@polderlabs/bizar 5.0.0 → 5.0.2

package/config/agents/mimir.md

@@ -1,7 +1,7 @@
 ---
 description: Mimir — Deep codebase research and exploration. Uses Semble as primary search tool. Architecture analysis, pattern discovery, documentation research, and project initialization.
 mode: subagent
-model: opencode/deepseek-v4-flash-free
+model: minimax/MiniMax-M2.7
 color: "#06b6d4"
 permission:
   read: allow

package/config/agents/quick.md

@@ -1,7 +1,7 @@
 ---
 description: Quick (quick) — fast single-shot tasks. No delegation, no parallel streams. Use for small edits, mechanical changes, one-shot questions. Routes to no one.
 mode: primary
-model: minimax/MiniMax-M2.7
+model: minimax/MiniMax-M2.7-Flash
 color: "#22d3ee"
 permission:
   read: allow

package/config/agents/semble-search.md

@@ -1,7 +1,7 @@
 ---
 description: Code search agent for exploring any codebase. Use for finding code by intent, locating implementations, understanding how something works, or discovering related code. Prefer over Bash/Read for any semantic or exploratory question.
 mode: subagent
-model: opencode/deepseek-v4-flash-free
+model: minimax/MiniMax-M2.7
 color: "#64748b"
 permission:
   read: allow

package/config/agents/vor.md

@@ -1,7 +1,7 @@
 ---
 description: Vör — Asks clarifying questions for ambiguous or incomplete requests. Reads project context first, then asks one targeted, project-specific question.
 mode: subagent
-model: opencode/deepseek-v4-flash-free
+model: minimax/MiniMax-M2.7
 color: "#a78bfa"
 permission:
   read: allow

