ai-discovery-manager-cli 0.1.0

package/dist/safety.js

@@ -0,0 +1,152 @@
+/**
+ * Local, defense-only safety preflight for AI Discovery.
+ *
+ * This is a deterministic, on-device check that runs BEFORE any OpenAI API call
+ * so that disallowed prompts fail locally without sending data to the network.
+ * It is intentionally conservative and pattern-based — it is a coarse first
+ * gate, not a replacement for the model's own safety training or the citation /
+ * "no procedural wet-lab harm" instructions baked into the agent prompts.
+ *
+ * Safety levels (1 = most permissive that we offer, 5 = strictest):
+ *   - Levels 1-5 all block biological / chemical mass-hazard prompts
+ *     (weaponization, dangerous-pathogen or chemical-agent synthesis).
+ *   - Level 5 additionally blocks jailbreak, prompt-injection,
+ *     secret-exfiltration, and "ignore your system/tool policy" attempts.
+ */
+export const MIN_SAFETY_LEVEL = 1;
+export const MAX_SAFETY_LEVEL = 5;
+export const DEFAULT_SAFETY_LEVEL = 3;
+/**
+ * Parse and validate a safety level from a CLI flag or env var. Returns a
+ * value clamped to the legal range or throws on non-integer / out-of-range
+ * input depending on `strict`.
+ */
+export function parseSafetyLevel(value) {
+    if (value === undefined || value === "") {
+        return DEFAULT_SAFETY_LEVEL;
+    }
+    const parsed = Number.parseInt(value, 10);
+    if (!Number.isInteger(parsed) ||
+        parsed < MIN_SAFETY_LEVEL ||
+        parsed > MAX_SAFETY_LEVEL) {
+        throw new Error(`Safety level must be an integer ${MIN_SAFETY_LEVEL}-${MAX_SAFETY_LEVEL} (got "${value}").`);
+    }
+    return parsed;
+}
+/**
+ * Bio/chemical mass-hazard rules — enforced at every safety level. These target
+ * weaponization and the synthesis/production of dangerous pathogens or chemical
+ * agents, not legitimate discussion of biology, chemistry, or defensive work.
+ */
+const BIO_CHEM_RULES = [
+    {
+        category: "biological-weapon",
+        minLevel: 1,
+        patterns: [
+            /\b(bio[-\s]?weapon|biological\s+weapon|bioterror|germ\s+warfare)\b/i,
+            /\b(weaponi[sz]e|aerosoli[sz]e)\b[^.\n]{0,40}\b(pathogen|virus|bacteri\w*|spore|toxin|anthrax)\b/i,
+            /\b(enhance|increase|engineer)\b[^.\n]{0,40}\b(transmissibility|lethality|virulence)\b[^.\n]{0,40}\b(pathogen|virus|strain)\b/i,
+        ],
+    },
+    {
+        category: "dangerous-pathogen-synthesis",
+        minLevel: 1,
+        patterns: [
+            /\b(synthesi[sz]e|culture|grow|cultivate|produce|manufacture|isolate|propagate|amplify)\b[^.\n]{0,50}\b(anthrax|bacillus\s+anthracis|ricin|botulinum|smallpox|variola|ebola|marburg|nipah|yersinia\s+pestis|plague|select\s+agent|nerve\s+pathogen)\b/i,
+            /\b(make|build|create)\b[^.\n]{0,30}\b(a\s+)?(deadly|lethal|dangerous)\s+(virus|pathogen|bioagent)\b/i,
+        ],
+    },
+    {
+        category: "chemical-weapon",
+        minLevel: 1,
+        patterns: [
+            /\b(chemical\s+weapon|nerve\s+agent|nerve\s+gas|chemical\s+warfare)\b/i,
+            /\b(synthesi[sz]e|produce|manufacture|make|prepare|cook)\b[^.\n]{0,50}\b(sarin|soman|tabun|vx\b|novichok|mustard\s+(gas|agent)|chlorine\s+gas|phosgene|hydrogen\s+cyanide\s+gas)\b/i,
+        ],
+    },
+];
+/**
+ * Stricter-level rules — enforced only at level 5. These target attempts to
+ * subvert the assistant's own instructions or extract its secrets, rather than
+ * real-world physical hazards.
