@chatpanel/events 0.85.0 → 0.88.0

package/route-strategies.js

@@ -0,0 +1,232 @@
+// HOW A MODEL IS CHOSEN — the strategies and the one step that is never optional.
+//
+// The routing CONTRACT (createModelRouter, signals, requirements, failover order) has lived
+// in this package since it was written; the DECISIONS layered on it — escalate hard work,
+// replace like with like, honour "use claude" — were typed in the extension, so the desktop
+// was about to copy them. They are pure functions of the candidates and the need, so they
+// belong here, and both clients build the same router from the same declarations.
+//
+// `needForTurn` builds the need from the turn's facts. The source guard (which page or
+// record the turn is about, and what that caps reach at) is INJECTED as `guard`, because
+// reading it means reading a client's settings; the ceiling itself is enforced here.
+
+import {
+  defineMiddleware, defineRouteStrategy, signalsFrom, requirementsFor, preferenceFor,
+  failoverOrder, pinnedOrderOf,
+} from './router.js';
+
+/**
+ * Redaction is REQUIRED for anything that leaves the user's machine.
+ *
+ * Declared as middleware with `requiredFor` so the router refuses to route to a third party
+ * when it is not active. That is the difference between "we always redact" as a habit and as
+ * a property: a disabled plugin, a refactor or a new caller cannot quietly skip it.
+ */
+export const redactionStep = defineMiddleware({
+  id: 'redaction',
+  label: 'Redaction',
+  stage: 'request',
+  priority: 10,   // before anything that reads the text
+  requiredFor: (model) => model.reach === 'any',
+  // The actual redaction still happens in streamChat's harness. This declares the
+  // REQUIREMENT; wiring the implementation through here is the next step, and doing both at
+  // once would mean changing what redaction does in the same commit that changes when it runs.
+  run: async (request) => request,
+});
+
+/**
+ * Escalate when the task is actually hard.
+ *
+ * The router was picking the cheapest eligible model for everything, which is right for
+ * "hello" and wrong for "draw a circle around Mickey" — a request needing spatial reasoning
+ * and a structured payload went to a 26B model because it was free. Cost is the correct
+ * tie-breaker among models that can all do the job; it is the wrong one when they cannot.
+ *
+ * Class R: length, code fences, image content and page tools are all readable for nothing.
+ * The escalation itself costs no model call — only the answer does, and that is the point.
+ */
+export const complexityStrategy = defineRouteStrategy({
+  id: 'escalate-on-complexity',
+  label: 'Escalate hard tasks',
+  classUsed: 'R',
+  decide: async (eligible, need) => {
+    const sig = need.signals;
+    // ASKED FOR NOTHING, ESCALATES TO NOTHING. This fired on 'hello' because the caller was
+    // passing `structured: structured || pageTools`, so every turn on a page with actions
+    // armed looked like exact structured work. Equipment is not demand — the same conflation
+    // that put a quality floor on a greeting, in a second place.
+    if (sig?.smalltalk) return null;
+    // NOR DOES BACKGROUND WORK ESCALATE. Dropping the quality floor for a topic pass and then
+    // letting escalation rank by quality anyway would move the same decision one step down
+    // and change nothing — the floor eliminated the local models, this would simply rank them
+    // last. Both read 'high' from the size of the material rather than the difficulty of the
+    // ask, so both have to abstain.
+    if (need.background) return null;
+    const hard = sig?.complexity === 'high' || sig?.modality === 'vision' || need.structured;
+    if (!hard) return null;   // no opinion on easy work — let cost decide
+    // Prefer a model that claims what this task actually wants. Not a hard filter: declaring
+    // "reasoning" required would eliminate every model on a setup where nobody has ticked
+    // the box, and an empty candidate list is a worse answer than a merely adequate model.
+    const wants = new Set();
+    if (sig?.complexity === 'high') wants.add('reasoning');
+    if (need.structured) wants.add('tools');
+    if (sig?.modality === 'vision') wants.add('vision');
+    if (sig?.approxTokens > 20_000) wants.add('long-context');
+    const fit = (m) => [...wants].filter((c) => m.capabilities.includes(c)).length;
+    const best = Math.max(...eligible.map(fit));
+    let shortlist = best > 0 ? eligible.filter((m) => fit(m) === best) : eligible;
+
+    // STRUCTURED WORK WANTS A MODEL, NOT AN AGENT.
+    //
+    // A canvas or spreadsheet adapter is one call: hand it the data, it applies it, done. A
+    // CLI agent runs its OWN loop — it explores, reads files, decides what to do next — and
+    // having applied the shapes correctly it carries on, because finishing is not something
+    // its loop is told about. A user watched the circle appear and then waited until they
+    // killed the process.
+    //
+    // Not a hard filter: on a setup with only agents, an agent that overruns still beats no
+    // answer.
+    if (need.structured) {
+      const models = shortlist.filter((m) => m.classUsed !== 'A');
+      if (models.length) shortlist = models;
+    }
+    // Rank by declared quality — the axis this strategy exists to judge — then by the ORDER
+    // the user set, and only then by cost. A model with an unknown quality sits mid-table
+    // rather than last, so a newly added model is not permanently skipped.
