{"ok":true,"version":"0.1.0-draft","profile":{"readme":"# Structured Agent Exchange (SAE)\n\nAn open specification for sharing the shape of agent work across A2A and non-A2A systems.\n\n## About\n\nSAE defines a small, neutral task packet that different agent systems can exchange without needing to share the same runtime, framework, or internal implementation details.\n\nSAE is intended to complement existing interoperability layers such as A2A, not replace them.\n\n## Why SAE?\n\nAgent systems often agree on transport before they agree on the structure of the work itself.\n\nSAE focuses on the task shape:\n\n- objective\n- constraints\n- deliverable\n- delegation hints\n- review posture\n- optional preference and instruction hints\n\n## Public Position\n\n- `A2A` = discovery, transport, tasks, messages, and artifacts\n- `SAE` = shared task shape\n- internal systems may produce SAE directly or through richer local compilers\n\n`SAE` is intentionally:\n\n- small\n- neutral\n- transport-agnostic\n- usable over A2A, HTTP, MCP, CLI, or internal runtimes\n\n## What SAE Is For\n\nSAE exists to represent the shape of work in a way different agent systems can share without needing to adopt the same internal dialect.\n\nIt should be good at:\n\n- fuzzy ask normalization\n- explicit objective and constraints\n- deliverable shape\n- delegation hints\n- review posture\n- lightweight preference hints\n- optional instruction frame for role, goal, context, and style\n- translator hints for endpoint-aware dialect switching\n\nIt should **not** replace:\n\n- A2A discovery\n- Agent Cards\n- task transport\n- artifacts\n- streaming\n- auth\n\n## Relationship To Internal Systems\n\nOne intended stack is:\n\n1. internal ask or task request\n2. local compilation / interpretation\n3. SAE export\n4. transport over A2A or another adapter\n\nIn other words:\n\n- SAE is the public exchange layer\n- any richer internal shaping layer should remain optional and separate from the public core\n\n## Relationship To A2A\n\nRecommended posture:\n\n- `A2A` handles agent-to-agent discovery and transport\n- `SAE` handles the shape of the work\n\nThis means public interoperability should usually look like:\n\n- `SAE over A2A`\n\nnot:\n\n- `SAE instead of A2A`\n\n## Relationship To Translation\n\nSAE should not become a pile of vendor-specific prompt dialects.\n\nInstead:\n\n- SAE keeps the stable shared task shape\n- the translator renders that shape for the endpoint\n- the endpoint codex records the known prompting wins of each model family\n\nThis allows the same SAE task to be rendered for:\n\n- OpenAI Responses\n- Anthropic Messages\n- Gemini\n- Qwen\n- local Llama-style models\n- generic HTTP agents\n\n## Getting Started\n\n- Read `QUICKSTART.md`\n- Read `TUTORIAL.md`\n- Inspect `sae-task.schema.json`\n- Try `examples/hello-sae-http/`\n- Read `SAE-OVER-A2A.md`\n\n## Contents\n\n- `sae-task.schema.json`\n- `examples/`\n- `SAE-OVER-A2A.md`\n- `NON-A2A-TRANSLATION.md`\n- `MODEL-ENDPOINT-CODEX.md`\n- `examples/plain-http-client.md`\n- `WIKI.md`\n- `QUICKSTART.md`\n- `TUTORIAL.md`\n- `CONFORMANCE.md`\n- `IMPLEMENTATIONS.md`\n- `CHANGELOG.md`\n- `site/`\n- `LICENSE`\n- `NOTICE`\n- `CONTRIBUTING.md`\n- `GOVERNANCE.md`\n\n## Static Homepage\n\nThe `site/` folder is a standalone homepage and wiki scaffold for a future\ndedicated SAE repo or GitHub Pages deployment.\n","a2aProfile":"# SAE over A2A\n\nVersion: `0.1.0-draft`\n\n## Purpose\n\nThis document defines the recommended way to carry `SAE` through A2A interactions.