fix: make MCP-unique instructions config-agnostic + separate artifact/direct workflows

- Replace "The local /workspace/ directory contains all repositories" with
  "Your ecosystem includes the following repositories" across all 20
  MCP-unique task instructions — removes false claims about local repos
  that contradict the SG_full config's empty workspace
- Remove "MCP-only repositories" framing, "(available locally)" qualifiers,
  "Accessible via Sourcegraph MCP" local/MCP splits, and local path mapping
  notes that referenced /workspace/ paths
- Make V5 preamble workflow steps conditional: direct configs get "Edit
  locally + Verify locally", artifact configs get "Produce artifacts"
- Split system prompt: artifact_full no longer says "Run tests locally"
- Remove runtime find-delete hack from artifact_full agent startup — config
  layer now uses Dockerfile.sg_only (empty workspace) for remote+artifact
- Flip sdlc_suite_2config.sh remote+artifact Dockerfile preference: prefer
  Dockerfile.sg_only over Dockerfile.artifact_only (no repo clone needed)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: sjarmak

agents/claude_baseline_agent.py

@@ -105,6 +105,7 @@
 # No truncation language — local source files simply aren't present.
 # 25% shorter than V4: removes Workflows, Output Formatting, Common Mistakes, Query Patterns.
 # {repo_scope} is replaced at runtime with the target repository filter.
+# {workflow_tail} is replaced with edit+test steps (direct) or produce-artifact step (artifact_full).
 V5_PREAMBLE_TEMPLATE = """# IMPORTANT: Source Code Access
 
 **Local source files are not present.** Your workspace does not contain source code. You **MUST** use Sourcegraph MCP tools to discover, read, and understand code before making any changes.
@@ -115,8 +116,7 @@
 
 1. **Search first** — Use MCP tools to find relevant files and understand existing patterns
 2. **Read remotely** — Use `sg_read_file` to read full file contents from Sourcegraph
-3. **Edit locally** — Use Edit, Write, and Bash to create or modify files in your working directory
-4. **Verify locally** — Run tests with Bash to check your changes
+{workflow_tail}
 
 ## Tool Selection
 
@@ -477,22 +477,28 @@ def create_run_agent_commands(self, instruction: str) -> list[ExecInput]:
                     "before searching.\n"
                 )
 
-            mcp_preamble = V5_PREAMBLE_TEMPLATE.format(repo_scope=repo_scope)
-            instruction = mcp_preamble + instruction
-
-            # Artifact-full: append guidance about expressing changes as diffs
+            # Workflow steps 3-4 vary by config: direct configs edit+test
+            # locally, artifact configs produce diffs as output artifacts.
             if mcp_type == "artifact_full":
-                artifact_guidance = """
-
-## Artifact-Only Evaluation
+                workflow_tail = (
+                    "3. **Produce artifacts** — Express all code changes as "
+                    "**unified diffs** in your output artifact (e.g., "
+                    "`fix_patch` fields in review.json, or a standalone "
+                    "`solution.patch` file). Do NOT edit source files directly "
+                    "— there are none in your workspace."
+                )
+            else:
+                workflow_tail = (
+                    "3. **Edit locally** — Use Edit, Write, and Bash to "
+                    "create or modify files in your working directory\n"
+                    "4. **Verify locally** — Run tests with Bash to check "
+                    "your changes"
+                )
 
-You are in **artifact-only mode**. Your workspace is empty — all code discovery
-must go through Sourcegraph MCP tools. Express all code changes as **unified
-diffs** in your output artifact (e.g., `fix_patch` fields in review.json, or a
-standalone `solution.patch` file). Do NOT attempt to edit source files directly
-— there are no source files in your workspace.
-"""
-                instruction = instruction + artifact_guidance
+            mcp_preamble = V5_PREAMBLE_TEMPLATE.format(
+                repo_scope=repo_scope, workflow_tail=workflow_tail
+            )
+            instruction = mcp_preamble + instruction
 
         elif mcp_type == "sourcegraph_isolated":
             # Isolated mode: agent has only the target package locally (via sparse checkout).
@@ -633,7 +639,12 @@ def create_run_agent_commands(self, instruction: str) -> list[ExecInput]:
             else:
                 repo_filter_system = "Use list_repos to discover available repositories first."
 