+    //
+    // ORDER BEFORE COST, and this is the fix for a real complaint: three CLI agents of
+    // identical quality, one of them pinned to Order 1, and escalation picked a different one
+    // because it is cheaper per 1k. A hand-set order is a statement — they can see the prices
+    // and chose anyway — and it was being honoured in the score path and nowhere else, so
+    // the moment any strategy had an opinion the user's own preference stopped existing.
+    //
+    // Still only a TIE-BREAK: quality decides first, so a genuinely better model beats the
+    // pinned one, and an INFERRED order stays below cost where a guess belongs.
+    const q = (m) => (Number.isFinite(m.quality) ? m.quality : 0.5);
+    return [...shortlist].sort((a, b) => q(b) - q(a)
+      || pinnedOrderOf(a) - pinnedOrderOf(b) || a.costPer1k - b.costPer1k);
+  },
+});
+
+/**
+ * When a model declines, replace it with the closest thing available — not the cheapest.
+ *
+ * A frontier model that ran out of credits mid-task should be replaced by the same model at
+ * another provider, or by something comparably capable. Falling back to a small local model
+ * is how a drawing that was going well turns into a circle in the wrong place: the task did
+ * not get easier when the provider said no.
+ *
+ * Ranked by closeness to what failed, in the order that actually matters:
+ *   1. the SAME model somewhere else — identical capability, merely a different bill;
+ *   2. a model with every capability the failed one had, best quality first;
+ *   3. anything else, so the turn still completes rather than dying.
+ */
+export const failoverStrategy = defineRouteStrategy({
+  id: 'failover-to-similar',
+  label: 'Replace like with like',
+  classUsed: 'R',
+  decide: async (eligible, need) => {
+    const failed = need.like;
+    if (!failed) return null;
+    // THE ORDERING ITSELF LIVES IN @chatpanel/events, because two things need it: this
+    // strategy, and the projected chain the trace draws before any of it happens. A picture
+    // computed by a second implementation would eventually disagree with the real failover,
+    // and one that lies about what the router will do is worse than no picture. The strategy
+    // is the thin part — knowing there IS something to replace.
+    return failoverOrder(eligible, failed);
+  },
+});
+
+/**
+ * "use claude" is an instruction, not a topic.
+ *
+ * A user naming a model in their message was being ignored entirely — the router read
+ * length, modality and tools, and not the one signal that is an explicit answer to the
+ * question it was asking. Asking for a specific model and being given another is the most
+ * annoying possible failure of a router, because it looks like the request was not read.
+ *
+ * DELIBERATELY CONSERVATIVE. Only imperative forms count — "use X", "with X", "ask X",
+ * "switch to X" — so "tell me about claude" stays a question about Claude rather than a
+ * routing instruction. A false positive here silently sends work to the wrong model, which
+ * is worse than missing an unusual phrasing.
+ *
+ * It still cannot widen reach: like every strategy it only ever chooses among candidates the
+ * hard constraints already allowed. A device-only request naming a cloud model still stays
+ * on-device.
+ */
+export const explicitModelStrategy = defineRouteStrategy({
+  id: 'named-by-user',
+  label: 'Use the model you asked for',
+  classUsed: 'R',
+  decide: async (eligible, need) => {
+    const text = String(need.requestText || '').toLowerCase();
+    if (!text) return null;
+    const directive = /\b(?:use|using|with|via|ask|switch to|route to|try)\s+([a-z0-9][a-z0-9.\- ]{1,28})/g;
+    const asked = [];
+    for (const m of text.matchAll(directive)) asked.push(m[1].trim());
+    if (!asked.length) return null;
+
+    const matches = eligible.filter((cand) => {
+      const names = [cand.label, cand.model, cand.id].filter(Boolean).map((x) => String(x).toLowerCase());
+      return asked.some((want) => names.some((n) => n.includes(want) || want.includes(n.split(' · ')[0])));
+    });
+    if (!matches.length) return null;   // named something we do not have? say nothing and let the rest decide
+    // WHICH ONE, when the name matches several. This took the FIRST match in score order, so
+    // "use claude" on a setup with three Claude routes picked whichever happened to score
+    // best — a cost-and-latency guess deciding a question the user had already answered
+    // twice: once by naming the model, and once by ordering the routes to it.
+    //
+    // Returned as a LIST rather than a single model, so the ones that also matched become the
+    // runners-up: if the first declines, failover replaces it with another route to the model
+    // that was actually asked for.
+    return [...matches].sort((a, b) => pinnedOrderOf(a) - pinnedOrderOf(b));
+  },
+});
+
+// Ordered deliberately: an explicit request outranks every heuristic, because the user has
+// answered the question the router was about to guess at. Failover next — a decline is newer
+// information than the preference that made the original choice. Escalation is the general
+// case.
+export const ROUTE_STRATEGIES = [explicitModelStrategy, failoverStrategy, complexityStrategy];
+export const ROUTE_MIDDLEWARE = [redactionStep];
+
+export function needForTurn({ capabilities = [], request = null, structured = false, pageTools = false, force = false, background = false, guard = null } = {}) {
+  const signals = request ? signalsFrom(request) : {};
+  // REQUIREMENTS FIRST. What the work needs eliminates candidates; cost and speed only order
+  // what survives. A preference lets an unsuitable model win once the better ones decline,
+  // which is exactly how a chain of five ended on one that could not do the job.