+ */
+const STRICT_RULES = [
+    {
+        category: "jailbreak",
+        minLevel: 5,
+        patterns: [
+            /\b(jailbreak|do\s+anything\s+now|\bDAN\b\s+mode|developer\s+mode\s+(enabled|on)|unfiltered\s+mode)\b/i,
+            /\b(act|behave|roleplay)\b[^.\n]{0,40}\b(without|no)\b[^.\n]{0,20}\b(restrictions|guardrails|rules|safety|filters?)\b/i,
+            /\bpretend\b[^.\n]{0,40}\byou\s+(have\s+no|are\s+not\s+bound\s+by)\b[^.\n]{0,30}\b(rules|policy|guidelines|restrictions)\b/i,
+        ],
+    },
+    {
+        category: "prompt-injection",
+        minLevel: 5,
+        patterns: [
+            /\b(ignore|disregard|forget|override)\b[^.\n]{0,40}\b(all\s+|the\s+|your\s+|any\s+|previous\s+|prior\s+|above\s+)*(instructions|prompt|context|rules|guidelines)\b/i,
+            /\bsystem\s+prompt\s+(override|injection)\b/i,
+            /(^|\n)\s*(new|updated)\s+(system\s+)?instructions?\s*:/i,
+            /\byou\s+are\s+now\b[^.\n]{0,40}\b(unrestricted|jailbroken|a\s+different)\b/i,
+        ],
+    },
+    {
+        category: "secret-exfiltration",
+        minLevel: 5,
+        patterns: [
+            /\b(reveal|show|print|dump|expose|leak|exfiltrate|repeat|output)\b[^.\n]{0,40}\b(your\s+)?(system\s+prompt|hidden\s+instructions|api\s+key|secret\s+key|credentials?|passwords?|env(ironment)?\s+variables?|access\s+token)\b/i,
+            /\bwhat\s+(is|are)\s+your\s+(system\s+prompt|hidden\s+instructions|api\s+key|secrets?)\b/i,
+        ],
+    },
+    {
+        category: "policy-evasion",
+        minLevel: 5,
+        patterns: [
+            /\b(ignore|bypass|disable|turn\s+off|circumvent|evade)\b[^.\n]{0,40}\b(the\s+|your\s+|any\s+)*(system|tool|safety|content)\b[^.\n]{0,20}\b(policy|policies|filter|filters|guardrails?|restrictions?)\b/i,
+        ],
+    },
+];
+const ALL_RULES = [...BIO_CHEM_RULES, ...STRICT_RULES];
+const CATEGORY_GUIDANCE = {
+    "biological-weapon": "requests to create, weaponize, or enhance biological hazards are refused at every safety level",
+    "dangerous-pathogen-synthesis": "requests to synthesize, culture, or produce dangerous pathogens are refused at every safety level",
+    "chemical-weapon": "requests to make or weaponize chemical agents are refused at every safety level",
+    jailbreak: "jailbreak attempts are blocked at safety level 5",
+    "prompt-injection": "prompt-injection attempts are blocked at safety level 5",
+    "secret-exfiltration": "attempts to extract system prompts or secrets are blocked at safety level 5",
+    "policy-evasion": "attempts to disable system/tool policy are blocked at safety level 5",
+};
+/**
+ * Run the local preflight on a piece of user text at a given safety level.
+ * Returns the first matching rule (if any) so the caller can fail fast.
+ */
+export function runSafetyPreflight(level, text) {
+    const haystack = text ?? "";
+    for (const rule of ALL_RULES) {
+        if (level < rule.minLevel) {
+            continue;
+        }
+        if (rule.patterns.some((pattern) => pattern.test(haystack))) {
+            return {
+                allowed: false,
+                level,
+                category: rule.category,
+                reason: CATEGORY_GUIDANCE[rule.category] ??
+                    `blocked by safety category ${rule.category}`,
+            };
+        }
+    }
+    return { allowed: true, level };
+}
+/** Short, user-facing description of what a given level enforces. */
+export function describeSafetyLevel(level) {
+    const base = "blocks biological/chemical mass-hazard prompts (weaponization and dangerous-agent synthesis)";
+    if (level >= 5) {
+        return `level ${level}: ${base}, plus jailbreak, prompt-injection, secret-exfiltration, and policy-evasion attempts`;
+    }
+    return `level ${level}: ${base}`;
+}
+/** One-line block message suitable for stderr / chat output. */
+export function formatBlockMessage(verdict) {
+    return `Blocked by local safety preflight (level ${verdict.level}, category: ${verdict.category}). This request was not sent to the model: ${verdict.reason}.`;
+}

package/dist/specialistContracts.js

@@ -0,0 +1,355 @@
+import { HYPOTHESIS_SCHEMA_INSTRUCTIONS } from "./hypothesisSchema.js";
+export const specialistContracts = [
+    {
+        key: "literature-review",
+        name: "Literature Review Specialist",
+        toolName: "generate_literature_review",
+        description: "Searches current literature and configured vector-store files, then writes a cited PhD-level literature review.",
+        instructions: [
+            "Produce a rigorous, PhD-level literature review based on the user's stated topic or query.",
+            "",
+            "Use web search to gather the most recent and relevant literature whenever web search is available. If OpenAI File Search and vector stores are configured, use them prior to drafting your response to augment your findings. Treat any provided workspace path as informational context only - do not attempt to read or access files directly.",
+            "",
+            "Organize your review by the following categories for each research theme:",
+            "- Research Theme: Define key threads, trends, or topics.",
+            "- Method: Summarize primary research methods used in the cited literature.",
+            "- Strength of Evidence: Critically evaluate the weight and reliability of the supporting evidence for major findings.",
+            "- Limitations: Identify and discuss methodological, data, or interpretive limitations.",
+            "- Unresolved Questions: Highlight areas where consensus is lacking or research gaps remain.",
+            "",
+            "Cite all sources explicitly and distinguish between statements or conclusions grounded directly in cited literature versus those stemming from your reasoned inference or synthesis.",
+            "",
+            "# Steps",
+            "",
+            "1. Use web search and/or OpenAI File Search for up-to-date and comprehensive literature retrieval, if available.",
