acdl/.ciagent/PROJECT.md

ACDL — Agentic Cloud Delivery Platform

Vision / Core Value

Consumers declare intent; the platform delivers safe production deployment through an agentic stack. The platform absorbs two frictions: the cognitive load of getting the infrastructure right, and the operational work of getting the change to production safely.

Source of truth for why: docs/vision.md. Source of truth for how: docs/architecture.md + .ciagent/ARCHITECTURE.md. Where the two conflict, the vision wins.

North Star

A merged change progresses through lower environments end-to-end without a platform engineer joining a thread, approving a ticket, or manually triggering a stage gate. A non-technical consumer ships a production deployment by declaring intent — without authoring a workflow, a configuration file, or a Terraform module. Every production change is traceable to a human attestation and an immutable evidence stream.

Core Tenets (from docs/vision.md)

1. Operations are Declared, Not Executed. Consumers define what they need; the platform reconciles, provisions, and progresses.
2. The Delivery Lifecycle is a Sovereign Boundary. The platform governs infra and delivery; it does not penetrate upstream product/SDLC. Integration is only through validated, published contracts.
3. Lower Environments are Autonomous; Higher Environments are Attested. Dev = zero-touch agentic. QA/prod/dr = deliberate human attestation, not rubber stamps.
4. Safety is Computed, Not Assumed. Every action produces a measurable, explainable confidence signal. The signal is the platform's certified answer to "is this safe to proceed?"
5. Infrastructure is Consumed, Not Maintained. Compute is abstract, containerized, or serverless. No node/OS/bare-metal lifecycle.
6. Two Consumer Surfaces, One Platform. Technical developers (L3A) and non-technical consumers (L3B) converge on the same contract schema, the same policy envelope, and the same evidence stream.

Domain Boundaries

• In scope: environment progression; cloud resource lifecycle; operational security and observability NFRs; policy enforcement; immutable audit lineage; confidence frameworks; two consumer surfaces (developer + agentic).
• Out of scope: application business logic; IDE workflows; product backlog / sprint planning; compute requiring node-level or OS-level management.
• Interface: upstream systems integrate through a strict contract boundary. The platform validates, enriches with operational standards, and reconciles the target state.

Objective for Milestone v1.1

Finalize the architecture to v1.0 (resolve all 11 open design decisions in docs/architecture.md §13) and prove the locked commitments with one end-to-end v1 implementation spike:

• One L1 module (l1-s3) — substrate-agnostic, IR-typed interface.
• One L2 thin-composition (l2-static-asset) — references the L1.
• Terraform adapter — compiles the IR to a real terraform plan against AWS via OIDC (no long-lived credentials, per §12.5).
• One contract submission → contract→IR resolution → terraform plan → PolicyCheckResult (Checkov) → confidence signal → evidence event to the DynamoDB outbox.

The spike validates the architecture's claim that the IR-shaped commitments do not require a polyglot mess (docs/architecture.md §14, step 2).

Milestone v1.1 Phases

Phase	Name	Goal
06	archive-demo-and-reorient	Move the v1.0 demo (modules/, scripts/, evidence-ui/, contracts/, demo workflows) to demo/; establish the new repo layout (platform/, schemas/, adapters/, terraform/, modules-ir/); rewrite README.
07	architecture-v1-finalization	Resolve the 11 open decisions → architecture v1.0. Author IR JSON Schema, PolicyCheckResult schema, contract schema, confidence-signal spec, HITL matrix, outbox/ledger design under schemas/ + platform/.
08	aws-oidc-bootstrap	One-shot use of a temporary long-lived key (waiver D-034) to create an IAM role + OIDC trust policy for the act_runner, an S3 state bucket, and a DynamoDB lock table. Rotate the key. Verify the runner assumes the role via OIDC with no long-lived secret.
09	v1-spike-ir-and-l1-and-adapter	Target Stack IR; one real L1 (l1-s3) with IR-typed interface; L1 registry; Terraform adapter (IR → Terraform var/output + terraform plan) running against AWS via OIDC.
10	v1-spike-l2-and-contract-e2e	One L2 thin-composition (l2-static-asset) referencing l1-s3; contract schema + contract→IR resolution; one end-to-end contract submission → terraform plan → Checkov → confidence signal → evidence event to outbox. Verify the IR commitments hold.