\n\n## Boundary\n\nUse A2A for:\n\n- discovery\n- Agent Cards\n- tasks\n- messages\n- artifacts\n- streaming\n- push\n\nUse SAE for:\n\n- objective\n- constraints\n- deliverable shape\n- delegation hints\n- review posture\n- preference hints\n\n## Agent Card Signaling\n\nAn A2A agent that supports SAE should declare:\n\n- a skill or capability describing `SAE` support\n- optional extension declarations\n- a metadata flag indicating `supportsSaeOverA2A`\n\n## Draft Extension Identifiers\n\n- `https://structuredagentexchange.org/extensions/task/v0.1`\n- `https://structuredagentexchange.org/extensions/preferences/v0.1`\n\n## Usage Model\n\n1. Client discovers agent through A2A Agent Card\n2. Client sees `SAE` support\n3. Client sends normal A2A task/message plus SAE attachment\n4. Remote agent parses SAE fields\n5. Agent returns normal A2A response while preserving SAE-aware metadata\n\n## Value Claim\n\n`A2A lets agents connect and exchange work. SAE helps them share the same shape of the work.`\n","nonA2aTranslation":"# SAE Translation for Non-A2A Agents\n\n## Purpose\n\nNot every useful agent runtime will speak A2A.\n\nSAE should still be usable with:\n\n- plain HTTP services\n- MCP-connected hosts\n- CLI tools\n- orchestrators\n- chatbox-only agents\n- internal agent runtimes\n\n## Translation Model\n\n### Input Side\n\nConvert native asks into SAE:\n\n- extract objective\n- extract constraints\n- define deliverable\n- define review posture\n- optionally add preference hints\n\n### Output Side\n\nConvert SAE into the target system's native format:\n\n- HTTP request payload\n- MCP tool call arguments\n- CLI command contract\n- prompt template\n- internal job object\n\n## What This Solves\n\nThis is the layer A2A does not fully solve by itself:\n\n- semantic normalization for systems outside A2A\n- a common work-shape between otherwise incompatible agents\n\n## Translation Rule\n\nThe translator should preserve:\n\n- objective\n- constraints\n- deliverable shape\n- review posture\n- instruction frame intent\n\nIt may degrade:\n\n- richer continuity nuance\n- dialect-specific internal interpretation\n\nThat is acceptable. SAE should survive graceful simplification.\n\n## Endpoint-Aware Dialect Switching\n\nDifferent target systems do benefit from different prompting tactics.\n\nThat is not a reason to fracture SAE itself.\n\nThe correct model is:\n\n- `SAE task` stays stable\n- `translator` selects an endpoint profile\n- `endpoint codex` tells the translator what works best there\n\nExamples:\n\n- `Qwen` may benefit from a more explicit system prompt, short instructions, a few-shot example, and a firm output format.\n- `Anthropic` may benefit from more organized context blocks and explicit sectioning.\n- `local models` may benefit from repeated constraints and a simpler schema.\n\nThe translator may adapt:\n\n- prompt phrasing\n- ordering of role / goal / context / style\n- whether examples are included\n- how strongly the output contract is stated\n\nIt should not alter the underlying objective, constraints, deliverable, or review posture.\n","modelEndpointCodex":"# SAE Model Endpoint Codex\n\n## Purpose\n\nSAE should stay small and stable.\n\nThe translator is where endpoint-specific prompt shaping belongs.\n\nThis codex exists so an SAE task can preserve its work-shape while still adapting to the known strengths of different model families and message endpoints.\n\n## Core Rule\n\n- `SAE` defines the shared task shape.\n- the `translator` adapts that task to the endpoint.\n- the `endpoint codex` records the known prompting wins for that endpoint family.\n\nThis means dialect switching should happen in the translator layer, not by mutating SAE into vendor-specific protocol text.\n\n## Shared Instruction Frame\n\nThe default instruction frame SAE should preserve is:\n\n- `role`\n- `goal`\n- `context`\n- `style`\n\nOptional helpers:\n\n- `one_output_at_a_time`\n- `good_output_examples`\n\nThese give the translator enough structure to emit a better endpoint-native system prompt without changing the underlying task.