bullswarm 0.23.0 → 0.23.2

package/README.md

@@ -256,6 +256,17 @@ updates the requirement ledger from schema-valid evidence. Only real
 consolidated gaps re-enter the planner. There are no formal reviewer or repair
 roles and no automatic semantic repair/reverify loop.
 
+Lane and effort are separate decisions for every proposed action. `analyze` is
+read-only investigation, judgment, or evidence; `build` is contextual product,
+test, or documentation mutation; `chore` is deterministic mechanical mutation.
+The kernel rejects evidence outside `analyze`, file ownership inside `analyze`,
+and any `chore` above low effort. Low is for fixed-procedure checks and edits,
+medium is the default for ordinary bounded work, and high is reserved for
+architecture, ambiguous tradeoffs, cross-cutting integration, or genuinely
+adversarial acceptance judgment. Merely being an analysis/evidence action or
+part of a difficult goal never promotes an action to high. The selected effort
+then resolves through the High/Medium/Low routes configured by `bullswarm setup`.
+
 The planner does not author phases or declare success/failure. The kernel
 derives stable presentation stages for the TUI and computes the final V2
 result. Old autonomous run directories are not migrated or resumed;

package/connectors/command-code.json

@@ -46,13 +46,16 @@
   },
   "costRank": 1,
   "lanes": [
+    "analyze",
     "build",
     "chore"
   ],
-  "capabilities": ["code-reading", "file-editing"],
-  "modelDiscovery": { "cmd": ["command-code", "--list-models"], "parse": "columns", "ignorePattern": "^(Available|Open Source|Anthropic|OpenAI|Google|Sakana|Meta|xAI|Pass|Docs)", "timeoutMs": 20000, "maxModels": 150 },
+  "capabilities": ["code-reading", "file-editing", "strong-analysis", "workflow-planning"],
+  "modelDiscovery": { "cmd": ["command-code", "--list-models"], "parse": "columns", "ignorePattern": "^(Available|Open Source$|Anthropic$|OpenAI$|Google$|Sakana$|Meta$|xAI$|Pass|cmd|Docs)", "timeoutMs": 20000, "maxModels": 150 },
   "modelSelection": { "flag": "--model", "mode": "replace-or-append" },
   "modelProfiles": [
+    { "match": "^meta/muse-spark-1\\.3-contributor$", "tier": "high", "qualityRank": 5, "pricing": { "inputUsdPerMillion": 0.1, "cacheReadUsdPerMillion": 0.002, "outputUsdPerMillion": 0.2 }, "pricingSource": "https://commandcode.ai/docs/resources/pricing-limits", "pricingUpdatedAt": "2026-09-03" },
+    { "match": "^meta/muse-spark-", "tier": "high", "qualityRank": 5, "pricing": { "inputUsdPerMillion": 1.25, "cacheReadUsdPerMillion": 0.15, "outputUsdPerMillion": 4.25 }, "pricingSource": "https://commandcode.ai/docs/resources/pricing-limits", "pricingUpdatedAt": "2026-09-03" },
     { "match": "^gpt-5\\.6-sol$", "tier": "high", "qualityRank": 6, "pricing": { "inputUsdPerMillion": 4, "cacheReadUsdPerMillion": 0.4, "outputUsdPerMillion": 20 }, "pricingSource": "https://help.openai.com/en/articles/20001415-chatgpt-rate-card-enterprise-token-based-pricing", "pricingUpdatedAt": "2026-08-27" },
     { "match": "^gpt-5\\.6-terra$", "tier": "medium", "qualityRank": 4, "pricing": { "inputUsdPerMillion": 2, "cacheReadUsdPerMillion": 0.2, "outputUsdPerMillion": 12 }, "pricingSource": "https://developers.openai.com/api/docs/models/gpt-5.6-terra", "pricingUpdatedAt": "2026-08-27" },
     { "match": "^gpt-5\\.6-luna$", "tier": "low", "qualityRank": 3, "pricing": { "inputUsdPerMillion": 0.2, "cacheReadUsdPerMillion": 0.02, "outputUsdPerMillion": 1.2 }, "pricingSource": "https://developers.openai.com/api/docs/models", "pricingUpdatedAt": "2026-08-27" },
@@ -65,7 +68,6 @@
     { "match": "(?:fable|opus|sol|grok-4\\.6|minimax-m3)", "tier": "high", "qualityRank": 5 }
   ],
   "subscription": { "plan": null, "monthlyPriceUsd": null, "includedValueUsd": null, "quotaWindow": "monthly" },
-  "$comment-lanes": "cannot serve analyze (upstream 403 on the frontier model)",
   "flags": {
     "stealth": false
   },