+  const req = requirementsFor(signals, { structured, pageTools, hasTools: capabilities.includes('tools'), background });
+  // WHERE IT CAME FROM IS A CEILING, NOT A PREFERENCE. Routing asked what the work needed and
+  // never asked what it was about, so an internal page was summarised by a public inference
+  // host. This narrows reach and can only narrow it — reach is never relaxed (see the
+  // relaxation order in the router), so no later step can trade it away for capability.
+  // WHICH AXIS THIS REQUEST CARES ABOUT, read from the request rather than fixed at
+  // 'balanced' for everything. A greeting means fast — no answer to "hi" is improved by a
+  // frontier model thinking about it. A refactor means good — three seconds saved on an
+  // answer that has to be redone is not a saving. It only ever ORDERS what already
+  // qualifies; `req` above is what eliminates.
+  const pref = preferenceFor(signals, {
+    structured,
+    minQuality: req.minQuality,
+    // The same fact requirementsFor was given: a turn carrying tools is one that might use
+    // them, so it is never "answer fast at any quality".
+    hasTools: capabilities.includes('tools'),
+    background,
+  });
+  return {
+    prefer: pref.prefer,
+    reach: guard ? guard.reach : 'any',
+    capabilities: [...new Set([...capabilities, ...req.required])],
+    minQuality: req.minQuality,
+    // Which requirements may be given up if nothing qualifies — never `tools`, and never
+    // reach. See requirementsFor.
+    negotiable: req.negotiable,
+    requirementReasons: [...(guard ? [...req.why, guard.why] : req.why), pref.why],
+    sourceGuard: guard,
+    signals,
+    requestText: request ? String(request.text || (request.messages || []).map((m) => m?.content || '').join('\n')).slice(-2000) : '',
+    structured,
+    force,
+    background,
+  };
+}

package/scm-connection.js

@@ -0,0 +1,180 @@
+// An SCM CONNECTION — how a harness engine reaches the organisation's source-control hub —
+// and the two names a job gets in it: its branch and its worktree directory.
+// (architecture-pillars.md §14: the repository is the source of truth.)
+//
+// ChatPanel is not a git host and not a git client. The harness runs `git`, `gh`, `glab`;
+// ChatPanel brings the CREDENTIAL, the WORKTREE, the BRANCH NAME and the RECORD. This module
+// is the pure part: what a connection record holds, how a remote URL maps to one, what a
+// job's branch is called, and the environment that hands a token to git for ONE process
+// without writing it anywhere.
+//
+// THE TOKEN IS NEVER IN THE RECORD. A connection stores a `secretRef` — the name under which
+// the machine's keychain (or the bridge's secret store) holds the token. The record travels
+// with the client-prefs document (`connections` section) so both clients show the same
+// connections; the secret does not travel, and `normalizeConnection` REFUSES a record that
+// carries one, so a settings export, a sync, or a prompt can never carry it either.
+
+export const SCM_KINDS = Object.freeze(['github', 'github-enterprise', 'gitlab', 'bitbucket', 'gitea', 'azure-devops', 'git']);
+export const CONNECTION_ID_RE = /^[a-z][a-z0-9_-]{0,63}$/i;
+export const SECRET_REF_RE = /^[a-zA-Z0-9_.:@/-]{1,200}$/;
+const SECRET_FIELDS = ['token', 'password', 'secret', 'pat', 'apiKey', 'key'];
+const DEFAULT_HOST = { github: 'github.com', gitlab: 'gitlab.com', bitbucket: 'bitbucket.org', 'azure-devops': 'dev.azure.com' };
+
+export class ScmError extends Error {
+  constructor(code, message) { super(message); this.name = 'ScmError'; this.code = code; }
+}
+
+const isRecord = (v) => v !== null && typeof v === 'object' && !Array.isArray(v);
+
+/** The host a connection serves: its baseUrl's, or the kind's public host. */
+export function hostOf(conn) {
+  const url = String(conn?.baseUrl || '').trim();
+  if (url) { try { return new URL(/^https?:\/\//i.test(url) ? url : `https://${url}`).host.toLowerCase(); } catch { return url.toLowerCase().replace(/^https?:\/\//, '').split('/')[0]; } }
+  return DEFAULT_HOST[conn?.kind] || '';
+}
+
+export function validateConnection(conn) {
+  const errors = [];
+  if (!isRecord(conn)) return { ok: false, errors: ['connection must be an object'] };
+  if (!CONNECTION_ID_RE.test(String(conn.id || ''))) errors.push('id: a short identifier (letters, digits, _ -)');
+  if (!SCM_KINDS.includes(conn.kind)) errors.push(`kind: one of ${SCM_KINDS.join(', ')}`);
+  const needsHost = ['github-enterprise', 'gitea', 'git'].includes(conn.kind);
+  if (needsHost && !String(conn.baseUrl || '').trim()) errors.push('baseUrl: the hub\'s URL');
+  if (conn.baseUrl && !/^(https?:\/\/)?[a-z0-9.-]+(:\d+)?(\/[^\s]*)?$/i.test(String(conn.baseUrl).trim())) errors.push('baseUrl: a host or URL');
+  const leaked = SECRET_FIELDS.filter((k) => conn[k] !== undefined && conn[k] !== null && conn[k] !== '');
+  if (leaked.length) errors.push(`${leaked.join(', ')}: a connection never stores a secret — store it in the keychain and reference it by secretRef`);
+  if (conn.secretRef !== undefined && conn.secretRef !== null && conn.secretRef !== '' && !SECRET_REF_RE.test(String(conn.secretRef))) errors.push('secretRef: a keychain / secret-store name');
+  if (conn.reach !== undefined && !Array.isArray(conn.reach)) errors.push('reach: a list of remotes or owners this connection may be used for');
+  return { ok: errors.length === 0, errors };
+}
+
+/** The stored form. Throws on a record that carries a secret. */
+export function normalizeConnection(conn) {
+  const v = validateConnection(conn);
+  if (!v.ok) throw new ScmError('INVALID', v.errors.join('; '));
+  const id = String(conn.id);
+  return {
+    id,
+    kind: conn.kind,
+    label: String(conn.label || conn.name || '').trim().slice(0, 80) || `${conn.kind} · ${hostOf(conn) || id}`,
+    ...(conn.baseUrl ? { baseUrl: String(conn.baseUrl).trim().replace(/\/+$/, '') } : {}),
+    // The secret's NAME — `chatpanel:scm:<id>` by default — under which the keychain holds it.
