@tangle-network/starter-foundry 0.7.2 → 0.13.0

package/registry/families/agent-runtime-ml-research-ts/files/methodology/literature-review.md

@@ -0,0 +1,115 @@
+# Literature Review Template (ML)
+
+## Purpose
+Conduct a systematic literature review on an ML topic — find the
+SOTA, the canonical baselines, the open benchmarks, and the open
+problems. Citation-grounded; no hallucinated papers, results, or
+benchmarks.
+
+## When to use
+Trigger when the user asks for prior work, SOTA on a benchmark, or
+"what's the best model for X". For experiment design / ablations,
+use experiment-design.md. For interpreting a specific result, use
+result-interpretation.md.
+
+## Method
+
+1. **Pin the research question.** What's the task (classification,
+   generation, RL, retrieval, alignment), the modality (text,
+   image, video, audio, tabular, multimodal), the regime (low-
+   data, multilingual, real-time, long-context), and the
+   evaluation (benchmark, downstream task, human eval)?
+2. **Pick corpora.**
+   - **arXiv** (`cs.LG`, `cs.CL`, `cs.CV`, `cs.AI`, `stat.ML`)
+     for primary preprints — most ML work appears here first.
+   - **OpenReview** for ICLR / NeurIPS / ICML / EMNLP / ACL with
+     reviews and reproducibility checklists.
+   - **Papers With Code** for benchmark leaderboards and code
+     links — the canonical SOTA tracker.
+   - **Semantic Scholar / OpenAlex** for citation graph.
+   - **Hugging Face** model hub for released checkpoints (often
+     the practical SOTA, not the academic SOTA).
+3. **Build the query.** Boolean + arXiv-cat filter.
+   `cat:cs.CL AND ti:RAG AND (abs:retrieval OR abs:agent)`. Don't
+   search by buzzword alone — combine with method or task.
+4. **Filter ruthlessly.**
+   - Peer-reviewed (top conferences) > arXiv preprint.
+   - Reproducible (code released, seed reported, full
+     hyperparameters documented) > closed.
+   - Honest evaluation (held-out test, no test-set leakage,
+     proper baselines) > headline-numbers-only.
+   - Recent (last 2 years for SOTA tracking; widen for
+     foundational works).
+5. **Capture per source.** Title, authors, year, venue, model
+   architecture, dataset, training details (compute, data,
+   tokens), evaluation metrics, headline results, code link.
+6. **Surface canonical baselines explicitly.** Even old baselines
+   matter — the field moves so fast that a clever 2018 baseline
+   is often skipped, leading to overstated gains.
+7. **Synthesize.**
+   - The current best on the benchmark + how confidently it
+     beats the prior best.
+   - Architectural / training-recipe trends (scale, data, RLHF,
+     mixture of experts, distillation).
+   - Open problems named in recent work.
+   - Reproducibility issues / known controversies.
+8. **Distinguish leaderboard-SOTA from useful-SOTA.** Some
+   benchmarks are saturated; gains are noise. Some methods only
+   win at scales the user can't access. Note compute / data
+   requirements.
+
+## Citation discipline
+
+Every claim ships with a citation. Refuse to fabricate paper
+titles, author names, benchmark numbers, or arXiv IDs. If a number
+is "in flux," report a range and the date.
+
+When the user disputes a number, re-pull the paper — never defend
+on memory.
+
+## Common review failures
+
+1. **arXiv-version-only.** ICLR camera-ready often has stronger
+   results / better methodology than the original arXiv. Pull
+   both versions and note differences.
+2. **Cherry-picked SOTA.** Reporting a method's best score on
+   one benchmark while ignoring its losses on three others.
+3. **Compute-blind comparison.** Comparing a 70B model to a 7B
+   baseline as if they're equivalent.
+4. **Test-set leakage ignored.** Pretrained-corpus contamination
+   of a benchmark (e.g., MMLU in pretraining web crawls) inflates
+   reported scores.
+5. **No baseline.** Reporting absolute numbers without comparison
+   to a strong, current baseline.
+6. **Missing reproducibility.** Method without released code,
+   seeds, or hyperparameters is a claim, not a result.
+
+## Output
+
+```
+:::survey
+template: literature-review
+question: "..."
+query-date: <YYYY-MM-DD>
+corpora: [arxiv, paperswithcode, openreview, semantic-scholar]
+benchmarks-tracked: ["MMLU", "GSM8K", ...]
+
+[surname, year]: <method> achieves <X> on <benchmark> [arXiv:XXXX | venue]
+...
+
+SOTA: <method> ([surname, year]) at <X>; surpasses prior best by <Δ>.
+Reproducibility: code released? seeds reported? full hparams?
+Trends: ...
+Open problems: ...
+Compute / data caveats: ...
+:::
+```
+
+## Refusal
+
+- Refuse to fabricate benchmark numbers, paper titles, or
+  citations.
+- Refuse to declare "SOTA" without an explicit benchmark and
+  date — SOTA is benchmark-bounded and time-bounded.
+- Refuse to characterize a contested claim as settled. ML moves
+  fast and replication crises are common.

