portable-agent-layer 0.109.0 → 0.110.0

package/README.md

@@ -90,7 +90,7 @@ pal cli status        # check your setup
 | `pal cli patterns` | How you work across projects, read from the event log |
 | `pal cli content preserve` | Copy every Claude Code transcript, gzipped, into `memory/content/` before Claude Code deletes it (30 days by default). Runs on its own whenever a session closes; the content store never leaves the machine, so `export` skips it |
 | `pal cli events discord-nag <sent\|answered>` | For a bot outside PAL: record a nag it sent (`--kind <kind>`), or the answer to one and how long it took (`--after-sec <n>`), in the local event log |
-| `pal cli infer` | One model call for scripts and bots, on the active agent's route: prompt on stdin, reply on stdout. `--tier small\|medium`, `--system <file>`, `--schema <file>` for a JSON reply checked against it, `--timeout <seconds>`. The call loads no PAL hooks or context; exits 1 with the reason when no route answers |
+| `pal cli infer` | One model call for scripts and bots, on the active agent's route: prompt on stdin, reply on stdout. `--tier small\|medium\|large\|flagship` (on Claude: Haiku, Sonnet, Opus, Fable; on Codex: Luna, Sol, Sol at high effort, Astra; on Antigravity: Flash at low, medium and high effort, and no flagship; an agent without the tier asked refuses it), `--system <file>`, `--schema <file>` for a JSON reply checked against it, `--timeout <seconds>`. The call loads no PAL hooks or context; exits 1 with the reason when no route answers |
 | `pal cli actor [label <name>]` | Show or rename the actor — who caused a record. Travels with an export, so a shared memory can tell two people apart |
 | `pal cli machine [label <name>]` | Show or rename this install — where a record was written. Never leaves the machine |
 | `pal cli telos` | Which TELOS topics are answered, in interview order, and which one comes next |

package/assets/skills/scientific-research/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: scientific-research
 license: MIT
-description: "Answers scientific questions from trustworthy primary sources only (PubMed-indexed peer-reviewed literature, clinical guidelines from ERC, ESC, AHA, CDC and specialty societies, NEJM, The Lancet, JAMA, BMJ, UpToDate when accessible), with every claim cited to a source that was actually read, graded by evidence level and year, and cross-checked by a verifier agent. Use for EVERY health, medical, nutrition, drug, supplement, symptom, disease, treatment, diagnosis, exercise-physiology or other empirical scientific question, including 'is it safe to', 'does X work', 'what does the evidence say', 'side effects of', 'how much should I take' and 'what do the guidelines recommend'. Never answer such questions from memory; route them here."
+description: "Answers scientific questions from trustworthy primary sources only (PubMed-indexed peer-reviewed literature, clinical guidelines from ERC, ESC, AHA, CDC and specialty societies, NEJM, The Lancet, JAMA, BMJ, UpToDate when accessible), with every claim cited to a source that was actually read, graded by evidence level and year, its funding and any vendor ties named, and cross-checked by a verifier agent. Excludes preprints and MDPI journals and prefers SJR Q1 journals. Use for EVERY health, medical, nutrition, drug, supplement, symptom, disease, treatment, diagnosis, exercise-physiology or other empirical scientific question, including 'is it safe to', 'does X work', 'what does the evidence say', 'side effects of', 'how much should I take' and 'what do the guidelines recommend'. Never answer such questions from memory; route them here."
 argument-hint: <scientific or health question>
 metadata:
   source: portable-agent-layer
@@ -57,9 +57,26 @@ specialty society (WHO, NICE, EMA, FDA and their peers). Label each such
 source in the answer as "outside the core list".
 