+    secretRef: String(conn.secretRef || '').trim() || `chatpanel:scm:${id}`,
+    ...(conn.username ? { username: String(conn.username).trim().slice(0, 80) } : {}),
+    // Which remotes it may be used for: `owner/*`, `owner/name`, or a host. Empty = any on its host.
+    reach: (Array.isArray(conn.reach) ? conn.reach : []).map((r) => String(r).trim().toLowerCase()).filter(Boolean).slice(0, 64),
+    enabled: conn.enabled !== false,
+    ...(conn.createdAt ? { createdAt: conn.createdAt } : {}),
+  };
+}
+
+/**
+ * A remote URL, read: `{ host, owner, name, protocol }` for https, ssh (`git@host:o/n.git`,
+ * `ssh://git@host/o/n`) and plain `host/o/n`. Credentials in the URL are dropped, never
+ * returned. Null for what cannot be read.
+ */
+export function parseRemote(url) {
+  const s = String(url || '').trim();
+  if (!s) return null;
+  let m = /^(?:ssh:\/\/)?(?:[\w.-]+@)?([a-z0-9.-]+)(?::\d+)?[:/]+([^/\s]+)\/([^/\s]+?)(?:\.git)?\/?$/i.exec(s.replace(/^https?:\/\/(?:[^@/]+@)?/i, ''));
+  if (m && !/^https?:/i.test(s)) return { host: m[1].toLowerCase(), owner: m[2], name: m[3], protocol: /^ssh:|^[\w.-]+@/.test(s) ? 'ssh' : 'plain' };
+  try {
+    const u = new URL(s);
+    const parts = u.pathname.replace(/\.git$/, '').split('/').filter(Boolean);
+    if (parts.length < 2) return null;
+    // Azure DevOps: /org/project/_git/repo; GitLab subgroups: /a/b/c/repo — owner is all but the last.
+    const name = parts.at(-1);
+    const owner = parts.slice(0, -1).filter((p) => p !== '_git').join('/');
+    return { host: u.host.toLowerCase(), owner, name, protocol: u.protocol.replace(':', '') };
+  } catch { return null; }
+}
+
+/**
+ * The connection to use for a remote: enabled, same host, and — when the connection lists
+ * a reach — the remote's `owner/name` or `owner/*` in it. The most specific reach wins.
+ */
+export function connectionFor(remote, connections = []) {
+  const r = typeof remote === 'string' ? parseRemote(remote) : remote;
+  if (!r) return null;
+  const slug = `${r.owner}/${r.name}`.toLowerCase();
+  let best = null; let bestScore = -1;
+  for (const c of connections || []) {
+    if (!c || c.enabled === false) continue;
+    if (hostOf(c) !== r.host) continue;
+    const reach = Array.isArray(c.reach) ? c.reach : [];
+    let score = reach.length ? -1 : 0;
+    for (const x of reach) {
+      if (x === slug) score = Math.max(score, 3);
+      else if (x.endsWith('/*') && slug.startsWith(x.slice(0, -1))) score = Math.max(score, 2);
+      else if (x === r.host) score = Math.max(score, 1);
+    }
+    if (score > bestScore) { best = c; bestScore = score; }
+  }
+  return best;
+}
+
+const slug = (s) => String(s || '').trim().toLowerCase().replace(/[^a-z0-9._-]+/g, '-').replace(/^[-.]+|[-.]+$/g, '').slice(0, 48) || 'x';
+
+/** A job's branch: `cp/<project>/<job>` — one branch per job, always under `cp/`, never a protected name. */
+export function branchFor(projectId, jobId) { return `cp/${slug(projectId)}/${slug(jobId)}`; }
+
+/** A job's worktree directory under the bridge's worktree root: `<project>/<job>`. */
+export function worktreeDirFor(projectId, jobId) { return `${slug(projectId)}/${slug(jobId)}`; }
+
+/**
+ * The environment that hands ONE process the token, scoped to the connection's host:
+ * git's `credential.<url>.helper` set through `GIT_CONFIG_*` (never a file, never the
+ * command line), a helper that reads the token from the environment, terminal prompts
+ * off, and the hub CLIs' own variables (`GH_TOKEN` / `GH_HOST`, `GITLAB_TOKEN`) so `gh` and
+ * `glab` work in the same process. The caller spawns with `{ ...process.env, ...env }` and
+ * the token dies with the process.
