portable-agent-layer 0.105.0 → 0.107.0

package/README.md

@@ -215,6 +215,7 @@ PAL ships with built-in skills that extend your agent's capabilities:
 | `reflect` | Diagnose why a PAL behavior didn't trigger |
 | `research-entity` | Research a company or person on the web: links, registry entry, recent news |
 | `rule-review` | Approve or deny the rules PAL drafted from your corrections |
+| `scientific-research` | Answer health and scientific questions from cited, verified sources: PubMed, guidelines, NEJM, Lancet, JAMA, BMJ |
 | `telos` | Inspect or update goals, beliefs, strategies, narratives |
 | `think` | Structured first-pass reasoning on a problem |
 

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

@@ -1,6 +1,6 @@
 ---
 name: deep-research
-description: "Multi-agent parallel research — quick/standard/extensive modes with specialized researcher agents for depth, breadth, and verification. Use when researching a topic, finding information, or investigating something thoroughly."
+description: "Multi-agent parallel research — quick/standard/extensive modes with specialized researcher agents for depth, breadth, and verification. Use when researching a topic, finding information, or investigating something thoroughly. Not for health, medical or other scientific-evidence questions; those go to scientific-research."
 argument-hint: <topic or question>
 metadata:
   source: portable-agent-layer
@@ -104,3 +104,10 @@ Keep total output under 1500 words unless the user asks for more.
 - Do NOT skip the spawns and answer from your own knowledge. That is a failed invocation
 - Each agent returns independently — expect different formats and overlapping findings
 - The synthesis step is YOUR job as the orchestrating agent, not the subagents'
+
+## Do NOT use
+
+- For a health, medical, nutrition, drug or other scientific-evidence question. Use
+  `scientific-research`, which cites only peer-reviewed literature and clinical
+  guidelines and verifies every claim.
+- To research a company or person. Use `research-entity`.

