Repository navigation
Conversation
Signed-off-by: frank.scholter_peres <frank.scholter_peres@mercedes-benz.com> Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
…ed lib Signed-off-by: frank.scholter_peres <frank.scholter_peres@mercedes-benz.com> Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Signed-off-by: frank.scholter_peres <frank.scholter_peres@mercedes-benz.com> Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Signed-off-by: frank.scholter_peres <frank.scholter_peres@mercedes-benz.com> Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Signed-off-by: frank.scholter_peres <frank.scholter_peres@mercedes-benz.com> Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Signed-off-by: frank.scholter_peres <frank.scholter_peres@mercedes-benz.com> Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
FScholPer
marked this pull request as draft
July 23, 2026 13:15
FScholPer
marked this pull request as ready for review
July 23, 2026 13:15
FScholPer
marked this pull request as draft
July 23, 2026 13:16
- Remove the agent-observation subsystem (local_observation.py, schema/observation.py, and its tests) to collapse to one mental model. - Persist an append-only experiences.jsonl and one derived weights.json instead of a per-run file tree; gitignore runtime artifacts/. - Key learning on edge classes (relation, source_type, target_type) so signals generalize across repositories instead of memorizing node ids. - Centralize every scoring/learning knob in harness/policy.yml (no magic numbers left in code); load via a validated Policy model. - Add a single score-ctx CLI (run / aggregate / demo). 'demo' shows a deterministic before/after where a decision record moves into the selection and the Lane A gate flips fail -> pass. - Make README self-contained (drop the dead design-doc link). Signed-off-by: frank.scholter_peres <frank.scholter_peres@mercedes-benz.com> Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Packages: - graphify-codegraph: Code structure queries (wraps graphify CLI) - context-discipline: Working memory + local learning (MCP server) Changes: - Simplify README (350→150 lines) for quick start - Create AGENTS.md for AI-generated packages - Create CONTRIBUTING.md for contributors - Create setup-context-discipline wizard - Fix setup-graphify with idempotent .gitignore handling - Fix initialize_session() to accept list + dict assumptions - Remove all context7-docs references (cloud service) - Add validation script for pre-commit checks Features: - Two independent setup wizards (./do setup-<pkg>) - Local-only observations (.score-local/) - Interactive setup with --verify and --help - All SPDX license headers in place Testing: - context-discipline: 6/6 integration tests passed - setup wizards: Tested in fresh repos - validation script: 29/29 checks passed
- Register graphify-codegraph and context-discipline in apm.yml marketplace - Update README with correct apm marketplace add workflow - Show two install paths: GitHub marketplace and local development - Clarify setup wizard usage post-compilation
- Create marketplace.json with package metadata - Register graphify-codegraph, context-discipline, context packages - Update README with correct marketplace add URL (https://) - Enable APM CLI to discover and install packages
APM CLI checks multiple paths: - marketplace.json (root) - .github/plugin/marketplace.json - .claude-plugin/marketplace.json Using -f to override gitignore for APM discovery.
APM should use the root marketplace.json file. Simplified configuration.
- Remove CONTRIBUTING.md - Update README references to point to CONTRIBUTION.md - Standardized naming convention
APM checks these locations: - marketplace.json (root) - .github/plugin/marketplace.json - .claude-plugin/marketplace.json All three must be present for APM to discover packages.
Changed name from 'Eclipse S-CORE APM Packages' to 'eclipse-score-packages' APM requires marketplace names to match [a-zA-Z0-9._-]+ pattern.
Use #local_harness_apm ref to specify the branch. APM now successfully discovers packages.
APM finds marketplace.json in the repository root. No need for duplicate copies.
- Show simple marketplace add command - Keep local path option for development - Removed confusing #ref pinning explanation
Removed packages array - APM finds packages from: - packages/*/apm.yml files in the repository - Not from marketplace.json package declarations This matches APM's expected marketplace discovery mechanism.