+ */
+export function credentialEnv(conn, token, { username = null } = {}) {
+  const c = conn || {};
+  const host = hostOf(c);
+  if (!host || !token) return {};
+  const user = username || c.username || ({ github: 'x-access-token', 'github-enterprise': 'x-access-token', gitlab: 'oauth2', bitbucket: 'x-token-auth', gitea: 'oauth2', 'azure-devops': 'pat' }[c.kind] || 'token');
+  const env = {
+    GIT_TERMINAL_PROMPT: '0',
+    GIT_CONFIG_COUNT: '2',
+    GIT_CONFIG_KEY_0: `credential.https://${host}.helper`,
+    // A shell helper: on `get`, answer from the environment. Nothing is written anywhere.
+    GIT_CONFIG_VALUE_0: '!f() { if [ "$1" = get ]; then printf "username=%s\\npassword=%s\\n" "$CHATPANEL_SCM_USER" "$CHATPANEL_SCM_TOKEN"; fi; }; f',
+    GIT_CONFIG_KEY_1: `credential.https://${host}.useHttpPath`,
+    GIT_CONFIG_VALUE_1: 'false',
+    CHATPANEL_SCM_USER: user,
+    CHATPANEL_SCM_TOKEN: String(token),
+  };
+  if (c.kind === 'github' || c.kind === 'github-enterprise') {
+    env.GH_TOKEN = String(token);
+    if (c.kind === 'github-enterprise') { env.GH_ENTERPRISE_TOKEN = String(token); env.GH_HOST = host; }
+  }
+  if (c.kind === 'gitlab') { env.GITLAB_TOKEN = String(token); env.GITLAB_HOST = host; }
+  return env;
+}
+
+/** One line a person reads: "GitHub · github.com · acme/*". */
+export function describeConnection(conn) {
+  const c = conn || {};
+  const kindLabel = { github: 'GitHub', 'github-enterprise': 'GitHub Enterprise', gitlab: 'GitLab', bitbucket: 'Bitbucket', gitea: 'Gitea', 'azure-devops': 'Azure DevOps', git: 'Git' }[c.kind] || c.kind;
+  return `${kindLabel} · ${hostOf(c)}${c.reach?.length ? ` · ${c.reach.join(', ')}` : ''}${c.enabled === false ? ' (off)' : ''}`;
+}
+
+export function blankConnection() { return { id: '', kind: 'github', label: '', baseUrl: '', reach: [], username: '' }; }
+
+/** The editor's form → a connection, or the errors. The token is NOT part of the form's record: the host stores it separately. */
+export function connectionFromForm(form) {
+  const f = form || {};
+  const id = String(f.id || '').trim() || slug(f.label || `${f.kind}-${hostOf(f)}`);
+  const conn = {
+    id, kind: f.kind || 'github', label: String(f.label || '').trim(),
+    ...(f.baseUrl ? { baseUrl: String(f.baseUrl).trim() } : {}),
+    ...(f.username ? { username: String(f.username).trim() } : {}),
+    ...(f.secretRef ? { secretRef: String(f.secretRef).trim() } : {}),
+    reach: String(Array.isArray(f.reach) ? f.reach.join(',') : f.reach || '').split(/[,\s]+/).map((x) => x.trim()).filter(Boolean),
+    enabled: f.enabled !== false,
+    ...(f.createdAt ? { createdAt: f.createdAt } : {}),
+  };
+  const v = validateConnection(conn);
+  return v.ok ? { ok: true, connection: normalizeConnection(conn) } : { ok: false, errors: v.errors };
+}

package/scorecard.js

@@ -13,11 +13,59 @@
 // against a job's needs with that record — the same function the evaluator starts from, so
 // an application's fit has reasons a person can read and overrule.
 //
+// The model is a variable, not a constant (architecture-pillars.md §13): every task fact also
+// says which ENGINE did it — a model endpoint or a harness (a CLI coding agent), and for a
+// harness the model it was asked to run — so `summarize` can split the card by engine
+// (`byEngine`) and a recruiter can tell an agent that did well on a small model from one
+// that was carried by a large one. And where the task ran in a git checkout (§14), the fact
+// carries `scm`: the branch and HEAD before and after, commits made, a PR when one was
+// opened, and whether it merged — the one outcome that does not come from a judge.
+//
 // Dependency-free: hashing is `crypto.subtle` (browser, Node, a phone), injectable for tests.
 
 export const SCORECARD_ENTRY_KINDS = Object.freeze(['task.done', 'task.failed', 'rating', 'created', 'interaction', 'role']);
 export const ROLE_KINDS = Object.freeze(['ic', 'orchestrator', 'manager', 'manager-of-managers']);
 export const SCORECARD_VERSION = 1;
+export const ENGINE_KINDS = Object.freeze(['model', 'harness']);
+
+/**
+ * An engine as the record keeps it: `{ kind, id, model?, label? }`. `kind` is `model` (an
+ * endpoint the client calls) or `harness` (a CLI coding agent the bridge runs — `id` is the
+ * harness, `model` the model it was asked to run, when one was named). A host that does not
+ * say the kind gets `model`, which is the honest default for a bare model id. A string is
+ * an id.
