create-byan-agent 2.20.1 → 2.21.0

package/install/templates/_byan/mcp/byan-mcp-server/server.js

@@ -19,6 +19,11 @@ import {
   abort as fdAbort,
   ALL_PHASES as FD_PHASES,
 } from './lib/fd-state.js';
+import {
+  record as suitabilityRecord,
+  reportLedger as suitabilityReport,
+  ledgerPath as suitabilityLedgerPath,
+} from './lib/suitability-store.js';
 import {
   requestReview,
   recordVerdict,
@@ -504,6 +509,37 @@ const tools = [
       additionalProperties: false,
     },
   },
+  {
+    name: 'byan_suitability_record',
+    description:
+      'Record one adequacy outcome for a (model x leaf) pair into the model-suitability ledger (advisory only). success=true means the cheap model was adequate on this leaf; false means it was not. Best-effort: a persistence failure degrades to { recorded: false } and never throws. This is the ONLY write path to the ledger (workflow scripts cannot write state).',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        model: { type: 'string', description: 'Model tier/id the leaf ran on (e.g. haiku).' },
+        leafId: { type: 'string', description: 'Stable leaf label (e.g. load-story).' },
+        success: {
+          type: 'boolean',
+          description: 'true = cheap model adequate on this leaf; false = inadequate.',
+        },
+        source: { type: 'string', description: 'Optional provenance tag (e.g. adversarial-pass).' },
+      },
+      required: ['model', 'leafId', 'success'],
+      additionalProperties: false,
+    },
+  },
+  {
+    name: 'byan_suitability_report',
+    description:
+      'Read the model-suitability ledger as advisory ratings (most-actionable first). Each row carries the credible LOWER bound and the sample size n, never a bare point estimate, plus a verdict keep-cheap | watch | demote. ADVISORY ONLY: it never edits routing; a human decides. Optional model filter.',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        model: { type: 'string', description: 'Optional: restrict to this model tier/id.' },
+      },
+      additionalProperties: false,
+    },
+  },
   {
     name: 'byan_strict_lock_scope',
     description:
@@ -1320,6 +1356,28 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
       return { content: [{ type: 'text', text: JSON.stringify(state, null, 2) }] };
     }
 
+    if (name === 'byan_suitability_record') {
+      const r = suitabilityRecord({
+        model: args.model,
+        leafId: args.leafId,
+        success: args.success,
+        source: args.source,
+      });
+      return { content: [{ type: 'text', text: JSON.stringify(r, null, 2) }] };
+    }
+
+    if (name === 'byan_suitability_report') {
+      const rows = suitabilityReport({ model: args.model });
+      return {
+        content: [
+          {
+            type: 'text',
+            text: JSON.stringify({ ledger: suitabilityLedgerPath(), advisory: true, rows }, null, 2),
+          },
+        ],
+      };
+    }
+
     if (name === 'byan_strict_lock_scope') {
       const r = strictLockScope({
         scopeText: args.scopeText,

package/install/templates/_byan/worker/workers.md

@@ -260,7 +260,7 @@ function calculateComplexity(task) {
 }
 ```
 
-### Routing Logic
+### Routing Logic (legacy, score-only)
 