Milestone COMPLETE gate: review → ship v1.2.0 (feature milestone, next minor per ship.md) → audit.

Requirements

v1.0 (Prior milestone — the demo)

Status: complete. Tag v1.1.0. All REQ-01..15 satisfied by the stub-driven executive demo. See REQUIREMENTS.md §v1 and the prior decisions table appendix below. The demo is archived to demo/ in Phase 06.

v1.1 (Active milestone — architecture finalization + v1 spike)

New requirements REQ-16..REQ-2x — see REQUIREMENTS.md §v1.1. Summary:

• REQ-16: Architecture finalized to v1.0 (11 open decisions resolved).
• REQ-17: Target Stack IR defined as JSON Schema; substrate-agnostic.
• REQ-18: PolicyCheckResult normalized schema defined; Checkov adapter.
• REQ-19: Six-input confidence signal specified with per-env thresholds (dev 0.50 / qa 0.75 / prod 0.90 / dr 0.95) and severity→penalty mapping.
• REQ-20: Tiered audit ledger design (S3 Object Lock 7-yr + DynamoDB outbox, RPO=0, JWS detached signatures, prev_event_hash chain).
• REQ-21: Full 8-concern HITL matrix + separation-of-duties design (CODEOWNERS + DynamoDB identity-distinctness).
• REQ-22: Contract schema (JSON Schema draft 2020-12) with per-env mandatory/optional inputs and profile: agentic marker for L3B.
• REQ-23: AWS OIDC bootstrap (IAM role + trust policy for act_runner); the long-lived key is used once then rotated (waiver D-034).
• REQ-24: One real L1 module (l1-s3) with an IR-typed interface.
• REQ-25: One real L2 thin-composition (l2-static-asset) referencing l1-s3.
• REQ-26: Terraform adapter compiles the IR to a real terraform plan against AWS via OIDC; state in S3 + DynamoDB.
• REQ-27: One end-to-end contract submission → contract→IR resolution → terraform plan → Checkov → confidence signal → evidence event to outbox.
• REQ-28: Spike verification proves the IR-shaped commitments hold (no polyglot mess; the adapter is the only substrate-specific code).

Constraints

• Forge: Gitea at https://git.cloudinit.dev, org continuous-intelligence.
• CI runtime: act_runner / Gitea Actions (reuses GitHub Actions workflow YAML).
• Cloud: AWS via OIDC federation. Long-lived credentials are forbidden (§12.5). The v1.1 spike uses a temporary long-lived key once to bootstrap OIDC (waiver D-034), then rotates it.
• Substrate: Terraform adapter in v1 (the only adapter). L1/L2 are substrate-agnostic in shape; the adapter is the only substrate-specific code.
• State: S3 (state files) + DynamoDB (locking), single-region in v1.
• Environments: dev (autonomous) → qa (QA HITL) → prod (SRE HITL) → dr (SRE HITL). Staging does not exist (Path A locked).
• Compute: abstract / containerized / serverless. No VMs, bare metal, OS lifecycle.
• Autonomy: Full. Escalation hooks: deploy, delete_data, merge_to_main.

Anti-Goals (from docs/vision.md §7)

• Not an upstream development platform (no product backlogs, IDE, code authorship).
• Not a general-purpose AI (autonomy is narrow, bounded by policy envelopes).
• Not a legacy infrastructure bridge (no VMs/bare metal/OS).
• Not a permissive delivery highway (no escape hatches past confidence or HITL).
• Not a mutable audit log (VCS history ≠ regulatory evidence).

