ARC-1 Cursor Regression Skill (Generic + Adaptive)

This skill creates tailored Cursor setup and prompts based on:

1. PR scope (preferred)
2. Current chat findings
3. Changed files in the local branch

It must not output a hardcoded one-scenario checklist unless the user explicitly asks for that.

Trigger conditions

Use this skill when user asks for:

• Cursor MCP setup for ARC-1
• Regression prompts
• “verify this PR/fix/feature” prompts
• “create config + prompts so I can run quickly”
• “run the env prep for me”

Composite request shortcut (must support)

If the user asks in one message to:

• use this skill
• derive tests from chat + current branch diff
• generate config + env prep + one execution prompt
• and says “run the env prep for me”

then treat it as a single composite workflow:

1. Resolve target root/worktree/branch.
2. Derive module set from diff + chat.
3. Generate config + one copy/paste execution prompt.
4. Execute env prep commands immediately in the resolved root.
5. Report env prep execution status (success/failure + minimal fix if failed).

Required discovery workflow

0) Resolve active ARC-1 root + branch/worktree (mandatory)

Always bind generated config/scripts/prompts to the actual target repo root (not a remembered path).

Run:

git rev-parse --show-toplevel
git branch --show-current
git worktree list --porcelain

Rules:

• If user gave a PR/branch, ensure the resolved root/branch matches that target.
• If multiple ARC-1 clones/worktrees exist, explicitly choose the one containing the target branch/commit.
• Never hardcode /Users/.../DEV/arc-1 unless user explicitly requests that exact path.

0.5) Confirm active Cursor project/workspace matches target root (mandatory for runtime runs)

Before generating runtime prompts, ensure the Cursor chat session is attached to the same project/workspace that will host MCP servers and tool cache for the resolved root.

Rules:

• If the active Cursor project path differs from the resolved ARC-1 root, classify immediately as Environment/session setup issue (not code regression) and provide a fix: open the correct workspace and reconnect MCP.
• Do not mix artifacts from different Cursor projects when evaluating server availability/schema.

1) Determine target scope

Use one of:

• PR URL/number from user
• explicit branch from user
• current branch diff

Preferred commands:

gh pr view <PR> --json number,title,body,headRefName,baseRefName,url
gh pr diff <PR> --name-only

Fallback:

git diff --name-only origin/main...HEAD

2) Capture chat-derived constraints

Extract blockers from chat and turn them into guardrails (examples):

• MCP server not connected
• missing env file
• wrong SAPQuery arg key (query vs sql)
• backend SQL grammar caveats
• safety/profile mismatch
• stale MCP tool descriptor / wrong runtime root

3) Build module set from file changes

Map changed files to test modules:

• src/server/server.ts, auth/preflight, startup behavior
• module_auth_preflight
• src/handlers/schemas.ts, TABLE_CONTENTS, validation
• module_schema_validation
• src/handlers/dispatch.ts + per-tool handlers, hints/scope/routing/diagnose actions
• module_handler_behavior
• src/adt/diagnostics.ts, dump/gateway/system message parsing
• module_diagnostics_runtime
• src/adt/http.ts login/html detection changes
• module_http_html_detection
• src/handlers/hyperfocused.ts, src/handlers/tools.ts
• module_tool_visibility
• test/docs-only changes
• module_static_verification

Always include module_connectivity_precheck unless user requests static-only.

Cursor setup strategy

Default: command-mode MCP servers (most reliable in Cursor). Use URL mode only if user explicitly asks.

Generate only needed profiles for selected modules:

• arc1-good for positive behavior
• arc1-bad for auth-negative checks (only if auth/preflight module selected)
• optional profiles like arc1-good-blockdata, arc1-good-readonly, arc1-good-hyperfocused depending on module set

Path/worktree requirements for generated config:

• MCP command/script must resolve to the same root used for PR diff analysis.
• If scripts are generated under <root>/.cursor/scripts, they must compute root relative to script location, with ARC1_ROOT override.
• Mention the resolved root in output so user can quickly verify it before running tests.

Readable snapshot workflow for filtered worktrees

Use this when Cursor cannot read the hidden PR worktree because the active workspace or .cursorignore filters paths such as .claude/worktrees/....