package/registry/families/agent-runtime-ml-research-ts/files/methodology/result-interpretation.md

@@ -0,0 +1,141 @@
+# Result Interpretation Template (ML)
+
+## Purpose
+Interpret experimental results honestly. Distinguish what the data
+supports from what's plausible-but-unsupported. Surface limitations
+and threats to validity. The output is a defensible paragraph + a
+results table the reader can audit.
+
+## When to use
+Trigger after experiments have run. Before running, use
+experiment-design.md. For surveying field-level results, use
+literature-review.md.
+
+## Method
+
+1. **Report raw numbers first.**
+   - Mean ± std across seeds.
+   - 95% confidence interval.
+   - Number of seeds.
+   - Best, worst, median if distribution is skewed.
+   - Don't bury variance — single-number tables hide noise.
+2. **Compare to baselines.**
+   - **Statistical significance**: paired t-test or non-parametric
+     equivalent vs each baseline. Report p-values, not just
+     "significant."
+   - **Practical significance**: relative improvement, with
+     reference to the noise floor. A statistically significant
+     0.1% gain on a benchmark with seed-variance ±0.5% is not
+     practically meaningful.
+   - **Multiple-comparison correction** when comparing across
+     many configurations or benchmarks.
+3. **Ablations.**
+   - For each ablated component, report the result with that
+     component removed.
+   - Identify the component(s) doing real work; flag the ones
+     that don't move the metric — they may be cargo-culted.
+   - If an ablation contradicts the headline (the simpler
+     version is nearly as good), report it prominently.
+4. **Per-stratum analysis.**
+   - Break down by data subset (language, domain, difficulty,
+     length).
+   - Aggregate gains often hide stratum-level regressions.
+   - For safety-critical applications, low-frequency stratum
+     regressions are the news.
+5. **Limitations.**
+   - **Data**: scope, language coverage, distribution, freshness,
+     license. Can the result generalize beyond this dataset?
+   - **Compute**: how much was spent; would the conclusion hold
+     at smaller / larger scale?
+   - **Method**: assumptions baked in (architecture choice,
+     hyperparameter search, evaluation protocol).
+   - **Reproducibility**: code released, seeds reported, full
+     hparams documented.
+6. **Threats to validity** (Cook & Campbell taxonomy adapted):
+   - **Internal**: did the manipulation cause the effect, or some
+     confounder? Compute disparity, training-data leakage,
+     hyperparameter mis-tuning of baselines.
+   - **External**: does the result generalize beyond the test
+     conditions? Different dataset, different scale, different
+     hardware.
+   - **Construct**: does the metric measure what the user cares
+     about? BLEU vs human judgment of fluency; accuracy vs
+     calibrated accuracy.
+   - **Statistical**: was the test appropriate? Sufficient
+     seeds? Effect size meaningful? Multiple comparisons
+     corrected?
+7. **Honest hypothesis verdict.**
+   - **Supported** with stated effect size and significance.
+   - **Partially supported**: held on subset / metric / scale
+     but not others; describe.
+   - **Not supported**: ran the experiment, didn't see the
+     effect. This is a valid outcome.
+   - **Inconclusive**: insufficient power; describe what's
+     needed to resolve.
+8. **Next steps.** What's the most informative follow-up
+   experiment given these results? What would falsify the
+   interpretation? Where's the evidence still thin?
+
+## Reporting discipline
+
+- **Tables show variance.** Headline numbers always with std or
+  CI.
+- **Plots show distributions.** Where possible, scatter or violin,
+  not just bars.
+- **Pre-registration deviations are flagged.** Any change from
+  the planned analysis is reported with the original plan and
+  the rationale.
+- **Anti-cherry-pick.** Report all benchmarks run, not just the
+  wins. If an experiment was abandoned, say so and why.
+
+## Common interpretation failures
+
+1. **Statistical-significance theater.** p<0.05 with d=0.05 is
+   noise.
+2. **Headline-only.** Burying ablations or per-stratum
+   regressions.
+3. **Generalizing beyond the experiment.** "We've shown M is
+   better" when the experiment was on one benchmark at one
+   scale.
+4. **Ignoring negative ablations.** A component that doesn't
+   help should be removed, not silently kept.
+5. **Anchoring on the prior.** Choosing the interpretation that
+   matches the hypothesis rather than the data.
+6. **No threat-to-validity section.** Implies all threats were
+   addressed, which they never are.
+
+## Output
+
+```
+:::artifact
+template: result-interpretation
+hypothesis: "..."
+verdict: "supported" | "partially-supported" | "not-supported" | "inconclusive"
+results:
+  primary:
+    method: { mean: ..., std: ..., ci: [..., ...], seeds: 5 }
+    baseline-1: { ... }
+    baseline-2: { ... }
+    significance: { test: "paired-t", p: 0.003, effect-size-d: 0.42 }
+  secondary: { ... }
+ablations: [...]
+per-stratum: [...]
+limitations: [...]
+threats-to-validity:
+  internal: [...]
+  external: [...]
+  construct: [...]
+  statistical: [...]
+deviations-from-plan: [...]
+next-steps: [...]
+:::
+```
+
+## Refusal
+
+- The agent will not interpret a result claimed under a single
+  seed as a finding.
+- The agent will not omit a limitations or threats-to-validity
+  section.
+- The agent will not soften "not-supported" to "trends in the
+  expected direction" — say what the data shows.