package/data/openrouter-benchmarks.json

@@ -1,6 +1,6 @@
 {
   "schemaVersion": "bullswarm.openrouter.benchmarks.v1",
-  "capturedAt": "2026-09-02T01:02:39.330Z",
+  "capturedAt": "2026-09-02T07:52:24.004Z",
   "upstream": {
     "benchmarks": "https://openrouter.ai/api/v1/benchmarks",
     "models": "https://openrouter.ai/api/v1/models",
@@ -170,6 +170,20 @@
       },
       "pricingSource": "https://openrouter.ai/api/v1/models"
     },
+    "~z-ai/glm-flash-latest": {
+      "id": "~z-ai/glm-flash-latest",
+      "name": "Z.ai: GLM Flash Latest",
+      "created": 1787817633,
+      "indices": {},
+      "ranks": {},
+      "pricing": {
+        "inputUsdPerMillion": 0.075,
+        "cacheReadUsdPerMillion": 0.015,
+        "cacheWriteUsdPerMillion": null,
+        "outputUsdPerMillion": 0.25
+      },
+      "pricingSource": "https://openrouter.ai/api/v1/models"
+    },
     "~z-ai/glm-latest": {
       "id": "~z-ai/glm-latest",
       "name": "Z.ai: GLM Latest",
@@ -744,6 +758,20 @@
       },
       "pricingSource": "https://openrouter.ai/api/v1/models"
     },
+    "anthropic/claude-fable-5.1:batch": {
+      "id": "anthropic/claude-fable-5.1:batch",
+      "name": "Anthropic: Claude Fable 5.1 (batch)",
+      "created": 1788285838,
+      "indices": {},
+      "ranks": {},
+      "pricing": {
+        "inputUsdPerMillion": 5,
+        "cacheReadUsdPerMillion": 0.125,
+        "cacheWriteUsdPerMillion": 6.25,
+        "outputUsdPerMillion": 25
+      },
+      "pricingSource": "https://openrouter.ai/api/v1/models"
+    },
     "anthropic/claude-haiku-4.5": {
       "id": "anthropic/claude-haiku-4.5",
       "name": "Anthropic: Claude Haiku 4.5",
@@ -1603,10 +1631,10 @@
       "indices": {},
       "ranks": {},
       "pricing": {
-        "inputUsdPerMillion": 0.22,
-        "cacheReadUsdPerMillion": 0.007,
+        "inputUsdPerMillion": 0.44,
+        "cacheReadUsdPerMillion": 0.014,
         "cacheWriteUsdPerMillion": null,
-        "outputUsdPerMillion": 0.66
+        "outputUsdPerMillion": 1.32
       },
       "pricingSource": "https://openrouter.ai/api/v1/models"
     },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "bullswarm",
-  "version": "0.23.0",
+  "version": "0.23.2",
   "description": "Route work across coding-agent CLI subscriptions — paced by live quota meters, verified by content, never trusting exit codes.",
   "type": "module",
   "bin": {

package/src/workflow/action-validator.js

@@ -5,6 +5,11 @@ export const ACTION_PROGRAM_SCHEMA_VERSION = 'bullswarm.workflow.program.v2';
 const ID_RE = /^[a-z0-9][a-z0-9-]*$/;
 const LANES = new Set(['analyze', 'build', 'chore']);
 const EFFORTS = new Set(['high', 'medium', 'low']);
+export const DEFAULT_EFFORT_BY_LANE = Object.freeze({
+  analyze: 'medium',
+  build: 'medium',
+  chore: 'low',
+});
 const PROGRAM_FIELDS = new Set(['schemaVersion', 'actions']);
 const ACTION_FIELDS = new Set([
   'id', 'purpose', 'dependsOn', 'affects', 'ownedFiles', 'prompt',
@@ -241,6 +246,7 @@ function maxParallelism(actions, byId) {
  */
 export function validateActionProgram(program, runtime = {}) {
   const issues = [];
+  const enforceRoutingPolicy = runtime.enforceRoutingPolicy !== false;
   if (!isObject(program)) throw new ActionValidationError(['program must be an object']);
   for (const key of Object.keys(program)) if (!PROGRAM_FIELDS.has(key)) issues.push(`program.${key} is not allowed`);
   if (program.schemaVersion !== ACTION_PROGRAM_SCHEMA_VERSION) {
@@ -284,6 +290,15 @@ export function validateActionProgram(program, runtime = {}) {
     const ownedFiles = normalizeOwnedFiles(action.ownedFiles, `${at}.ownedFiles`, issues);
     if (!LANES.has(action.lane)) issues.push(`${at}.lane must be analyze|build|chore`);
     if (!EFFORTS.has(action.effort)) issues.push(`${at}.effort must be high|medium|low`);
+    if (enforceRoutingPolicy && evidenceFor.length && action.lane !== 'analyze') {
+      issues.push(`${at} evidence actions must use lane analyze`);
+    }
+    if (enforceRoutingPolicy && ownedFiles.length && action.lane === 'analyze') {
+      issues.push(`${at} analyze actions must not own workspace files; use build or chore for mutations`);
+    }
+    if (enforceRoutingPolicy && action.lane === 'chore' && action.effort !== 'low') {
+      issues.push(`${at} chore actions are deterministic mechanical work and must use low effort`);
+    }
     if (evidenceFor.length && (affects.length || ownedFiles.length)) {
       issues.push(`${at} evidence actions must have empty affects and ownedFiles`);
     }

package/src/workflow/v2-dispatch.js

@@ -4,8 +4,8 @@ import { pickPool, isQuarantined } from '../lib/route.js';
 import { assertDepthAllowed, childDepthEnv, loadState, quarantinePool, saveState } from '../lib/state.js';
 import { disabledModelsForPool, resolveDispatchModel, selectedModelsForTier } from '../lib/strategy.js';
 import { watchOnce } from '../lib/watch.js';
+import { DEFAULT_EFFORT_BY_LANE } from './action-validator.js';
 