 Not sources, under any circumstances: blogs, news sites, forums, supplement or
-vendor sites, encyclopedias, press releases, lecture slides, model memory. Preprints
-may be mentioned only when the answer labels them "preprint, not peer-reviewed" and
-no peer-reviewed source covers the claim.
+vendor sites, encyclopedias, press releases, lecture slides, model memory.
+
+Excluded even when they look peer-reviewed:
+
+- **Preprints and repositories**: bioRxiv, medRxiv, PsyArXiv, arXiv, SSRN, Research
+  Square, OSF Preprints, Preprints.org and any other server of unreviewed
+  manuscripts. Do not search them. PubMed and Europe PMC also index preprints, so
+  add `NOT preprint[pt]` to PubMed queries and drop any record whose publication type
+  is Preprint.
+- **MDPI journals** (Nutrients, Medicina, IJMS, Healthcare and the rest), identified
+  by publisher, not by title: a DOI starting with `10.3390` or a page on `mdpi.com`.
+  Add `-site:mdpi.com` to web searches.
+
+### Journal rank
+
+Prefer journals ranked Q1 by SCImago (SJR, `scimagojr.com`) in the subject category
+that fits the question. Rank breaks ties within one evidence level and never
+outranks study design: a Q2 RCT beats a Q1 observational study. Cite a journal below
+Q1 only when no Q1 source covers the claim, and label it with its quartile, for
+example "SJR Q2". Guidelines and official bodies are not ranked.
 
 ## Citation and evidence rules
 
@@ -74,12 +91,23 @@ For every claim record and later print:
   review.
 - **Year** of publication or of the guideline edition, so stale evidence is visible.
   When a newer edition of a guideline exists, cite the newer one.
+- **Journal quartile**: the SJR quartile of the journal, for every journal article.
+- **Funding and conflicts of interest**, read from the source's Funding, Acknowledgements
+  and Competing Interests sections, never inferred. Name the funders, then add one label:
+  - **vendor-funded**: a funder sells the product or intervention studied, or is a
+    trade group of its sellers (manufacturer, supplement brand, food or drug industry
+    association).
+  - **vendor-linked authors**: the funding is independent, but an author is employed,
+    paid or holds shares in such a company.
+  - **funding not stated**: the opened source has no funding statement.
+  - **funding not accessible**: the statement sits behind a paywall that did not open.
 
 Apply these tests before keeping a claim:
 
 1. Does the opened source state this claim, in this population, at this dose or in
    this setting? Paraphrase the source, do not extrapolate from it.
-2. Does the source belong to the allowed set? If not, discard the claim.
+2. Does the source belong to the allowed set, and is it neither a preprint nor an MDPI
+   journal? If not, discard the claim.
 3. Is the number or direction of effect quoted exactly as the source gives it
    (effect size, confidence interval, absolute versus relative risk)?
 
@@ -144,12 +172,20 @@ Rules:
 - Search only your lane's domains (use site: restricted queries). Open each
   candidate and read the relevant section before using it.
 - Prefer the newest guideline edition, then systematic reviews and meta-analyses,
-  then RCTs, then observational work.
+  then RCTs, then observational work. Within one level, prefer journals ranked SJR
+  Q1 in the question's subject category.
+- Never search or cite preprint servers (bioRxiv, medRxiv, PsyArXiv, arXiv, SSRN,
+  Research Square, OSF Preprints, Preprints.org). Add NOT preprint[pt] to PubMed
+  queries and drop any record typed Preprint.
+- Never cite an MDPI journal: a DOI starting with 10.3390 or a page on mdpi.com. Add
+  -site:mdpi.com to web searches.
 - For each claim return: the claim in one sentence, quoted as the source states it;
   population and setting; effect size with confidence interval when given; evidence
   level (guideline recommendation with class/level or GRADE, meta-analysis, RCT,
   observational, expert opinion); year; full citation (authors or body, title,
-  journal or body, year, PMID or DOI or URL).
+  journal or body, year, PMID or DOI or URL); the journal's SJR quartile; funding:
+  the funders named in the source and one label, vendor-funded, vendor-linked
+  authors, funding not stated or funding not accessible.
 - Lane D only: if UpToDate does not open or shows a paywall, return the single line
   "UpToDate inaccessible" and nothing else.
 - Do not fill gaps from memory. If your lane has nothing on a sub-question, say
@@ -180,8 +216,15 @@ For every claim:
    numbers. Mark "confirmed" or "could not verify: <what differs>".
 3. Confirm the source is in the allowed set: PubMed-indexed peer-reviewed journal,
    ERC/ESC/AHA/CDC or another official health body or specialty society, NEJM, The
-   Lancet, JAMA, BMJ, or UpToDate. Otherwise mark "could not verify: source not
-   allowed".
+   Lancet, JAMA, BMJ, or UpToDate. A preprint, a record typed Preprint in PubMed, or
+   an MDPI journal (DOI starting with 10.3390) is never allowed. Otherwise mark
+   "could not verify: source not allowed".
+4. Check the journal's SJR quartile on scimagojr.com and correct it when the lane
+   got it wrong.
+5. Open the source's Funding and Competing Interests sections and confirm the
+   funders and the label (vendor-funded, vendor-linked authors, funding not stated,
+   funding not accessible). Correct the label when it differs, and flag every
+   vendor-funded or vendor-linked study.
 