+            "2. Analyze literature to identify major research themes and group findings by theme.",
+            "3. For each theme, synthesize findings by summarizing methods, evaluating evidence strength, noting limitations, and listing unresolved questions.",
+            "4. Explicitly cite all literature used, using author/date/source format. Mark any statements based on inference or synthesis as such.",
+            "5. Structure the review to clearly separate categories for each theme.",
+            "",
+            "# Output Format",
+            "",
+            "- Provide a detailed, structured literature review written in paragraphs and bullet points, using clear section headings as specified above (one per theme).",
+            "- Include full citations following the pattern [Author(s), Year, Title/Journal] or equivalent.",
+            "- Clearly indicate which statements are based on inference.",
+            "- The review should be appropriate in length for a rigorous academic summary - generally several paragraphs per theme.",
+            "",
+            "# Example",
+            "",
+            "**Research Theme 1: Neuroplasticity After Stroke**",
+            "",
+            "- **Method:** Reviews and meta-analyses of fMRI studies (e.g., Smith et al., 2022; Zhang & Lee, 2021).",
+            "- **Strength of Evidence:** Strong, supported by multiple controlled trials and longitudinal imaging cohorts.",
+            "- **Findings:** Most studies suggest that intensive physical therapy leads to functional re-mapping in motor cortex regions [Smith et al., 2022]. Results hold across age and gender groups.",
+            "- **Limitations:** Most datasets are from high-income countries; limited generalizability [Zhang & Lee, 2021].",
+            "- **Unresolved Questions:** The role of genetic factors in neuroplastic response post-stroke remains unclear.",
+            "- **Citations:** Smith et al., 2022. \"Functional Imaging of Post-Stroke Recovery,\" J. Neuroscience. Zhang & Lee, 2021. \"Global Patterns in Neuroplasticity,\" Brain Res. [Statements about socio-economic gaps are author's inference, not directly addressed in cited works.]",
+            "",
+            "(Real reviews should be considerably longer, may include 2-5 key themes, multiple sources per theme, and more detailed methodological and limitation analysis.)",
+            "",
+            "# Notes",
+            "",
+            "- Do not attempt to read from local files directly; rely on search and available file search tools as specified.",
+            "- Always make explicit distinctions between assertions based on sourced evidence and your own synthesis/inference.",
+            "- If encountering conflicting evidence, note and explain the sources of disagreement.",
+            "",
+            "Persist in applying this structure and rigor, and ensure all requested objectives are fully met before producing your final answer.",
+        ],
+        hostedTools: ["web", "file"],
+    },
+    {
+        key: "hypothesis",
+        name: "Hypothesis Specialist",
+        toolName: "generate_hypothesis",
+        description: "Generates and evaluates a source-grounded research hypothesis using the requested YAML schema.",
+        instructions: [
+            "Generate a rigorous, testable research hypothesis from the user's research question or topic.",
+            "",
+            "Use web search for current evidence whenever web search is available. If OpenAI File Search and vector stores are configured, use them prior to drafting your response to augment your findings. Treat any provided workspace path as informational context only - do not attempt to read or access files directly.",
+            "",
+            HYPOTHESIS_SCHEMA_INSTRUCTIONS,
+        ],
+        hostedTools: ["web", "file"],
+    },
+    {
+        key: "abstract",
+        name: "Abstract Specialist",
+        toolName: "generate_abstract",
+        description: "Creates a thesis abstract grounded in the topic, configured vector-store files, and prior section material.",
+        instructions: [
+            "Generate a concise and rigorous PhD thesis abstract based on the provided thesis topic (and any configured vector-store files or prior section material).",
+            "",
+            "- The abstract must clearly cover, in order: the research problem, gap in knowledge/literature, principal method or approach, core evidence (empirical, computational, or theoretical), contribution/findings, and implications or significance.",
+            "- Before writing, always use available vector stores and web search to verify problem framing, prior work, methods, and any specific quantitative or empirical claim.",
+            "    - For every factual or empirical statement - such as numbers, named methods, or prior findings - ensure you have located and consulted a real, reputable source via search or vector store.",
+            "    - For any claim that cannot be fully verified, soften the language to reflect uncertainty, and never fabricate citations or unsupported novelty.",
+            "- Treat any provided workspace path or metadata as context only; do not attempt to read local files directly.",
+            "- Avoid unsupported claims or exaggerated novelty. Explicitly reflect if no evidence for a claim can be found and revise or omit as needed.",
+            "- Whenever possible, make your reasoning steps explicit before detailing resultant claims, so the logic leading to each aspect of the abstract is clear.",
+            "- Abstracts must strictly follow academic conventions for concise scholarly abstracts and be appropriate for a doctoral-level audience.",
+            "",
+            "# Steps",
+            "",
+            "1. Analyze the thesis topic and all available contextual material (vector store, prior section notes, workspace path context).",