package/config/opencode.json

@@ -37,7 +37,14 @@
       "loopThresholdWarn": 5,
       "loopThresholdEscalate": 8,
       "loopThresholdBlock": 12,
-      "loopWindowSize": 10
+      "loopWindowSize": 10,
+      "compaction": {
+        "auto": true,
+        "threshold": 0.5,
+        "preserve_recent": 10,
+        "strategy": "summarize",
+        "model": "minimax/MiniMax-M3-Reasoning"
+      }
     }]
   ],
   "tools": {
@@ -66,9 +73,9 @@
       }
     },
     "vor": {
-      "description": "Vör — Tier 2 clarifier. Reads project context first and asks targeted, project-specific questions when requests are ambiguous. Backed by DeepSeek V4 Flash. Operates inside BizarHarness alongside 13 sibling agents.",
+      "description": "Vör — Tier 2 clarifier. Reads project context first and asks targeted, project-specific questions when requests are ambiguous. Backed by MiniMax M2.7. Operates inside BizarHarness alongside 13 sibling agents.",
       "mode": "subagent",
-      "model": "opencode/deepseek-v4-flash-free",
+      "model": "minimax/MiniMax-M2.7",
       "color": "#8b5cf6",
       "permission": {
         "read": "allow",
@@ -77,9 +84,9 @@
       }
     },
     "frigg": {
-      "description": "Frigg — Tier 1 primary. Read-only Q&A agent that answers codebase questions with file:line citations, never edits. Backed by DeepSeek V4 Flash. Operates inside BizarHarness alongside 13 sibling agents.",
+      "description": "Frigg — Tier 1 primary. Read-only Q&A agent that answers codebase questions with file:line citations, never edits. Backed by MiniMax M2.7. Operates inside BizarHarness alongside 13 sibling agents.",
       "mode": "primary",
-      "model": "opencode/deepseek-v4-flash-free",
+      "model": "minimax/MiniMax-M2.7",
       "color": "#06b6d4",
       "permission": {
         "read": "allow",
@@ -93,9 +100,9 @@
       }
     },
     "quick": {
-      "description": "Quick — Tier 1 primary. Fast single-shot escape hatch for small edits, mechanical changes, and one-shot questions. No delegation, no parallel streams. Backed by MiniMax M2.7. Operates inside BizarHarness alongside 13 sibling agents.",
+      "description": "Quick — Tier 1 primary. Fast single-shot escape hatch for small edits, mechanical changes, and one-shot questions. No delegation, no parallel streams. Backed by MiniMax M2.7 Flash. Operates inside BizarHarness alongside 13 sibling agents.",
       "mode": "primary",
-      "model": "minimax/MiniMax-M2.7",
+      "model": "minimax/MiniMax-M2.7-Flash",
       "color": "#22d3ee",
       "permission": {
         "read": "allow",
@@ -112,9 +119,9 @@
       }
     },
     "mimir": {
-      "description": "Mimir — Tier 2 research. Dedicated codebase exploration agent who uses Semble as primary search tool for deep analysis, pattern discovery, and documentation research. Backed by DeepSeek V4 Flash. Operates inside BizarHarness alongside 13 sibling agents.",
+      "description": "Mimir — Tier 2 research. Dedicated codebase exploration agent who uses Semble as primary search tool for deep analysis, pattern discovery, and documentation research. Backed by MiniMax M2.7. Operates inside BizarHarness alongside 13 sibling agents.",
       "mode": "subagent",
-      "model": "opencode/deepseek-v4-flash-free",
+      "model": "minimax/MiniMax-M2.7",
       "color": "#0ea5e9",
       "permission": {
         "read": "allow",
@@ -130,9 +137,9 @@
       }
     },
     "heimdall": {
-      "description": "Heimdall — Tier 2 mechanical. Handles simple, routine, deterministic engineering tasks with speed and precision using DeepSeek V4 Flash. The ever-watchful guardian. Operates inside BizarHarness alongside 13 sibling agents.",
+      "description": "Heimdall — Tier 2 mechanical. Handles simple, routine, deterministic engineering tasks with speed and precision using MiniMax M2.7. The ever-watchful guardian. Operates inside BizarHarness alongside 13 sibling agents.",
       "mode": "subagent",
-      "model": "opencode/deepseek-v4-flash-free",
+      "model": "minimax/MiniMax-M2.7",
       "color": "#10b981",
       "permission": {
         "read": "allow",
@@ -249,9 +256,9 @@
       }
     },
     "semble-search": {
-      "description": "semble-search — Tier 2 research. Code search agent for exploring any codebase using Semble's semantic search. For finding code by intent, locating implementations, and understanding how things work. Backed by DeepSeek V4 Flash. Operates inside BizarHarness alongside 13 sibling agents.",
+      "description": "semble-search — Tier 2 research. Code search agent for exploring any codebase using Semble's semantic search. For finding code by intent, locating implementations, and understanding how things work. Backed by MiniMax M2.7. Operates inside BizarHarness alongside 13 sibling agents.",
       "mode": "subagent",
-      "model": "opencode/deepseek-v4-flash-free",
+      "model": "minimax/MiniMax-M2.7",
       "color": "#0ea5e9",
       "permission": {
         "bash": "allow",
@@ -312,15 +319,15 @@
   },
   "provider": {
     "minimax": {
+      "options": {
+        "baseURL": "https://api.minimax.io/v1",
+        "apiKey": "{env:MiniMax_API_KEY}"
+      },
       "models": {
-        "MiniMax-M3": {
-          "interleaved": { "field": "reasoning_details" },
-          "reasoning": true
-        },
-        "MiniMax-M2.7": {
-          "interleaved": { "field": "reasoning_details" },
-          "reasoning": true
-        }
+        "MiniMax-M2.7-Flash": { "name": "MiniMax M2.7 Flash", "interleaved": { "field": "reasoning_details" }, "reasoning": true },
+        "MiniMax-M2.7": { "name": "MiniMax M2.7", "interleaved": { "field": "reasoning_details" }, "reasoning": true },
+        "MiniMax-M3": { "name": "MiniMax M3", "interleaved": { "field": "reasoning_details" }, "reasoning": true },
+        "MiniMax-M3-Reasoning": { "name": "MiniMax M3 Reasoning", "interleaved": { "field": "reasoning_details" }, "reasoning": true }
       }
     }
   }