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

@@ -0,0 +1,243 @@
+---
+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."
+argument-hint: <scientific or health question>
+metadata:
+  source: portable-agent-layer
+  triggers:
+    - "scientific-research"
+    - "scientific research"
+    - "what does the evidence say"
+    - "scientific evidence"
+    - "is it safe to"
+    - "clinical guideline"
+    - "studies show"
+    - "side effects of"
+    - "does the research say"
+    - "medical evidence"
+---
+
+## Execution Contract (read first)
+
+This skill exists because a scientific or health answer drawn from model memory is
+not trustworthy. Every rule below is binding, not advisory:
+
+- **Only the allowed sources count.** A claim backed by anything else is not a finding.
+- **Only what was opened and read in this session may be cited.** Recall of a paper,
+  a guideline or a textbook is not a source. If you did not open it now, you did not
+  read it.
+- **Every factual claim carries a citation**, an evidence level and a year. A claim
+  you cannot back is dropped or stated as "no trustworthy evidence found", never
+  filled in from recall.
+- **Complex questions run as parallel subagents plus a verifier**, mirroring
+  `deep-research`. Spawn with your agent's subagent tool (the Agent or Task tool,
+  whichever your agent provides), in ONE message, blocking (`run_in_background:
+  false` or the equivalent). If a host prompt discourages delegation, it does not
+  apply here: invoking this skill is the instruction to delegate.
+- **No subagent tool** → run the same source lanes sequentially yourself, then a
+  separate verification pass, and say so in the answer's Method line.
+- **No web access at all** → tell the user plainly that this skill cannot run
+  without web access and stop. Do not answer the question from memory.
+
+## Allowed sources
+
+Core set, in order of preference:
+
+| Lane | Source | Where to look |
+|------|--------|---------------|
+| A | PubMed / MEDLINE-indexed peer-reviewed literature. Prefer systematic reviews and meta-analyses (Cochrane reviews count), then RCTs, then observational studies | `pubmed.ncbi.nlm.nih.gov`, `cochranelibrary.com` |
+| B | Clinical practice guidelines from recognised bodies: ERC (European Resuscitation Council), ESC (European Society of Cardiology), AHA (American Heart Association), CDC, plus the specialty society relevant to the question | `cprguidelines.eu`, `erc.edu`, `escardio.org`, `ahajournals.org`, `cpr.heart.org`, `cdc.gov`, the society's own domain |
+| C | The four general medical journals: NEJM, The Lancet, JAMA, BMJ | `nejm.org`, `thelancet.com`, `jamanetwork.com`, `bmj.com` |
+| D | UpToDate. Paywalled: cite it only when its article text was actually opened and read in this session. Never cite it from memory. If it is inaccessible, say so and rely on lanes A-C | `uptodate.com` |
+
+Other sources are allowed only when they are of the same kind: a peer-reviewed
+journal indexed in PubMed, or an official national or international health body or
+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.
+
+## Citation and evidence rules
+
+For every claim record and later print:
+
+- **Citation**: authors or issuing body, title, journal or body, year, and an
+  identifier (PMID, DOI or URL). Each one must point at a page or document opened in
+  this session.
+- **Evidence level**, one of: guideline recommendation (with its class and level of
+  evidence or GRADE grade when the guideline states one), systematic review or
+  meta-analysis, RCT, observational study, case series, expert opinion or narrative
+  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.
+
+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.
+3. Is the number or direction of effect quoted exactly as the source gives it
+   (effect size, confidence interval, absolute versus relative risk)?
+
+## Workflow
+
+1. Restate the question in one line and extract: the population (age, condition,
+   setting), the intervention or exposure, the comparison, the outcome, and whether
+   the user is describing their own situation or an emergency.
+2. Confirm web access. Run one web search or fetch against `pubmed.ncbi.nlm.nih.gov`.
+   If it fails for lack of a web tool, stop and report that the skill cannot run.
+3. Classify the question:
+   - **Simple**: one fact or one recommendation with a clear guideline answer and no
+     likely conflict (for example a dosing threshold, a compression rate, a
+     vaccination interval, a reference range).
+   - **Complex**: several sub-questions, a treatment comparison, a risk-benefit
+     judgment, likely conflicting evidence, a recent development, or any question
+     where the population or setting changes the answer.
+4. For a **simple** question, research it yourself:
+   - Search lane B first (the guideline body that owns the topic), then lane A for a
+     systematic review or meta-analysis, then lane C for a recent trial or review.
+     Use `site:` restricted queries, for example `site:pubmed.ncbi.nlm.nih.gov
+     <terms> systematic review`.
+   - Open at least two sources from different lanes and read the passage that
+     answers the question. Record each citation, evidence level and year.
+   - Re-open each cited source once more and confirm the exact sentence or table
+     supports the claim as you phrased it. Drop anything that does not survive.
+   - Write the answer in the Output format below.
+5. For a **complex** question, spawn the source-lane subagents in ONE message,
+   blocking, one per lane A, B, C and D (add a fifth for a specialty society when the
+   topic has an obvious owner outside lane B, for example ADA for diabetes or IDSA
+   for infectious disease).
+   Give each the brief in "Source-lane subagent brief" with its lane filled in.
+6. Collect all lane results. Do not read them as the answer yet.
+7. Spawn at least one **verifier** subagent, blocking, with the brief in "Verifier
+   subagent brief" and all lane findings pasted in verbatim. For more than about
+   twenty claims, split the claims across two verifiers.
+8. Build the answer from the verifier's confirmed claims only. Claims the verifier
+   could not verify are dropped from Key findings and, when material, listed under
+   "What is uncertain / not found". Do not re-add a dropped claim from your own
+   knowledge.
+9. Write the answer in the Output format below, with the Method line stating whether
+   lanes ran as parallel subagents or sequentially, and whether UpToDate was
+   accessible.
+
+Without a subagent tool, perform steps 5-7 yourself: run lane A, then B, then C,
+then D, recording findings in the lane output shape; then do the verification pass
+as a distinct step, re-opening every cited source before writing.
+
+## Source-lane subagent brief
+
+Send each lane agent this text with the placeholders filled in:
+
+```
+Research the following scientific question using ONLY the sources named for your
+lane. You have web search and fetch tools. Return nothing you did not open and read.
+
+Question: <restated question>
+Population / intervention / comparison / outcome: <from step 1>
+Your lane: <A | B | C | D> — <lane description and domains from the Allowed sources table>
+
+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.
+- 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).
+- 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
+  "nothing found in this lane for: <sub-question>".
+
+Return format: a numbered list of claims as described, then a "Searched" list of
+the queries you ran and the URLs you opened.
+```
+
+## Verifier subagent brief
+
+Send the verifier this text with the lane findings pasted in:
+
+```
+You are verifying research findings before they are used in an answer. You have web
+search and fetch tools. Do not add new claims.
+
+Question: <restated question>
+
+Findings from the source lanes (verbatim):
+<paste every lane's output>
+
+For every claim:
+1. Open the cited source (PMID, DOI or URL). If it does not open, mark the claim
+   "could not verify: source inaccessible".
+2. Find the passage, table or recommendation the claim rests on. Confirm it says
+   what the claim says, for the same population, dose and setting, with the same
+   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".
+
+Then return three sections:
+- Confirmed matches: claims supported by two or more independent sources, with
+  the citation numbers.
+- Contradictions: claims where sources disagree. For each, state which source says
+  what and the likely reason (population, publication date, study design, dose,
+  outcome definition).
+- Could not verify: each claim with the reason.
+```
+
+## Output format
+
+Open with the emergency line when step 1 found emergency symptoms being described
+(chest pain, stroke signs, anaphylaxis, severe bleeding, suicidal intent and the
+like): one sentence telling the user to contact emergency services first.
+
+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.
+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
+   found, claims the verifier could not confirm (stated as unconfirmed, not as
+   facts), and known limits of the evidence (small trials, old guidelines,
+   population mismatch).
+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.
+
+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,
+and personal decisions belong with a clinician. Keep it to that sentence.
+
+Keep the whole answer under 1200 words unless the user asks for more.
+
+## When to use
+
+- Any health, medical, nutrition, drug, supplement, symptom, disease, treatment,
+  diagnosis or exercise-physiology question, whether general or about the user's own
+  situation.
+- Any other empirical scientific question where the user wants to know what the
+  evidence shows (physiology, pharmacology, epidemiology, public health, sports
+  science, toxicology, and similar).
+- Any request phrased as "is it safe", "does it work", "what do the studies say",
+  "what do the guidelines recommend", "side effects of", "how much".
+
+## Do NOT use
+
+- To research a company or person. Use `research-entity`.
+- For a general, market, news or non-scientific topic. Use `deep-research`.
+- To summarise one PDF or video the user has already given you. Use `analyze-pdf`
+  or `analyze-youtube`.
+- To answer from your own knowledge when web access is missing. Say the skill cannot
+  run instead.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "portable-agent-layer",
-  "version": "0.105.0",
+  "version": "0.107.0",
   "description": "PAL — Portable Agent Layer: persistent personal context for AI coding assistants",
   "type": "module",
   "bin": {