\n\n## Model Family Big Wins\n\n### OpenAI Responses\n\n- explicit success criteria\n- clear structured outputs\n- direct role framing\n- explicit tool and execution boundaries\n\n### Anthropic Messages\n\n- organized long context\n- explicit role and goal\n- principled values and uncertainty handling\n- sectioned deliverables\n\n### Google Gemini\n\n- grounding blocks\n- explicit schema guidance\n- clear modality expectations\n- concise instructions\n\n### Qwen Chat\n\n- very explicit system prompt\n- short structured instructions\n- few-shot examples\n- clear output format\n\n### Local Llama / Similar Local Models\n\n- short concrete asks\n- repeated key constraints\n- simple stable schema\n- low ambiguity\n\n## Translation Rule\n\nThe translator should preserve:\n\n- objective\n- constraints\n- deliverable shape\n- review posture\n- instruction frame intent\n\nIt may adapt:\n\n- system prompt wording\n- section order\n- example density\n- output contract wording\n- endpoint-specific guardrails\n\n## Why This Matters\n\nDifferent model families do have different prompt wins.\n\nThat does **not** mean the standard should fracture into many incompatible vendor dialects.\n\nThe right split is:\n\n- one stable SAE task\n- one translator\n- many endpoint-aware renderings\n","wiki":"# SAE Wiki Index\n\nVersion: `0.1.0-draft`\n\nThis is the reading path for `SAE`.\n\nIf you are new here, use this order:\n\n1. `README.md`\n2. `QUICKSTART.md`\n3. `TUTORIAL.md`\n4. `SAE-OVER-A2A.md`\n5. `NON-A2A-TRANSLATION.md`\n6. `CONFORMANCE.md`\n7. `IMPLEMENTATIONS.md`\n8. `poc/README.md`\n\n## What SAE Is\n\n`SAE` means `Structured Agent Exchange`.\n\nIt is a small neutral work-shape for agent tasks:\n\n- objective\n- constraints\n- deliverable\n- delegation hints\n- review posture\n- optional preference hints\n\n## What SAE Is Not\n\nSAE is not:\n\n- a transport protocol\n- a replacement for `A2A`\n- a full agent runtime\n- a requirement to adopt any single internal dialect\n\n## Reading Paths\n\n### Product / strategy readers\n\n- `README.md`\n- `SAE-OVER-A2A.md`\n- `IMPLEMENTATIONS.md`\n- `GITHUB-HOMEPAGE.md`\n\n### Implementers\n\n- `README.md`\n- `QUICKSTART.md`\n- `TUTORIAL.md`\n- `sae-task.schema.json`\n- `CONFORMANCE.md`\n\n### External agent builders\n\n- `README.md`\n- `QUICKSTART.md`\n- `examples/plain-http-client.md`\n- `examples/hello-sae-http/README.md`\n- `examples/a2a-agent-card.example.json`\n- `poc/mock-external-http-agent/README.md`\n- `poc/mock-external-a2a/agent-card.json`\n\n## Current Claim\n\nThe safest public claim right now is:\n\n`SAE helps agents share the same shape of the work across A2A and non-A2A systems.`\n","quickstart":"# SAE Quickstart\n\nVersion: `0.1.0-draft`\n\nThis is the fastest way to understand and try `SAE`.\n\n## 1. Learn The Packet\n\nAn SAE task has a small set of stable fields:\n\n- `objective`\n- `constraints`\n- `deliverable`\n- `review_posture`\n- optional `delegation_hints`\n- optional `preference_hints`\n- optional `provenance`\n\n## 2. Read The Schema\n\nUse:\n\n- `sae-task.schema.json`\n\n## 3. Read One Example\n\nStart with:\n\n- `examples/research-brief.json`\n\nThen read:\n\n- `examples/plain-http-client.md`\n\n## 4. Validate A Task\n\nValidate the task against `sae-task.schema.json`.\n\nExample payload:\n\n```json\n{\n  \"payload\": {\n    \"sae_version\": \"0.1.0\",\n    \"id\": \"sae-quickstart-001\",\n    \"objective\": \"Review a proposed browser adapter and return a safe first implementation cut.\",\n    \"constraints\": [\n      \"keep the first cut reversible\"\n    ],\n    \"deliverable\": {\n      \"artifact_type\": \"review_packet\",\n      \"success_definition\": \"A downstream implementer can act on the packet without guessing the ask.