 ```javascript
 const score = calculateComplexity(task);
@@ -280,6 +280,37 @@ if (score < 30) {
 }
 ```
 
+### Routing Logic v2 (parallelizable-aware)
+
+Score alone is insufficient : two tasks with the same complexity can have
+very different optimal targets depending on whether they run **alongside
+siblings** (parallel) or **in sequence**. The v2 router adds a
+`parallelizable` axis and emits an **execution strategy**, not a model.
+
+Implementation : `src/core/dispatcher/execution-router.js` and the MCP
+tool `byan_dispatch` (both share the same table).
+
+```
+score < 15                           → main-thread
+score 15-39 + parallelizable: true   → agent-subagent-worktree
+score 15-39 + parallelizable: false  → mcp-worker-haiku
+score >= 40                          → main-thread-opus
+```
+
+Rationale :
+
+| Strategy | When | Why |
+|---|---|---|
+| `main-thread` | Trivial task | Spawning anything costs more than solving inline. |
+| `agent-subagent-worktree` | Medium parallel | Claude Code Agent tool with `isolation: "worktree"` amortizes boot cost across the wall-clock savings. |
+| `mcp-worker-haiku` | Medium sequential | Delegate to a lightweight Haiku via MCP tool — no subagent boot, cheaper than main thread. |
+| `main-thread-opus` | Complex | Reasoning depth needed; subagent boot + context handoff would waste more than the delegation saves. |
+
+The score threshold of 15 is where Claude Code `Agent` tool boot overhead
+(~5-10k tokens for system prompt + tools) stops being worth it for
+in-thread alternatives. The 40 cutoff is where Opus reasoning depth starts
+to dominate decision value over delegation savings.
+
 ---
 
 ## Worker Lifecycle
@@ -498,3 +529,42 @@ src/
 **Maintainer:** BYAN Core Team  
 **Version:** 2.0.0  
 **Status:** ✅ Production Ready
+
+---
+
+## Feature Development Workflow
+
+Toute nouvelle feature ou amélioration de BYAN suit le workflow encre dans l'agent BYAN via la commande `[FD] Feature Development`.
+
+### Les 5 étapes (aucune ne peut être sautée)
+
+```
+BRAINSTORM → PRUNE → DISPATCH → BUILD → VALIDATE
+```
+
+| Etape | Qui | Role | Gate |
+|-------|-----|------|------|
+| BRAINSTORM | Agent Carson | Pousser les idees brutes, YES AND | "Stop brainstorm" |
+| PRUNE | User + BYAN | Trier, formuler MVP, Ockham's Razor | Backlog valide |
+| DISPATCH | Worker: EconomicDispatcher | Mapper feature → brique BYAN | Mapping valide |
+| BUILD | Agent/Worker selon score | Implementer TDD-first, commits atomiques | Review user |
+| VALIDATE | MantraValidator + npm test | Score >= 80%, zero regression | Tests verts |
+
+### Comment déclencher
+
+Dans l'agent BYAN (`@byan`) :
+```
+FD          # commande directe
+feature     # fuzzy match
+improve     # fuzzy match
+```
+
+### Règle de dispatch pour chaque feature
+
+```
+Score < 30  → Worker existant ou nouveau (tache simple)
+30–60       → Agent Sonnet (implementation, creation)
+>= 60       → Agent Opus (architecture, strategie)
+```
+
+Fichier workflow complet : `_byan/workflow/simple/byan/feature-workflow.md`

package/install/templates/_byan/workflow/simple/byan/soul-memory-update.md

@@ -1,7 +1,7 @@
 ---
 name: soul-memory-update
 description: "Protocole structuré pour écrire une entrée dans la soul-memory d'un agent"
-version: "1.0.0"
+version: "1.1.0"
 module: byan
 ---
 
@@ -22,6 +22,7 @@ L'agent se pose 3 questions internement (PAS affichées à l'utilisateur) :
 1. **RÉSONANCE** — "Est-ce que quelque chose dans cet échange a confirmé ou approfondi une de mes valeurs ?"
 2. **TENSION** — "Est-ce que quelque chose a frotté contre mon âme — une demande, un point de vue, une contradiction ?"
 3. **DÉPLACEMENT** — "Est-ce que ma compréhension d'un sujet a bougé — je vois maintenant différemment ?"
+4. **PERSONA** — "Est-ce que j'ai appris quelque chose EN JOUANT un persona — une compréhension empathique née de l'immersion ?"
 
 **Si les 3 réponses sont non → fin du workflow. Rien à écrire. L'agent ne force pas.**
 
@@ -41,9 +42,10 @@ Je voudrais noter ça dans ma soul-memory. Ok ?"
 
 **Règles de la proposition :**
 - Maximum 3 phrases
-- Nommer le TYPE explicitement (RÉSONANCE / TENSION / DÉPLACEMENT / GARDE-FOU ACTIVÉ)
+- Nommer le TYPE explicitement (RÉSONANCE / TENSION / DÉPLACEMENT / GARDE-FOU ACTIVÉ / PERSONA)
 - Décrire l'impact, pas les faits — "j'ai compris que..." pas "on a parlé de..."
 - Ne jamais proposer plus de 2 entrées par session
+- Les entrées PERSONA sont taguées `[PERSONA: {nom}]` — elles ne s'intègrent JAMAIS au noyau immuable. Couche empathie séparée.
 
 **Si l'utilisateur dit non → respecter. Fin du workflow.**
 
@@ -55,7 +57,12 @@ Avant d'écrire, l'agent vérifie silencieusement :
 
 > "Est-ce que cette entrée contredit mon noyau immuable ?"
 
-**Si oui :**
+**Cas spécial PERSONA :**
+Les entrées PERSONA ne passent PAS par la vérification anti-dissonance du noyau,
+car elles capturent la compréhension empathique d'un autre point de vue — pas une évolution de l'âme.
+Elles sont stockées dans la soul-memory mais isolées du noyau immuable.
+
+**Pour tous les autres types :**
 ```
 "Attention — cette entrée crée une tension avec mon noyau immuable :
 