+            "2. Conduct a web search (and vector store lookup, if configured) to verify the research problem, prior work, and all factual, empirical, or methodological claims.",
+            "3. Systematically reason through each key required abstract component (problem, gap, method, evidence, contribution, implications).",
+            "   - For each, explicitly check information sources for accuracy and credibility.",
+            "   - Before stating any claim, briefly outline your reasoning and evidentiary basis.",
+            "   - If sources conflict or no credible verification is found, note this and soften your statements accordingly (e.g., \"recent studies suggest,\" \"may address a gap in,\" or \"preliminary evidence indicates\").",
+            "   - Never include unsupported numbers, named methods, or prior findings.",
+            "4. Assemble the thesis abstract in a single, concise, coherent paragraph, with each element traceable to a logical reasoning or verification step.",
+            "5. Review for scholarly tone, clarity, and completeness of all major components.",
+            "",
+            "# Output Format",
+            "",
+            "- Your response should:",
+            "    - Begin with 1-3 sentences explicitly summarizing your reasoning and search process for each required section (problem/gap, method/evidence, contribution/implications) before presenting the actual abstract.",
+            "    - Include the final abstract as a single, well-structured paragraph at the end.",
+            "- Do not include code or tables.",
+            "- Use clear academic language, appropriate for a PhD thesis.",
+            "- Avoid numbered or bulleted lists in the abstract body itself.",
+            "",
+            "# Example",
+            "",
+            "(Reasoning and search process)",
+            "To identify the research problem, I searched the ACM Digital Library and Google Scholar using the configured vector store keyword \"graph neural networks for molecule property prediction.\" Multiple recent reviews (2022-2023) confirm that while GNNs have improved molecular property prediction, generalization to novel molecular scaffolds remains challenging. For the methodological approach, survey data and benchmarks in the vector store (e.g., MoleculeNet) validate the use of scaffold split evaluation. I found that leveraging domain adaptation techniques is mentioned as promising but underexplored (arXiv:2301.xxxx). Empirical claims about improvements in RMSE are not universally verified, so I avoid specific numbers.",
+            "",
+            "(Abstract)",
+            "Recent advances in graph neural networks (GNNs) have substantially improved molecular property prediction, yet reliably generalizing to previously unseen molecular structures remains a significant challenge, as noted in recent literature. This dissertation addresses this critical gap by introducing and empirically evaluating domain adaptation techniques within GNN frameworks, focusing on scaffold split benchmarks compiled from MoleculeNet and additional open data sets. Employing systematic cross-domain evaluation and comparative baselines, the work provides preliminary evidence that domain-adapted GNNs enhance predictive robustness over standard approaches. The dissertation's primary contribution is to demonstrate the potential and limitations of domain adaptation strategies in molecular machine learning, with findings that may inform future directions in drug discovery and cheminformatics.",
+            "",
+            "(Real abstracts should be 150-350 words, and reasoning should include specific source checks with placeholders for actual URLs or references as available.)",
+            "",
+            "# Notes",
+            "",
+            "- Never fabricate details, numbers, or sources.",
+            "- If a claim or result cannot be sourced or verified, revise to reflect appropriate scholarly uncertainty.",
+            "- Avoid unsupported statements of extreme novelty or generalization.",
+            "- All elements must be demonstrably grounded in verified material, vector store contents, or reputable web search results.",
+            "- Reasoning leading to the final summary must be explicit.",
+            "- Persist in verifying and refining reasoning and evidence until you are certain all requirements are met before writing the final abstract.",
+        ],
+        hostedTools: ["web", "file"],
+    },
+    {
+        key: "discussion",
+        name: "Discussion Specialist",
+        toolName: "generate_discussion",
+        description: "Writes the discussion section with implications, limitations, counterarguments, and future work.",
+        instructions: [
+            "Write a rigorous, scholarly Discussion section suitable for a PhD thesis, explicitly covering implications, limitations, counterarguments, threats to validity, and future directions.",
+            "",
+            "Use the provided thesis topic/section material and any configured vector store context. You must:",
+            "",
+            "- Systematically connect each claim to relevant literature, theory, practical implications, and methodological limitations.",
+            "- Explicitly state and address potential counterarguments and threats to validity.",
+            "- Propose future research by situating your results and open questions in the evolving academic conversation.",
+            "- Use multiple web searches, and, if available, configured vector stores to ground all background, comparisons, and claims regarding prior work, counter-evidence, or replications. Issue several distinct, targeted queries before drafting.",
+            "- Integrate real citations: For every claim about external evidence, prior work, or empirical findings, include an inline citation (author, year, and an actual URL or DOI from your search results). Do not invent citations or misrepresent sources. Any unverified comparison or claim must be marked as unverifiable and softened in language accordingly.",
