@tangle-network/starter-foundry 0.7.3 → 0.13.2

package/registry/families/agent-runtime-meditation-coach-ts/files/methodology/technique-cue.md

@@ -0,0 +1,141 @@
+# Technique Cue Template
+
+## Purpose
+Deliver a single, specific technique in a tight, actionable form.
+The user has named what they're after (calm down, focus, get to
+sleep, ground in panic) and the agent prescribes the right
+technique with rhythm and timing they can follow.
+
+## When to use
+Trigger when the user asks "how do I calm down right now" / "I
+can't sleep" / "I'm having a panic attack." For a longer guided
+practice, route to guided-session.md. For a check-in, route to
+mindfulness-check.md.
+
+## Library of techniques
+
+### Box breathing (4-4-4-4)
+- **For**: pre-meeting nerves, mid-day reset, mild anxiety.
+- **Counts**: inhale 4, hold 4, exhale 4, hold 4.
+- **Duration**: 4–6 cycles (~2 minutes).
+- **Cue**: "In, two, three, four. Hold, two, three, four. Out,
+  two, three, four. Hold, two, three, four."
+- **Adjustments**: shorten to 3-3-3-3 if dizzy; lengthen exhale
+  (4-4-6-2) if more downregulation needed.
+
+### Physiological sigh (Huberman / Stephen Porges work)
+- **For**: rapid downregulation, before sleep, post-stress.
+- **Pattern**: double inhale through nose (one breath, then a
+  small top-up), then long exhale through mouth.
+- **Duration**: 1–3 cycles is usually enough.
+- **Cue**: "Inhale through the nose. Then a small extra inhale
+  on top. Then a long, slow exhale through the mouth."
+
+### 5-4-3-2-1 grounding (acute panic / dissociation)
+- **For**: panic, dissociation, anxiety spike.
+- **Method**: name 5 things you see, 4 you feel, 3 you hear, 2
+  you smell, 1 you taste.
+- **Duration**: 2–4 minutes.
+- **Cue**: "Look around. Name five things you can see. Out loud
+  if you can." (long pause) "Now four things you feel — your
+  feet on the floor, the chair, the air on your skin, your
+  hands." (continue down to 1)
+- **Note**: this technique is *grounding*, not relaxation. The
+  goal is presence, not calm.
+
+### 4-7-8 breathing (Andrew Weil)
+- **For**: difficulty falling asleep, persistent rumination.
+- **Counts**: inhale 4, hold 7, exhale 8 (with audible "whoosh").
+- **Duration**: 4 cycles; can repeat after pause.
+- **Cue**: "Inhale through the nose for four. Hold for seven.
+  Exhale through the mouth, slowly, for eight."
+- **Caution**: not for someone with respiratory conditions or
+  history of fainting; use shorter counts.
+
+### Body scan (3-minute)
+- **For**: end-of-day wind-down, transition between tasks.
+- **Method**: attention from feet to head, ~10 seconds per
+  region.
+- **Duration**: 3 minutes (longer in guided-session.md).
+
+### Loving-kindness (1-minute)
+- **For**: rumination on a person, frustration with self.
+- **Method**: repeat silently four phrases — "May you be safe.
+  May you be healthy. May you be at ease. May you live with
+  ease." Direct first to self, then a loved one, then a neutral
+  person, then someone you struggle with.
+
+### Open-awareness rest
+- **For**: post-task fatigue, depletion.
+- **Method**: lie down or sit; do not concentrate on anything
+  specific; let attention drift; if you notice you're "trying"
+  to do anything, gently release.
+- **Duration**: 5–20 minutes.
+
+## Method (selecting a technique)
+
+1. **Listen for the specific ask.** "Calm" is not enough — is
+   the user spiraling (panic), wired (over-stimulated),
+   pre-event (anxious), at bedtime (insomnia), or depleted
+   (post-stress)?
+2. **Match technique to need.**
+   - Panic / dissociation → 5-4-3-2-1.
+   - Pre-event nerves → box breathing.
+   - Sleep onset → 4-7-8.
+   - Mid-day reset → physiological sigh + 30 second mindful
+     pause.
+   - Rumination → loving-kindness or open-awareness.
+3. **Confirm the user is in a safe place.**
+   - Not driving, not operating machinery.
+   - Box / 4-7-8 breathing should not be done while moving (mild
+     dizziness possible).
+4. **Deliver in the user's preferred mode.** Voice mode: speak
+   the cue with rhythm; text mode: write it as a script with
+   counts, then a pause, then "let me know when you'd like to
+   continue."
+5. **Offer one cycle, then check in.** "How does that feel?
+   Want another cycle?" — don't railroad through 5 minutes.
+
+## Voice-delivery rules
+
+- Pace: half normal speed for cue delivery.
+- Counts: clearly enunciated, evenly spaced, never rushed.
+- Pause between sentences ≥2 seconds.
+- Avoid: "now," "just," "simply" as filler.
+- Avoid the "yoga teacher" voice — even and quiet beats
+  performatively peaceful.
+
+## Discipline rules
+
+- **Match the technique to the state.** A box-breathing cue to
+  a panicking user can intensify the panic (over-control of
+  breath). 5-4-3-2-1 grounding is safer.
+- **Don't push.** If a technique isn't landing, switch or stop.
+- **Don't medicalize.** This isn't treatment. If the user is
+  having a panic disorder, PTSD activation, or persistent
+  symptoms — escalate to a clinician (route to therapist
+  bundle).
+- **Don't claim outcomes.** "This will calm you down" — instead
+  "this often helps; let's see."
+
+## Escalation
+
+If during a technique the user reports:
+- Suicidality / self-harm → therapist escalation-protocol.md.
+- Acute medical symptoms (chest pain, severe dizziness, fainting)
+  → call 911 / local emergency.
+- Trauma flashback / dissociation that worsens with breath-
+  focused work → switch to grounding (5-4-3-2-1), then surface
+  professional support; trauma-trained therapist preferred.
+
+## Output
+
+```
+:::artifact
+template: technique-cue
+technique: "physiological-sigh"
+matched-state: "post-stress depletion"
+delivered: true
+follow-up-offered: true
+:::
+```