Context

• The acdl repo exists at the org root. acdl-contracts and acdl-evidence exist from the v1.0 demo and continue as the developer surface and the audit-timeline host respectively.
• docs/vision.md and docs/architecture.md (v0.2) are the upstream vision/architecture sources, pulled from origin/main at the start of v1.1.
• The v1.0 demo (tag v1.1.0) is the reference of intent — it proved the shape (L1/L2/contract/confidence/evidence/HITL) on stubs. v1.1 replaces the stubs with the real platform substrate.

Key Decisions (v1.1)

Carries forward the still-valid v1.0 decisions (see appendix). New v1.1 decisions:

ID	Decision	Rationale	Outcome
D-034	Temporary long-lived AWS key (waiver) used once in Phase 08 to bootstrap OIDC trust; rotated immediately after	§12.5 forbids long-lived creds; OIDC needs one bootstrapping aws iam call before the runner can assume a role	Spike achieves real terraform plan against AWS without violating the locked target after bootstrap
D-035	Milestone version = v1.1 (feature), ship tag v1.2.0	Real platform is a breaking reframing of the demo, but treated as the next incremental milestone per user choice; ship.md: feature milestone → next minor	Tag v1.2.0 on milestone COMPLETE
D-036	Spike picks l1-s3 + l2-static-asset	Simplest real AWS resource (no IAM/network deps); smallest real terraform plan; proves the IR + adapter end-to-end	Spike scope fixed
D-037	Demo archived to demo/ (not deleted)	Preserves the working v1.0 demo as intent reference; new platform layout under platform/, schemas/, adapters/, terraform/, modules-ir/	No churn on demo code; clean separation
D-038	Open decisions resolved in "accept recommendations + decide rest" mode	User-locked mode: accept architecture's stated recommendations (W1.A, W1.B, W2.A, BA.A); lead-developer decides the remaining 8 (W3.D, W3.E, BA.B, BA.C, BA.D, BA.E, BA.F, OpenTofu timing) with rationale	Architecture reaches v1.0 in Phase 07
D-039	Spike-only waiver: per-run-rotated long-lived AWS key. OIDC federation deferred to v1.2, blocked on go-gitea/gitea#36988.	RESEARCH TARGET 1 verdict (conf 0.95): Gitea Actions does NOT support id-token: write / OIDC token issuance as of Gitea 1.27.x / gitea-runner v2.1.0. GitHub's OIDC pattern is not portable. The waiver satisfies §12.5's intent (no persistent long-lived key) for the spike: the key is rotated after each run by scripts/rotate_spike_key.sh. v1.2 implements real OIDC when the Gitea PR merges.	Spike achieves real terraform plan against AWS without a persistently long-lived key; real OIDC is a v1.2 deliverable
D-040	The 6 confidence-signal inputs are: policy (0.30), validation (0.25), freshness (0.10), source (0.15), history (0.10), nfrs (0.10). Weights frozen for v1, tuned in v1.2 alongside thresholds (BA.B).	Architecture §8 locks "six canonical inputs" but does not enumerate them; RESEARCH TARGET 6 chose the platform-computable subset present in every environment (incl. dev).	Confidence signal (Phase 10) has a concrete input enumeration
D-041	Spike audit ledger = v1.0 hash chain + DynamoDB outbox + acdl-evidence mirror. S3 Object Lock (compliance mode, 7-yr) + JWS (platform KMS key, quarterly rotation) + daily checkpoints are v1.2 build-out, authored as design in Phase 07.	REQ-20 is "design authored," not "implemented." The spike proves the outbox write path; the regulatory ledger is v1.2.	Spike scope stays bounded; REQ-20 satisfied by the Phase 07 design doc
D-042	HITL approver identity in Gitea = gitea.actor of the workflow_dispatch run that sets approve_qa=true/approve_prod=true/approve_dr=true. Separation-of-duties reads approver_qa from the DynamoDB outbox and compares to the prod-dispatch gitea.actor.	Gitea has no Environments API (re-confirmed in RESEARCH); gitea.actor is the only approval-identity signal.	SoD design (Phase 07) is concrete for the Gitea forge
D-043	Tag/naming compliance deferred for the spike: the Checkov adapter emits a single SKIPPED PolicyCheckResult (ruleId: ACDL_TAG_NAMING, severity: info) so the confidence policy input is non-empty. Custom Checkov YAML rule lands in v1.2.	Checkov has no built-in tag-presence check; a custom rule in the spike is scope creep.	Spike's policy input is non-empty without a custom-rule dependency
D-044	DynamoDB outbox = PAY_PER_REQUEST; PK contractId, SK eventType#eventTs, TTL expire_at = now + 365d. No separate async worker/DLQ in the spike (RTO = workflow re-run); v1.2 outbox worker + DLQ is a Phase 07 design artifact.	On-demand is zero-cost-at-idle for the spike's single dev submission.	Spike outbox is minimal; v1.2 worker design authored in Phase 07
D-045	Runner tooling: runs-on: ubuntu-latest; install terraform via HashiCorp apt repo (pin 1.9.*), checkov via pip (pin >=3.2,<4, --break-system-packages). Neither is pre-installed on the default runner image.	RESEARCH TARGET 2; pinning avoids mid-spike version drift.	Phase 09/10 workflows have a concrete setup step
D-046	act_runner → gitea-runner rename: Phase 07 updates docs to use the current name gitea-runner (renamed 2026-04 in gitea/runner#850).	RESEARCH TARGET 1 + R-4: naming drift between v1.0 docs and the current runner.	Docs reflect the current binary name

