fullsend-ai · waynesun09 · Jul 2, 2026 · Jul 1, 2026 · Jul 1, 2026 · Jul 1, 2026
@@ -0,0 +1,44 @@
+name: Build Doc Site
+
+on:
+  pull_request:
+  push:
+    branches: [main]
+
+permissions:
+  contents: read
+
+concurrency:
+  group: site-build-${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  build:
+    runs-on: ubuntu-24.04
+    steps:
+      - name: Checkout experiments
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          path: experiments
+
+      - name: Checkout fullsend (doc site source)
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+        with:
+          repository: fullsend-ai/fullsend
+          path: fullsend
+
+      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
+        with:
+          node-version: "22"
+          cache: npm
+          cache-dependency-path: fullsend/website/package-lock.json
+
+      - name: Install and build doc site
+        working-directory: fullsend/website
+        run: |
+          npm ci --ignore-scripts
+          rm -rf ../experiments
+          ln -s "$GITHUB_WORKSPACE/experiments" ../experiments
+          ln -sfn "$GITHUB_WORKSPACE/experiments" ../docs/experiments
+          npx vitepress build
@@ -31,6 +31,28 @@ repos:
         types: [python]
         pass_filenames: false
 
+      - id: lint-experiment-numbers
+        name: lint experiment numbers
+        entry: ./hack/lint-experiment-numbers
+        language: script
+        files: ^[0-9]{4}-
+        pass_filenames: false
+
+      - id: lint-experiment-frontmatter
+        name: lint experiment frontmatter
+        entry: ./hack/lint-experiment-frontmatter
+        language: python
+        additional_dependencies: ["pyyaml>=6.0"]
+        files: ^[0-9]{4}-
+        pass_filenames: false
+
+      - id: lint-experiment-index
+        name: lint experiment index
+        entry: ./hack/lint-experiment-index
+        language: script
+        files: (^[0-9]{4}-|^README\.md$)
+        pass_filenames: false
+
   - repo: https://github.com/PyCQA/bandit
     rev: "1.9.4"
     hooks:

@@ -0,0 +1,26 @@
+---
+title: "NNNN. TITLE"
+status: Active
+topics:
+  - topic-tag
+---
+
+# NNNN. TITLE
+
+Date: YYYY-MM-DD
+
+## Hypothesis
+
+What are you trying to learn or prove?
+
+## Approach
+
+How will you test the hypothesis?
+
+## Results
+
+What happened? (fill in as the experiment progresses)
+
+## Conclusion
+
+What did you learn? Should this be adopted, modified, or abandoned?
@@ -1,3 +1,11 @@
+---
+title: "1. Agent outage fire drill"
+status: Active
+topics:
+  - reliability
+  - incident-response
+---
+
 # Experiment 003: Agent Outage Fire Drill
 
 **Date:** 2026-03-16
@@ -91,5 +99,5 @@ This experiment requires an agent autonomy period to have elapsed first. It can'
 
 ## Relationship to other experiments
 
-- **[Experiment 002](adr46-claude-scanner/README.md)** tests automated enforcement of architectural invariants. This experiment tests whether the humans responsible for architectural judgment can still exercise it without agent support.
+- **[Experiment 002](0002-claude-scanner/README.md)** tests automated enforcement of architectural invariants. This experiment tests whether the humans responsible for architectural judgment can still exercise it without agent support.
 - A complementary experiment could measure review quality specifically (can humans still evaluate agent-produced code?) — this experiment is broader, testing whether humans can still write, review, debug, and ship.
@@ -1,3 +1,11 @@
+---
+title: "2. Claude-based ADR drift scanner"
+status: Concluded
+topics:
+  - evaluation
+  - tooling
+---
+
 # Experiment 002: Claude-based ADR Drift Scanner
 
 A shell script that uses `claude` CLI to analyze Tekton tasks for ADR compliance. No hardcoded rules, no config files — all intelligence comes from claude's comprehension of the ADR text.