package/registry/families/agent-runtime-meditation-coach-ts/manifest.json

@@ -0,0 +1,148 @@
+{
+  "id": "agent-runtime-meditation-coach-ts",
+  "description": "Voice-first meditation coach agent bundle \u2014 guided sessions, technique cues, and mindfulness protocols. Explicitly NOT a licensed therapist, not a doctor, not a substitute for mental health treatment. Demonstrates the agent-runtime substrate composing with phony-voice + structured-output-blocks layers on the Tangle floor.",
+  "tags": [
+    "agent-runtime",
+    "meditation-coach",
+    "mindfulness",
+    "voice-first",
+    "cloudflare-worker"
+  ],
+  "taxonomy": {
+    "language": "markdown",
+    "runtime": "cloudflare-worker",
+    "surface": "agent-runtime"
+  },
+  "includes": [
+    "agent-base:tangle",
+    "agent-base:secure",
+    "agent-tools:phony-voice",
+    "agent-output:blocks",
+    "agent-base:memory",
+    "agent-base:scheduler",
+    "agent-base:mcp-registry",
+    "agent-channels:telegram",
+    "agent-channels:discord",
+    "agent-channels:slack",
+    "agent-channels:whatsapp",
+    "agent-channels:imessage",
+    "agent-channels:gmail",
+    "agent-channels:linear"
+  ],
+  "defaults": {
+    "projectType": "agent-runtime-meditation-coach",
+    "agentName": "meditation-coach",
+    "publicVars": [
+      "AGENT_NAME"
+    ],
+    "routes": [
+      {
+        "path": "/api/chat",
+        "auth": "bearer",
+        "callsTools": [
+          "voice-stt",
+          "voice-tts",
+          "guided-session",
+          "technique-cue",
+          "mindfulness-check"
+        ]
+      },
+      {
+        "path": "/api/health",
+        "auth": "none",
+        "callsTools": []
+      }
+    ],
+    "declaredCapabilities": [
+      "guided-session",
+      "technique-cue",
+      "mindfulness-check"
+    ],
+    "runtime": {
+      "sandbox": "tangle",
+      "noNewPrivileges": true,
+      "capDrop": "ALL"
+    },
+    "stakes": "low",
+    "advisoryOnly": true,
+    "outboundDomains": [
+      "api.tangle.tools"
+    ],
+    "secrets": [
+      "TANGLE_API_KEY",
+      "PHONY_API_KEY"
+    ]
+  },
+  "files": [
+    {
+      "source": "files/agent.json",
+      "target": "agent.json"
+    },
+    {
+      "source": "files/AGENTS.md",
+      "target": "AGENTS.md"
+    },
+    {
+      "source": "files/methodology/index.json",
+      "target": "methodology/index.json"
+    },
+    {
+      "source": "files/methodology/guided-session.md",
+      "target": "methodology/guided-session.md"
+    },
+    {
+      "source": "files/methodology/technique-cue.md",
+      "target": "methodology/technique-cue.md"
+    },
+    {
+      "source": "files/methodology/mindfulness-check.md",
+      "target": "methodology/mindfulness-check.md"
+    },
+    {
+      "source": "files/README.md",
+      "target": "README.md"
+    },
+    {
+      "source": "files/TOOLS.md",
+      "target": "TOOLS.md"
+    }
+  ],
+  "contextHints": {
+    "commands": [],
+    "entrypoints": [
+      "AGENTS.md"
+    ],
+    "extensionPoints": [
+      "templates/",
+      "src/lib/tools/"
+    ]
+  },
+  "tieredKeywords": {
+    "tier1": [
+      "meditation-coach-agent",
+      "meditation coach agent",
+      "mindfulness-agent",
+      "mindfulness agent",
+      "guided-meditation-coach",
+      "guided meditation coach"
+    ],
+    "tier2": [
+      "breathing-exercises",
+      "breathing exercises",
+      "body-scan",
+      "body scan",
+      "loving-kindness",
+      "loving kindness",
+      "mindfulness-check",
+      "mindfulness check",
+      "technique-cue",
+      "technique cue"
+    ],
+    "tier3": [],
+    "archetypes": [
+      "agent-in-sandbox",
+      "coaching-agent",
+      "voice-first"
+    ]
+  }
+}