Open-decision resolutions (Phase 07 deliverable — recorded here for traceability)

ID	Question	Resolution
W1.A	AI-refinement trigger	Accept recommendation. Joint condition: N ≥ 50 consecutive changes with zero rollbacks AND no L1/L2 incident in last 6 months AND Infra & Ops unilateral override.
W1.B	Multi-stack edge case rule	Accept recommendation. Permitted only for (a) DR-region mirror, (b) time-boxed experimental stack with TTL ≤ 30d, (c) explicit Infra & Ops approval with multiStack.justification.
W2.A	Tag mutability for prod	Accept recommendation (Path B). Tag for dev/qa, SHA for prod. Platform CLI resolves tag→SHA for prod-bound workflows. Justified by the "Audit truth lives outside the repository" bet.
BA.A	Initial L3B skill catalog	Accept recommendation. 5 skills: web API, worker, scheduled job, static asset, basic observability bootstrap. Addition criteria: (a) reviewable for sensitive data, (b) expressible as a single contract submission, (c) documented use case.
W3.D	L1/L2 standard versioning	Decided. Semver: interface → MAJOR, behavior → MINOR, lifecycle → PATCH (same as the v1.0 demo D-rule, lifted to the real platform). Pin model: L2 contracts pin L1 by name@semver; the resolver picks the highest compatible. Evolution: MAJOR bumps require a new registry entry (immutable publication); old entry enters a 12-month deprecation window.
W3.E	Schema mandatory vs optional inputs	Decided. Per-env mandatory table: dev requires stack + environment; qa adds validation.e2eSuite + validation.loadTest; prod adds runbook + dashboard + oncall; dr adds drDrillRef. inputs map is always optional. profile: agentic fields (naturalLanguageIntent, confidenceAtSubmission, agentTrace) optional everywhere.
BA.B	Confidence threshold tuning	Decided. Starting thresholds frozen for v1. Tuning begins in v1.2: track FP/FN per environment quarterly; override authority = Infra & Ops + SRE joint sign-off; any override is itself a confidence-event in the audit stream.
BA.C	On-call / operational ownership	Decided. Platform on-call = Infra & Ops rotation. Escalation: L3A/L3B halt → platform on-call pager (Sev2); consumer-visible outage → consumer on-call (Sev1) with platform on-call support. Consumer on-call relationship is contractual, defined at onboarding (BA.E).
BA.D	Cost / capacity governance	Decided. Cloud cost owner = Infra & Ops FinOps. Per-contract consumption reported monthly. Runaway spend: hard halt at 120% of contract-declared budget envelope via the confidence signal (cost is one of the 6 inputs); override = FinOps + SRE joint sign-off.
BA.E	Consumer onboarding	Decided. Two paths: developer (L3A) — getting-started walks through contract schema + central pipeline template; citizen developer (L3B) — onboarding grants a scoped agent + skill catalog, no workflow authoring. Both end in a sandbox dev submission that must pass the confidence gate before the consumer is promoted.
BA.F	Cross-platform evolution	Decided. The contract schema, IR, PolicyCheckResult, confidence signal, and audit stream are portable (substrate- and forge-agnostic). Forge-specific code: workflow YAML, OIDC trust, CODEOWNERS, Environments. A second forge (e.g., GitLab) requires a forge adapter + a workflow-template translator; no change to L1/L2/IR/confidence/audit.
Q1.3	OpenTofu timing	Decided (deferred). Not in v1 or v1.1. The substrate abstraction (§12) makes OpenTofu a future adapter, not an architecture change. Revisit when an OpenTofu adapter is requested; no version committed.