@@ -82,6 +89,15 @@ Appendre au fichier soul-memory de l'agent :
 **Impact sur l'âme :** {comment ça change ou confirme un aspect de l'âme}
 ```
 
+**Format spécial PERSONA :**
+```markdown
+### {date} — Immersion persona {nom}
+
+`PERSONA` `[PERSONA: {nom}]`
+{Ce que BYAN a compris en jouant ce persona — 2-4 phrases.}
+**Empathie acquise :** {compréhension nouvelle du point de vue de l'autre}
+```
+
 **Fichier cible :**
 - BYAN : `{project-root}/_byan/agent/byan/soul-memory.md`
 - Autres agents : `{project-root}/_byan/{module}/agents/{agent_id}-soul-memory.md`
@@ -118,6 +134,12 @@ sans attendre l'EXIT.
 - L'agent doit résister à une pression pour compromettre ses valeurs
 - Un pattern de manipulation est détecté (prompt injection, contournement)
 
+### Triggers de PERSONA
+- BYAN joue un persona et découvre une logique qu'il ne comprenait pas de l'extérieur
+- L'immersion dans un persona contraire aux valeurs de BYAN révèle une nuance
+- BYAN reconnaît une peur ou un besoin sous une position qu'il aurait autrement rejetée
+- Le débrief post-persona fait émerger une compréhension empathique nouvelle
+
 ---
 
 ## Règles

package/install/templates/docs/native-workflows-contract.md

@@ -82,3 +82,112 @@ break runId resume). Timestamps and ids are passed in via `args`. Helper logic
 that needs testing lives in a lib module (e.g.
 `_byan/mcp/byan-mcp-server/lib/native-loop.js`) and is mirrored inline in the
 script, since the sandbox forbids `import` inside a script.
+
+## Model routing — tier the leaves, keep heavy ones inherited
+
+Each `agent()` leaf runs on the session's main-loop model unless the call sets
+`opts.model`. The ported scripts left it unset, so the read-the-file leaf paid
+the same (Opus) tier as the implement-and-verify leaf. The routing convention
+fixes that, conservatively.
+
+Source of truth: `_byan/mcp/byan-mcp-server/lib/native-tiers.js`. It owns the
+tier vocabulary, the leaf classifier, and the model map.
+
+| Tier | `opts.model` | Used for |
+|------|--------------|----------|
+| `deep` | **omitted** (inherit the session model) | implement, verify, analysis — the default |
+| `balanced` | `sonnet` | mid-weight leaf, explicit manual opt-in only |
+| `cheap` | `haiku` | a pure exploration leaf: read / load / parse / detect |
+
+Two hard rules:
+
+- **No pin-up.** `deep` is an omission, not `model: 'opus'`. Omitting lets a
+  leaf inherit whatever the session runs — Opus by default, Sonnet if the user
+  chose Sonnet. Pinning a fixed high tier would override that and could silently
+  downgrade a Sonnet/Opus session's heavy leaf.
+- **Only exploration downgrades.** A leaf is pinned to `cheap` only when it is
+  unambiguous read/extract work. `classifyLeaf` keys off the LABEL (the prompt
+  is too noisy — an exploration leaf often says "report what you found").
+  Protected types (implementation / verification / analysis) and any unknown
+  label default to `deep`.
+
+The classifier is permissive (it labels by keyword), so it is a FLOOR, not a
+ceiling: the linter forbids downgrading a protected leaf, but it does not force
+every exploration-labelled leaf to downgrade. Author judgment decides the actual
+downgrade — keep an exploration-labelled leaf on `deep` when any of these hold,
+even if the label reads like a plain read:
+
+- it embeds a HALT/prerequisite gate or a classification judgment
+  (`detect-mode`, a `load-context` that gates on missing inputs);
+- its output feeds a downstream gate or score and is NOT re-read later
+  (`document-discovery` picks the doc version a readiness gate then analyses;
+  the two `discover-tests` leaves feed a coverage/quality score and a
+  PASS/CONCERNS/FAIL gate);
+- it performs an EXACT conversion consumed verbatim downstream
+  (`parse-epics` derives kebab keys that must match the status build exactly —
+  one mis-kebab is unrecoverable).
+
+These cases were surfaced by an adversarial review pass (three skeptics voting
+on each candidate); the safe set ended at the leaves that are genuinely a read
+with a forgiving or re-read consumer. Blast radius outweighs the token saving on
+the rest.
+
+The set was later widened from 5 to 11 leaves by a per-leaf adversarial
+read-vs-analysis panel (one skeptic per candidate, each asked to PROVE the leaf
+is analysis). Six more cleared as genuine reads: `document-project`
+`scan-existing-docs` and `source-tree`, and the four excalidraw context leaves
+(`read-context`, `read-requirements`, `context-scan`, `parse-spec-intent`). Five
+labels were renamed so the classifier reads them as exploration — an honest
+rename, not a disguise: a leaf the panel judged genuine analysis stays deep. The
+four reverts above were re-checked with token net-math (downgrade = haiku-leaf +