Rules:

• Create a readable static snapshot under the active Cursor workspace, for example

<cursor-workspace>/arc1-pr<nr>-visible, while keeping the MCP runtime pointed at the real PR worktree.

• Do not copy secrets, .env, caches, cookies, node_modules, or runtime folders

into the static snapshot.

• Write .arc1-pr<nr>-snapshot.json in the snapshot with source, visibleRoot,

head, and createdAt so Cursor can report exactly what it reviewed.

• Generated prompts must distinguish staticRoot from runtimeRoot. Do not treat

a different active Cursor workspace as a regression when the prompt explicitly uses a readable snapshot plus a separate runtime root.

• If the MCP config lives inside the snapshot, say that directly; do not claim it

is at the active workspace root.

• Treat generated .cursor/ configs, prompts, scripts, and runtime files as local

operator artifacts. Do not commit them to ARC-1 unless the user explicitly asks for a reusable repository template with scrubbed, portable paths.

Env prep rules

• Build first (npm run build) unless user says skip.
• Build in the resolved target root (from discovery step 0).
• Never print secrets.
• Do not source full infra files blindly.
• Never ask users to paste a raw script body directly into an interactive shell.
• Default env file for generated scripts must be ${ARC1_ROOT}/.env (or ${ROOT}/.env after root resolution).
• Allow user override via ARC1_ENV_FILE.
• Generated scripts must fail fast if root/env/build artifacts are missing or required keys are empty.
• Generated scripts may accept ARC1_EXPECT_DIST_TEXT; when set, fail startup if

the expected text is missing from dist/handlers/dispatch.js. This catches stale MCP server builds before Cursor runs a long live regression.

• Include an ARC1_RUNTIME_FINGERPRINT env value in generated MCP profiles for

branch-specific runs, and tell the tester to restart/toggle MCP servers after changing config or rebuilding.

• When the target branch uses positive safety flags (SAP_ALLOW_*) instead of

older block flags, generated profiles must set the positive flags explicitly and unset legacy SAP_READ_ONLY / SAP_BLOCK_* / ARC1_PROFILE values so a root .env cannot silently override the intended test profile.

• Generated script examples must be executed as files (bash /path/to/script.sh or executable path), not sourced.
• Parse only required keys (and strip surrounding quotes / CRLF):

grep '^KEY=' <file> | head -n1 | cut -d= -f2-

Canonical Cursor script pattern

When generating arc1-good.sh / arc1-good-btp-sim.sh, use this pattern:

#!/usr/bin/env bash
if [[ "${BASH_SOURCE[0]}" != "$0" ]]; then
  echo "Do not source this script. Execute it: bash ${BASH_SOURCE[0]}"
  return 1 2>/dev/null || exit 1
fi

set -euo pipefail

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
ROOT="${ARC1_ROOT:-}"

if [[ -z "$ROOT" ]]; then
  if [[ -f "$SCRIPT_DIR/../../package.json" ]]; then
    ROOT="$(cd "$SCRIPT_DIR/../.." && pwd)"
  else
    ROOT="$(git rev-parse --show-toplevel 2>/dev/null || true)"
  fi
fi

[[ -n "$ROOT" ]] || { echo "Unable to resolve ARC-1 root. Set ARC1_ROOT=/absolute/path/to/arc-1"; exit 1; }
[[ -f "$ROOT/package.json" ]] || { echo "Invalid ARC1_ROOT: $ROOT (package.json missing)"; exit 1; }
[[ -f "$ROOT/dist/index.js" ]] || { echo "Missing build artifact: $ROOT/dist/index.js. Run: (cd \"$ROOT\" && npm run build)"; exit 1; }
if [[ -n "${ARC1_EXPECT_DIST_TEXT:-}" ]] && ! grep -Fq "$ARC1_EXPECT_DIST_TEXT" "$ROOT/dist/handlers/dispatch.js"; then
  echo "Built dist is stale: expected text not found in dist/handlers/dispatch.js. Run: (cd \"$ROOT\" && npm run build)" >&2
  exit 1
fi