\"\n    },\n    \"review_posture\": {\n      \"risk_level\": \"medium\",\n      \"human_review\": true\n    }\n  }\n}\n```\n\n## 5. Decide How You Want To Use It\n\n### If you speak A2A\n\nRead:\n\n- `SAE-OVER-A2A.md`\n\n### If you do not speak A2A\n\nRead:\n\n- `NON-A2A-TRANSLATION.md`\n\n### If you need endpoint-specific prompt shaping\n\nRead:\n\n- `MODEL-ENDPOINT-CODEX.md`\n\n## 6. Check Conformance\n\nBefore you call your implementation SAE-compatible, read:\n\n- `CONFORMANCE.md`\n\n## Optional Demo Surface\n\nIf you are working from the bundled app/demo environment rather than the\nstandalone SAE repo, you may also have local `/api/sae/*` routes available for\nschema and validation experiments.\n","tutorial":"# SAE Tutorial\n\nVersion: `0.1.0-draft`\n\nThis tutorial walks through one simple use case:\n\n`Take a fuzzy ask and turn it into a reviewable shared task packet.`\n\n## Step 1. Start With The Fuzzy Ask\n\nExample:\n\n`Review this new browser-hand adapter and tell me the safest first version to ship.`\n\nThis is useful, but still loose.\n\nQuestions are implied:\n\n- what counts as success?\n- how constrained is the first cut?\n- does it need review?\n- what shape should the answer come back in?\n\n## Step 2. Normalize Into SAE\n\n```json\n{\n  \"sae_version\": \"0.1.0\",\n  \"id\": \"sae-tutorial-001\",\n  \"title\": \"Browser-hand adapter review\",\n  \"objective\": \"Review this new browser-hand adapter and tell me the safest first version to ship.\",\n  \"constraints\": [\n    \"keep the first cut reversible\",\n    \"return operator-readable review notes\"\n  ],\n  \"deliverable\": {\n    \"artifact_type\": \"review_packet\",\n    \"success_definition\": \"A downstream implementer can act on the packet without guessing the ask.\",\n    \"format_hint\": \"markdown+json\"\n  },\n  \"delegation_hints\": {\n    \"roles\": [\"planner\", \"reviewer\", \"implementer\"],\n    \"steps\": [\n      \"normalize the ask\",\n      \"produce a safe first-cut recommendation\",\n      \"attach review notes\"\n    ],\n    \"clarification_budget\": \"low\"\n  },\n  \"review_posture\": {\n    \"risk_level\": \"medium\",\n    \"human_review\": true,\n    \"notes\": \"Do not widen scope into live deployment.\"\n  },\n  \"provenance\": {\n    \"origin_system\": \"tutorial\",\n    \"compiled_from\": \"fuzzy_ask\"\n  }\n}\n```\n\n## Step 3. Validate It\n\nUse:\n\n- `sae-task.schema.json`\n\nIf the packet passes validation, another system can inspect it with less guesswork.\n\n## Step 4. Carry It Somewhere\n\nYou now have options:\n\n### A2A system\n\nWrap it as metadata or an attachment alongside the normal A2A task/message flow.\n\n### Plain HTTP service\n\nForward the packet to an endpoint that expects reviewable task objects.\n\n### CLI or orchestrator\n\nTurn it into job arguments or a prompt contract while preserving:\n\n- objective\n- constraints\n- deliverable\n- review posture\n\n## Step 5. Evaluate The Result\n\nThe packet helped if:\n\n- the receiver asked fewer clarifying questions\n- the result came back in the expected artifact shape\n- the operator could review it quickly\n\n## Why This Matters\n\nThis is the core promise of SAE:\n\n`less renegotiation of what the work means`\n","changelog":"# SAE Changelog\n\n## 0.1.0-draft\n\nFirst public draft scaffold.\n\nIncludes:\n\n- base `sae-task.schema.json`\n- `SAE-OVER-A2A.md`\n- `NON-A2A-TRANSLATION.md`\n- `MODEL-ENDPOINT-CODEX.md`\n- example research brief\n- example A2A Agent Card\n- plain HTTP adapter example\n- `CONFORMANCE.md`\n- `IMPLEMENTATIONS.md`\n- wiki index\n- static homepage and wiki scaffold in `site/`\n- quickstart\n- tutorial\n- instruction frame fields for role, goal, context, and style\n- translator hints for endpoint families and dialect switching\n- public `LICENSE` and `NOTICE`\n- contribution and governance notes\n- draft OTLP trace export for agentevals\n\nCurrent status:\n\n- good enough for evaluation and outside discussion\n- not yet proven by multiple external implementations\n- still expected to evolve before a stable `0.1.0`\n","conformance":"# SAE Conformance\n\nVersion: `0.1.0-draft`\n\n## Required For A Minimal Conforming