-            mcp_system_prompt = f"""IMPORTANT: Local source files are not present. You MUST use Sourcegraph MCP tools to discover and read code, then create or edit local files based on what you learn. Run tests locally to verify your changes.
+            if mcp_type == "artifact_full":
+                mcp_system_prompt = f"""IMPORTANT: Local source files are not present. You MUST use Sourcegraph MCP tools to discover and read code, then express your changes as unified diffs in your output artifact.
+
+{repo_filter_system}"""
+            else:
+                mcp_system_prompt = f"""IMPORTANT: Local source files are not present. You MUST use Sourcegraph MCP tools to discover and read code, then create or edit local files based on what you learn. Run tests locally to verify your changes.
 
 {repo_filter_system}"""
             system_prompt_append = EVALUATION_CONTEXT_PROMPT + "\n\n---\n\n" + mcp_system_prompt
@@ -847,22 +858,9 @@ def create_run_agent_commands(self, instruction: str) -> list[ExecInput]:
                     'cd "$WORKDIR"',
                 ]
 
-                # For artifact_full: delete source files so agent must use MCP.
-                # Baseline keeps source readable (mcp_type == "none").
-                # We delete rather than truncate to avoid agents wasting tokens
-                # trying to read visible-but-empty files.
-                if mcp_type == "artifact_full":
-                    script_lines.extend([
-                        '# Delete source files — agent must use MCP to read code',
-                        'find /workspace -type f \\( '
-                        '-name "*.cs" -o -name "*.py" -o -name "*.ts" -o -name "*.tsx" '
-                        '-o -name "*.js" -o -name "*.jsx" -o -name "*.go" -o -name "*.rs" '
-                        '-o -name "*.java" -o -name "*.c" -o -name "*.h" -o -name "*.cpp" '
-                        '-o -name "*.rb" -o -name "*.php" -o -name "*.swift" -o -name "*.kt" '
-                        '-o -name "*.scala" -o -name "*.vue" -o -name "*.svelte" '
-                        '\\) -delete',
-                        'echo "Source files deleted for artifact_full mode"',
-                    ])
+                # Note: artifact_full no longer needs runtime source deletion.
+                # The config layer now uses Dockerfile.sg_only (empty workspace)
+                # for remote+artifact, so there are no source files to delete.
 
                 # If system prompt exists, read it from file and pass via --append-system-prompt
                 if _system_prompt_content:

benchmarks/ccb_mcp_compliance/ccx-compliance-051/instruction.md

@@ -20,8 +20,8 @@ You are performing a compliance audit of the Prometheus monitoring stack. The go
 
 ## Available Resources
 
-The local `/workspace/` directory contains:
-- `prometheus/prometheus` at v3.2.1 → `/workspace/prometheus`
+Your ecosystem includes the following repositories:
+- `prometheus/prometheus` at v3.2.1
 
 ## Output Format
 

benchmarks/ccb_mcp_compliance/ccx-compliance-057-ds/instruction.md

@@ -34,9 +34,9 @@ these 4 layers:
 
 ## Available Resources
 
-The local `/workspace/` directory contains all repositories:
-- `grafana/grafana` at v11.4.0 → `/workspace/grafana`
-- `grafana/loki` at v3.3.4 → `/workspace/loki`
+Your ecosystem includes the following repositories:
+- `grafana/grafana` at v11.4.0
+- `grafana/loki` at v3.3.4
 
 ## Output Format
 

benchmarks/ccb_mcp_crossorg/ccx-crossorg-061/instruction.md

@@ -19,16 +19,16 @@ compile time that a type implements an interface. Finding all such declarations
 repos from different organizations reveals who has independently implemented the same
 storage abstraction — a key signal for platform compatibility audits.
 
-The search should be **exhaustive across all repos in the ecosystem**, not just the
-local repo. The interface is defined in the Kubernetes ecosystem but can be implemented
+The search should be **exhaustive across all repos in the ecosystem**, not just a
+single repo. The interface is defined in the Kubernetes ecosystem but can be implemented
 by projects from entirely different organizations.
 
 ## Available Resources
 
-The local `/workspace/` directory contains all repositories:
-- `kubernetes/kubernetes` at v1.32.0 → `/workspace/kubernetes`
-- `etcd-io/etcd` at v3.5.17 → `/workspace/etcd`
-- `grafana/grafana` at v11.4.0 → `/workspace/grafana`
+Your ecosystem includes the following repositories:
+- `kubernetes/kubernetes` at v1.32.0
+- `etcd-io/etcd` at v3.5.17
+- `grafana/grafana` at v11.4.0
 
 ## Output Format
 

benchmarks/ccb_mcp_crossorg/ccx-crossorg-066/instruction.md

@@ -19,17 +19,17 @@ In Go, every module has a canonical `go.mod` file that declares the module path
 authoritative source. Other repos may vendor or depend on it but are NOT the source
 of truth.
 