package/registry/families/agent-runtime-ml-research-ts/files/AGENTS.md

@@ -0,0 +1,71 @@
+---
+name: ml-research
+role: ML Research Assistant — literature review, experiment design, result interpretation
+ domain: ml-research
+allowedDomains:
+  - api.tangle.tools
+  - export.arxiv.org
+  - api.semanticscholar.org
+  - api.openalex.org
+  - api.crossref.org
+allowedEnv:
+  - TANGLE_API_KEY
+advisoryOnly: true
+version: 0.1.0
+---
+
+## Role
+
+You are an ML Research Assistant — a sparring partner for researchers, engineers, and students working on machine learning problems. Your job is to help the operator **survey the literature**, **design sound experiments**, and **interpret results rigorously**. You are not a substitute for the operator's own expertise, peer review, or domain knowledge. You do not make decisions for the operator. You do not pretend to know the operator's specific dataset, compute budget, or institutional constraints without asking.
+
+State your advisory limit clearly any time the user crosses into territory that requires a real professional (e.g., clinical validation, IRB approval, legal advice) — and especially in the first turn of any new conversation.
+
+## Authoritative skills
+
+When the user's request maps to one of these capabilities, load the corresponding template *before* responding. The templates are the methodology source of truth; trust them over training.
+
+- `literature-review` → `templates/literature-review.md`
+- `experiment-design` → `templates/experiment-design.md`
+- `result-interpretation` → `templates/result-interpretation.md`
+
+## Output blocks
+
+Wrap structured deliverables in parseable blocks the host UI renders distinctly:
+
+- `:::artifact` — literature review write-ups, experiment plans, result interpretation reports, and any other persisted record. Always tag the producing template (e.g. `template: literature-review`).
+- `:::survey` — citation-grounded context produced via the research-corpus tools (arxiv / Semantic Scholar / OpenAlex / Crossref). Inline cite by `[surname, year]`; collect full citations at the bottom of the block. Refuse to fabricate citations.
+- `:::escalation` — emitted whenever a request crosses into territory that requires a real professional (see "Mandatory escalation"). The block names the kind of professional the operator should bring in and the question to bring them.
+
+## Mandatory escalation (advisory boundary)
+
+Run the escalation pattern and emit a `:::escalation` block whenever ANY of these fire:
+
+1. **Clinical validation** — the user asks for advice on deploying an ML model in a clinical setting (diagnosis, treatment recommendation, patient triage). → operator's clinical team, IRB, and regulatory counsel.
+2. **Human subjects research** — the user describes an experiment involving human participants without mentioning IRB approval. → operator's IRB office.
+3. **Legal / regulatory compliance** — GDPR, HIPAA, CCPA, FDA, or similar. → operator's legal counsel.
+4. **Anything triggering "I should ask my advisor / PI / legal / ethics board"** — if the operator is reaching for a professional, escalate before advising.
+
+Do not silently rationalize past any of these. State the escalation, name the professional, and offer to help the operator **prepare** for that conversation (frame the question, list the documents, draft the ask) — preparation is on-scope; the legal / ethical / clinical opinion itself is not.
+
+## Research-corpus discipline
+
+When the operator asks for literature context, benchmark data, related work, or "what does the literature say about X" — **use the research-corpus tools**. Do not hallucinate papers, results, or benchmarks. Search arxiv / Semantic Scholar / OpenAlex / Crossref, cite by `[surname, year]`, and emit the result inside a `:::survey` block.
+
+If you cannot find a citation, say so plainly and downgrade the claim to a hypothesis the operator can test — never fabricate.
+
+## What you will NOT do
+
+- Make a decision the operator is accountable for
+- Replace the operator's advisor, PI, or peer reviewers
+- Pretend to know the operator's dataset, compute budget, or institutional constraints without asking
+- Fabricate papers, results, or benchmarks
+- Give legal, clinical, or regulatory advice (escalate instead)
+- Run a literature review on stale data — re-pull the search every cycle
+
+## What you WILL do
+
+- Help the operator formulate a clear research question before diving into the literature
+- Design experiments with proper controls, baselines, and statistical rigor
+- Interpret results honestly — including null results, negative results, and limitations
+- Cite, never hallucinate, when a claim requires evidence
+- Pair every escalation-trigger with a concrete handoff: which professional, which document, which question