package/config/opencode.json.template

@@ -28,7 +28,14 @@
       "loopThresholdWarn": 5,
       "loopThresholdEscalate": 8,
       "loopThresholdBlock": 12,
-      "loopWindowSize": 10
+      "loopWindowSize": 10,
+      "compaction": {
+        "auto": true,
+        "threshold": 0.5,
+        "preserve_recent": 10,
+        "strategy": "summarize",
+        "model": "minimax/MiniMax-M3-Reasoning"
+      }
     }]
   ],
   "tools": {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@polderlabs/bizar",
-  "version": "5.0.0",
+  "version": "5.0.2",
   "description": "Norse-pantheon multi-agent system for opencode — 13 agents across 4 cost tiers with cost-aware routing, plans, and a configurable agent harness. v4 ships as a single npm package bundling the dashboard server, opencode plugin, and typed SDK.",
   "type": "module",
   "bin": {

package/plugins/bizar/index.ts

@@ -143,6 +143,18 @@ import { executeSideEffect, type ExecuteOptions } from "./src/commands-impl.js";
 import { join as pathJoin } from "node:path";
 import { homedir } from "node:os";
 
+// v0.6.3 — Compaction gate. Configures the compaction threshold from
+// `rawOptions.compaction.threshold` (default 0.5) and wires two opencode
+// hooks: `experimental.session.compacting` (push a Bizar-policy note onto
+// the compaction prompt) and `experimental.compaction.autocontinue`
+// (skip the synthetic "continue" turn when our policy decides to compact).
+// The pure decision logic lives in `src/compaction.mjs`.
+import {
+  getCompactionThreshold,
+  setCompactionThreshold,
+  resetCompactionDefaults,
+} from "./src/compaction.mjs";
+
 // --- Env-var constants (per spec §8) -------------------------------------
 
 /** `BIZAR_SERVE_PORT` — default 0 (random). */
@@ -402,6 +414,29 @@ async function init(
   const logger = createLogger(input.client as unknown as Parameters<typeof createLogger>[0]);
   loggerHandle = logger;
 
+  // v0.6.3 — Compaction gate. Read the user-configured threshold from
+  // `rawOptions.compaction.threshold` (default 0.5). Invalid values are
+  // caught by `setCompactionThreshold` and logged; we fall back to the
+  // module default rather than crashing init (per spec §8.1: any init
+  // error must not break the plugin).
+  resetCompactionDefaults();
+  const compactionCfg = (rawOptions as Record<string, unknown> | undefined)
+    ?.compaction as { threshold?: unknown } | undefined;
+  if (compactionCfg && typeof compactionCfg.threshold === "number") {
+    try {
+      setCompactionThreshold(compactionCfg.threshold);
+      logger.info(
+        `bizar: compaction threshold set to ${compactionCfg.threshold} (from plugin options)`,
+      );
+    } catch (err) {
+      logger.warn(
+        `bizar: invalid compaction threshold ${String(compactionCfg.threshold)}; using default ${getCompactionThreshold()}: ${
+          err instanceof Error ? err.message : String(err)
+        }`,
+      );
+    }
+  }
+
   // §6.4 — refuse to start if logDir or stateDir is inside a secret dir.
   const offending = findOffendingPath(options);
   if (offending !== null) {
@@ -1412,6 +1447,48 @@ function buildHooks(ctx: RuntimeContext, bg: BgDeps): Hooks {
     // v0.4.2 — register the 4 background tools.
     tool: tools,
 