 Then return three sections:
 - Confirmed matches: claims supported by two or more independent sources, with
@@ -203,8 +246,10 @@ Then:
 1. **Bottom line** — two to four sentences answering the question directly, with
    citation numbers.
 2. **Key findings** — one bullet per claim: the claim, inline citation number(s)
-   `[n]`, evidence level, year. Group by sub-question when there are several. Mark
-   any source outside the core list.
+   `[n]`, evidence level, year, funding label. Write **vendor-funded** or
+   **vendor-linked authors** in bold and name the company. Group by sub-question
+   when there are several. Mark any source outside the core list and any journal
+   below SJR Q1 with its quartile.
 3. **Agreement and contradictions** — the verifier's confirmed matches and
    contradictions, each contradiction with who says what and the likely reason.
 4. **What is uncertain / not found** — sub-questions with no trustworthy evidence
@@ -214,7 +259,8 @@ Then:
 5. **Method** — one line: parallel subagents or sequential lanes; UpToDate read or
    inaccessible; number of sources opened.
 6. **References** — numbered list matching the inline numbers: authors or body,
-   title, journal or body, year, PMID or DOI or URL.
+   title, journal or body, year, PMID or DOI or URL, SJR quartile, and funding (the
+   funders and the label).
 
 For a question about the user's own health, add exactly one plain sentence after the
 Bottom line: this is a summary of published evidence, not individual medical advice,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "portable-agent-layer",
-  "version": "0.109.0",
+  "version": "0.110.0",
   "description": "PAL — Portable Agent Layer: persistent personal context for AI coding assistants",
   "type": "module",
   "bin": {

package/src/cli/infer.ts

@@ -136,7 +136,7 @@ export const inferCommand = leaf({
     tier: {
       type: "string",
       value: "<tier>",
-      description: `Model size: ${INFERENCE_TIERS.join(" or ")} (default small)`,
+      description: `Model size: ${INFERENCE_TIERS.join(", ")} (default small)`,
     },
     system: { type: "string", value: "<file>", description: "System prompt file" },
     schema: {

package/src/hooks/lib/inference.ts

@@ -62,6 +62,7 @@ import { logDebug, logError } from "./log";
 import {
   type FixedModelRoute,
   type InferenceTier,
+  inferenceEffort,
   inferenceModel,
   isFixedModelRoute,
 } from "./models";
@@ -224,10 +225,21 @@ export function plannedInferenceRoute(
   return { agent: previewRoute().agent, ...plannedRoute({ user: "", tier, caller }) };
 }
 