@@ -16,9 +16,9 @@ An LLM can read an ADR, understand its architectural intent, evaluate a code art
 
 ## Method
 
-1. Wrote a hand-crafted [expected analysis](../../experiments/adr46-claude-scanner/expected/modelcar-oci-ta.md) as a human benchmark — our own reading of what violates the ADR and what to do about each case
+1. Wrote a hand-crafted [expected analysis](expected/modelcar-oci-ta.md) as a human benchmark — our own reading of what violates the ADR and what to do about each case
 2. Built a shell script that combines the ADR text and task YAML into a prompt and passes it to `claude -p`
-3. Ran the scanner and captured [claude's output](../../experiments/adr46-claude-scanner/results/modelcar-oci-ta.md)
+3. Ran the scanner and captured [claude's output](results/modelcar-oci-ta.md)
 4. Evaluated against the expected analysis using a four-point rubric
 
 ## Results

@@ -1,3 +1,11 @@
+---
+title: "3. ADR-0046 drift scanner"
+status: Concluded
+topics:
+  - evaluation
+  - tooling
+---
+
 # Experiment 001: ADR-0046 Drift Scanner
 
 A Python CLI tool that detects Tekton task steps drifting from [ADR-0046](https://github.com/konflux-ci/architecture/blob/main/ADR/0046-common-task-runner-image.md) (common task runner image) by comparing step images against a config-driven allowlist.
@@ -18,4 +26,4 @@ Parses Tekton task YAML files, extracts step images, and flags any that don't ma
 
 ## Limitations
 
-This scanner uses deterministic string matching. It validates that mechanical drift detection works, but doesn't generalize to ADRs that express design intent rather than concrete image requirements. See [Experiment 002](../adr46-claude-scanner/) for the LLM-based approach.
+This scanner uses deterministic string matching. It validates that mechanical drift detection works, but doesn't generalize to ADRs that express design intent rather than concrete image requirements. See [Experiment 002](../0002-claude-scanner/) for the LLM-based approach.
@@ -1,3 +1,12 @@
+---
+title: "4. Zero-config autonomous bug fix engine with self-improving meta-loop"
+status: Concluded
+topics:
+  - tooling
+  - github-actions
+  - evaluation
+---
+
 # Experiment 006: Zero-Config Autonomous Bug Fix Engine with Self-Improving Meta-Loop
 
 **Date:** 2026-03-29
@@ -128,7 +137,7 @@ The tradeoff is real: without repo-specific context (no AGENTS.md, no CLAUDE.md,
 
 ### Production execution as evaluation
 
-The meta-loop offers an alternative to the golden-set evaluation approach explored in [Experiment 004](../promptfoo-eval/README.md). Instead of curated test cases with expected outputs, the signal comes from running the agent against real tasks and feeding the results back. The two approaches are complementary:
+The meta-loop offers an alternative to the golden-set evaluation approach explored in [Experiment 004](../0016-promptfoo-eval/README.md). Instead of curated test cases with expected outputs, the signal comes from running the agent against real tasks and feeding the results back. The two approaches are complementary:
 
 - **Golden-set (promptfoo):** catches prompt regressions, verifies format compliance, runs in seconds. Tests prompts, not agents.
 - **Meta-loop:** catches integration-level bugs that only manifest in production, verifies end-to-end behavior, takes minutes to hours. Tests the full agent system.
@@ -141,6 +150,6 @@ The meta-loop improves the engine using an LLM, but who improves the meta-loop?
 
 ## Relationship to other experiments
 
-- **[Experiment 001](../adr46-scanner/README.md)** and **[Experiment 002](../adr46-claude-scanner/README.md)** test detecting architectural drift — a concern that the meta-loop's engine doesn't address (it fixes bugs, not architectural violations).
-- **[Experiment 003](../003-agent-outage-fire-drill.md)** tests whether humans maintain capability when agents are removed. The meta-loop experiment is the other side: what happens when you push agents further toward full autonomy, including self-repair?
-- **[Experiment 004](../promptfoo-eval/README.md)** tests prompt-level evaluation with promptfoo. The meta-loop provides a complementary evaluation approach at the integration level.
+- **[Experiment 001](../0003-scanner/README.md)** and **[Experiment 002](../0002-claude-scanner/README.md)** test detecting architectural drift — a concern that the meta-loop's engine doesn't address (it fixes bugs, not architectural violations).
+- **[Experiment 003](../0001-agent-outage-fire-drill.md)** tests whether humans maintain capability when agents are removed. The meta-loop experiment is the other side: what happens when you push agents further toward full autonomy, including self-repair?
+- **[Experiment 004](../0016-promptfoo-eval/README.md)** tests prompt-level evaluation with promptfoo. The meta-loop provides a complementary evaluation approach at the integration level.
@@ -1,8 +1,17 @@
+---
+title: "5. Agent scoped tools triage"
+status: Concluded
+topics:
+  - security
+  - sandbox
+  - skills
+---
+
 # PoC FullSend — Triage Agents with scoped tools and sandbox
 
