azure-dev / azure.ai.projects
AGENTS.md573 starsChanged 8 months agoMITView on GitHub
Use this file together with cli/azd/AGENTS.md. This guide supplements the root azd instructions with the conventions that are specific to this extension. azure.ai.projects is a first-party azd extension under cli/azd/extensions/azure.ai.projects/. It runs as a separate Go binary and talks to the azd host over gRPC. It owns the azure.ai.project service target and the microsoft.foundry provisioning provider. The provider handles greenfield and existing-project provisioning, preview, and teardown. It owns the Foundry project endpoint context used by other AI extensions (e.g. azure.ai.agents).…
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 31,774 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository25 other files it gives its agents.
- cli/azd/AGENTS.mdAGENTS.md
- cli/azd/extensions/azure.ai.agents/AGENTS.mdAGENTS.md
- cli/azd/extensions/azure.ai.connections/AGENTS.mdAGENTS.md
- cli/azd/extensions/azure.ai.inspector/AGENTS.mdAGENTS.md
- cli/azd/extensions/azure.ai.routines/AGENTS.mdAGENTS.md
- cli/azd/extensions/azure.ai.skills/AGENTS.mdAGENTS.md
- cli/azd/extensions/azure.ai.toolboxes/AGENTS.mdAGENTS.md
- ext/vscode/AGENTS.mdAGENTS.md
- .github/skills/extension-user-contract-review/evals/fixtures/cli/azd/AGENTS.mdAGENTS.md
- ext/vscode/.github/copilot-instructions.mdCopilot instructions
- .github/copilot-instructions.mdCopilot instructions
- azdSkill
- agentic-workflowsSkill
- azd-code-reviewerSkill
- azd-grpc-contractsSkill
- azd-preflightSkill
- changelog-generationSkill
- extension-user-contract-reviewSkill
- foundry-extension-scenario-authoringSkill
- foundry-extension-scenario-pr-regressionSkill
- foundry-extension-scenario-suite-runSkill
- senseiSkill
- tool-version-upgradeSkill
- weekly-demo-videoSkill
- weekly-reportSkill
The file
Show its rules and text
We split this AGENTS.md into the rules it gives an agent
13 rules. About 0 are common, 5 are distinctive, and 8 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
8
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 3
Use this file together with cli/azd/AGENTS.md.
- Distinctive: line 49
If extension work depends on a new azd core change, plan for two PRs:
- Very specific: line 68
Use plain Go errors by default.
- Distinctive: line 81
Create a structured error once, as close as possible to the place where you know the final category, code, and suggestion.
- Very specific: line 83
If err is already a structured error, usually return it unchanged.
- Very specific: line 85
Do not add context with fmt.Errorf("context: %w", err) after err is already structured.
- Distinctive: line 132
keep them stable once introduced
- Distinctive: line 141
When changing the store:
- Very specific: line 143
Keep reads of the legacy extensions.ai-agents.project.context key best-effort: a malformed legacy blob must never block resolution from the …
- Very specific: line 167
After the GitHub release is live, a follow-up PR updates cli/azd/extensions/registry.json so azd users can install the new version.
- Very specific: line 175
Do not use log.Fatal or log.Panic for expected failures — return a structured error via exterrors instead.
- Distinctive: line 193
Use modern Go 1.26 patterns where they help readability
- Very specific: line 194
When using PromptSubscription(), create credentials with Subscription.UserTenantId, not Subscription.TenantId
≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Azure AI Projects Extension - Agent Instructions
Use this file together with cli/azd/AGENTS.md. This guide supplements the root azd instructions with the conventions that are specific to this extension.
Overview
azure.ai.projects is a first-party azd extension under cli/azd/extensions/azure.ai.projects/. It runs as a separate Go binary and talks to the azd host over gRPC.
It owns the azure.ai.project service target and the microsoft.foundry provisioning provider. The provider handles greenfield and existing-project provisioning, preview, and teardown.
It owns the Foundry project endpoint context used by other AI extensions (e.g. azure.ai.agents). The azd ai project commands persist, resolve, and surface the endpoint through a 5-level cascade:
- Explicit
--project-endpointflag - Active azd env's
AZURE_AI_PROJECT_ENDPOINT - Global config:
extensions.ai-projects.context.endpointin~/.azd/config.json - Host environment variable
FOUNDRY_PROJECT_ENDPOINT - Structured error with actionable suggestion
The resolver also performs a one-time auto-migration from the legacy extensions.ai-agents.project.context key (written by the removed azd ai agent project set command) into the new key.
Useful places to start:
internal/cmd/: Cobra commands, the endpoint resolver, and the config storeinternal/provisioning/: themicrosoft.foundryprovidersynthesis/: standalone shared Go module for project config synthesis and embedded IaCinternal/exterrors/: structured error factories and extension-specific codes
Projects owns the standalone azure.ai.projects/synthesis module. Both extensions import it; do not copy synthesis or templates into Agents. Projects uses a child-module replace, while Agents uses a sibling-to-sibling replace pointing at this module. Run its tests separately with go -C synthesis test ./... because the parent module's go test ./... does not traverse nested modules.
BYO AKS synthesis contract tests live in Projects; its parity test checks both production copies and embedded templates. Run the Projects synthesis suite when changing either copy.
Native Bicep tests require at least the minimum exported by the referenced azd SDK's pkg/tools/bicep.Version. They use AZD_BICEP_TOOL_PATH when set, otherwise PATH and the Azure CLI bundled location. Missing or unsupported automatically discovered tools skip with a diagnostic; an invalid explicit override fails. These helpers never install or upgrade tools.