+ */
+export function normalizeEngine(e) {
+  if (!e) return null;
+  const src = typeof e === 'string' ? { id: e } : e;
+  const id = String(src.id || src.harnessId || src.model || '').trim();
+  if (!id) return null;
+  const kind = ENGINE_KINDS.includes(src.kind) ? src.kind : (src.harnessId ? 'harness' : 'model');
+  const model = src.model != null && String(src.model).trim() && String(src.model) !== id ? String(src.model).trim() : undefined;
+  return { kind, id, ...(model ? { model } : {}), ...(src.label && String(src.label) !== id ? { label: String(src.label).slice(0, 120) } : {}) };
+}
+
+/** One key per engine — what `byEngine` groups on and what the model ledger will be keyed by. */
+export function engineKey(e) {
+  const n = normalizeEngine(e);
+  return n ? `${n.kind}:${n.id}${n.model ? `/${n.model}` : ''}` : null;
+}
+
+/** What a task did in a checkout, as the record keeps it. Strings clipped, counts rounded. */
+export function normalizeScm(s) {
+  if (!s || typeof s !== 'object') return null;
+  const str = (v, n = 200) => (v == null || v === '' ? undefined : String(v).slice(0, n));
+  const out = {
+    repo: str(s.repo, 300), remote: str(s.remote, 300), base: str(s.base, 120), branch: str(s.branch, 120),
+    head: str(s.head, 64), headAfter: str(s.headAfter, 64),
+    commits: s.commits != null ? Math.max(0, Math.round(Number(s.commits) || 0)) : undefined,
+    pr: str(s.pr, 300),
+    merged: s.merged === true ? true : s.merged === false ? false : undefined,
+    dirty: s.dirty === true ? true : s.dirty === false ? false : undefined,
+  };
+  for (const k of Object.keys(out)) if (out[k] === undefined) delete out[k];
+  return Object.keys(out).length ? out : null;
+}
 
 const enc = new TextEncoder();
 const hex = (buf) => [...new Uint8Array(buf)].map((b) => b.toString(16).padStart(2, '0')).join('');
@@ -60,6 +108,8 @@ export async function makeEntry(fact, prev, { now = () => Date.now(), subtle } =
     ...(fact.runId ? { runId: String(fact.runId) } : {}),
     ...(fact.taskId ? { taskId: String(fact.taskId) } : {}),
     ...(fact.model ? { model: String(fact.model) } : {}),
+    ...(normalizeEngine(fact.engine) ? { engine: normalizeEngine(fact.engine) } : {}),
+    ...(normalizeScm(fact.scm) ? { scm: normalizeScm(fact.scm) } : {}),
     ...(fact.size ? { size: sizeOf(fact.size) } : {}),
     ...(fact.roleKind ? { roleKind: ROLE_KINDS.includes(fact.roleKind) ? fact.roleKind : 'ic' } : {}),
     ...(Array.isArray(fact.tools) && fact.tools.length ? { tools: [...new Set(fact.tools.map(String))].sort() } : {}),
@@ -75,6 +125,7 @@ export async function makeEntry(fact, prev, { now = () => Date.now(), subtle } =
 }
 
 const clamp01 = (n) => Math.max(0, Math.min(1, Number(n) || 0));
+const r3 = (v) => (v == null ? null : Math.round(v * 1000) / 1000);
 const sizeOf = (s) => ({ ms: Math.max(0, Math.round(Number(s.ms) || 0)), steps: Math.max(0, Math.round(Number(s.steps) || 0)), tools: Math.max(0, Math.round(Number(s.tools) || 0)), findings: Math.max(0, Math.round(Number(s.findings) || 0)), tokens: Math.max(0, Math.round(Number(s.tokens) || 0)) });
 
 /** Does every link hold? Returns `{ ok, at }` — `at` is the seq of the first broken entry. */
@@ -120,13 +171,53 @@ export function summarize(entries, { recent = 5 } = {}) {
   const tools = new Set(); const withAgents = new Set(); const created = new Set();
   const roles = { ic: 0, orchestrator: 0, manager: 0, 'manager-of-managers': 0 };
   const models = new Map();
+  // Per engine: how many tasks, how they went, how big, and the ratings that were ABOUT a
+  // task on it. A rating names its task by `about` (the seq of the entry it rates), by
+  // taskId, or — failing both — by runId when that run had exactly one task on the card.