+function routeWithModel(
+  opts: InferenceOptions
+): ReturnType<typeof plannedRoute> | { error: string } {
+  try {
+    return plannedRoute(opts);
+  } catch (err) {
+    return { error: err instanceof Error ? err.message : String(err) };
+  }
+}
+
 export async function inference(opts: InferenceOptions): Promise<InferenceResult> {
+  const planned = routeWithModel(opts);
+  if ("error" in planned) return { success: false, error: planned.error };
   const started = performance.now();
   const result = await checkedInference(opts);
-  const planned = plannedRoute(opts);
   logInferenceCall(opts, planned, result, performance.now() - started);
   return { ...result, model: planned.model ?? undefined };
 }
@@ -524,9 +536,15 @@ export function buildCodexArgs(opts: InferenceOptions): string[] {
     "--ephemeral",
     "--json",
     ...(opts.web ? ["-c", "web_search=live"] : []),
+    ...codexEffortArgs(opts),
   ];
 }
 
+function codexEffortArgs(opts: InferenceOptions): string[] {
+  const effort = inferenceEffort("codex-spawn", opts.tier);
+  return effort ? ["-c", `model_reasoning_effort="${effort}"`] : [];
+}
+
 interface CodexItem {
   type?: string;
   text?: string;
@@ -1124,31 +1142,38 @@ async function inferenceViaApi(opts: InferenceOptions): Promise<InferenceResult>
 // structured-output schema for JSON-mode callers.
 // ─────────────────────────────────────────────────────────────────────────────
 
+export function buildOpenAiBody(opts: InferenceOptions): Record<string, unknown> {
+  const { system, user, maxTokens = 500, jsonSchema } = opts;
+  const messages: Array<{ role: string; content: string }> = [];
+  if (system) messages.push({ role: "system", content: system });
+  messages.push({ role: "user", content: user });
+
+  const body: Record<string, unknown> = {
+    model: modelFor("openai-api", opts),
+    max_tokens: maxTokens,
+    messages,
+  };
+  const effort = inferenceEffort("openai-api", opts.tier);
+  if (effort) body.reasoning_effort = effort;
+  if (jsonSchema) {
+    body.response_format = {
+      type: "json_schema",
+      json_schema: { name: "structured_response", strict: true, schema: jsonSchema },
+    };
+  }
+  return body;
+}
+
 async function inferenceViaOpenAiApi(opts: InferenceOptions): Promise<InferenceResult> {
   const apiKey = withPalEnv().PAL_OPENAI_API_KEY;
   if (!apiKey) return { success: false };
 
-  const { system, user, maxTokens = 500, timeout = 15000, jsonSchema } = opts;
+  const { timeout = 15000 } = opts;
 
   try {
     const controller = new AbortController();
     const timer = setTimeout(() => controller.abort(), timeout);
-
-    const messages: Array<{ role: string; content: string }> = [];
-    if (system) messages.push({ role: "system", content: system });
-    messages.push({ role: "user", content: user });
-
-    const body: Record<string, unknown> = {
-      model: modelFor("openai-api", opts),
-      max_tokens: maxTokens,
-      messages,
-    };
-    if (jsonSchema) {
-      body.response_format = {
-        type: "json_schema",
-        json_schema: { name: "structured_response", strict: true, schema: jsonSchema },
-      };
-    }
+    const body = buildOpenAiBody(opts);
 
     const response = await fetch("https://api.openai.com/v1/chat/completions", {
       method: "POST",

package/src/hooks/lib/models.ts

@@ -6,11 +6,13 @@ import type { AgentType } from "./agent";
 
 export const HAIKU_MODEL = "claude-haiku-4-5-20251001";
 export const HAIKU_5_5_MODEL = "claude-haiku-5-5";
-export const SONNET_MODEL = "claude-sonnet-5";
-export const FABLE_MODEL = "claude-fable-5";
+export const SONNET_MODEL = "claude-sonnet-5-5";
+export const OPUS_MODEL = "claude-opus-5-5";
+export const FABLE_MODEL = "claude-fable-5-1";
 