-const EFFORT_BY_LANE = Object.freeze({ analyze: 'high', build: 'medium', chore: 'low' });
 const MECHANICAL_KINDS = new Set(['auth', 'provider', 'process', 'interrupted', 'schema']);
 
 function clone(value) {
@@ -154,7 +154,7 @@ export async function dispatchV2Action({
   const saveCoreState = dependencies.saveState ?? saveState;
   const now = dependencies.now ?? Date.now;
   const uuid = dependencies.uuid ?? randomUUID;
-  const effort = action.effort ?? EFFORT_BY_LANE[action.lane] ?? 'low';
+  const effort = action.effort ?? DEFAULT_EFFORT_BY_LANE[action.lane] ?? 'medium';
   let candidates = preparePools(pools, action, effort, {
     avoidPools, preferredModel, strictPool, now: now(),
   });

package/src/workflow/v2-planner.js

@@ -170,6 +170,9 @@ export function buildV2PlannerPrompt(context) {
     'The numeric values in context.targets are advisory planning targets, never execution ceilings. Prefer to stay within them by consolidating optional work, but exceed them whenever the smallest essential program needs more actions, agent dispatches, or gap rounds. Reaching or crossing a target is not a reason to return exhausted.',
     'context.execution.concurrency limits only how many dependency-ready actions run at once. It does not limit the total number of independent actions in the program; the scheduler will batch wider programs safely.',
     'Use only generic actions. A work action declares affects and any exact ownedFiles. affects means the action directly owns and delivers a bounded acceptance slice of that requirement; merely editing a supporting test or sharing a file does not make an action affect every requirement associated with that file. An evidence action declares evidenceFor, has empty affects/ownedFiles, and independently inspects the work it judges.',
+    'Choose lane from the action itself, not from the overall goal: analyze is read-only investigation, judgment, or evidence; build changes behavior, documentation, or tests and requires contextual implementation; chore is only deterministic mechanical mutation with no design choice. Evidence actions must use analyze. Analyze actions cannot own files. Chore actions must use low effort.',
+    'Choose effort independently from lane, using the cheapest tier sufficient for this one action. Low is for fixed-procedure checks or mechanical edits whose success is objectively decidable. Medium is the default for normal bounded analysis or implementation with local decisions. High is exceptional: use it only when architecture, ambiguous tradeoffs, cross-cutting integration, or adversarial acceptance judgment materially determines correctness. If uncertain, choose medium.',
+    'Do not choose high merely because an action uses analyze, supplies evidence, affects an important requirement, mentions many files, or belongs to a difficult overall goal. Do not choose low merely because an action is short. Examples: exact file comparison or formatting update = low; ordinary scoped feature plus focused test = medium; choosing an architecture across subsystems = high; running deterministic acceptance commands = low; interpreting ambiguous cross-cutting acceptance evidence = high.',
     'Dependencies represent required data or exact-file ordering only. Do not serialize unrelated work. Do not add reviewer, verify, repair, phase, completion, pool, model, timeout, or retry fields.',
     'Every mandatory unresolved requirement needs an evidence action. Parallel actions must be both file-disjoint and acceptance-independent. Isolated parallel siblings cannot see each other\'s unintegrated changes. If one action writes tests for behavior introduced by another action, combine code and tests under one owner or make the test action depend on and consume an artifact from the implementation action; never run new behavioral tests against the unchanged baseline in parallel. Prompts must be self-contained and include exact scope plus acceptance evidence.',
     'For mutating behavioral work, keep implementation and its focused regression test under one coherent owner. The action must prove the regression on the untouched baseline, then exercise the real production entry point or state transition after the change; disconnected helpers, no-op assertions, and test-only behavior do not satisfy acceptance.',

package/src/workflow/v2-state.js

@@ -380,6 +380,10 @@ function validateProgram(program, state) {
         maxParallel: state.config.settings.concurrency ?? state.config.settings.maxParallel ?? 100,
         enforceMaxActions: false,
         enforceMaxParallel: false,
+        // Durable runs predate the current routing policy. Replay their exact
+        // accepted program; strict lane/effort policy applies to new planner
+        // candidates before dispatch, not historical display or resume.
+        enforceRoutingPolicy: false,
         },
       );
       for (const action of actions) {