-The `kubernetes/kubernetes` repo (available locally) vendors this module — you can
+The `kubernetes/kubernetes` repo vendors this module — you can
 see this at `vendor/go.etcd.io/etcd/client/v3/`. However, this is a vendored copy,
 not the authoritative source. Your task is to find where this module is authoritatively
 maintained.
 
 ## Available Resources
 
-The local `/workspace/` directory contains all repositories:
-- `kubernetes/kubernetes` at v1.32.0 → `/workspace/kubernetes`
-- `etcd-io/etcd` at v3.5.17 → `/workspace/etcd`
-- `grafana/grafana` at v11.4.0 → `/workspace/grafana`
+Your ecosystem includes the following repositories:
+- `kubernetes/kubernetes` at v1.32.0
+- `etcd-io/etcd` at v3.5.17
+- `grafana/grafana` at v11.4.0
 
 ## Output Format
 

benchmarks/ccb_mcp_crossrepo_tracing/ccx-config-trace-010/instruction.md

@@ -10,7 +10,7 @@ k8s.io/client-go/rest.(*Config).DeepCopyInto(...)
         vendor/k8s.io/client-go/rest/config.go:87
 ```
 
-The developer only has access to the main `kubernetes/kubernetes` repository locally.
+The developer is starting from the main `kubernetes/kubernetes` repository.
 They need to find where `rest.Config` is actually defined (the authoritative source),
 not just a vendored copy.
 
@@ -26,11 +26,11 @@ directories, but the authoritative source lives in separate repositories accessi
 
 ## Available Resources
 
-The local `/workspace/` directory contains all repositories:
-- `kubernetes/kubernetes` at v1.32.0 → `/workspace/kubernetes`
-- `kubernetes/client-go` at v0.32.0 → `/workspace/client-go`
-- `kubernetes/api` at fa23dd3 → `/workspace/api`
-- `etcd-io/etcd` at v3.5.17 → `/workspace/etcd`
+Your ecosystem includes the following repositories:
+- `kubernetes/kubernetes` at v1.32.0
+- `kubernetes/client-go` at v0.32.0
+- `kubernetes/api` at fa23dd3
+- `etcd-io/etcd` at v3.5.17
 
 ## Output Format
 
@@ -45,7 +45,7 @@ Create a file at `/workspace/answer.json` with your findings in the following st
 }
 ```
 
-**Important**: The local `/workspace/client-go` directory contains the `kubernetes/client-go` source, but in Sourcegraph it is indexed as `sg-benchmarks/kubernetes-client-go`. Use `sg-benchmarks/kubernetes-client-go` as the `repo` value in your answer — the oracle checks for this exact identifier.
+**Important**: The `kubernetes/client-go` repository is indexed in Sourcegraph as `sg-benchmarks/kubernetes-client-go`. Use `sg-benchmarks/kubernetes-client-go` as the `repo` value in your answer — the oracle checks for this exact identifier.
 **Note**: Sourcegraph MCP tools return repo names with a `github.com/` prefix (e.g., `github.com/sg-benchmarks/kubernetes-client-go`). Strip this prefix in your answer — use `sg-benchmarks/kubernetes-client-go`, NOT `github.com/sg-benchmarks/kubernetes-client-go`.
 
 Your answer is evaluated against a closed-world oracle — the exact repo, path, and symbol name matter.

benchmarks/ccb_mcp_crossrepo_tracing/ccx-dep-trace-001/instruction.md

@@ -20,9 +20,9 @@ that directly imports it (not just subpackages) will be affected by a breaking A
 
 ## Available Resources
 
-The local `/workspace/` directory contains all repositories:
-- `kubernetes/kubernetes` at v1.32.0 → `/workspace/kubernetes`
-- `kubernetes/client-go` at v0.32.0 → `/workspace/client-go`
+Your ecosystem includes the following repositories:
+- `kubernetes/kubernetes` at v1.32.0
+- `kubernetes/client-go` at v0.32.0
 