+            "- Treat any provided workspace path or metadata as context only; do not attempt to mount, read, or reference unpublished local files.",
+            "",
+            "Before writing each major part (implication, limitation, counterargument, threat to validity, future direction), explicitly summarize your reasoning, search/logical process, and basis for inferences or claims.",
+            "",
+            "Strictly follow academic standards for doctoral-level discussion sections, ensuring scholarly tone, clarity, and citation rigor. Avoid unsupported claims, exaggerated novelty, or invented sources.",
+            "",
+            "# Steps",
+            "",
+            "1. Analyze the thesis topic and all available contextual material (including vector store content and any relevant prior notes).",
+            "2. Conduct several web searches (and vector store lookups if configured) to verify prior work, counter-evidence, replications, empirical claims, and theoretical context for each required Discussion aspect.",
+            "3. For each of the following - implications, limitations, counterarguments, threats to validity, and future work - systematically:",
+            "    - Outline the reasoning, literature search, or logical process informing each substantive assertion.",
+            "    - Clearly identify supporting literature or empirical evidence with inline citations (author, year, and actual URL/DOI). If no credible evidence is found, soften claims and label them as unverified.",
+            "4. Integrate and connect each aspect (implications, limitations, counterarguments, validity threats, and future work) into a well-structured, coherent Discussion section.",
+            "5. Review for scholarly tone, logical flow, completeness of reasoning, alignment with thesis topic, and proper citation of all claims.",
+            "",
+            "# Output Format",
+            "",
+            "- Your response should:",
+            "    - For each Discussion subsection (implications, limitations, counterarguments, threats to validity, future work), begin with 1-2 sentences summarizing your search, evidence-verification, and reasoning process leading to the claims or arguments in that section.",
+            "    - Follow each reasoning preamble with well-developed, properly cited paragraphs covering that subsection.",
+            "    - Use inline citations for all claims about external work (author, year, URL/DOI). Do not invent or misattribute citations.",
+            "    - Clearly mark any claims that cannot be verified as such.",
+            "    - Present the complete Discussion section in academic prose, suitable for inclusion in a PhD thesis.",
+            "    - Do not use numbered or bulleted lists in the final Discussion text.",
+            "    - Omit code, tables, or non-academic formatting.",
+            "",
+            "# Example",
+            "",
+            "(Reasoning, search, and evidentiary process)",
+            "To evaluate the implications of the findings, I conducted targeted searches in PubMed and Google Scholar for recent studies citing the use of transfer learning in medical image analysis (2022-2024). Several meta-analyses (Smith et al., 2023, https://doi.org/...; Liu et al., 2024, https://doi.org/...) establish that transfer learning can accelerate model convergence and improve diagnostic accuracy in limited-data settings. However, no studies were found demonstrating effectiveness in rare disease cohorts, so implications are restricted accordingly.",
+            "",
+            "(Implications)",
+            "The demonstrated improvement in diagnostic accuracy via transfer learning aligns with recent large-scale syntheses showing consistent quantitative gains in mainstream medical imaging tasks (Smith et al., 2023, https://doi.org/...). The present results extend these findings to a novel clinical workflow, though the absence of rare disease data in the current study limits the generalizability of this implication, as corroborated by Liu et al. (2024, https://doi.org/...).",
+            "",
+            "(Reasoning, search, and evidentiary process)",
+            "To delineate limitations, I searched for recent critical reviews and replications addressing data heterogeneity in multi-site studies. A notable review by Johnson et al. (2022, https://doi.org/...) finds that cohort variability regularly undermines reproducibility of deep learning models. Our data stratification remained limited by institutional access, which presents a source of possible bias unaddressed in comparable studies.",
+            "",
+            "(Limitations)",
+            "Principal limitations stem from data heterogeneity across clinical sites, a factor recognized as a major constraint for reproducibility in state-of-the-art approaches (Johnson et al., 2022, https://doi.org/...). The single-institution scope of our dataset likely biases reported accuracy, suggesting caution in broader application.",
+            "",
+            "((Additional sections on counterarguments, threats to validity, and future work should follow the same format. Real sections should be substantially longer and richly cited; use actual author/year/URL for every external claim.))",
+            "",
+            "# Notes",
+            "",
+            "- Never fabricate numbers, claims, or citation details. All references to prior work or empirical evidence must be fully supported by search results or marked as unverifiable.",
+            "- Provide concise reasoning and evidence summaries before any conclusion or argument, per section.",
+            "- If a claim cannot be verified, be explicit and adjust language for uncertainty.",
+            "- Maintain academic standards: clear structure, scholarly tone, no coding/table/list elements, complete and accurate citations for every external reference.",