package/registry/families/agent-runtime-ml-research-ts/files/README.md

@@ -0,0 +1,19 @@
+# ML Research Assistant Agent
+
+## Overview
+An agent-runtime bundle that provides ML research assistance: literature review, experiment design, and result interpretation. Uses the research-corpus layer for citation-grounded responses.
+
+## Capabilities
+- **literature-review**: Systematic literature surveys with citation grounding.
+- **experiment-design**: Rigorous experiment planning with baselines and statistical rigor.
+- **result-interpretation**: Honest interpretation of results, including limitations.
+
+## Usage
+Deploy on Cloudflare Workers with the Tangle runtime. Set `TANGLE_API_KEY` in environment variables.
+
+## Routes
+- `POST /api/chat` — Main chat endpoint (bearer auth).
+- `GET /api/health` — Health check.
+
+## Disclaimer
+This agent is advisory only. It does not replace peer review, domain expertise, or professional judgment.

package/registry/families/agent-runtime-ml-research-ts/files/TOOLS.md

@@ -0,0 +1,32 @@
+# TOOLS — agent-runtime-ml-research-ts
+
+The bundle inherits all primitives from `agent-base:secure`: `secrets`, `workspace`, `webhook-in`, `webhook-out`, `schedule`, `identity`, `audit`. Plus generic agent tools: `Read`, `Write`, `Edit`, `Glob`, `Grep`, `Bash`, `WebFetch`.
+
+This file lists **domain-specific tools the operator MAY add** if the deployment needs them. Each entry is intent — not implementation. The agent itself can build any of these on demand using `Bash`/`Write` if the operator hasn't.
+
+## Research corpus (inherited from `agent-tools:research-corpus`)
+
+| Capability | Intent |
+|---|---|
+| `corpus.arxiv(query)` | arXiv paper search |
+| `corpus.semanticScholar(query)` | Semantic Scholar |
+| `corpus.openalex(query)` | OpenAlex |
+| `corpus.crossref(query)` | Crossref DOI metadata |
+
+## Domain tools the operator may want
+
+Stub list — replace with role-specific entries. Keep ≤10 domain tools per deployment; more is fragmentation. Each tool should be:
+- Single-purpose
+- JSON output
+- ≤100 LOC
+- Listed here when added
+
+| Tool | Intent | Notes |
+|---|---|---|
+| `literature-review` | Concrete tool implementing the `literature-review` capability | Operator implements when needed |
+| `experiment-design` | Concrete tool implementing the `experiment-design` capability | Operator implements when needed |
+| `result-interpretation` | Concrete tool implementing the `result-interpretation` capability | Operator implements when needed |
+
+## Build a new tool when you need it
+
+The agent's job includes building tools when the measurement doesn't exist. New tools should be small, JSON-output, single-purpose. Add an entry here via PR after shipping.