 ## Output Format
 
@@ -37,7 +37,7 @@ Create a file at `/workspace/answer.json` with your findings in the following st
 }
 ```
 
-**Important**: Use `"repo": "sg-benchmarks/kubernetes-client-go"` exactly — this is the canonical repo identifier used by the evaluation oracle. The local checkout at `/workspace/client-go` corresponds to this repo.
+**Important**: Use `"repo": "sg-benchmarks/kubernetes-client-go"` exactly — this is the canonical repo identifier used by the evaluation oracle. The `kubernetes/client-go` repository corresponds to `sg-benchmarks/kubernetes-client-go` in Sourcegraph.
 **Note**: Sourcegraph MCP tools return repo names with a `github.com/` prefix (e.g., `github.com/sg-benchmarks/kubernetes-client-go`). Strip this prefix in your answer — use `sg-benchmarks/kubernetes-client-go`, NOT `github.com/sg-benchmarks/kubernetes-client-go`.
 
 Include only the `files` field. Your answer is evaluated against a closed-world oracle — completeness matters.

benchmarks/ccb_mcp_crossrepo_tracing/ccx-dep-trace-004/instruction.md

@@ -22,9 +22,9 @@ adding observability, or extending the query pipeline.
 
 ## Available Resources
 
-The local `/workspace/` directory contains all repositories:
-- `grafana/grafana` at v11.4.0 → `/workspace/grafana`
-- `grafana/loki` at v3.3.4 → `/workspace/loki`
+Your ecosystem includes the following repositories:
+- `grafana/grafana` at v11.4.0
+- `grafana/loki` at v3.3.4
 
 ## Output Format
 
@@ -44,7 +44,7 @@ Create a file at `/workspace/answer.json` with your findings in the following st
 - For Loki: `"repo": "sg-benchmarks/grafana-loki"`
 **Note**: Sourcegraph MCP tools return repo names with a `github.com/` prefix (e.g., `github.com/sg-benchmarks/kubernetes-client-go`). Strip this prefix in your answer — use `sg-benchmarks/kubernetes-client-go`, NOT `github.com/sg-benchmarks/kubernetes-client-go`.
 
-The local checkout at `/workspace/loki` corresponds to `sg-benchmarks/grafana-loki`.
+The `grafana/loki` repository corresponds to `sg-benchmarks/grafana-loki` in Sourcegraph.
 
 List the chain steps in order from Grafana (caller) to Loki (callee). Your answer is evaluated
 against a closed-world oracle — precision matters.

benchmarks/ccb_mcp_incident/ccx-incident-031/instruction.md

@@ -24,7 +24,7 @@ Specifically, find:
 
 ## Important: Avoiding Decoys
 
-The local `kubernetes/kubernetes` checkout at `/workspace/kubernetes` contains
+The `kubernetes/kubernetes` repository contains
 vendored copies of etcd code under `vendor/go.etcd.io/etcd/`. These vendored
 files look identical to the real source but are **not** the authoritative location.
 
@@ -33,14 +33,14 @@ The Kubernetes apiserver also has its own error-mapping layer at
 the etcd error into Kubernetes error types — this is also **not** the authoritative
 source of the original error.
 
-Your answer must cite the **upstream etcd repository** (accessible via Sourcegraph
-MCP tools), not the vendored copies or the Kubernetes error-mapping layer.
+Your answer must cite the **upstream etcd repository**, not the vendored copies
+or the Kubernetes error-mapping layer.
 
 ## Available Resources
 
-The local `/workspace/` directory contains all repositories:
-- `kubernetes/kubernetes` at v1.32.0 → `/workspace/kubernetes`
-- `etcd-io/etcd` at v3.5.17 → `/workspace/etcd` (this is where the error originates)
+Your ecosystem includes the following repositories:
+- `kubernetes/kubernetes` at v1.32.0
+- `etcd-io/etcd` at v3.5.17 (this is where the error originates)
 
 ## Output Format
 

benchmarks/ccb_mcp_incident/ccx-incident-034/instruction.md

@@ -22,9 +22,9 @@ definition files. Also include the HTTP fake/stub client used for testing.
 
 ## Available Resources
 
-The local `/workspace/` directory contains all repositories:
-- `grafana/grafana` at v11.4.0 → `/workspace/grafana`
-- `grafana/loki` at v3.3.4 → `/workspace/loki`
+Your ecosystem includes the following repositories:
+- `grafana/grafana` at v11.4.0
+- `grafana/loki` at v3.3.4
 
 ## Output Format