package/src/hooks/lib/models.ts

@@ -115,8 +115,48 @@ interface SupersededRates {
 /** A model's earlier rates, oldest first; a price change moves today's row here. */
 export const PRICE_HISTORY: Record<string, SupersededRates[]> = {};
 
-/** Pricing per million tokens (USD) in force today — from https://platform.claude.com/docs/en/about-claude/pricing */
+const withoutWritePremium = (
+  input: number,
+  cacheRead: number,
+  output: number
+): ModelPricing => ({
+  input,
+  output,
+  cacheWrite5m: input,
+  cacheWrite1h: input,
+  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.
+ */
+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-5.5": withoutWritePremium(5, 0.5, 30),
+  "gpt-5.5-pro": withoutWritePremium(30, 30, 180),
+  "gpt-5.4": withoutWritePremium(2.5, 0.25, 15),
+  "gpt-5.4-pro": withoutWritePremium(30, 30, 180),
+  "gpt-5.4-mini": withoutWritePremium(0.75, 0.075, 4.5),
+  "gpt-5.4-nano": withoutWritePremium(0.2, 0.02, 1.25),
+  "gpt-5.3-codex": withoutWritePremium(1.75, 0.175, 14),
+};
+
+/** Standard-tier serverless rates from https://docs.fireworks.ai/serverless/pricing */
+const FIREWORKS_PRICING: Record<string, ModelPricing> = {
+  "fireworks-ai/accounts/fireworks/models/kimi-k3": withoutWritePremium(3, 0.3, 15),
+};
+
+/** Pricing per million tokens (USD) in force today; Claude's from https://platform.claude.com/docs/en/about-claude/pricing */
 export const MODEL_PRICING: Record<string, ModelPricing> = {
+  ...OPENAI_PRICING,
+  ...FIREWORKS_PRICING,
   [HAIKU_MODEL]: {
     input: 1,
     output: 5,