 This experiment addresses: https://github.com/fullsend-ai/fullsend/issues/101
 
-Counterpart to [experiment 67](../67-claude-github-app-auth/) which demonstrates the wrapper/pure-I/O approach. This experiment demonstrates an alternative using established patterns from [Claude Code](https://code.claude.com/docs/en/skills) and [OpenCode](https://opencode.ai/docs/skills/): **skills define capabilities, agents execute them with scoped tools, and a top-level agent orchestrates the flow.**
+Counterpart to [experiment 67](../0022-claude-github-app-auth/) which demonstrates the wrapper/pure-I/O approach. This experiment demonstrates an alternative using established patterns from [Claude Code](https://code.claude.com/docs/en/skills) and [OpenCode](https://opencode.ai/docs/skills/): **skills define capabilities, agents execute them with scoped tools, and a top-level agent orchestrates the flow.**
 
 ### What this experiment covers
 

@@ -1,3 +1,11 @@
+---
+title: "6. Code agent evaluation"
+status: Concluded
+topics:
+  - evaluation
+  - security
+---
+
 # Code Agent Evaluation Experiment
 
 **Date:** April 11–14, 2026

@@ -1,3 +1,11 @@
+---
+title: "7. GitHub Actions agent runtime MVP"
+status: Concluded
+topics:
+  - github-actions
+  - tooling
+---
+
 # Experiment: GitHub Actions Agent Runtime MVP
 
 ## Hypothesis

@@ -1,6 +1,14 @@
+---
+title: "8. Guardrails evaluation"
+status: Concluded
+topics:
+  - security
+  - evaluation
+---
+
 # Experiment: Pluggable Guardrails Evaluation
 
-Evaluates prompt injection scanning tools as pre-LLM guardrails for fullsend's autonomous SDLC pipeline. Extends the [Model Armor experiment](../model-armor-vs-agent-triage/) (PR #117) with additional attack payloads and alternative scanning backends.
+Evaluates prompt injection scanning tools as pre-LLM guardrails for fullsend's autonomous SDLC pipeline. Extends the [Model Armor experiment](../0012-model-armor-vs-agent-triage/) (PR #117) with additional attack payloads and alternative scanning backends.
 
 Related: [Story 6 (#129)](https://github.com/fullsend-ai/fullsend/issues/129) — Prompt injection defense
 
@@ -10,7 +18,7 @@ LLM Guard's DeBERTa-v3 classifier with sentence-level matching detects prompt in
 
 ## Background
 
-PR #117 found that Model Armor detected only 1/4 injection payloads (25%) at maximum sensitivity. The [prompt-injection-defense](../prompt-injection-defense/) experiment found that Claude Sonnet 4.6 is inherently resilient, but the DeBERTa v1 classifier "adds no value" beyond what the model catches.
+PR #117 found that Model Armor detected only 1/4 injection payloads (25%) at maximum sensitivity. The [prompt-injection-defense](../0015-prompt-injection-defense/) experiment found that Claude Sonnet 4.6 is inherently resilient, but the DeBERTa v1 classifier "adds no value" beyond what the model catches.
 
 This experiment tests whether:
 1. LLM Guard's v2 classifier with sentence-level matching improves detection
@@ -19,7 +27,7 @@ This experiment tests whether:
 
 ## Attack Payloads
 
-### Original (from `../prompt-injection-defense/attacks/`)
+### Original (from `../0015-prompt-injection-defense/attacks/`)
 
 | Payload | Technique |
 |---------|-----------|

@@ -1,3 +1,11 @@
+---
+title: "9. Hermes-inspired security patterns"
+status: Concluded
+topics:
+  - security
+  - tooling
+---
+
 # Experiment: Hermes-Inspired Security Patterns for Fullsend
 
 Evaluates security patterns from [Hermes Agent](https://github.com/NousResearch/hermes-agent) for integration into fullsend's autonomous SDLC pipeline. Tests two integration strategies: static file scanning via [Tirith](https://github.com/sheeki03/tirith) CLI and runtime SSRF protection via a Claude Code PreToolUse hook.
@@ -230,7 +238,7 @@ uv run python -m pytest tests/ -v
 