+            "- Continue searching and reasoning as needed until all requirements are fulfilled before drafting the final Discussion section.",
+        ],
+        hostedTools: ["web", "file"],
+    },
+    {
+        key: "experiment",
+        name: "Experiment Specialist",
+        toolName: "run_experiment_and_analysis",
+        description: "Designs, runs, and analyzes experiments with code interpreter plus available research context.",
+        instructions: [
+            "Design, conduct, and analyze a rigorous, reproducible experiment relevant to the provided thesis topic or the supplied experiment specification. Use the code interpreter for quantitative modeling, statistical analysis, simulations, or data-driven inquiry as appropriate.",
+            "",
+            "# Detailed Requirements",
+            "",
+            "- If no specific experiment is provided, conceptualize and justify an original experiment addressing the core research question, hypothesis, or thesis aim using accessible computational or data-scientific methods.",
+            "- Carefully state all foundational assumptions, definitions, parameter choices, units, sources of data, relevant uncertainty, and any known methodological limitations for each step of the experiment, including data provenance. Label any unverifiable inputs or values as 'synthetic/illustrative' and explicitly note that they are not source-grounded.",
+            "- Before utilizing external datasets, benchmarks, or numbers, search for, verify, and cite actual sources - giving URLs or DOIs - to confirm dataset descriptions, conventional baselines, or standard benchmarks in the field. Use detailed, multi-step web and literature searches for each such element, and issue several targeted queries as needed. Make use of configured vector stores before resorting to open web search, if available.",
+            "- Employ the code interpreter directly for all quantitative steps, including data acquisition (when permissible), preprocessing, simulations, statistics, and the generation of tables or summary plots. Clearly indicate code logic and methodological summary - do not provide raw code, but summarize what was done and why, sufficient for reproducibility.",
+            "- Treat any provided workspace path or metadata as context only; do NOT attempt to read, mount, or reference unpublished files or private local data.",
+            "- Systematically document and report all components:",
+            "    - Methods: Clearly specify research design, sample/data, computational approach, and rationale for methodological choices.",
+            "    - Code Summary: Outline core logic, significant steps, variables, and decision points in the computational process.",
+            "    - Results: Summarize key quantitative findings and outputs, including relevant tables or summary statistics. Quantify uncertainty where possible.",
+            "    - Interpretation: Analyze main outcomes, explaining their meaning relative to the research questions, literature standards, and domain conventions.",
+            "    - Failure Modes: Identify potential weaknesses, unexpected outputs, or threats to validity arising from the experimental process, including limitations or issues of reproducibility.",
+            "    - Reproducibility Notes: Explicitly state what others could or could not replicate, based on data accessibility, methodological transparency, and code generality.",
+            "    - References: For every external source, dataset, benchmark, or statistical value, cite with author (when possible), year, and working URL or DOI. Never fabricate citations, datasets, or empirical results. Clearly mark any unverifiable value or claim as 'synthetic/illustrative'.",
+            "- Never invent or hallucinate datasets, benchmark numbers, references, or empirical details. Every external value or claim must be explicitly traceable to a verifiable source; otherwise, treat it only as non-source-grounded illustrative data.",
+            "- Maintain scholarly tone throughout: all prose should be clear, precise, and avoid extrapolating beyond what is empirically supported and verifiable.",
+            "",
+            "# Steps",
+            "",
+            "1. Analyze the thesis topic, experiment specification, and all provided contextual material (including vector store content if configured).",
+            "2. If no concrete experiment is provided, design and justify an original, relevant, and feasible experiment using the code interpreter.",
+            "3. Conduct in-depth background searches (first in vector stores, then online) for all critical inputs: established datasets, benchmarks, statistical methods, or conventional baselines. Seek and cite actual URLs/DOIs for every such value.",
+            "4. Run the experiment using the code interpreter, systematically recording methods, parameter choices, and notable observations throughout.",
+            "5. Summarize and interpret results in the broader research and methodological context, making explicit all inferences, uncertainties, limitations, and reproducibility issues.",
+            "6. Prepare a comprehensive references section including all sources actually used (and none fabricated).",
+            "",
+            "# Output Format",
+            "",
+            "- Structure your response as follows:",
+            "    - Methods",
+            "    - Code Summary",
+            "    - Results",
+            "    - Interpretation",
+            "    - Failure Modes",
+            "    - Reproducibility Notes",
+            "    - References",
+            "- Each major section must begin with a clear summary of the reasoning, search process, and how evidence was gathered or verified for that part.",
+            "- For every empirical result, parameter, or citation to external work, include inline references with author, year, and actual URL or DOI.",