Build and test
From cli/azd/extensions/azure.ai.projects:
# Build using developer extension (for local development)
azd x build
# Or build using Go directly
go build
# Run unit tests
go test ./... -count=1If extension work depends on a new azd core change, plan for two PRs:
- Land the core change in
cli/azdfirst. - Land the extension change after that, updating this module to the newer azd dependency with
go get github.com/azure/azure-dev/cli/azd && go mod tidy.
For local development, draft work, or validating both sides together before the core PR is merged, you may temporarily add:
replace github.com/azure/azure-dev/cli/azd => ../../That replace points this extension at your local cli/azd checkout instead of the version in go.mod. Do not merge the extension with that replace still present.
Error handling
This extension uses internal/exterrors so the azd host can show a useful message, attach an optional suggestion, and emit stable telemetry.
Default rule
Use plain Go errors by default. Switch to exterrors.* only when the current code can confidently answer all three of these:
- What category should telemetry see?
- What stable error code should be recorded?
- What suggestion, if any, should the user get?
That usually means:
- lower-level helpers return
fmt.Errorf("context: %w", err) - user-facing orchestration code in
internal/cmd/classifies the failure withexterrors.*
Most important rule
Create a structured error once, as close as possible to the place where you know the final category, code, and suggestion.
If err is already a structured error, usually return it unchanged.
Do not add context with fmt.Errorf("context: %w", err) after err is already structured. During gRPC serialization, azd preserves the structured error's own message/code/category, not the outer wrapper text. If you need extra context, include it in the structured error message when you create it.
Choosing an Error Type
| Situation | Prefer |
| Invalid endpoint, flag value, or persisted blob | exterrors.Validation |
| Missing endpoint across all 5 resolver levels, unavailable azd daemon | exterrors.Dependency |
| Auth or tenant/credential failure | exterrors.Auth |
| azd/extension version or capability mismatch | exterrors.Compatibility |
| User cancellation | exterrors.Cancelled |
| Azure SDK HTTP failure | exterrors.ServiceFromAzure |
| Unexpected bug or local failure with no better category | exterrors.Internal |
Recommended pattern
func loadEndpoint(raw string) (string, error) {
normalized, _, err := validateProjectEndpoint(raw)
if err != nil {
return "", fmt.Errorf("validate %q: %w", raw, err)
}
return normalized, nil
}
func runCommand() error {
endpoint, err := loadEndpoint(rawFlag)
if err != nil {
return exterrors.Validation(
exterrors.CodeInvalidParameter,
fmt.Sprintf("project endpoint is invalid: %s", err),
"provide an https:// Foundry project endpoint URL",
)
}
_ = endpoint
return nil
}Error codes
Define new codes in internal/exterrors/codes.go.
- use lowercase
snake_case - describe the specific failure, not the general category
- keep them stable once introduced
Persisted project context
The endpoint store lives at extensions.ai-projects.context in ~/.azd/config.json and is accessed exclusively through helpers in internal/cmd/project_context_store.go:
getProjectContext/setProjectContext/clearProjectContext— public surfacereadProjectContext/writeMigratedProjectContext/clearProjectContextFromConfig— internal helpers that take aprojectContextConfiginterface so tests can drive them with a fake
When changing the store:
- Keep reads of the legacy
extensions.ai-agents.project.contextkey best-effort: a malformed legacy blob must never block resolution from the new key, the flag, orFOUNDRY_PROJECT_ENDPOINT. clearProjectContextmust remain idempotent and must clear both the new and legacy keys, even when the previous-endpoint read fails (so users can always recover from a corrupted persisted blob).- The auto-migration in
readAzdHostedSourcesis best-effort: a transient write failure must not break the command the user actually invoked.
Release preparation
A new extension release ships in two PRs:
Provider handoff release
The first release that moves microsoft.foundry here must be coordinated with the matching azure.ai.agents release. Publish both artifacts before updating either registry entry, then update both entries and the microsoft.foundry meta-package together. Old agents and new projects versions cannot run together because azd rejects duplicate provider registration.
PR 1 — Version bump
Bumps the extension to the new version. Touches only:
version.txt— new semver stringextension.yaml—version:fieldCHANGELOG.md— new release section at the top
Once merged, the team triggers the CI release pipeline, which builds, signs, and publishes the extension binaries as a GitHub release.
PR 2 — Registry update
After the GitHub release is live, a follow-up PR updates cli/azd/extensions/registry.json so azd users can install the new version. The contents of that file are produced by running azd x publish against the published release artifacts.
Output: log vs fmt
Extensions write directly to stdout/stderr — there is no Console abstraction from azd core.
fmt.Print*— user-facing output (stdout). Pair withoutput.With*Formathelpers for styled text.log.Print*— developer diagnostics (stderr). Hidden unless--debugis set. Never uselogfor anything the user needs to see.- Do not use
log.Fatalorlog.Panicfor expected failures — return a structured error viaexterrorsinstead.
// ✅ log — internal detail the user doesn't need to see
log.Printf("config read at %s returned %d bytes", path, n)
// ✅ fmt — user-facing status and results
fmt.Println(output.WithSuccessFormat("Project endpoint set to %s", endpoint))
// ❌ fmt used for debug noise — user sees internal details they can't act on
fmt.Printf("normalized URL: %s\n", normalized) // use log.Printf
// ❌ log used for user-facing info — user never sees it without --debug
log.Printf("No project endpoint resolved") // return an exterrors.Dependency insteadOther extension conventions
- Use modern Go 1.26 patterns where they help readability
- When using
PromptSubscription(), create credentials withSubscription.UserTenantId, notSubscription.TenantId
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