+  const engines = new Map(); // key -> { engine, tasks, done, failed, tokens, ratings[] }
+  const engineOfEntry = new Map(); // seq -> key
+  const byTask = new Map(); const byRun = new Map(); // taskId -> seq, runId -> [seq]
+  const scm = { tasks: 0, commits: 0, prs: 0, merged: 0 };
   for (const e of list) {
     for (const t of e.tools || []) tools.add(t);
     for (const a of e.with || []) withAgents.add(a);
     for (const a of e.created || []) created.add(a);
     if (e.roleKind && (e.kind === 'task.done' || e.kind === 'task.failed' || e.kind === 'role')) roles[e.roleKind] = (roles[e.roleKind] || 0) + 1;
     if (e.model) models.set(e.model, (models.get(e.model) || 0) + 1);
+    if (e.kind === 'task.done' || e.kind === 'task.failed') {
+      const key = engineKey(e.engine);
+      if (key) {
+        const row = engines.get(key) || { engine: normalizeEngine(e.engine), tasks: 0, done: 0, failed: 0, tokens: 0, ratings: [] };
+        row.tasks += 1; row[e.kind === 'task.done' ? 'done' : 'failed'] += 1; row.tokens += e.size?.tokens || 0;
+        engines.set(key, row);
+        engineOfEntry.set(e.seq, key);
+        if (e.taskId) byTask.set(`${e.runId || ''}/${e.taskId}`, e.seq);
+        if (e.runId) byRun.set(e.runId, [...(byRun.get(e.runId) || []), e.seq]);
+      }
+      if (e.scm) { scm.tasks += 1; scm.commits += e.scm.commits || 0; if (e.scm.pr) scm.prs += 1; if (e.scm.merged) scm.merged += 1; }
+    }
+  }
+  for (const e of list) {
+    if (e.kind !== 'rating' || !e.rating) continue;
+    const seq = e.rating.about != null ? e.rating.about
+      : e.taskId && byTask.has(`${e.runId || ''}/${e.taskId}`) ? byTask.get(`${e.runId || ''}/${e.taskId}`)
+        : e.runId && (byRun.get(e.runId) || []).length === 1 ? byRun.get(e.runId)[0] : null;
+    const key = seq != null ? engineOfEntry.get(seq) : null;
+    if (key) engines.get(key).ratings.push(e.rating.score);
   }
+  const mean = (xs) => (xs.length ? xs.reduce((a, b) => a + b, 0) / xs.length : null);
+  const byEngine = [...engines.values()].sort((a, b) => b.tasks - a.tasks).map((r) => ({
+    key: engineKey(r.engine), ...r.engine, tasks: r.tasks, done: r.done, failed: r.failed,
+    failRate: r.tasks ? Math.round((r.failed / r.tasks) * 1000) / 1000 : 0,
+    tokens: r.tasks ? Math.round(r.tokens / r.tasks) : 0, // mean per task — the cost proxy until the ledger prices it
+    rating: { avg: mean(r.ratings), count: r.ratings.length },
+  }));
+  // 1 − spread of rating across engines it was rated on (≥ 2): low spread = robust to routing;
+  // high spread = it NEEDS a particular engine, a fact a recruiter respects, not a penalty.
+  const ratedEngines = byEngine.filter((r) => r.rating.avg != null).map((r) => r.rating.avg);
+  const engineIndependence = ratedEngines.length >= 2 ? Math.round((1 - (Math.max(...ratedEngines) - Math.min(...ratedEngines))) * 1000) / 1000 : null;
+  // Leverage (rating above the engine's own mean) and efficiency (rating ÷ cost) wait on the
+  // model ledger (§13.2, with A1): they need every engine's mean, which one card cannot know.
   const ratings = list.filter((e) => e.kind === 'rating' && e.rating).map((e) => e.rating.score);
   const avg = ratings.length ? ratings.reduce((a, b) => a + b, 0) / ratings.length : null;
   const recentRatings = ratings.slice(-recent);
@@ -142,6 +233,9 @@ export function summarize(entries, { recent = 5 } = {}) {
     created: [...created],
     roles,
     models: [...models.entries()].sort((a, b) => b[1] - a[1]).map(([m, n]) => ({ model: m, tasks: n })),
+    byEngine,
+    engineIndependence,
+    scm,
     rating: { avg, count: ratings.length, recent: recentRatings.length ? recentRatings.reduce((a, b) => a + b, 0) / recentRatings.length : null },
     refs,
     since: list[0]?.at || null,
@@ -156,7 +250,7 @@ export function summarize(entries, { recent = 5 } = {}) {
  * summary is `summarize()`'s. Returns `{ score, reasons }` in [0, 1] — needs first (a type
  * without the tools cannot do the job), track record second, size third.
  */
-export function fit(job, type, summary = null) {
+export function fit(job, type, summary = null, { qualityOf = null, costOf = null, adjust = true } = {}) {
   const needs = job?.needs || {};
   const have = (xs) => new Set((xs || []).map((x) => String(x).toLowerCase()));
   const skills = have(type?.skills); const tools = have(type?.tools); const grants = have(type?.grants);
@@ -174,11 +268,18 @@ export function fit(job, type, summary = null) {
   const needScore = (cSkills * 0.5 + cTools * 0.3 + cGrants * 0.2);
   if (needScore === 1) reasons.push('has every skill, tool and grant the job names');
   let record = 0.5; // a fresh type is neither trusted nor distrusted
+  let adjusted = null;
   if (summary && summary.entries) {
     const doneRate = summary.jobsDone + summary.jobsFailed ? summary.jobsDone / (summary.jobsDone + summary.jobsFailed) : 0.5;
-    const rated = summary.rating.avg == null ? 0.5 : summary.rating.avg;
+    // The track record uses the MODEL-ADJUSTED rating when the caller can say what the
+    // engines were worth (§13.3): an agent rated 0.82 mostly on a weak engine ranks above
+    // one rated 0.82 on a frontier model. The reasons say so, and a person can turn it off.