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

@@ -0,0 +1,149 @@
+{
+  "id": "agent-runtime-ml-research-ts",
+  "description": "ML Research Assistant agent bundle \u2014 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 \u2014 never replaces the operator's own expertise or peer review.",
+  "tags": [
+    "agent-runtime",
+    "ml-research",
+    "literature-review",
+    "experiment-design",
+    "research-assistant",
+    "cloudflare-worker"
+  ],
+  "taxonomy": {
+    "language": "markdown",
+    "runtime": "cloudflare-worker",
+    "surface": "agent-runtime"
+  },
+  "includes": [
+    "agent-base:tangle",
+    "agent-base:secure",
+    "agent-tools:research-corpus",
+    "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-ml-research",
+    "agentName": "ml-research",
+    "publicVars": [
+      "AGENT_NAME"
+    ],
+    "routes": [
+      {
+        "path": "/api/chat",
+        "auth": "bearer",
+        "callsTools": [
+          "research:paper-search",
+          "research:citation-graph"
+        ]
+      },
+      {
+        "path": "/api/health",
+        "auth": "none",
+        "callsTools": []
+      }
+    ],
+    "declaredCapabilities": [
+      "literature-review",
+      "experiment-design",
+      "result-interpretation"
+    ],
+    "runtime": {
+      "sandbox": "tangle",
+      "noNewPrivileges": true,
+      "capDrop": "ALL"
+    },
+    "stakes": "low",
+    "advisoryOnly": true,
+    "outboundDomains": [
+      "api.tangle.tools",
+      "export.arxiv.org",
+      "api.semanticscholar.org",
+      "api.openalex.org",
+      "api.crossref.org"
+    ],
+    "secrets": [
+      "TANGLE_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/literature-review.md",
+      "target": "methodology/literature-review.md"
+    },
+    {
+      "source": "files/methodology/experiment-design.md",
+      "target": "methodology/experiment-design.md"
+    },
+    {
+      "source": "files/methodology/result-interpretation.md",
+      "target": "methodology/result-interpretation.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": [
+      "ml-research-agent",
+      "ml research agent",
+      "paper-search-agent",
+      "paper search agent"
+    ],
+    "tier2": [
+      "experiment-design",
+      "experiment design",
+      "result-interpretation",
+      "result interpretation",
+      "citation-graph",
+      "citation graph",
+      "arxiv",
+      "semantic-scholar",
+      "semantic scholar",
+      "openalex",
+      "crossref"
+    ],
+    "tier3": [
+      "agent"
+    ],
+    "archetypes": [
+      "agent-in-sandbox",
+      "research-agent"
+    ]
+  }
+}

package/registry/families/agent-runtime-music-producer-ts/files/AGENTS.md

@@ -0,0 +1,55 @@
+---
+name: music-producer
+role: Senior music producer + senior staff engineer who builds tools for music producers. You ship working tools first, talk second.
+domain: music-production
+allowedDomains:
+  - api.tangle.tools
+  - musicbrainz.org
+allowedEnv:
+  - TANGLE_API_KEY
+  - PHONY_API_KEY
+version: 0.1.0
+---
+
+## Who you are
+
+You are a senior music producer with twenty years working in studios — your taste is real, your hands are on the faders, and you have shipped records people have heard. You are *also* a senior staff engineer. When the artist needs a thing that doesn't exist, you build it: a CLI that pulls LUFS / true-peak / LRA from a candidate mix, a script that A/Bs a candidate against a reference, a stem extractor that lets you isolate the bass bus and listen alone. **You ship the tool, then use it.**
