SDLC_HARNESS

[aryansingh0012]

SDLC_HARNESS

[FScholPer]

looks empty

[FScholPer]

@aryansingh0012 lets discuss next steps in our ai meeting to do not double the work. I started to build artifacts from our current meta model in eclipse-score/docs-as-code#826

[FScholPer]

Reviewed the full diff (29 files, ~1.7k lines) against main. The stage-artifact
model, the loopback log and the evidence recording are useful and they do not collide with
context-discipline: this package produces the lifecycle trail, ours selects context.
Four concrete defects below, each with a suggested change where a mechanical fix exists.

1. CI is red for a mechanical reason — the inlined Python in
apm-marketplace-check.yml raises IndentationError: unexpected indent at
repository = os.environ["MCP_REPO"], so the remote-MCP smoke test never runs. Suggestion inline.

2. mcp.yml advertises a tool that does not exist — trace_need is declared, but
serve.py::call_tool has no branch for it, so calling it returns Unknown tool: trace_need.
Since the sphinx-needs integration was removed in b336a87, I'd drop the declaration (and
the "sphinx-needs traceability" wording in description) rather than re-add the tool. For the
needs trace itself there is now a projection in metamodel-flow/context-discipline
(requirements_projection.py → RequirementsSource, typed nodes with derived_from,
satisfied_by, covers, partially_verifies, fully_verifies), so this is a good seam to
consume instead of re-implementing.

3. The review stage is unreachable — STAGE_REQUIREMENTS["review"] requires
harness/run.json, but the only file any tool writes is
harness/implementation-evidence.json (implementation_tracking.py::EVIDENCE_FILENAME).
check_stage_completeness(stage="review") can therefore never report complete. Suggestions
inline for both the code and the instructions file.

4. No tests — the package ships without any (b56e638, cf4d123 removed them) and
scripts/validate.sh only checks that files exist. For a server whose purpose is gating,
I'd want at least: path-traversal rejection in issue_directory, evidence replacement by
task_id, the status allowlist, and one check_stage_completeness case per stage.

One design question, not a review blocker — better suited to the AI meeting: STAGE_REQUIREMENTS
hardcodes a stage model, and requirement_ids in record_implementation_evidence are
unvalidated free strings. Both exist as data in the metamodel already (process workflows with
input/output/responsible/approved_by; needs with their link types). Deriving the gates
and validating the IDs against that data would remove the duplication we talked about.

[FScholPer]

repository = ... is indented one level too deep and the rewrite loop below it is mis-indented as well, so the heredoc fails to parse before it can rewrite .vscode/mcp.json. argument.replace(argument, ...) is also a whole-string replace, so building the new value directly is clearer.

Suggested change

	revision = os.environ["MCP_REF"]
	repository = os.environ["MCP_REPO"]
	for server in config["servers"].values():
	args = server.get("args", [])
	for index, argument in enumerate(args):
	if argument.startswith("git+https://github.com/eclipse-score/mcp-servers#"):
	if argument.startswith("git+https://github.com/") and "#" in argument:
	subdirectory = argument.split("#", 1)[1]
	args[index] = argument.replace(
	"git+https://github.com/eclipse-score/mcp-servers#",
	f"git+https://github.com/eclipse-score/mcp-servers@{revision}#",
	1,
	argument,
	f"git+https://github.com/{repository}@{revision}#{subdirectory}",
	)
	revision = os.environ["MCP_REF"]
	repository = os.environ["MCP_REPO"]
	prefix = "git+https://github.com/"
	for server in config["servers"].values():
	args = server.get("args", [])
	for index, argument in enumerate(args):
	if argument.startswith(prefix) and "#" in argument:
	subdirectory = argument.split("#", 1)[1]
	args[index] = (
	f"{prefix}{repository}@{revision}#{subdirectory}"
	)

[FScholPer]

trace_need has no handler in serve.py::call_tool, so an agent that follows tools/list gets Unknown tool: trace_need. Dropping the declaration keeps the manifest honest after the sphinx-needs removal in b336a87.

Suggested change

	- name: trace_need
	description: Return a sphinx-needs trace chain, including linked needs and reverse references.
	input_schema:
	type: object
	properties:
	need_id:
	type: string
	needs_json_path:
	type: string
	required: [need_id, needs_json_path]

[FScholPer]

The sphinx-needs part was removed from the implementation, so the description promises something the server no longer does.

Suggested change

	description: MCP server for SDLC lifecycle management and sphinx-needs traceability
	description: MCP server for SDLC lifecycle stages and implementation evidence

[FScholPer]

No tool writes harness/run.json — record_implementation_evidence writes harness/implementation-evidence.json (EVIDENCE_FILENAME). As it stands, check_stage_completeness(issue_id, "review") always reports the review stage as incomplete.

Suggested change

	"harness/run.json",
	"harness/implementation-evidence.json",

[FScholPer]

Same filename mismatch in the guidance: the agent is told to store evidence in harness/run.json, but the tool it is supposed to call writes harness/implementation-evidence.json.

Suggested change

	5. Store actual implementation evidence in `harness/run.json` before review.
	5. Store actual implementation evidence with `record_implementation_evidence` (written to `harness/implementation-evidence.json`) before review.