4 },',
         '  "autoRefine": { "enabled": true, "everyTurns": 100, "gate": true, "commit": false },',
         '  "outcomeImportance": { "enabled": true, "bump": 0.03 },',
-        '  "classifier": { "model": "typesafe/jev-latest" }',
+        '  "classifier": { "auto": true }',
         "}",
-        `Guidance: gate (needs classifier.model) spends one cheap yes/no question before each auto-refine, so cadence becomes signal. Durable writes are autonomous in this config: with autoImport on, the layers re-export on every store change (that is what makes /refine output survive into new sessions); commit only adds a flush at each auto-run and matters when autoImport is off. Prefer autoImport off + manual /harness export until you trust the refinements, then opt in. This nudge fires every ${state.everyTurns} completed turns and shows a guide every ${state.detailEvery}th reminder (about every ${state.everyTurns * state.detailEvery} completed turns).`,
+        `Guidance: gate spends one cheap yes/no question before each auto-refine, so cadence becomes signal. With classifier.auto (default true) and no explicit classifier.model, the hosted github-copilot/microsoft-decision-1 resolves automatically on pi >= 1.5.1 with Copilot auth; otherwise set classifier.model to any classifier catalog entry (e.g. a local llama.cpp classifier). Durable writes are autonomous in this config: with autoImport on, the layers re-export on every store change (that is what makes /refine output survive into new sessions); commit only adds a flush at each auto-run and matters when autoImport is off. Prefer autoImport off + manual /harness export until you trust the refinements, then opt in. This nudge fires every ${state.everyTurns} completed turns and shows a guide every ${state.detailEvery}th reminder (about every ${state.everyTurns * state.detailEvery} completed turns).`,
     ].join("\n"),
 ];
 /** Rotation slot for the ordinal-th detailed reminder (1-based, pure). */

package/dist/setup.js

@@ -99,7 +99,7 @@ export const PRESETS = {
         title: "durable sync + awareness reminders",
         description: "autoImport on (refinements survive into new sessions) paired with the importRemind awareness reminders.",
         notes: [
-            "Gate-ready: to make refinement signal-driven later, set classifier.model + autoRefine.gate (pi >= 0.99). This preset leaves the gate off.",
+            "Gate-ready: to make refinement signal-driven later, set autoRefine.gate (pi >= 0.99; a classifier resolves automatically on pi >= 1.5.1 with Copilot, else set classifier.model). This preset leaves the gate off.",
         ],
         deltas: recommendedDeltas(),
     },

package/dist/status.js

@@ -125,11 +125,15 @@ function applicableHints(state, config, impact) {
     // Classifier/gate not fully configured → the cheapest activation lever is
     // invisible to the user. `classifier.model` ALONE changes nothing gating-wise
     // (the gate needs `autoRefine.gate: true` too — review #27), so the hint
-    // stays until both are set and the copy names both keys.
-    if (!config.classifier?.model || config.autoRefine?.gate !== true) {
+    // stays until the gate is on and the copy names the keys. Since auto-
+    // discovery (0.22.0) a classifier may resolve with `classifier.model` unset
+    // (hosted default on pi ≥ 1.5.1 + Copilot auth), so the copy treats the
+    // model as optional.
+    if (config.autoRefine?.gate !== true ||
+        (!config.classifier?.model && config.classifier?.auto === false)) {
         hints.push({
             id: "classifier-gate",
-            text: 'the auto-refine gate is available: set `classifier.model` and `autoRefine: { "gate": true }` in harness.json so cadenced refines skip when nothing durable happened',
+            text: 'the auto-refine gate is available: set `autoRefine: { "gate": true }` in harness.json so cadenced refines skip when nothing durable happened (a classifier resolves automatically on pi ≥ 1.5.1 with GitHub Copilot; otherwise also set `classifier.model`)',
         });
     }
     // Zero citations ever and the loop is off → the outcome half of the fitness

package/harness.schema.json

@@ -71,12 +71,17 @@
     },
     "classifier": {
       "type": "object",
-      "description": "Classifier model for the cheap gated paths.",
+      "description": "Classifier for the cheap gated paths.",
       "additionalProperties": true,
       "properties": {
         "model": {
           "type": "string",
           "description": "\"provider/id\" or a bare id of a classifier catalog entry (e.g. \"typesafe/jev-latest\" or a local llama.cpp classifier). Unset/unresolvable → every classifier feature degrades to its no-classifier behavior."
+        },
+        "auto": {
+          "type": "boolean",
+          "default": true,
+          "description": "When model is unset (default), fall back to a hosted zero-setup classifier when the registry lists one — pi ≥ 1.5.1 ships github-copilot/microsoft-decision-1 with GitHub Copilot auth. Set false to require an explicit model."
         }
       }
     },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-continual-harness",
-  "version": "0.21.0",
+  "version": "0.22.0",
   "description": "Online self-improvement layer for pi. A unified, in-trajectory harness-state store (prompt notes, memory, skill descriptions, sub-agent specs) with a manual /refine that proposes evidence-backed structured CRUD deltas. Composes with pi-reflect (offline) and pi-mem (storage); owns only the online optimizer layer. Inspired by Continual Harness (arXiv 2605.09998) and ACE (arXiv 2510.04618).",
   "keywords": [
     "pi-package",
@@ -44,9 +44,9 @@
   },
   "devDependencies": {
     "@biomejs/biome": "^2.6.4",
-    "@earendil-works/pi-ai": "^1.5.0",
-    "@earendil-works/pi-coding-agent": "^1.5.0",
-    "@earendil-works/pi-schema": "^1.5.0",
+    "@earendil-works/pi-ai": "^1.5.2",
+    "@earendil-works/pi-coding-agent": "^1.5.2",
+    "@earendil-works/pi-schema": "^1.5.2",
     "@types/node": "^22.0.0",
     "@vitest/coverage-v8": "^3.2.7",
     "typebox": ">=1.0.0",