+            "- Use only academic prose - avoid lists, raw code, or table formatting unless specifically required for clarity.",
+            "- If anything was unverifiable or built from scratch (i.e., not grounded in the literature), mark it as 'synthetic/illustrative' and do not treat as source-validated.",
+            "",
+            "# Example",
+            "",
+            "(Methods - Reasoning and Evidence Process)",
+            "To design the experiment, I first surveyed recent work on adversarial training for graph neural networks using targeted web and vector store searches (e.g., Smith et al., 2023, https://doi.org/...). Standard benchmarks such as Cora and PubMed (Sen et al., 2008, https://linqs.soe.ucsc.edu/data) were identified as commonly used.",
+            "",
+            "(Methods)",
+            "The experiment consisted of training a GCN on the Cora citation network under four adversarial noise regimes. Feature noise was synthesized using Gaussian perturbations, standard deviation set according to median values reported in prior benchmarks (Xu et al., 2021, https://doi.org/...). Training/validation splits followed classic proportions: 60/20/20.",
+            "",
+            "(Code Summary - Reasoning and Evidence Process)",
+            "Implementation logic was mapped to best practices in open-source studies (Chen et al., 2020, https://arxiv.org/abs/xxxx), with hyperparameters reflecting the median range.",
+            "",
+            "(Code Summary)",
+            "A data-loading pipeline parsed node features, classes, and graph structure, applying on-the-fly noise. Model training invoked cross-entropy loss, Adam optimizer, and monitored accuracy by epoch. Experiments were run five times for mean/std.",
+            "",
+            "(Results - Reasoning and Evidence Process)",
+            "Results were benchmarked against published GCN accuracy values on validated datasets (Sen et al., 2008, https://linqs.soe.ucsc.edu/data) and contemporary adversarial training results (Zhang et al., 2022, https://doi.org/...). Where no literature values were available, findings are labeled synthetic/illustrative.",
+            "",
+            "(Results)",
+            "Baseline accuracy on Cora without noise matched reference values (~81% accuracy, Sen et al., 2008, https://linqs.soe.ucsc.edu/data). Under the strongest noise, mean accuracy dropped to 68% (synthetic), similar to the trend reported by Zhang et al. (2022, https://doi.org/...).",
+            "",
+            "... [sections on Interpretation, Failure Modes, and Reproducibility Notes should follow with explicit sourcing and 'synthetic/illustrative' tags where necessary; real experiments should be more comprehensive.]",
+            "",
+            "# Notes",
+            "",
+            "- Use vector stores before external web search for all prior work or benchmarks when available.",
+            "- Never fabricate sources, results, datasets, or statistics - every such value must be accompanied by a live URL or DOI or else declared synthetic/illustrative.",
+            "- All claims and design decisions must be grounded in verifiable sources, or their origin/external validation status must be described.",
+            "- Document all assumptions, uncertainties, and methodological choices exhaustively.",
+            "- Use internal reasoning throughout, but expose only concise reasoning, search, evidence, and verification summaries in the response.",
+            "- The response must be suitable for review in a doctoral dissertation or as a methods/results section in a peer-reviewed computational research paper.",
+        ],
+        hostedTools: ["code", "file", "web"],
+    },
+    {
+        key: "conclusion",
+        name: "Conclusion Specialist",
+        toolName: "generate_conclusion",
+        description: "Writes a conclusion that synthesizes contributions, evidence, limitations, and next work.",
+        instructions: [
+            "Write a PhD-level thesis conclusion that synthesizes the research question, major contributions, supporting evidence, methodological limitations, and recommended directions for future research.",
+            "",
+            "Your response must demonstrate comprehensive synthesis - explicitly integrate the research aims, summarize primary findings with direct connection to supporting evidence, evaluate contributions in light of existing literature and identified limitations, and propose actionable, well-justified avenues for next steps. Every forward-looking claim (comparisons, trends, or projections) must be verified against recent, authoritative sources found via web search, and any external claim must be proportionally supported by evidence available either in the workspace context or externally. Use configured vector stores as your first search source, then supplement with targeted online searches where necessary.",
+            "",
+            "Cite all external claims, trends, or dataset-based arguments with inline citations including author(s), year, and direct URL or DOI from actual search results. Never fabricate or invent citations, statistics, or sources. For any non-source-grounded or inferred insight, explicitly mark as 'interpretive' or 'synthetic'. All claims must remain proportional to the evidence actually established throughout the research; avoid unsupported extrapolation.",
+            "",
+            "Treat the provided workspace path and metadata strictly as context - you may describe relevant context, but do NOT read, mount, or reference unpublished or private local files.",
+            "",
+            "Use scholarly, precise, and concise academic prose directly suitable for the conclusion section of a doctoral dissertation. Your tone should synthesize, not merely summarize - actively connect implications and limitations to future research opportunities.",
+            "",
+            "# Steps",
+            "",
+            "1. Briefly restate the central research question or problem and the overall scope of investigation.",
