[czprz]

Summary

Replaces the Profile/Service/Dep + Docker Compose/Kubernetes rendering architecture with the flat Task-based manifest and scheduler described in docs/design/task-executor-refactor.md (Phase 1 of 3).

Motivation

The old model's strong coupling to Compose/k8s schema versions made adding capability feel like "a new feature every time," and the design doc (reviewed and refined across several prior review cycles — see design/task-executor-refactor-doc branch, PR #8) proposes a simpler, Make-like flat task graph with run:/deps:/ready: instead. This PR builds that core.

Changes

• internal/config: new Manifest/Task/ReadyCheck schema — task names are top-level devx.yaml keys directly (no profiles:/services:/deps: wrapper). includes: resolves and namespaces another file's tasks (infra:db), with per-file vars: scoping. New Validate/ValidateTools/ValidateSetup.
• internal/graph: generalized to map[string]config.Task; added Closure() for partial up/run <name> invocations.
• internal/task (new package — the architectural core): ${VAR} substitution + run:-text inference (ports/names/detached-ness), process and container drivers, os:/sources:/generates: inference helpers, a .devx/tasks.json state registry so status/logs/exec/down can reconstruct what's running across separate CLI invocations, and a levelized parallel scheduler (Up/Run) with --dry-run/--keep-going/--group.
• internal/doctor: dropped k8s/registry checks, added requires: binary checks, rewired port-collision checking to the new Task model.
• CLI: up/down/run/status/logs/exec/validate/doctor/tasks/new/init rewritten for the Task model. devx <name> is shorthand for devx run <name>. render/lock/providers/export now print a clear "not supported" error (no compose/k8s artifact to render anymore).
• devx down correctness fix: the state registry now records each detached process task's process-group PID, and down signals it directly (SIGTERM, then SIGKILL after stopTimeout) — previously a no-op for anything that wasn't a container.
• Deleted: internal/compose, internal/k8s, internal/providers, internal/lock, and internal/ai/export.go (GenerateDeployment depended on the deleted Profile type; its only caller was already a stub).
• Docs: docs/manifest.md and schemas/devx.schema.json rewritten for the Task schema; examples/basic/devx.yaml rewritten (db/cache/migrate/api/worker tasks).

Known gaps, flagged explicitly (not silently dropped)

• The design doc assumed internal/process.Supervisor already existed — it didn't, anywhere in the codebase. internal/task's process/container drivers were built from scratch.
• The old container: auto|true|false convenience (auto-dispatch a run into the devcontainer) has no Task-model equivalent yet.
• --watch is 1s mtime polling, not fsnotify; run: executes via the system shell, not an embedded shell interpreter — both intentionally deferred per the design doc's phasing.
• Cobra CLI migration and the workspace/agents feature (parallel per-branch workspaces + auto-started AI coding agent) are out of scope for this PR — tracked as later phases.

Testing

• Unit tests added / updated — internal/config, internal/graph, internal/task (new: resolve_test.go, scheduler_test.go), internal/doctor all pass.
• Integration tests added / updated — not added in this PR; tests/integration/ still targets the old schema and will need a follow-up rewrite.
• Tested manually — steps:
  • go build ./..., go vet ./..., go test ./..., golangci-lint run ./... all pass clean.
  • Ran the full lifecycle against real Docker on examples/basic/devx.yaml: devx validate (clean), devx up -d (db/cache/migrate/api/worker all start and become ready), devx status (shows all 4 containers running), curl localhost:8080/ (API responds), devx down (all 4 containers stop and are removed).

Checklist

• go build ./... passes
• go test ./... passes
• Docs updated (docs/manifest.md, schemas/devx.schema.json, examples/basic/devx.yaml)
• CHANGELOG.md updated under [Unreleased]

[czprz]

In-depth review

Ran a thorough code review (concurrency, correctness, dependency-resolution logic, ${VAR} substitution precedence, YAML unmarshalling edge cases, schema/docs drift) over the full diff. Found and fixed 4 real issues:

#	Severity	Finding	Fix
1	High	-p HOST:CONTAINER port inference captured the container port, not the host-reachable one, whenever they differ (e.g. -p 8080:80 → inferred 80). Only host==container mappings worked by accident.	Rewrote portPattern/added hostPortFromMapping to correctly parse [ip:]hostPort:containerPort[/proto]. Added regression test.
2	High	An included file's own deps: between two of its own tasks were never namespaced during includes: merging — broke any reusable module with internal dependencies.	Added namespaceDeps to rewrite intra-include deps to their merged name. Added regression test.
3	High	devx logs --follow (and attached devx up) busy-spun a full CPU core with no backoff when a tailed file had no new data.	Added the same 300ms poll interval already used by internal/task's followReader.
4	Medium	stopTimeout: was documented and validated but never actually read by devx down — always used a hardcoded 10s.	devx down now resolves each task's own stopTimeout: (skipped under --force, which is always immediate).

Also flagged, not fixed (correctly out of scope for this phase, now documented as such rather than silently implied as working): restart: never\|on-failure\|always is validated/persisted but devx has no supervisor loop watching a detached task for exit — a crashed task with restart: always just stays dead. Marked "reserved for a later phase" in docs/manifest.md/schemas/devx.schema.json, same treatment as agents:.

All fixes verified: go build/vet/test/lint clean, plus re-ran the full real-Docker lifecycle against examples/basic/devx.yaml (whose api task's -p ${API_PORT}:80 mapping is exactly finding #1's bug scenario) — validate → up -d → status → curl localhost:8080/ → down all behave correctly.

See commit da314a2 for details.