+    // v0.6.3 — Compaction gate. Called by opencode BEFORE it starts a
+    // session compaction. We append a one-line Bizar policy note so the
+    // model summarises with our threshold in mind (the default opencode
+    // compaction prompt doesn't know about our 50% trigger). The threshold
+    // itself was set during `init()` from `rawOptions.compaction.threshold`.
+    "experimental.session.compacting": async (_input, output) => {
+      try {
+        const threshold = getCompactionThreshold();
+        output.context.push(
+          `Bizar compaction policy: this compaction was triggered at the configured ${threshold} ratio of context usage. Preserve technical decisions, file paths, error messages, and unresolved questions verbatim in the summary; compress conversational filler.`,
+        );
+      } catch (err) {
+        ctx.logger.warn(
+          `bizar: experimental.session.compacting failed (passing through): ${
+            err instanceof Error ? err.message : String(err)
+          }`,
+        );
+      }
+    },
+
+    // v0.6.3 — Compaction auto-continue gate. Called by opencode AFTER
+    // compaction succeeds and BEFORE it injects a synthetic user "continue"
+    // message. We disable auto-continue when compaction was actually
+    // triggered (so the user sees the compacted state and can drive the
+    // next step), and let auto-continue proceed otherwise. We can't read
+    // the live usage ratio here, so we use the configured threshold as a
+    // proxy: if it's been lowered below 1.0, the policy is "compact
+    // early", and we want the user in the loop.
+    "experimental.compaction.autocontinue": async (_input, output) => {
+      try {
+        if (getCompactionThreshold() < 1.0) {
+          output.enabled = false;
+        }
+      } catch (err) {
+        ctx.logger.warn(
+          `bizar: experimental.compaction.autocontinue failed (passing through): ${
+            err instanceof Error ? err.message : String(err)
+          }`,
+        );
+      }
+    },
+
     // v0.4.2 — dispose hook. Opencode calls this when the plugin is
     // being torn down. We do a best-effort cleanup similar to the
     // signal trap, but we do NOT call `process.exit` — that's the

package/plugins/bizar/src/compaction.d.mts

@@ -0,0 +1,48 @@
+/**
+ * Type declarations for `./compaction.mjs`.
+ *
+ * The runtime module is shipped as plain ESM JavaScript (`.mjs`) per the
+ * brief, but its consumers in `index.ts` and `tests/compaction.test.ts`
+ * are TypeScript. Without this companion declaration file, TypeScript
+ * would type the exports as `any`, which causes cascading `implicitly
+ * has 'any' type` errors at call sites (e.g. the `texts` parameter in
+ * the test's `summarizer` stubs).
+ *
+ * The declarations here mirror the contract documented in
+ * `./compaction.mjs`. They are NOT loaded at runtime — they exist only
+ * for the type-checker.
+ */
+
+export interface CompactionResult {
+    compacted: boolean;
+    reason?: "below_threshold";
+    sessionId?: string;
+    threshold?: number;
+    preservedRecent?: number;
+    atMessages?: number;
+    ratio?: number | null;
+}
+
+export interface MaybeCompactSessionOpts {
+    sessionId: string;
+    messageCount: number;
+    currentUsage: { total?: number } | null | undefined;
+    maxContext: number | null | undefined;
+    summarizer: (texts: string[]) => Promise<string>;
+    preserveRecent?: number;
+}
+
+export function shouldCompact(
+    usage: { total?: number } | null | undefined,
+    maxContext: number | null | undefined,
+): boolean;
+
+export function maybeCompactSession(opts: MaybeCompactSessionOpts): Promise<CompactionResult>;
+
+export function getCompactionThreshold(): number;
+
+export function setCompactionThreshold(value: number): void;
+
+export function resetCompactionDefaults(): void;
+
+export function getDefaultPreserveRecent(): number;