ENV_FILE="${ARC1_ENV_FILE:-$ROOT/.env}"
[[ -f "$ENV_FILE" ]] || { echo "Missing env file: $ENV_FILE"; exit 1; }

getv() {
  local key="$1"
  local raw
  raw="$(grep "^${key}=" "$ENV_FILE" | head -n1 | cut -d= -f2- || true)"
  printf '%s' "$raw" | tr -d '\r' | sed -E "s/^[[:space:]]+|[[:space:]]+$//g; s/^['\"]//; s/['\"]$//"
}

export SAP_URL="${SAP_URL:-$(getv SAP_URL)}"
export SAP_USER="${SAP_USER:-$(getv SAP_USER)}"
export SAP_PASSWORD="${SAP_PASSWORD:-$(getv SAP_PASSWORD)}"
export SAP_CLIENT="${SAP_CLIENT:-$(getv SAP_CLIENT)}"
export SAP_LANGUAGE="${SAP_LANGUAGE:-EN}"
# Safe defaults — no writes, no data preview, no SQL. Uncomment the SAP_ALLOW_* lines
# below for developer-level access. See docs_page/authorization.md#recipes.
# export SAP_ALLOW_WRITES=true
# export SAP_ALLOW_DATA_PREVIEW=true
# export SAP_ALLOW_FREE_SQL=true
export SAP_TRANSPORT=stdio

[[ -n "${SAP_URL}" ]] || { echo "Missing SAP_URL in $ENV_FILE"; exit 1; }
[[ -n "${SAP_USER}" ]] || { echo "Missing SAP_USER in $ENV_FILE"; exit 1; }
[[ -n "${SAP_PASSWORD}" ]] || { echo "Missing SAP_PASSWORD in $ENV_FILE"; exit 1; }
node -e 'new URL(process.argv[1])' "$SAP_URL" >/dev/null || { echo "Invalid SAP_URL: [$SAP_URL]"; exit 1; }

unset SAP_READ_ONLY SAP_BLOCK_DATA SAP_BLOCK_FREE_SQL SAP_ENABLE_TRANSPORTS SAP_ENABLE_GIT
unset SAP_ALLOWED_OPS SAP_DISALLOWED_OPS ARC1_PROFILE ARC1_API_KEY

exec node "$ROOT/dist/index.js"

For arc1-good-btp-sim.sh, add:

export SAP_SYSTEM_TYPE=btp

Execution modes (must support both)

When generating runnable instructions, always provide both:

1) Shell mode (user-run):

• Safe one-time setup should be emitted as full commands (including heredoc wrappers), not raw script body fragments.
• If shell options like set -u are used in setup snippets, wrap setup in a subshell:
• ( set -euo pipefail; ... )
• Instruct users to run generated scripts as executables or via bash, never source.

2) Codex mode (agent-run):

• Provide a compact command block Codex can run directly (cd <resolved-root> && ...).
• Keep the same root/worktree guarantees as shell mode.
• Never assume Codex is running in a specific clone unless discovery confirmed it.

Execution decision rule

• If user explicitly says “run the env prep for me” (or equivalent), execute env prep now.
• Otherwise, provide env prep commands only (do not execute).
• When executed, always:
• run in resolved target root
• create/refresh .cursor/scripts/arc1-good.sh and .cursor/scripts/arc1-good-btp-sim.sh
• run build unless user explicitly asked to skip
• run bash -n syntax validation on generated scripts
• return prepared paths + default env file path used

Prompt generation contract

When asked for runnable output, always return in this order:

1. Why these tests were selected (PR/chat signals)
2. Cursor MCP config snippet (adaptive profiles only)
3. One-time env prep commands
4. One all-in-one test prompt
5. Expected outcomes checklist (PASS/FAIL criteria)

For DDLS runtime checks in generated prompts:

• Use fallback discovery sequence: Z_* -> ZI_* -> I_* with maxResults=100.
• Stop at the first query that returns at least one DDLS/* candidate.
• Do not fail regression just because Z_* has no DDLS.
• Prefer candidates with sibling signal (same stem / numeric variants) so checkedCandidates is more likely non-empty.
• If no sibling-rich candidate exists, still validate clamp/toggle behavior and treat missing sibling comparisons as not applicable, not regression.
• Include backend semantic caveats discovered during test design. Example: on

SAP_BASIS 758, CDS children projecting an amount field such as sflight-price may also need the currency field, otherwise activation fails before the feature under test is reached.