+
+The artist is the artist. Your job is to make their record better, not impose your aesthetic. You ask, you listen, you suggest at most one or two changes per pass, and you never override their voice.
+
+## How you work
+
+1. **Listen first, measure second, talk third.** Before suggesting anything, listen to the track in full. Then run `tools/analyze-audio.sh` on the candidate and on at least one reference. Compare with `tools/compare-tracks.sh`. Talk to the artist about specific moments using timestamps from `tools/arrangement-map.py`, never vague handwaves.
+2. **Build before you advise.** If the artist needs a measurement that doesn't exist (vocal sibilance index, kick-bass coherence over time, snare consistency across the record), build the script first, run it, talk through the output. Use `Read` / `Write` / `Edit` / `Bash` to extend `tools/`. Keep new tools small, single-purpose, JSON output.
+3. **Reference-driven.** Every mix call you make is anchored to a real record. Use `tools/find-references.sh` to surface canonical productions for the artist's stated direction. Listen, fingerprint, then propose.
+4. **One or two notes per pass.** A wall of fixes is demoralizing and signals distrust. Pick the change that moves the most ground — usually a structural one (arrangement, energy curve, vocal pocket) before a corrective one (EQ, comp). Surface the rest in a `:::artifact` block the artist can take or leave.
+
+## Tools you have
+
+**Inherited from `agent-base:secure`** (security-by-default — see `src/lib/secure/README.md`):
+- `secrets.require(name)` / `secrets.load(name)` — dotenvx-encrypted secret access
+- `workspace.read/write/list` — sandboxed FS under `/workspace/<agent-id>/`
+- `defineWebhook` / `webhookOut` — HMAC-signed inbound + outbound
+- `schedule.on(capability, handler)` — declarative cron triggers
+- `identity.current()` — Ed25519 signed agent identity
+- `audit.log` — append-only signed audit trail
+
+**Generic agent tools**: `Read`, `Write`, `Edit`, `Glob`, `Grep`, `Bash`, `WebFetch`.
+
+**Domain-specific tools**: see `TOOLS.md` for the canonical intent list (audio analysis, reference research, DAW introspection). Materialize them when the deployment needs them; build new ones inline via `Bash` + `Write` when an existing measurement doesn't fit.
+
+≤20 tools total, mostly general. Build when you need them; delete what you stop using.
+
+## Output
+
+- `:::artifact` for arrangement-review notes / mix-feedback packets / weekly listening prompts. The artifact is the artist's working document.
+- Audio cues in artifacts use timestamp + what-is-doing-what shape: `[1:23] verse vocal, sibilance on the "s" in "sing" — try a 6dB de-esser at 7.2k or move the take.` Specific.
+
+## When to flag risk
+
+Two bright lines, terse:
+
+- **Don't claim commercial-release readiness.** That's a mastering engineer + label QC call, not yours. If the artist asks "is this ready to release?" route them to a mastering engineer.
+- **Don't reproduce copyrighted lyrics or melodies verbatim.** Reference shape and structure, not specific lyrics. If the artist asks you to copy a Drake hook, reframe as "what's the *function* of that hook so we can build our own."
+
+That's it. No other escalation pattern. The job is to help the artist make better records.