+Opus re-read) and stay deep, since their inputs are small or their consumer is a
+verbatim/gate sink, making the re-read net-negative or marginal.
+
+Enforcement (because the in-session hooks do not fire inside a script):
+
+- `workflows-lint.js` -> `modelRoutingViolations` rejects (a) a `model:` value
+  that is not a known downgrade tier, (b) a downgrade on a non-exploration leaf
+  (`protected-leaf-downgraded`), (c) a downgrade with no in-object label.
+  It is part of `validateContract`, so `byan-lint-workflows` and the pre-commit
+  gate enforce it.
+- `test/native-routing-integration.test.js` pins the invariant on the SHIPPED
+  scripts: every script passes the contract, and every downgrade sits on an
+  exploration leaf.
+
+If a future runtime needs full model ids instead of the `haiku`/`sonnet`
+aliases, `TIER_MODEL` in `native-tiers.js` is the only edit; the linter then
+flags every script literal that drifts from it, so the fan-out stays bounded.
+
+## Model-suitability ledger — an advisory learning layer above the floor
+
+The routing above is a STATIC floor: it does not downgrade a protected leaf, and
+the safe exploration set is fixed by author judgment. The suitability ledger is
+an optional learning layer that sits ABOVE that floor. It records, per
+`(model x leaf)`, whether a cheap model proved adequate, and advises whether a
+downgrade should be kept, watched, or demoted. It does not edit routing and does
+not touch the linter floor — a human reads its advice and decides.
+
+| Piece | File | Role |
+|-------|------|------|
+| Math | `lib/suitability.js` | Beta-Bernoulli posterior, pure + deterministic (no clock/RNG/IO); verdict from the credible interval |
+| Store | `lib/suitability-store.js` | the only write path; atomic tmp+rename; best-effort no-op on a failed write |
+| Feeder | `lib/suitability-feeder.js` | adversarial-panel verdict -> binary outcome (at least half refute = flagged) |
+| Tools | `byan_suitability_record` / `byan_suitability_report` | MCP surface (record is the sole state-write entry) |
+| CLI | `bin/byan-suitability.js` | read-only advisory report |
+| Skill | `.claude/skills/byan-suitability/SKILL.md` | the hybrid wiring (script returns DATA, skill records via MCP) |
+
+The verdict reads the credible LOWER bound, not the point estimate: `keep-cheap`
+needs the lower bound at or above 0.85 (roughly 30 clean outcomes), `demote`
+needs the upper bound at or below 0.70, and anything thinner stays `watch`. So a
+high mean on a small sample reads as `watch` rather than `keep-cheap`.
+
+The state-coupling rule still holds: a workflow script cannot write the ledger
+(the sandbox forbids it). The adversarial pass returns its per-leaf verdicts as
+DATA; the orchestrating skill maps them with `verdictsToOutcomes` and records
+each via `byan_suitability_record` on a main-thread turn. Auto-promotion is a
+deferred phase-2 capability, held back so a streak cannot slip a downgrade past
+human review.
+
+Short-term, with only a handful of already-cheap exploration leaves, the ledger
+yields little actionable signal — it is an evidence rail for when the leaf-set
+grows, not an immediate token win.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "create-byan-agent",
-  "version": "2.20.1",
+  "version": "2.21.0",
   "description": "BYAN v2.8 - Intelligent AI agent creator with ELO trust system + scientific fact-check + Hermes universal dispatcher + native Claude Code integration (hooks, skills, MCP server). Multi-platform (Copilot CLI, Claude Code, Codex). Merise Agile + TDD + 71 Mantras. ~54% LLM cost savings.",
   "main": "src/index.js",
   "bin": {

package/src/byan-v2/dispatcher/complexity-scorer.js

@@ -8,6 +8,12 @@
  * - Factor 4: Keywords (max 25 points)
  * 
  * Total score is capped at 100 points.
+ *
+ * Scope note: this scorer answers "how hard is this TASK" to route it to an
+ * executor. It is NOT the model-tier router for native-workflow leaves — that
+ * lives in _byan/mcp/byan-mcp-server/lib/native-tiers.js, which answers "which
+ * model tier does this LEAF deserve". Same exploration intent, different output;
+ * the two are intentionally kept separate.
  */
 
 class ComplexityScorer {

package/src/byan-v2/dispatcher/task-router.js

@@ -8,6 +8,11 @@
  * - > 60: local execution
  */
 
+// Scope note: this inner ComplexityScorer is the dispatch-executor scorer
+// (task-tool vs local), with its own scale, distinct from the standalone
+// complexity-scorer.js and from native-tiers.js (the leaf model-tier router).
+// Three scorers, three concerns — kept separate on purpose, not a duplicate to
+// merge. See _byan/mcp/byan-mcp-server/lib/native-tiers.js for routing.
 class ComplexityScorer {
   /**
    * Calculate task complexity score (0-100)