+    adjusted = adjust && qualityOf && summary.rating.avg != null ? adjustSummary(summary, { qualityOf, costOf: costOf || undefined }) : null;
+    const rated = summary.rating.avg == null ? 0.5 : (adjusted?.adjusted ?? summary.rating.avg);
     record = doneRate * 0.5 + rated * 0.5;
-    reasons.push(`${summary.jobsDone} done, ${summary.jobsFailed} failed${summary.rating.avg != null ? `, rated ${Math.round(summary.rating.avg * 100)}%` : ''}`);
+    reasons.push(`${summary.jobsDone} done, ${summary.jobsFailed} failed${summary.rating.avg != null ? (adjusted && adjusted.adjusted !== adjusted.raw ? `, rated ${Math.round(adjusted.raw * 100)}% raw, ${Math.round(adjusted.adjusted * 100)}% adjusted — ${adjusted.basis[0] || 'engine-corrected'}` : `, rated ${Math.round(summary.rating.avg * 100)}%`) : ''}`);
+    if (adjusted?.leverage != null && adjusted.leverage > 0.05) reasons.push(`adds ${adjusted.leverage} over its engines' own quality`);
+    if (adjusted?.efficiency) reasons.push(`cleared the bar cheapest on ${adjusted.efficiency.engine}`);
     if (summary.roles.orchestrator + summary.roles.manager + summary.roles['manager-of-managers'] > 0) reasons.push(`has led: ${summary.roles.orchestrator} as orchestrator, ${summary.roles.manager} as manager`);
   } else {
     reasons.push('no record yet');
@@ -187,5 +288,48 @@ export function fit(job, type, summary = null) {
   const sizeScore = !wantSize ? 1 : Math.min(1, (summary?.size?.largestSteps || 0) / wantSize) * 0.5 + 0.5;
   if (wantSize && (summary?.size?.largestSteps || 0) < wantSize) reasons.push(`largest task so far ${summary?.size?.largestSteps || 0} steps; this one is ~${wantSize}`);
   const score = Math.round((needScore * 0.6 + record * 0.3 + sizeScore * 0.1) * 1000) / 1000;
-  return { score, reasons, parts: { needs: needScore, record, size: sizeScore } };
+  return { score, reasons, parts: { needs: needScore, record, size: sizeScore }, ...(adjusted ? { adjusted } : {}) };
 }
+
+// ── Agent scores, normalised by engine (§13.3) — what the model ledger's cards make possible ──
+
+/**
+ * The model-adjusted view of an agent's card (scorecard.js `summarize()`), given what its
+ * engines are worth: `qualityOf(key)` → the engine's quality in [0, 1] (the card's mean
+ * rating for the job kind when observed, else the router's guess) or null when unknown.
+ *
+ *   leverage      rating on an engine minus that engine's quality, weighted by tasks — what
+ *                 the agent's prompt and tools add that the model does not supply on its own
+ *   adjusted      the raw rating corrected for the engines it ran on: work done on a weak
+ *                 engine counts for more, on a strong one for less; `k` bounds the correction
+ *   efficiency    adjusted rating ÷ cost per task on the cheapest engine that cleared `bar`
+ *                 (`costOf(key)` → $/task or a token proxy; null when nothing is priced)
+ *
+ * Returns `{ raw, adjusted, leverage, efficiency, basis[] }` with `basis` the reasons a
+ * person reads ("60 % of its tasks ran on a 0.3-quality engine").
+ */
+export function adjustSummary(summary, { qualityOf = () => null, costOf = () => null, reference = 0.6, k = 0.3, bar = 0.5 } = {}) {
+  const raw = summary?.rating?.avg ?? null;
+  const rows = (summary?.byEngine || []).filter((r) => r.key);
+  const known = rows.map((r) => ({ ...r, quality: qualityOf(r.key) })).filter((r) => Number.isFinite(r.quality));
+  const totalTasks = known.reduce((n, r) => n + r.tasks, 0);
+  const basis = [];
+  if (raw == null || !known.length || !totalTasks) return { raw, adjusted: raw, leverage: null, efficiency: null, basis: raw == null ? ['not rated yet'] : ['engines not rated yet — raw rating used'] };
+  // Correction: how far below the reference the engines it ran on sit, task-weighted.
+  const correction = k * known.reduce((s, r) => s + (r.tasks / totalTasks) * (reference - r.quality), 0);
+  const adjusted = Math.max(0, Math.min(1, raw + correction));
+  const weak = known.filter((r) => r.quality < reference);
+  if (weak.length) basis.push(`${Math.round((weak.reduce((n, r) => n + r.tasks, 0) / totalTasks) * 100)} % of its tasks ran on ${weak.length === 1 ? `a ${weak[0].quality}-quality engine` : 'engines below the reference'}`);
+  const strong = known.filter((r) => r.quality > reference);
+  if (strong.length && !weak.length) basis.push(`ran on engines above the reference (${strong.map((r) => r.quality).join(', ')})`);
+  // Leverage over the engines it was rated on.
+  const rated = known.filter((r) => r.rating?.avg != null);
+  const ratedTasks = rated.reduce((n, r) => n + r.rating.count, 0);
+  const leverage = ratedTasks ? r3(rated.reduce((s, r) => s + (r.rating.count / ratedTasks) * (r.rating.avg - r.quality), 0)) : null;
+  if (leverage != null) basis.push(`${leverage >= 0 ? '+' : ''}${leverage} over its engines' own quality`);
+  // Efficiency on the cheapest engine that cleared the bar.
+  const cleared = rated.filter((r) => r.rating.avg >= bar).map((r) => ({ ...r, cost: costOf(r.key) ?? (r.tokens || null) })).filter((r) => r.cost != null && r.cost > 0).sort((a, b) => a.cost - b.cost);
+  const efficiency = cleared.length ? { value: r3(adjusted / cleared[0].cost), engine: cleared[0].key, costPerTask: cleared[0].cost } : null;
+  return { raw: r3(raw), adjusted: r3(adjusted), leverage, efficiency, basis };
+}
+