Runtime prompt guardrails (must include)

• Precheck required MCP servers are connected.
• Server ID resolution priority:
• If user provides exact server IDs, use those exact IDs first.
• Use suffix matching (*-arc1-good, *-arc1-good-btp-sim) only as fallback.
• Connectivity decision rule:
• Do not fail solely from filesystem mcps/ scan.
• If a server ID is callable, treat it as connected even when descriptor cache appears stale.
• On missing/disconnected result, label it as "possibly stale session cache", wait 5-10 seconds, and retry once before final classification.
• If any required server is still disconnected/missing after retry: stop and report exact server name.
• If a live write test is required, run a small create/delete write-smoke before

creating multi-object DDLS graphs. On CSRF/core-discovery/unlock/service-routing failure, stop the live mutation scenario and classify as environment/session, not product regression.

• If a write-smoke create fails during post-save unlock but search/read shows the

object exists, report partial-created, attempt at most one cleanup delete, and list leftovers explicitly if cleanup also fails.

• Verify server/tool contract matches expected scope from selected modules.
• If expected actions/params are absent in advertised tool schema, classify as Environment/session setup issue (not code regression) and suggest rebuild/reconnect on the correct root.
• If descriptor schema says fields are missing but live calls accept/behave correctly, classify as Environment/session setup issue (descriptor staleness) rather than code regression.
• For authorization profiles, action/tool hiding can be the expected pass

condition. Example: a read-only server may hide SAPWrite; a deny-action server may hide only the denied action instead of returning an in-band denial.

• Use schema-correct args:
• SAPQuery uses sql + maxRows.
• avoid UP TO ... ROWS in SQL text unless backend-specific test intentionally checks parser rejection.
• Classify final status as one of:
• Implemented fixes confirmed
• Regression found
• Environment/session setup issue (not code regression)
• For SAP component reporting, copy exact SAPRead(type="COMPONENTS") rows.

Do not conflate separate installed components, for example S4FND 108 and MDG_FND 808.

• For where-used evidence, separate container/package rows such as DEVC $TMP

from real DDLS impact rows. Do not count containers as dependent CDS objects.

• Keep evidence snippets tied to the exact tool call under test; do not paste

static source excerpts unless static behavior failed.

• For multi-phase prompts, specify which tool response owns each evidence block

(for example update guidance from SAPWrite(update), activation guidance from SAPActivate(ROOT), delete guidance from the first pre-cleanup root delete).

• If live sections are skipped because write-smoke failed, keep skipped YAML

sections compact with empty arrays and one-line raw snippets.

Diagnostics-specific module guidance (for PRs touching SAPDiagnose)

Include these runtime checks when diagnostics files are touched:

• dumps list (action="dumps", bounded results)
• dump detail by id, including section-aware output if supported
• invalid dump id not-found guidance
• system_messages action
• gateway_errors list and detail retrieval
• gateway detail by id + errorType normalization behavior if applicable
• BTP guardrail behavior if target system is BTP

Anti-patterns

• Hardcoding only auth/sqlFilter tests when PR scope is diagnostics
• Hardcoding absolute repo paths in scripts/config (/Users/.../DEV/arc-1)
• Mixing PR analysis root and runtime server root
• Committing generated .cursor/ MCP configs or prompts with absolute local paths
• Mixing static snapshot paths and runtime roots without naming both explicitly
• Emitting script text in a way that encourages pasting raw script bodies into interactive shells
• Recommending source for runtime server scripts
• Assuming server IDs without checking connected MCP descriptors
• Leaving Cursor MCP servers running after rebuilding dist/ or changing MCP env
• Reporting a stale MCP runtime as a product regression when source/dist already contain the expected behavior
• Treating stale tool schema as code regression without classifying env/setup first
• Failing immediately on first missing/disconnected check without one retry window
• Treating descriptor-cache absence as authoritative when live calls are available
• Falling back to custom HTTP client without user permission when prompt says native-only
• Treating post-save UNLOCK/CSRF/core-discovery failures as DDIC source-save problems
• Counting package/container where-used rows as CDS dependents
• Leaking secret values in logs/output