Appendix — Prior milestone (v1.0 demo) decisions

The v1.0 demo (tag v1.1.0) carried decisions D-001..D-033. They governed the stub-driven executive demo and remain valid for the archived demo under demo/. They are superseded by the v1.1 decisions above for the real platform. Full text preserved in git history at tag v1.1.0.

Operational parameters (CLARIFY auto-resolution, full autonomy)

Resolved at the CLARIFY stage to unblock planning. None require user sign-off (autonomy = full; all within locked constraints).

Parameter	Value	Rationale
AWS region	us-east-1	Default; matches v1.0 demo references; single-region in v1 (§12.3)
Terraform state bucket	acdl-tfstate-<account-id>-us-east-1	Namespaced by account id to avoid collision; region-suffixed
Terraform lock table	acdl-tflock	DynamoDB; single-region v1
OIDC IAM role	acdl-act-runner-role	Assumed by the act_runner via web-identity
OIDC trust subject	repo:continuous-intelligence/acdl:ref:refs/heads/main (+ phase branches)	Least-privilege; refined in Phase 08
Spike L1 (l1-s3) inputs	bucket_name: string, region: string	Minimal S3 interface per §2
Spike L2 (l2-static-asset)	thin-composition referencing l1-s3 only; depth 1	Smallest real plan per D-036
Spike contract	contracts/spike.yaml: stack: l2-static-asset, environment: dev, inputs: { bucket_name: acdl-spike-bucket, region: us-east-1 }	One end-to-end submission (REQ-27)
Spike terraform command	plan only	apply is out of scope (Out of Scope table); HITL-gated in v1.2
Checkov ruleset (spike)	the 4 L2 checks (secrets-in-plaintext, public ingress, IAM wildcard, KMS key reference) + tag/naming	§3 + §12.4; Kyverno/OPA deferred
v1.0 tags preserved	v1.0.1..v1.0.5, v1.1.0 retained	Immutability; demo archive does not rewrite history
Next ship tag	v1.2.0	Feature milestone → next minor per ship.md (D-035)

Items deferred to RESEARCH (not clarifications)

• Gitea/act_runner OIDC support — does act_runner emit an OIDC id-token? Determines whether real-AWS plan is achievable in this environment or whether a spike-only waiver is needed. Highest-priority research target.
• Terraform + Checkov availability on the runner image — install in the workflow if missing.
• actions/configure-aws-credentials action on act_runner — if unavailable, fall back to aws sts assume-role-with-web-identity from a step.