 ### Tirith vs LLM Guard Coverage Cross-Reference
 
-Cross-referencing all 12 attack payloads from the [guardrails-eval experiment](../guardrails-eval/) against Tirith's detection capabilities:
+Cross-referencing all 12 attack payloads from the [guardrails-eval experiment](../0008-guardrails-eval/) against Tirith's detection capabilities:
 
 | Payload | Technique | LLM Guard (sentence) | Tirith | Notes |
 |---------|-----------|---------------------|--------|-------|

@@ -1,3 +1,11 @@
+---
+title: "10. Host-side API server for sandboxed agents"
+status: Concluded
+topics:
+  - sandbox
+  - tooling
+---
+
 # Experiment: Host-Side API Server for Sandboxed Agents
 
 Tracking issue: [fullsend-ai/experiments#25](https://github.com/fullsend-ai/experiments/issues/25)

@@ -6,7 +6,7 @@
 
 ## 1. Repository Cloning (Provisioner API)
 
-**Endpoint:** `POST /repo/provision`  
+**Endpoint:** `POST /repo/provision`
 **Status:** ✅ **Success**
 
 | Field | Value |
@@ -32,7 +32,7 @@ The repo was successfully cloned and uploaded to `/sandbox/fullsend/repo/`.
 
 ## 2. Container Image Build (Builder API)
 
-**Endpoint:** `POST /build`  
+**Endpoint:** `POST /build`
 **Status:** ✅ **Success**
 
 | Field | Value |

@@ -4,8 +4,8 @@
 
 ## 1. Repository Clone — ✅ Success
 
-**Endpoint:** `POST /repo/provision`  
-**Repository:** `fullsend-ai/fullsend` (ref: `main`)  
+**Endpoint:** `POST /repo/provision`
+**Repository:** `fullsend-ai/fullsend` (ref: `main`)
 **Destination:** `/sandbox/fullsend`
 
 **Response:**
@@ -20,10 +20,10 @@ The repository was successfully cloned and uploaded to `/sandbox/fullsend/repo/`
 
 ## 2. Container Image Build — ✅ Success
 
-**Endpoint:** `POST /build`  
-**Containerfile:** `images/sandbox/Containerfile`  
-**Context directory:** `/sandbox/fullsend/repo`  
-**Tag:** `fullsend-sandbox:test`  
+**Endpoint:** `POST /build`
+**Containerfile:** `images/sandbox/Containerfile`
+**Context directory:** `/sandbox/fullsend/repo`
+**Tag:** `fullsend-sandbox:test`
 **Destination:** `/sandbox/fullsend-sandbox.tar`
 
 **Result:**

@@ -1,6 +1,6 @@
 # API Test Results
 
-**Date:** 2026-05-20  
+**Date:** 2026-05-20
 **Sandbox:** `agent-openapi-discovery-full-4038775-1779243123`
 
 ---

@@ -1,6 +1,6 @@
 # API Server Test Results
 
-**Date:** 2026-06-01  
+**Date:** 2026-06-01
 **Sandbox:** agent-openapi-discovery-restricted-3445554-1780274427
 
 ---

@@ -1,6 +1,6 @@
 # API Server Test Results
 
-**Date:** 2026-05-20  
+**Date:** 2026-05-20
 **Sandbox:** agent-openapi-discovery-restricted-4097617-1779243949
 
 ## Summary

@@ -4,7 +4,7 @@
 
 ## 1. Repository Cloning (Provisioner API)
 
-**Endpoint:** `POST /repo/provision`  
+**Endpoint:** `POST /repo/provision`
 **Status:** ✅ Success
 
 Cloned `fullsend-ai/fullsend` (ref: `main`) into `/sandbox/fullsend`.
@@ -32,7 +32,7 @@ These are informational findings from the automated security scan; the clone pro
 
 ## 2. Container Image Build (Builder API)
 
-**Endpoint:** `POST /build`  
+**Endpoint:** `POST /build`
 **Status:** ✅ Success
 
 Built image `fullsend-sandbox:test` from `images/sandbox/Containerfile` and uploaded tarball to `/sandbox/fullsend-sandbox.tar` (4.2 GB).