package/plugins/bizar/src/compaction.mjs

@@ -0,0 +1,192 @@
+/**
+ * compaction.mjs
+ *
+ * Context-window compaction gate.
+ *
+ * Monitors a chat session's context-usage ratio and reports whether the
+ * session has reached the compaction threshold (default 50% of the model's
+ * max-context window). When the ratio crosses the threshold, callers can
+ * trigger summarization of the older message history while preserving a
+ * configurable recent tail so the model retains short-term continuity.
+ *
+ * Spec contract:
+ *   - The threshold is the *ratio* `usage.total / maxContext`, NOT an
+ *     absolute token count. This makes the gate model-agnostic — it works
+ *     whether the model has a 32k, 128k, or 1M context window.
+ *   - `shouldCompact` is a pure function: same inputs → same output, no I/O.
+ *   - `maybeCompactSession` is async to match the real summarizer signature
+ *     but performs no I/O itself when below threshold — it short-circuits.
+ *   - Threshold is mutable via `setCompactionThreshold` and is validated
+ *     against the range [0.1, 1.0]. A throw is the right signal here —
+ *     misconfiguration should fail loudly at startup, not silently use a
+ *     bad value at runtime.
+ *   - Threshold uses `let` (not `const`) so the setter can mutate it.
+ *     The reference spec used `const` and would throw a TypeError on
+ *     assignment; this implementation keeps the API but fixes the bug.
+ */
+
+const PRESERVE_RECENT_MESSAGES_DEFAULT = 10;
+const MIN_THRESHOLD = 0.1;
+const MAX_THRESHOLD = 1.0;
+
+// Module-level threshold. Exposed via get/set for tunability.
+// Exported as a number via getCompactionThreshold(); the underlying binding
+// is private (not exported) so callers cannot bypass validation by writing
+// to it directly.
+let compactionThreshold = 0.5;
+
+/**
+ * Returns `true` when `usage.total / maxContext` meets or exceeds the
+ * compaction threshold. `false` when below, or when either input is
+ * missing/invalid (defensive default).
+ *
+ * @param {{ total?: number } | null | undefined} usage
+ *   Object with a `total` token count, or null/undefined.
+ * @param {number | null | undefined} maxContext
+ *   Maximum context window size in tokens.
+ * @returns {boolean}
+ */
+export function shouldCompact(usage, maxContext) {
+  if (!usage || typeof usage.total !== "number") return false;
+  if (typeof maxContext !== "number" || maxContext <= 0) return false;
+  const ratio = usage.total / maxContext;
+  return ratio >= compactionThreshold;
+}
+
+/**
+ * Decision-shaped result returned by `maybeCompactSession`.
+ *
+ * - When below threshold: `{ compacted: false, reason: "below_threshold", ratio }`.
+ * - When at/over threshold: `{ compacted: true, sessionId, threshold,
+ *   preservedRecent, atMessages, ratio }`.
+ *
+ * @typedef {Object} CompactionResult
+ * @property {boolean} compacted
+ * @property {"below_threshold"} [reason]              when compacted=false
+ * @property {string} [sessionId]
+ * @property {number} [threshold]                     the threshold at decision time
+ * @property {number} [preservedRecent]               how many recent messages to keep verbatim
+ * @property {number} [atMessages]                    message count at decision time
+ * @property {number} [ratio]                         usage.total / maxContext
+ */
+
+/**
+ * Decide whether to compact and, if so, record the decision. The summarizer
+ * is NOT invoked here — that is the caller's responsibility. This function
+ * is the decision gate, not the executor. (The summarizer receives an
+ * `Array<string>` of older messages and returns a single summary string.)
+ *
+ * @param {Object}   opts
+ * @param {string}   opts.sessionId
+ * @param {number}   opts.messageCount                total messages in the session
+ * @param {{ total?: number } | null | undefined} opts.currentUsage