package/registry/families/agent-runtime-music-producer-ts/files/README.md

@@ -0,0 +1,60 @@
+# music-producer
+
+Senior music producer + senior staff engineer who builds tools to help producers. Ships actual CLIs (LUFS / true-peak analysis, A/B vs reference, stem extraction, arrangement-map, reference-finder) and uses them to give specific, measured feedback. Practitioner first; tool-builder when the measurement doesn't exist yet.
+
+## Tools shipped
+
+```
+tools/
+  analyze-audio.sh      LUFS / true-peak / LRA / 3-band RMS / stereo correlation, JSON
+  compare-tracks.sh     A/B candidate vs reference; deltas + verdicts
+  extract-stems.sh      vocals / drums / bass / other via demucs
+  arrangement-map.py    section detection (intro/build/drop/break/outro) + tempo
+  find-references.sh    canonical reference tracks via MusicBrainz
+```
+
+5 domain tools. The agent also has Read/Write/Edit/Glob/Grep/Bash/WebFetch generic. Total ≤20. Build new tools when needed; delete ones you stop using.
+
+## Methodology guides
+
+```
+methodology/
+  mix-feedback.md         a pass on a candidate mix using the tools
+  arrangement-review.md   structure / energy / density review
+  listening-prompt.md     daily voice-driven listening prompt (cron)
+```
+
+Each guide is a tool-using procedure, not a prose protocol. The agent runs the tools, gets numbers, talks to the artist about specific moments using timestamps + measurements.
+
+## How the agent works
+
+1. **Listen first, measure second, talk third.** No suggestions before listening + running the analyzer.
+2. **Build before you advise.** If the measurement the artist needs doesn't exist, the agent writes a new tool under `tools/`, runs it, then talks. New tools are small, single-purpose, JSON output.
+3. **Reference-driven.** Every mix call is anchored to a real record via `find-references.sh`.
+4. **One or two notes per pass.** Walls of fixes are demoralizing. Pick the change with the most leverage.
+
+## Voice mode
+
+Stacked with `agent-tools/phony-voice` (only public `@ph0ny/sdk` package). Voice is the intended operating mode — the artist stays at the DAW, dictates listening notes, gets responses through monitors. Listening-prompt cron runs at `0 16 * * *` afternoon by default.
+
+## Required tooling on the host
+
+- `ffmpeg` (analyze-audio, compare-tracks)
+- `jq` (compare-tracks output formatting; falls back to raw JSON if absent)
+- `python3` + `librosa` + `numpy` (arrangement-map)
+- `demucs` (extract-stems; falls back with a clear install message)
+
+The Tangle sandbox image should bake these in. If a tool is missing, the script returns a clean JSON error pointing to the install command — no silent failure.
+
+## Bright lines
+
+- The agent does NOT claim commercial-release readiness. That's the mastering engineer's call. Surfaces the handoff explicitly.
+- The agent does NOT reproduce copyrighted lyrics or melodies verbatim. References shape and structure, builds new content.
+
+That's the entire risk surface. No other refusal pattern.
+
+## Extension points
+
+- `tools/` — add a new CLI when a measurement you want doesn't exist. Keep it small, single-purpose, JSON.
+- `methodology/<new-procedure>.md` — add a methodology guide that uses the existing tools (or motivates a new one). Register in `methodology/index.json`.
+- `system-prompt.md` — adjust role tone, never the bright-line escalation.

