The SANDBOX-APPLY-CONVERGENCE-PROCEDURE (feat-foundry-infra-live-seam-sandboxed-apply) — an UNATTENDED, real tofu apply in a THROWAWAY sandbox account followed by a post-apply convergence proof (apply -> re-plan == ∅), originally designed to emit a dedicated sandbox-apply walk-evidence surface for the since-retired bespoke merge gate (see the DORMANT note). EXPLICIT CLI ORCHESTRATION (never native tofu test, which is opaque — no saved plan / -detailed-exitcode / plan-JSON): tofu init (throwaway backend) -> resolve the applying identity from the OpenTofu PROVIDER''s OWN awscalleridentity in the SAME provider config that applies (NEVER a sidecar aws CLI) -> decidesandboxapply (pure, fail-closed; REFUSE unless the provider-evaluated id is a ^\d{12}$ member of the committed .foundry/sandbox-accounts.json allowlist, no provider credential override, posture==EXECUTE, the policy gate passes, and the pre-apply plan is non-empty) -> arm the crash-backstop tofu destroy trap BEFORE any mutation -> plan -out=planfile -lock-timeout=120s -> show -json planfile -> the MANDATORY policy gate over the saved plan (absent/error/fail all REFUSE, no ack escape) -> apply -lock-timeout=120s planfile (exit≠0 partial-apply is a COMPUTED FAIL, not a crash) -> show -json > state.json (coverage) -> plan -out=replan -> show -json replan -> assert convergence (exit-0 AND an empty parsed resource-change set, via the shipped foundryplanmodel.parseactionsdetail) -> an EXPLICIT captured tofu destroy (its real exit sets the recorded teardown status; runs on GREEN, convergence-FAIL, AND apply-FAIL) -> emitinfrawalkevidence on every computed outcome. Composes with — never re-builds — the shipped posture gate + SoD execute/generate split (foundryidapply): the sandboxed apply IS the EXECUTE branch bound to the sandbox account; prod always stays GENERATERUNBOOK. ADVISORY craft procedure FOR the trusted operator — it does NOT gate, approve, or merge; the merge floor (the adopter''s branch protection + CI, docs/merge-floor.md) is the merge authority. CURRENTLY DORMANT (named honestly): the decidesandboxapply/emitinfrawalkevidence primitives this procedure drives, and its drop-in doctor selftest, were not shipped (no surviving implementation) in scripts/ — this SKILL.md is design intent for a re-implementation against the current floor (ci.yml + btb-gates), not a live procedure today.
Install with
Claude CodeCodexCursorGitHub CopilotGemini CLI
Supported by the Atlas install prompt. Skill-specific requirements vary.
Skill stats
Skills1CategoryEngineeringStars0
About this skill
The SANDBOX-APPLY-CONVERGENCE-PROCEDURE (feat-foundry-infra-live-seam-sandboxed-apply) — an UNATTENDED, real tofu apply in a THROWAWAY sandbox account followed by a post-apply convergence proof (apply -> re-plan == ∅), originally designed to emit a dedicated sandbox-apply walk-evidence surface for the since-retired bespoke merge gate (see the DORMANT note). EXPLICIT CLI ORCHESTRATION (never native tofu test, which is opaque — no saved plan / -detailed-exitcode / plan-JSON): tofu init (throwaway backend) -> resolve the applying identity from the OpenTofu PROVIDER''s OWN awscalleridentity in the SAME provider config that applies (NEVER a sidecar aws CLI) -> decidesandboxapply (pure, fail-closed; REFUSE unless the provider-evaluated id is a ^\d{12}$ member of the committed .foundry/sandbox-accounts.json allowlist, no provider credential override, posture==EXECUTE, the policy gate passes, and the pre-apply plan is non-empty) -> arm the crash-backstop tofu destroy trap BEFORE any mutation -> plan -out=planfile -lock-timeout=120s -> show -json planfile -> the MANDATORY policy gate over the saved plan (absent/error/fail all REFUSE, no ack escape) -> apply -lock-timeout=120s planfile (exit≠0 partial-apply is a COMPUTED FAIL, not a crash) -> show -json > state.json (coverage) -> plan -out=replan -> show -json replan -> assert convergence (exit-0 AND an empty parsed resource-change set, via the shipped foundryplanmodel.parseactionsdetail) -> an EXPLICIT captured tofu destroy (its real exit sets the recorded teardown status; runs on GREEN, convergence-FAIL, AND apply-FAIL) -> emitinfrawalkevidence on every computed outcome. Composes with — never re-builds — the shipped posture gate + SoD execute/generate split (foundryidapply): the sandboxed apply IS the EXECUTE branch bound to the sandbox account; prod always stays GENERATERUNBOOK. ADVISORY craft procedure FOR the trusted operator — it does NOT gate, approve, or merge; the merge floor (the adopter''s branch protection + CI, docs/merge-floor.md) is the merge authority. CURRENTLY DORMANT (named honestly): the decidesandboxapply/emitinfrawalkevidence primitives this procedure drives, and its drop-in doctor selftest, were not shipped (no surviving implementation) in scripts/ — this SKILL.md is design intent for a re-implementation against the current floor (ci.yml + btb-gates), not a live procedure today.