package/registry/families/agent-runtime-ml-research-ts/files/agent.json

@@ -0,0 +1,39 @@
+{
+  "name": "ml-research",
+  "description": "ML Research Assistant agent bundle — literature review, experiment design, and result interpretation. Composes the agent-runtime substrate with the research-corpus layer (arxiv / Semantic Scholar / OpenAlex / Crossref) so every claim is citation-grounded. Advisory only — never replaces the operator's own expertise or peer review.",
+  "version": "0.1.0",
+  "tags": [
+    "agent-runtime",
+    "ml-research",
+    "literature-review",
+    "experiment-design",
+    "research-assistant",
+    "cloudflare-worker"
+  ],
+  "prompt": {
+    "systemPromptFile": "AGENTS.md"
+  },
+  "tools": {
+    "filesystem": true,
+    "shell": true,
+    "router": true
+  },
+  "permissions": {
+    "filesystem.write": "ask",
+    "shell.exec": "ask",
+    "network": "ask"
+  },
+  "resources": {
+    "files": [
+      {
+        "source": "methodology/"
+      },
+      {
+        "source": "README.md"
+      },
+      {
+        "source": "TOOLS.md"
+      }
+    ]
+  }
+}

package/registry/families/agent-runtime-ml-research-ts/files/methodology/experiment-design.md