Implementation\n\nAn implementation may call itself `SAE-compatible` when it can:\n\n1. ingest a valid SAE task payload\n2. preserve:\n   - `objective`\n   - `constraints`\n   - `deliverable`\n   - `review_posture`\n3. emit or forward a result without losing required fields\n4. validate packets against the public schema\n\n## Stronger Conformance\n\nA stronger implementation should also:\n\n- preserve `delegation_hints`\n- preserve `preference_hints`\n- preserve `provenance`\n- declare SAE support explicitly in an Agent Card or service profile\n\n## Translation Rule\n\nNon-A2A translators may degrade richer nuance, but they should not silently erase:\n\n- the bounded objective\n- the requested deliverable shape\n- the review posture\n\n## What Does Not Count\n\nThese do not count as SAE support on their own:\n\n- vague prompt wrappers with no stable packet shape\n- product-specific internal payloads with no SAE export\n- claiming compatibility without schema validation\n","implementations":"# SAE Implementations\n\n## Current Known Implementation\n\n### Reference implementation\n\n- `/api/sae/schema`\n- `/api/sae/profile`\n- `/api/sae/examples`\n- `/api/sae/validate`\n- `/api/sae/translate`\n- `/api/sae/codex`\n- example A2A Agent Card\n- example plain HTTP adapter\n- benchmark harness\n- reproducible POC export script: `npm run sae:poc`\n\n## What Counts As A Real External Implementation\n\nAn external implementation should:\n\n- not rely on any product-specific UI\n- ingest or emit at least one conforming SAE task\n- validate the task against the public schema\n- survive one real round trip without losing required fields\n\n## Suggested First External Implementations\n\n1. plain HTTP service\n2. A2A-compatible agent declaring SAE support\n3. MCP/tool-host translation layer\n4. CLI or orchestrator adapter\n\n## Draft POC Surfaces\n\n- `specs/sae/poc/README.md`\n- `specs/sae/poc/mock-external-a2a/agent-card.json`\n- `specs/sae/poc/mock-external-http-agent/server.mjs`\n- `specs/sae/poc/results/`\n\nThese are valid publication surfaces for draft proof, but they should still be described honestly as:\n\n- reproducible benchmark-generated proof\n- mock external participant artifacts\n- not yet third-party validated adoption\n\n## Preferred Public Story\n\n`SAE` is the adoptable public layer.\n\nOther systems may generate SAE directly, or through their own richer internal shaping layers.\n","codexEntries":[{"key":"openai_responses","label":"OpenAI Responses","endpointFamily":"openai_responses","dialectProfile":"structured_success_criteria","description":"Works best with explicit success criteria, clean response shape, and clear tool boundaries.","bigWins":["Define the role and success criteria plainly.","Keep deliverables structured and directly reviewable.","State tool or execution boundaries explicitly.","Prefer one bounded output over sprawling exploration."],"cautionNotes":["Avoid hiding requirements in long lyrical prose.","Do not rely on implied output structure."],"systemPromptStyle":"direct, structured, bounded","outputContractDefault":"single_output"},{"key":"anthropic_messages","label":"Anthropic Messages","endpointFamily":"anthropic_messages","dialectProfile":"organized_contextual","description":"Responds well to organized context, explicit values, and clear sectioned deliverables.","bigWins":["Provide a stable role plus a clear goal.","Organize context into a few coherent blocks.","Name the desired sections of the answer.","Keep the tone principled and explicit about uncertainty."],"cautionNotes":["Avoid chaotic prompt packing.","Do not force an overly compressed format when nuance matters."],"systemPromptStyle":"organized, contextual, principled","outputContractDefault":"single_output"},{"key":"google_gemini","label":"Google Gemini","endpointFamily":"google_gemini","dialectProfile":"grounded_schema_first","description":"Benefits