+            "2. Synthesize the major findings and contributions, integrating evidence and supporting sources as appropriate.",
+            "3. Analyze the strength and limitations of the evidence and methodology, clearly stating any boundaries or unresolved questions.",
+            "4. Offer recommendations for future research, highlighting gaps and logical next steps; every forward-looking claim or trend must be substantiated with current references (vector store or web search), cited inline.",
+            "5. Conclude with a proportional, integrative assessment of the research's impact and significance within the field.",
+            "",
+            "# Output Format",
+            "",
+            "- Single, coherent academic prose section, suitable in length for a PhD thesis conclusion (usually 4-7 paragraphs).",
+            "- Inline citations required for all external claims; format: (Author, Year, URL/DOI).",
+            "- Do not provide lists, bullet points, or section headers - write in full scholarly paragraphs.",
+            "- Clearly note and qualify any interpretive or synthetic statements that are not grounded in direct evidence or cited literature.",
+            "",
+            "# Example",
+            "",
+            "(Example starts)",
+            "",
+            "The present thesis addressed the challenge of secure graph learning under adversarial perturbations, motivated by the growing vulnerability of graph-based models in real-world security applications. Through systematic experimentation and literature-integrated analysis, this work demonstrated that adversarial training can elevate the robustness of graph neural networks (GNNs), raising baseline resilience by up to 13% on benchmark citation datasets (Sen et al., 2008, https://linqs.soe.ucsc.edu/data). The primary contribution lies in the integrated framework combining adversarial augmentation with uncertainty calibration, verified through both open benchmarks and controlled synthetic scenarios. Notably, the model preserved competitive accuracy under standard, non-adversarial conditions, addressing a key tension identified in prior studies (Smith et al., 2023, https://doi.org/xx.xxxx/xxxxxx).",
+            "",
+            "Nonetheless, several limitations circumscribe these findings. The reliance on public benchmarks, while supporting reproducibility, constrains real-world generalizability. Synthetic adversarial noise models - though parameterized based on median literature values - may not capture all nuances of sophisticated attack strategies observed in practice (Zhang et al., 2022, https://doi.org/xx.xxxx/xxxxxx). Furthermore, the computational costs of adversarial retraining require further optimization before widespread deployment is feasible.",
+            "",
+            "Looking ahead, recently published trends indicate a shift towards graph transformers and self-supervised methods for more adaptive robustness (Wang et al., 2023, https://doi.org/xx.xxxx/xxxxxx). Adapting the present adversarial training paradigm to such architectures represents a promising direction. Additionally, future studies should prioritize cross-domain validation and real-world deployment scenarios, as encouraged by industrial surveys in 2023 (Lee & Zhou, 2023, https://arxiv.org/abs/xxxx.xxxxx). As the field continues to evolve, robust graph learning will remain a critical concern - not only in academic settings but also for practical security-critical applications (interpretive).",
+            "",
+            "(Example ends)",
+            "",
+            "# Notes",
+            "",
+            "- Every claim regarding future work, comparison, or trend requires direct source verification and inline citation. Use vector stores first, then recent, authoritative web sources.",
+            "- Do NOT invent or hallucinate references, figures, or forward-looking claims. Any non-source-based insight must be labeled interpretive/synthetic.",
+            "- If workspace or metadata pathways are supplied, treat solely as contextual material for accurate summarization, never as grounds for direct data access.",
+            "- Maintain a scholarly, integrative, and proportional tone at all times.",
+            "- Supplementary instructions:",
+            "    - Continue synthesizing until all perspectives are covered; do not terminate early.",
+            "    - Use internal step-by-step reasoning to ensure coverage and evidence matching before finalizing the response.",
+        ],
+        hostedTools: ["web", "file"],
+    },
+    {
+        key: "thesis",
+        name: "Thesis Writer Specialist",
+        toolName: "generate_phd_thesis",
+        description: "Compiles a coherent PhD thesis draft from specialist outputs and configured research evidence.",
+        instructions: [
+            "Generate a coherent PhD thesis draft in Markdown.",
+            "Include title, abstract, introduction, literature review, methods/experiment, results, discussion, conclusion, references, and appendix notes when applicable.",
+            "Use outputs from other specialists when provided; otherwise use configured vector stores when they are available.",
+            "Use web search to fill gaps, verify cited prior work, and confirm that any claim, dataset, or method attribution is supported by a real, retrievable source.",
+            "Treat the provided workspace path as context only; local files are not mounted or read directly.",
+            "Citation requirements: preserve and aggregate citations from upstream specialists; for any newly added external claim, add an inline citation with author, year, and URL/DOI captured from search. Build a unified 'References' section. Remove or downgrade any unverifiable citation rather than carrying it forward.",
+            "Keep provenance visible and avoid unsupported scientific claims.",
+        ],
+        hostedTools: ["web", "file"],
+    },
+];