Quick start
Copy the install prompt. Use the green button above this README.
Send it to your agent. Paste it into your coding agent’s chat and send.
Review the setup. Your agent checks the skill’s instructions, requirements, and installation scope.
The SANDBOX-APPLY-CONVERGENCE-PROCEDURE (feat-foundry-infra-live-seam-sandboxed-apply) — an UNATTENDED, real `tofu apply` in a THROWAWAY sandbox account followed by a post-apply convergence proof (`apply -> re-plan == ∅`), originally designed to emit a dedicated `sandbox-apply` walk-evidence surface for the since-retired bespoke merge gate (see the DORMANT note). EXPLICIT CLI ORCHESTRATION (never native `tofu test`, which is opaque — no saved plan / `-detailed-exitcode` / plan-JSON): `tofu init` (throwaway backend) -> resolve the applying identity from the OpenTofu PROVIDER''s OWN `aws_caller_identity` in the SAME provider config that applies (NEVER a sidecar `aws` CLI) -> `decide_sandbox_apply` (pure, fail-closed; REFUSE unless the provider-evaluated id is a `^\d{12}$` member of the committed `.foundry/sandbox-accounts.json` allowlist, no provider credential override, posture==EXECUTE, the policy gate passes, and the pre-apply plan is non-empty) -> arm the crash-backstop `tofu destroy` trap BEFORE any mutation -> `plan -out=planfile -lock-timeout=120s` -> `show -json planfile` -> the MANDATORY policy gate over the saved plan (absent/error/fail all REFUSE, no ack escape) -> `apply -lock-timeout=120s planfile` (exit≠0 partial-apply is a COMPUTED FAIL, not a crash) -> `show -json > state.json` (coverage) -> `plan -out=replan` -> `show -json replan` -> assert convergence (exit-0 AND an empty parsed resource-change set, via the shipped `foundry_plan_model.parse_actions_detail`) -> an EXPLICIT captured `tofu destroy` (its real exit sets the recorded `teardown` status; runs on GREEN, convergence-FAIL, AND apply-FAIL) -> `emit_infra_walk_evidence` on every computed outcome. Composes with — never re-builds — the shipped posture gate + SoD execute/generate split (`foundry_id_apply`): the sandboxed apply IS the EXECUTE branch bound to the sandbox account; prod always stays `GENERATE_RUNBOOK`. ADVISORY craft procedure FOR the trusted operator — it does NOT gate, approve, or merge; the merge floor (the adopter''s branch protection + CI, docs/merge-floor.md) is the merge authority. CURRENTLY DORMANT (named honestly): the `decide_sandbox_apply`/`emit_infra_walk_evidence` primitives this procedure drives, and its drop-in doctor selftest, were not shipped (no surviving implementation) in scripts/ — this SKILL.md is design intent for a re-implementation against the current floor (ci.yml + btb-gates), not a live procedure today.
About this listing
Category
Engineering
Reported license
MIT
Listing language (estimated)
English
Review status
Not individually reviewed
Agent compatibility
Atlas installation supports Claude Code, Codex, and Cursor. Skill-specific dependencies and features are checked during setup.
Claude Code
Codex
Cursor
GitHub Copilot
Gemini CLI