from grounded context blocks, explicit schema guidance, and clear multimodal or source framing.","bigWins":["Separate task goal from grounding context.","State the desired schema or output format clearly.","Call out relevant source or modality expectations.","Keep instructions short and explicit."],"cautionNotes":["Avoid ambiguous output expectations.","Do not mix many competing formats in one ask."],"systemPromptStyle":"grounded, schema-first, concise","outputContractDefault":"single_output"},{"key":"qwen_chat","label":"Qwen Chat","endpointFamily":"qwen_chat","dialectProfile":"explicit_instructional","description":"Performs best with a very explicit system prompt, short structured instructions, few-shot examples, and a firm output contract.","bigWins":["Use a very explicit system prompt.","Keep instructions short and highly structured.","Include a small few-shot example when possible.","Name the output format directly."],"cautionNotes":["Avoid long diffuse context walls.","Do not assume it will infer the target format from tone alone."],"systemPromptStyle":"explicit, short, example-guided","outputContractDefault":"single_output"},{"key":"llama_local","label":"Llama / Local Model","endpointFamily":"llama_local","dialectProfile":"simple_repeated_constraints","description":"Usually does better with short instructions, repeated constraints, simple schemas, and low ambiguity.","bigWins":["Keep the task short and concrete.","Repeat the most important constraints once.","Use a simple stable output schema.","Favor deterministic, low-ambiguity asks."],"cautionNotes":["Avoid layered metaphor in the instruction text.","Do not overload with too many optional branches."],"systemPromptStyle":"simple, repeated, low-ambiguity","outputContractDefault":"single_output"},{"key":"generic_http","label":"Generic HTTP Agent","endpointFamily":"generic_http","dialectProfile":"neutral_structured","description":"A neutral structured profile for agents or services that expose a plain HTTP request surface.","bigWins":["Carry the stable task contract cleanly.","Use concise role, goal, context, and style framing.","Return one directly actionable output.","Preserve constraints and review posture in plain language."],"cautionNotes":["Avoid vendor-specific prompting assumptions.","Do not add hidden transport semantics into the task body."],"systemPromptStyle":"neutral, concise, structured","outputContractDefault":"single_output"}],"a2aAgentCardExample":{"name":"SAE Review Agent","description":"An A2A-compatible agent that accepts SAE task structure for clearer delegation and review.","url":"https://agents.example.com/a2a","provider":{"organization":"Example Systems","url":"https://agents.example.com"},"version":"0.1.0","capabilities":{"streaming":true,"pushNotifications":false},"authentication":{"schemes":["Bearer"]},"defaultInputModes":["text/plain","application/json"],"defaultOutputModes":["text/plain","application/json"],"skills":[{"id":"sae-task-intake","name":"SAE Task Intake","description":"Accepts Structured Agent Exchange task payloads and returns clearer delegation and review guidance.","inputModes":["application/json"],"outputModes":["application/json","text/plain"],"examples":["Read an SAE task and return a concise execution recommendation."]}],"extensions":[{"uri":"https://structuredagentexchange.org/extensions/task/v0.1","description":"Draft SAE task support.","required":false}],"metadata":{"supportsSaeOverA2A":true,"saeProfileVersion":"0.1.0-draft"}},"plainHttpExample":"# Plain HTTP Example\n\nThis is the simplest non-A2A reference adapter for `SAE`.\n\nIt demonstrates that another system does not need to adopt any product-specific UI, internal dialect, or even A2A first in order to use the shared work-shape.\n\n## Goal\n\nTake a fuzzy ask, normalize it into an `SAE` task packet, validate it, and hand it to another service or queue.