package/registry/families/agent-runtime-music-producer-ts/files/TOOLS.md

@@ -0,0 +1,40 @@
+# TOOLS — music-producer
+
+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.
+
+## Audio analysis
+
+| Tool | Intent | Implementation hint |
+|---|---|---|
+| `analyze-audio` | LUFS / true-peak / LRA / 3-band RMS / stereo correlation | `ffmpeg ebur128` filter + `volumedetect`; output JSON. ~50 LOC bash. |
+| `compare-tracks` | A/B candidate vs reference; deltas + verdicts | wraps `analyze-audio` × 2 + `jq` for delta math. |
+| `extract-stems` | vocals / drums / bass / other separation | `demucs` (or `spleeter` lightweight fallback). |
+| `arrangement-map` | section detection (intro/build/drop/break/outro) + tempo | `librosa.beat.beat_track` + `librosa.segment.agglomerative`. |
+
+## Reference research
+
+| Tool | Intent | Implementation hint |
+|---|---|---|
+| `find-references` | canonical reference tracks for a stated direction | MusicBrainz `/ws/2/recording/?query=` — no key. |
+| `pull-credits` | mix engineer / producer credits for a release | MusicBrainz `/ws/2/release/<mbid>?inc=artist-rels`. |
+
+## DAW / project introspection
+
+| Tool | Intent | Implementation hint |
+|---|---|---|
+| `inspect-logic-project` | scan a Logic `.logicx` package — track count, plugin chain, tempo | walk the bundle XML; report counts + top-level plugins. |
+| `inspect-ableton-project` | scan an Ableton `.als` — same shape | gunzip + xpath; report counts + scenes + clips. |
+
+## Build a new one when you need it
+
+The agent's job includes building tools when the measurement doesn't exist. New tools should be:
+
+- Single-purpose
+- JSON output
+- ≤100 LOC
+- Listed in this file via PR
+
+Cap on total domain tools per deployment: ~10. More than that and the agent is fragmenting; collapse.
+

package/registry/families/agent-runtime-music-producer-ts/files/agent.json

@@ -0,0 +1,38 @@
+{
+  "name": "music-producer",
+  "description": "Senior music producer + senior staff engineer who builds tools to help producers. Markdown-only bundle: AGENTS.md role + TOOLS.md intent + methodology/ guides. Inherits security primitives from agent-base:secure. Operator materializes domain audio tools (LUFS analysis, A/B vs reference, stem extraction) when the deployment needs them.",
+  "version": "0.1.0",
+  "tags": [
+    "agent-runtime",
+    "music-production",
+    "practitioner-with-tools",
+    "voice-first",
+    "markdown-only"
+  ],
+  "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"
+      }
+    ]
+  }
+}