- Use package@marketplace syntax - Show marketplace name check command - APM discovers packages from apm.yml files in package directories
Changed name from 'Eclipse S-CORE' to 'eclipse-score' APM validates marketplace names with [a-zA-Z0-9._-]+
- Add packages array to marketplace.json with apm paths - Update README to use correct marketplace name 'mcp-servers' - Ensures APM can discover and browse all 3 packages - Install command: apm install graphify-codegraph@mcp-servers context-discipline@mcp-servers
rge branch 'local_harness_apm' of https://github.com/eclipse-score/mcp-servers into local_harness_apm
FScholPer
added a commit
that referenced
this pull request
Oct 21, 2026
Fixes all 4 issues from PR #49 review: Issue #1 (P1 - HIGH): Windows platform compatibility - Move fcntl import to try/except with graceful fallback - Add platform-compatible locking for Windows support - Update build() to handle missing fcntl module Issue #2 (P2 - MEDIUM): Extraction output directory - Clear GRAPHIFY_OUT env var in graphify subprocess - Prevents interference with --out flag - Ensures correct output location Issue #3 (P2 - MEDIUM): Cache path sharing with context-discipline - Add resolve_graph_path() public function - Update MergedGraph.build() to check user cache - Update _graph_setup_status() to check user cache - Both consumers now find cached graphs Issue #4 (P2 - MEDIUM): Registry seeding race condition - Add locking around registry seeding in prepare() - Recheck graph existence after lock acquisition - Prevents concurrent cold starts from corrupting cache Add comprehensive test coverage: - 7 new tests for resolve_graph_path() - Test for GRAPHIFY_OUT env clearing - Test for locking race condition prevention - Total: 19 tests passing Update documentation: - Add resolve_graph_path() usage example to README - Document programmatic access API All code passes: - Ruff linter (clean) - Pyright type checker (0 errors) - Pytest test suite (19/19 passing)
FScholPer
added a commit
that referenced
this pull request
Oct 22, 2026
* Serve graphify from a user cache seeded by central graphs graphify-codegraph now prepares the graph before starting the server: an existing graphify-out in the repository is used as is, otherwise the graph lives under ~/.local/share/score-agent. For eclipse-score repos an empty cache is seeded from ghcr.io/eclipse-score/mcp-servers/graphs (merge-base, then main) with sha256 digest check, and a locked code-only build refreshes it in the background when HEAD moved. Co-Authored-By: frank.scholter_peres <frank.scholter_peres@mercedes-benz.com> * fix(graphify-codegraph): address PR review issues Fixes all 4 issues from PR #49 review: Issue #1 (P1 - HIGH): Windows platform compatibility - Move fcntl import to try/except with graceful fallback - Add platform-compatible locking for Windows support - Update build() to handle missing fcntl module Issue #2 (P2 - MEDIUM): Extraction output directory - Clear GRAPHIFY_OUT env var in graphify subprocess - Prevents interference with --out flag - Ensures correct output location Issue #3 (P2 - MEDIUM): Cache path sharing with context-discipline - Add resolve_graph_path() public function - Update MergedGraph.build() to check user cache - Update _graph_setup_status() to check user cache - Both consumers now find cached graphs Issue #4 (P2 - MEDIUM): Registry seeding race condition - Add locking around registry seeding in prepare() - Recheck graph existence after lock acquisition - Prevents concurrent cold starts from corrupting cache Add comprehensive test coverage: - 7 new tests for resolve_graph_path() - Test for GRAPHIFY_OUT env clearing - Test for locking race condition prevention - Total: 19 tests passing Update documentation: - Add resolve_graph_path() usage example to README - Document programmatic access API All code passes: - Ruff linter (clean) - Pyright type checker (0 errors) - Pytest test suite (19/19 passing) * fix: resolve CI formatting and type checking issues - Apply ruff formatting to 4 Python files - Fix unused variable warnings (document, graph1, graph2) - Add type annotation for seeded list * fix: prioritize in-repo graph over cache and report correct path When no graph exists, report the in-repo path (graphify-out/graph.json) instead of the user cache path. This matches user expectations and the graphify-codegraph prepare() behavior. Fixes test failures in test_initialize_session_reports_missing_graph_setup and test_query_graph_reports_missing_graph_setup.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Phase 1 walking skeleton for the Agent Context Attention Layer. Makes the contract and harness ideas concrete and runnable end-to-end, kept minimal but real (no stubs).
Stacked on the Phase 0 branch (
context-attention-phase0-schema) so this diff is Phase 1 only.The thread it demonstrates
A task seeds a single node (
pr__mcp_servers__42);get_contextreaches the affectedcontract(viaaffects) and the relevantdec_rec(viadiscussed_in) by traversing typed edges — nothing hardcoded. The emptyBaselineHarnessfails the gate; theContextHarnesscandidate passes. That contrast is the point.Contract (stable operating surface)
eclipse-score/docs-as-code@4bc0fbf(docs-as-code#628):repo-manifest.schema.json,agent-card.schema.json,adapter_contract_v0_1.{md,schema.json}. Provenance recorded in.github/references/README.mdandharness/contract/README.md..github/score/repo-manifest.json(python; builduv sync, testuv run pytest, lintuv run ruff check ., typecheckuv run pyright;mcp.tools: [search]) and.github/score/agent-card.json, both validated against the vendored schemas in tests.Harness (deterministic evaluation seam)
AssuranceHarness.get_context(task_spec) -> strmirrored from upstream;BaselineHarness(returns"") +ContextHarness(calls the engine).adapter.pyimplements thevalidate | run | reportenvelope fromadapter_contract_v0.1: issue-first (issue_id), artifacts under.stage/ISSUE-<n>/…,iso26262profile, full error-code set. CLI:uv run python -m score_context.harness harness/spec/request_run.json.Engine slice (
libs/score-context, still MCP-free)graph.py: fragment loader + deterministic compose with natural-key conflict detection, on the Phase 0Node/Edgemodels.context.py:get_context(task, repo, role, top_n)returns structured output (selected / decisions / contracts / repositories / blocking_tasks) + rendered text. Scoring = seeds + relation weights + neighbor degree + provenance freshness. Explicitly marked a first-cut placeholder for personalized PageRank; the seam and output shape are stable, the ranking math is not.APM wrap (first real example)
packages/context/.apm/context/attention-task-001.mdregenerated from a real candidate run;packages/context/.apm/skills/query-context/SKILL.mdis now a real skill.apm pack --offlinestaysok: true.Impact Analysis
.stage/and generated APM output are gitignored.Validation
reuse lintcompliant (54/54 files)uv run ruff check ./ruff format --check .cleanuv run pyright— 0 errorsuv run pytest— 9 passedapm pack --offline --json—ok: truepass, baseline verdictfail; artifact written to.stage/ISSUE-1/harness/run.jsonactionlintwas unavailable in the environment and skipped.