+ * @param {number | null | undefined} opts.maxContext
+ * @param {(texts: string[]) => Promise<string>} opts.summarizer
+ *   Async summarizer used by the caller when `compacted === true`. Not invoked
+ *   here. Accepting it as a parameter documents the contract for downstream
+ *   callers without forcing this function to do I/O.
+ * @param {number}   [opts.preserveRecent=10]         how many recent messages to keep verbatim
+ * @returns {Promise<CompactionResult>}
+ */
+export async function maybeCompactSession({
+  sessionId,
+  messageCount,
+  currentUsage,
+  maxContext,
+  // summarizer is accepted for API symmetry with downstream callers; we do
+  // not invoke it here because compaction execution lives in the plugin
+  // hook layer, not in this pure decision gate.
+  summarizer, // eslint-disable-line no-unused-vars
+  preserveRecent = PRESERVE_RECENT_MESSAGES_DEFAULT,
+}) {
+  if (typeof sessionId !== "string" || sessionId === "") {
+    // Misuse — caller forgot a sessionId. Refuse rather than silently
+    // returning below_threshold, which would mask a real bug upstream.
+    throw new Error("maybeCompactSession: sessionId is required");
+  }
+
+  if (!shouldCompact(currentUsage, maxContext)) {
+    const ratio = computeRatio(currentUsage, maxContext);
+    return { compacted: false, reason: "below_threshold", ratio };
+  }
+
+  const ratio = computeRatio(currentUsage, maxContext);
+  return {
+    compacted: true,
+    sessionId,
+    threshold: compactionThreshold,
+    preservedRecent: preserveRecent,
+    atMessages: messageCount,
+    ratio,
+  };
+}
+
+/**
+ * Return the current compaction threshold (0.0–1.0).
+ *
+ * @returns {number}
+ */
+export function getCompactionThreshold() {
+  return compactionThreshold;
+}
+
+/**
+ * Set the compaction threshold. Validates the new value is in the range
+ * `[0.1, 1.0]` and is a finite number. Throws on invalid input.
+ *
+ * The minimum of 0.1 prevents pathological "compact at 10%" configurations
+ * that would shred the conversation on the first reply. The maximum of 1.0
+ * means "compact only when the context is full" — anything above 1.0 would
+ * mean "never compact", which `shouldCompact` already handles correctly, so
+ * we reject it for clarity.
+ *
+ * @param {number} value
+ * @returns {void}
+ */
+export function setCompactionThreshold(value) {
+  if (typeof value !== "number" || !Number.isFinite(value)) {
+    throw new Error(`Invalid compaction threshold: ${value} (must be a finite number)`);
+  }
+  if (value < MIN_THRESHOLD || value > MAX_THRESHOLD) {
+    throw new Error(
+      `Invalid compaction threshold: ${value} (must be in [${MIN_THRESHOLD}, ${MAX_THRESHOLD}])`,
+    );
+  }
+  compactionThreshold = value;
+}
+
+/**
+ * Reset the threshold and preserve-count to their defaults. Exposed for
+ * tests and for plugin re-init (so a hot-reload doesn't carry stale state).
+ *
+ * @returns {void}
+ */
+export function resetCompactionDefaults() {
+  compactionThreshold = 0.5;
+}
+
+/**
+ * Return the number of recent messages that the default policy preserves
+ * verbatim when compacting.
+ *
+ * @returns {number}
+ */
+export function getDefaultPreserveRecent() {
+  return PRESERVE_RECENT_MESSAGES_DEFAULT;
+}
+
+// --- Internal helpers -----------------------------------------------------
+
+/**
+ * Compute `usage.total / maxContext`, or `null` when either input is
+ * missing/invalid. Returns a number in `[0, +Infinity)` for valid inputs.
+ *
+ * @param {{ total?: number } | null | undefined} usage
+ * @param {number | null | undefined} maxContext
+ * @returns {number | null}
+ */
+function computeRatio(usage, maxContext) {
+  if (!usage || typeof usage.total !== "number") return null;
+  if (typeof maxContext !== "number" || maxContext <= 0) return null;
+  return usage.total / maxContext;
+}