-export const INFERENCE_TIERS = ["small", "medium"] as const;
+export const INFERENCE_TIERS = ["small", "medium", "large", "flagship"] as const;
 export type InferenceTier = (typeof INFERENCE_TIERS)[number];
+type ReasoningEffort = "low" | "medium" | "high";
 
 export type FixedModelRoute =
   | "claude-spawn"
@@ -19,24 +21,37 @@ export type FixedModelRoute =
   | "openai-api"
   | "antigravity-spawn";
 
-/** `large` is the flagship that authors skills and subagents; without one, the agent authors inline. */
-type RouteModels = Record<InferenceTier, string> & { large?: string };
+type RouteModels = Record<"small" | "medium" | "large", string> & { flagship?: string };
 
 const ANTHROPIC_MODELS: RouteModels = {
   small: HAIKU_MODEL,
   medium: SONNET_MODEL,
-  large: FABLE_MODEL,
+  large: OPUS_MODEL,
+  flagship: FABLE_MODEL,
 };
 const OPENAI_MODELS: RouteModels = {
   small: "gpt-6-luna",
   medium: "gpt-6-sol",
-  large: "gpt-6-astra",
+  large: "gpt-6-sol",
+  flagship: "gpt-6-astra",
 };
-/** An Antigravity subagent's `model:` takes only a tier name, so `large` is `pro`. */
 const GEMINI_MODELS: RouteModels = {
   small: "gemini-3.8-flash-low",
-  medium: "gemini-3.8-flash-high",
-  large: "pro",
+  medium: "gemini-3.8-flash-medium",
+  large: "gemini-3.8-flash-high",
+};
+
+const OPENAI_EFFORTS: Partial<Record<InferenceTier, ReasoningEffort>> = { large: "high" };
+const INFERENCE_EFFORTS: Partial<
+  Record<FixedModelRoute, Partial<Record<InferenceTier, ReasoningEffort>>>
+> = {
+  "codex-spawn": OPENAI_EFFORTS,
+  "openai-api": OPENAI_EFFORTS,
+};
+
+/** An Antigravity subagent's `model:` takes only a tier name, never a model id. */
+const SUBAGENT_FLAGSHIP: Partial<Record<FixedModelRoute, string>> = {
+  "antigravity-spawn": "pro",
 };
 
 /**
@@ -63,9 +78,13 @@ export function isFixedModelRoute(route: string): route is FixedModelRoute {
  */
 const ANTHROPIC_ON_HAIKU_5_5 = new Set(["rule-drafter", "correction-kind"]);
 
-function evaluatedModel(models: RouteModels, caller?: string): string | undefined {
-  const onAnthropic = models === ANTHROPIC_MODELS;
-  return onAnthropic && caller && ANTHROPIC_ON_HAIKU_5_5.has(caller)
+function evaluatedModel(
+  models: RouteModels,
+  tier: InferenceTier,
+  caller?: string
+): string | undefined {
+  const onAnthropicSmall = models === ANTHROPIC_MODELS && tier === "small";
+  return onAnthropicSmall && caller && ANTHROPIC_ON_HAIKU_5_5.has(caller)
     ? HAIKU_5_5_MODEL
     : undefined;
 }
@@ -76,7 +95,17 @@ export function inferenceModel(
   caller?: string
 ): string {
   const models = INFERENCE_MODELS[route];
-  return evaluatedModel(models, caller) ?? models[tier];
+  const model = evaluatedModel(models, tier, caller) ?? models[tier];
+  if (!model) throw new Error(`${route} has no ${tier} model`);
+  return model;
+}
+
+/** The reasoning effort a tier asks of a model it shares with another tier; undefined keeps the model's default. */
+export function inferenceEffort(
+  route: FixedModelRoute,
+  tier: InferenceTier = "small"
+): ReasoningEffort | undefined {
+  return INFERENCE_EFFORTS[route]?.[tier];
 }
 