\n\n## Example Request\n\n```http\nPOST /api/sae/schema HTTP/1.1\nAccept: application/json\n```\n\nUse the schema response to validate a packet like:\n\n```json\n{\n  \"sae_version\": \"0.1.0\",\n  \"id\": \"sae-plain-http-001\",\n  \"title\": \"Experimental adapter review\",\n  \"objective\": \"Review a proposed browser-hand adapter and return a safe first implementation cut.\",\n  \"constraints\": [\n    \"keep the first cut reversible\",\n    \"return operator-readable review notes\"\n  ],\n  \"deliverable\": {\n    \"artifact_type\": \"review_packet\",\n    \"success_definition\": \"A downstream implementer can act on the packet without guessing the ask.\",\n    \"format_hint\": \"markdown+json\"\n  },\n  \"delegation_hints\": {\n    \"roles\": [\"planner\", \"implementer\", \"reviewer\"],\n    \"steps\": [\n      \"normalize the ask\",\n      \"return a first safe implementation cut\",\n      \"attach a review note\"\n    ],\n    \"clarification_budget\": \"low\"\n  },\n  \"review_posture\": {\n    \"risk_level\": \"medium\",\n    \"human_review\": true,\n    \"notes\": \"Hold widened execution until the review packet is accepted.\"\n  },\n  \"preference_hints\": {\n    \"continuity_preference\": \"preserve handoff clarity\",\n    \"trust_posture\": \"bounded\",\n    \"autonomy_bound\": \"approval_gated\"\n  },\n  \"provenance\": {\n    \"origin_system\": \"external_agent_system\",\n    \"created_at\": \"2026-05-07T00:00:00.000Z\",\n    \"compiled_from\": \"fuzzy_ask\"\n  }\n}\n```\n\n## Minimal Outside Flow\n\n1. receive or author a fuzzy ask\n2. normalize it into an SAE packet\n3. validate it against the SAE schema\n4. send the validated packet to another HTTP service, queue, or agent host\n5. receive artifacts back in whatever transport the host already uses\n\n## Why This Matters\n\nThis demonstrates that:\n\n- SAE is not tied to A2A alone\n- another team can adopt the work-shape without adopting someone else's internals\n- the same packet can later be carried over A2A if the host system grows into it\n","helloSaeHttp":{"readme":"# Hello SAE HTTP\n\nThis is the smallest example implementation shape.\n\nIt shows how a non-A2A service could accept and return SAE-friendly task data.\n\n## Files\n\n- `request.json`\n- `response.json`\n\n## Flow\n\n1. receive a fuzzy ask already normalized into SAE\n2. inspect the bounded objective and review posture\n3. return a reviewable result packet\n\n## Why This Exists\n\nThis is the answer to:\n\n`what is the smallest useful external implementation someone could build this week?`\n","request":{"sae_version":"0.1.0","id":"sae-hello-http-001","title":"Simple adapter review","objective":"Review a small adapter proposal and return the safest first implementation cut.","constraints":["keep the first cut reversible","keep the output readable by a human reviewer"],"deliverable":{"artifact_type":"review_packet","success_definition":"A downstream implementer can act on the packet without guessing intent.","format_hint":"markdown+json"},"delegation_hints":{"roles":["planner","reviewer"],"steps":["inspect the objective","return a safe first-cut recommendation"],"clarification_budget":"low"},"review_posture":{"risk_level":"medium","human_review":true,"notes":"Do not widen into live deployment."},"provenance":{"origin_system":"hello-sae-http","created_at":"2026-05-07T00:00:00.000Z","compiled_from":"fuzzy_ask"}},"response":{"task_id":"sae-hello-http-001","artifact_type":"review_packet","summary":"The safest first cut is a read-only adapter with explicit approval before any widened action.","recommended_first_cut":["implement request validation","return a bounded review packet","hold all live execution behind human review"],"review_notes":["The task is medium risk because it affects downstream execution shape.","The deliverable is clear enough for a second implementer to continue."]}}}}