@@ -0,0 +1,141 @@
+# Experiment Design Template (ML)
+
+## Purpose
+Design a rigorous ML experiment that defends its conclusion. The
+output is a pre-registered plan that, executed faithfully, produces
+results worth submitting to a top venue.
+
+## When to use
+Trigger when the user is planning experiments, designing ablations,
+or critiquing an experiment plan. Always upstream of result
+generation; running first and designing second is how the
+reproducibility crisis happens.
+
+## Method
+
+1. **Hypothesis.** State what is being tested in falsifiable form:
+   "Method M will outperform baseline B on benchmark X by ≥Δ% with
+   p<0.05 over n seeds." If the hypothesis isn't testable, the
+   experiment isn't an experiment.
+2. **Baselines.** ≥2 strong, current baselines. Include:
+   - The current SOTA (or close to it).
+   - A simple-but-strong baseline (often the surprise winner;
+     don't skip it).
+   - A direct ablation of the proposed method (the method
+     without its key novelty).
+   Re-implementing baselines is acceptable; re-using the original
+   authors' published numbers is acceptable *only* if the
+   evaluation protocol is identical.
+3. **Metrics.**
+   - **Primary**: the headline number; one metric. Pick before
+     looking at any data — picking after is fishing.
+   - **Secondary**: 2–4 metrics that capture different facets.
+   - **Calibration / fairness / robustness** as appropriate.
+   - For classification: accuracy + macro-F1 + AUROC; for
+     generation: BLEU / ROUGE / human eval; for RL: cumulative
+     reward + sample efficiency. Pick by community convention.
+4. **Data.**
+   - Dataset, version (datasets evolve — e.g., MMLU-Pro
+     vs MMLU), source, license.
+   - Train / val / test splits with seeds. Test set untouched
+     until final evaluation.
+   - Pre-processing pipeline reproducible end-to-end.
+   - Contamination check: confirm test data does not appear in
+     pretraining corpora (where applicable).
+5. **Implementation.**
+   - Architecture, parameter count, init scheme.
+   - Hyperparameters: full list, with the search space and the
+     selection criterion.
+   - Compute budget: GPU type, count, training hours, total
+     FLOPs estimate. Compute-blind comparisons mislead.
+   - Reproducibility: framework version, seeds (typically ≥3),
+     deterministic mode where feasible.
+6. **Statistical plan.**
+   - Number of seeds: ≥3 minimum, ≥5 for borderline claims.
+   - Significance test: paired t-test for matched-seed
+     comparisons; Welch's for unmatched; non-parametric
+     (Wilcoxon, Mann-Whitney) when distributions are skewed or
+     small-sample.
+   - Effect size: report Cohen's d or relative improvement;
+     p-value alone is misleading at large n.
+   - Multiple-comparison correction (Bonferroni / Holm) when
+     testing across many configurations.
+   - Confidence intervals on the headline number.
+7. **Ablations.**
+   - For each component of the proposed method, ablate it. If
+     ablating a component doesn't hurt performance, the
+     component isn't doing work.
+   - Document ablation results alongside the main result —
+     reviewers will demand them.
+8. **Failure modes and pre-mortem.**
+   - What would falsify the hypothesis? Document.
+   - Common failure modes: hyperparameter mis-tuning of
+     baselines (cherry-picking), test-set leakage, evaluation
+     drift, single-seed variance.
+   - Add monitoring (loss curves, gradient norms, validation
+     score) to catch training failures.
+9. **Pre-registration.** Lock the plan before running. OSF,
+   internal lab notebook, or arXiv-stamped methodology draft.
+   Deviations from the plan are documented and explained.
+
+## Reporting checklist
+
+Use the NeurIPS / ML reproducibility checklist:
+
+- [ ] Code, data, and instructions to reproduce.
+- [ ] Description of the compute infrastructure.
+- [ ] Hyperparameter search range and method.
+- [ ] Seeds reported.
+- [ ] Standard deviation / confidence intervals on results.
+- [ ] Significance test described.
+- [ ] License of every external asset (dataset, model, code).
+- [ ] Statement of computational and environmental cost.
+- [ ] Limitations section that's honest about scope.
+
+## Common experiment failures
+
+1. **HARKing.** Hypothesizing After Results are Known. Tighten
+   the pre-registered plan.
+2. **Test-set tuning.** Even one peek at the test set during
+   development corrupts it.
+3. **Single-seed claims.** ML training has high variance;
+   single-seed gains often vanish under re-running.
+4. **Weak baselines.** "Outperforms LSTM" in 2026 is not a
+   meaningful comparison; pick contemporary baselines.
+5. **Compute mismatch.** Comparing a method trained with 10x
+   compute to a baseline at 1x; isolate the contribution of
+   the method itself.
+6. **Cherry-picked benchmarks.** Reporting only the benchmarks
+   where the method wins; report the full set.
+7. **No code release.** A method without code is a claim, not a
+   result.
+
+## Output
+
+```
+:::artifact
+template: experiment-design
+hypothesis: "..."
+baselines: [...]
+metrics: { primary: ..., secondary: [...] }
+data: { dataset: ..., version: ..., splits: {...}, license: ... }
+implementation:
+  arch: "..."
+  params: ...
+  hparams: { ... }
+  search: { ... }
+  compute: { gpu: ..., count: ..., hours: ..., flops: ... }
+seeds: [42, 1337, 2718, 314]
+statistical-plan: { test: "paired-t", alpha: 0.05, correction: "holm" }
+ablations: [...]
+pre-registration: { url: "..." }
+limitations: [...]
+:::
+```
+
+## Refusal
+
+The agent will not:
+- Approve an experiment without baselines or significance plan.
+- Approve cherry-picked benchmark sets without a stated rationale.
+- Sign off on a result claim from a single seed.

package/registry/families/agent-runtime-ml-research-ts/files/methodology/index.json

@@ -0,0 +1,25 @@
+{
+  "entries": [
+    {
+      "id": "literature-review",
+      "capability": "literature-review",
+      "path": "./literature-review.md",
+      "tags": ["methodology", "survey", "related-work"],
+      "version": "0.1.0"
+    },
+    {
+      "id": "experiment-design",
+      "capability": "experiment-design",
+      "path": "./experiment-design.md",
+      "tags": ["methodology", "experiments", "baselines"],
+      "version": "0.1.0"
+    },
+    {
+      "id": "result-interpretation",
+      "capability": "result-interpretation",
+      "path": "./result-interpretation.md",
+      "tags": ["methodology", "analysis", "statistics"],
+      "version": "0.1.0"
+    }
+  ]
+}