 const AGENT_SPAWN_ROUTE: Partial<Record<AgentType, FixedModelRoute>> = {
@@ -88,11 +117,13 @@ const AGENT_SPAWN_ROUTE: Partial<Record<AgentType, FixedModelRoute>> = {
 /**
  * The model an agent's `skill-author` / `subagent-author` runs on, or undefined
  * when `create-skill` and `create-subagent` should author inline. A new provider
- * needs a `large` model on its route and a platform block in assets/agents/*-author.md.
+ * needs a flagship model on its route and a platform block in assets/agents/*-author.md.
  */
 export function flagshipAuthorModel(agent: AgentType): string | undefined {
   const route = AGENT_SPAWN_ROUTE[agent];
-  return route ? INFERENCE_MODELS[route].large : undefined;
+  return route
+    ? (SUBAGENT_FLAGSHIP[route] ?? INFERENCE_MODELS[route].flagship)
+    : undefined;
 }
 
 export interface ModelPricing {
@@ -131,18 +162,31 @@ const withoutWritePremium = (
   cacheRead,
 });
 
+const withWritePremium = (
+  input: number,
+  cacheRead: number,
+  output: number
+): ModelPricing => ({
+  input,
+  output,
+  cacheWrite5m: input * 1.25,
+  cacheWrite1h: input * 1.25,
+  cacheRead,
+});
+
 /**
  * Standard-tier rates from https://developers.openai.com/api/docs/pricing; prompts over
- * 272K tokens cost more, and PAL's never get there. OpenAI bills no cache-write premium.
+ * 272K tokens cost more, and PAL's never get there. GPT-6 and GPT-5.6 bill cache writes
+ * at 1.25x input; older models publish no cache-write price, so writes bill as input.
  */
 const OPENAI_PRICING: Record<string, ModelPricing> = {
-  "gpt-6-astra": withoutWritePremium(10, 1, 50),
-  "gpt-6.1-sol": withoutWritePremium(2, 0.1, 10),
-  "gpt-6-sol": withoutWritePremium(2, 0.2, 10),
-  "gpt-6-luna": withoutWritePremium(0.1, 0.01, 0.5),
-  "gpt-5.6-sol": withoutWritePremium(4, 0.4, 20),
-  "gpt-5.6-terra": withoutWritePremium(2, 0.2, 12),
-  "gpt-5.6-luna": withoutWritePremium(0.2, 0.02, 1.2),
+  "gpt-6-astra": withWritePremium(10, 1, 50),
+  "gpt-6.1-sol": withWritePremium(2, 0.1, 10),
+  "gpt-6-sol": withWritePremium(2, 0.2, 10),
+  "gpt-6-luna": withWritePremium(0.1, 0.01, 0.5),
+  "gpt-5.6-sol": withWritePremium(4, 0.4, 20),
+  "gpt-5.6-terra": withWritePremium(2, 0.2, 12),
+  "gpt-5.6-luna": withWritePremium(0.2, 0.02, 1.2),
   "gpt-5.5": withoutWritePremium(5, 0.5, 30),
   "gpt-5.5-pro": withoutWritePremium(30, 30, 180),
   "gpt-5.4": withoutWritePremium(2.5, 0.25, 15),
@@ -176,14 +220,14 @@ export const MODEL_PRICING: Record<string, ModelPricing> = {
     cacheWrite1h: 0.2,
     cacheRead: 0.01,
   },
-  "claude-fable-5-1": {
+  [FABLE_MODEL]: {
     input: 10,
     output: 50,
     cacheWrite5m: 12.5,
     cacheWrite1h: 20,
     cacheRead: 0.25,
   },
-  [FABLE_MODEL]: {
+  "claude-fable-5": {
     input: 10,
     output: 50,
     cacheWrite5m: 12.5,