Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Types of changes: Added, Changed, Deprecated, Removed, Fixed, Security

[Unreleased]

[5.1.0] - 2026-03-08

Added

• .github/workflows/generate-sbom.yaml: auto-generate SPDX 2.3 SBOM on push to main (pyproject.toml/uv.lock changes) and weekly; outputs docs/SBOM/sbom.spdx.json (machine-readable) and docs/SBOM/sbom.md (human-readable markdown table)
• .github/ISSUE_TEMPLATE/bug_report.md, question.md: GitHub issue templates for bug reports and questions
• .github/ISSUE_TEMPLATE/config.yaml: disable blank issues, force template usage
• .github/templates/llms.txt.additions.tpl: project-specific llms.txt additions (Best Practices, Optional links) separate from generic skeleton
• scripts/writeup/README.md: usage guide and pandoc/LaTeX gotchas (moved from AGENT_LEARNINGS)
• AGENT_LEARNINGS.md: patterns for gh pr edit GraphQL workaround, CC sandbox git restriction, PR squash merge API
• .claude/skills/enforcing-doc-hierarchy/: skill for auditing doc hierarchy, detecting broken references and chain breaks
• docs/analysis/ClaudeCode/: 13 CC feature analyses — Remote Control, Cloud Sessions, Memory System, Cowork & Plugins, Chrome Extension, Platform Docs, Plugin Packaging, Ralph Enhancement Research, Sandboxing, Skills Adoption, Changelog Feature Scan, Model/Provider Config, llms.txt
• docs/analysis/CC-adoption-plan.md: consolidated CC adoption plan (supersedes CC-skills-Ralph-adoption-plan.md); moved out of ClaudeCode/ subdirectory

Removed

• docs/analysis/ClaudeCode/: 17 CC feature analyses moved to standalone repo claude-code-research

Changed

• ralph/scripts/ralph.sh: per-story CLAUDE_CODE_EFFORT_LEVEL auto-computation, ad-hoc INSTRUCTION parameter, opt-in de-sloppify pass
• ralph/scripts/ralph.sh: worktree-namespaced /tmp paths via sha256sum-based RALPH_TMP_DIR; all hardcoded /tmp/claude/ralph_* paths replaced
• Makefile: ralph_run and ralph_run_worktree accept INSTRUCTION and DESLOPIFY parameters
• ralph/scripts/generate_prd_json.py: --check-overlaps flag for file-overlap detection between stories
• ralph/docs/templates/prompt.md: skill routing table replacing flat bullet list
• ralph/README.md: Remote Control monitoring tip in Configuration section
• docs/analysis/CC-adoption-plan.md: removed stale “Path-Scoped Rules Implementation” section; moved completed Tier 1 items to Done subsection
• ralph/TODO.md: marked 9 completed items as done (Remote Control, effort level, /tmp namespace, check-overlaps, de-sloppify, test dir consolidation, steering instructions, trigger table)
• ralph/scripts/ralph.sh: set CLAUDE_CODE_DISABLE_GIT_INSTRUCTIONS=1 for headless claude -p invocations
• .claude/rules/agent-patterns.md: new path-scoped rule for agent implementation (triggered by src/app/agents/, .claude/agents/, .claude/skills/)
• .claude/rules/testing.md: new path-scoped rule for test patterns (triggered by tests/)

Fixed

• Makefile: fix ralph_status recipe using legacy .passes field instead of .status == "passed"
• docs/analysis/CC-adoption-plan.md: repoint all CC doc links to claude-code-research repo with correct subdirectory paths; fix AGENTS.md and ralph/TODO.md relative path depth
• src/app/utils/run_context.py: wrap 101-char run_dir line to satisfy ruff E501
• .github/workflows/codeql.yaml: upgrade dismiss-alerts v1 → v2.0.2 to fix EISDIR error on SARIF directory input

Removed

• docs/analysis/ClaudeCode/: 17 CC feature analyses moved to standalone repo claude-code-research
• docs/analysis/RDI-Agentbeats/: moved AgentBeats_basics.md to docs/analysis/
• docs/analysis/CC-skills-Ralph-adoption-plan.md: superseded by ClaudeCode/CC-adoption-plan.md
• docs/analysis/agents-md-refactor-plan.md: consolidated into other docs

Changed

• .github/templates/llms.txt.tpl: stripped to universal skeleton (${PROJECT_NAME}, ${PROJECT_DESC}, ${BLOB} vars only)
• .github/workflows/write-llms-txt.yaml: append additions template, POSIX-safe validate, exclude docs/llms.txt from trigger to prevent CI cycle
• AGENT_LEARNINGS.md: DRY/YAGNI cleanup — removed LaTeX/pandoc one-offs, collapsed -X ours and PR merge API to cross-references
• ralph/docs/LEARNINGS.md: PR merge API block replaced with cross-reference to AGENT_LEARNINGS
• docs/analysis/ClaudeCode/, RDI-Agentbeats/: reorganized CC and AgentBeats analysis into dedicated subdirectories
• scripts/writeup/generate-plantuml-png.sh, run-pandoc.sh: POSIX-hardened (quoting, error handling, LaTeX \makeatletter fix)
• AGENT_LEARNINGS.md: new patterns — shell keyword collision in jq args (SC1010), stale test fixture cross-file pollution
• docs/howtos/peerread-agent-usage.md: slimmed to unique content
• README.md: restructured badge layout and navigation
• .github/workflows/pytest.yaml: pin uv cache to uv.lock hash for precise invalidation

[5.0.0] - 2026-03-03

Added

• scripts/batch_run.py: batch composition runner — runs make app_cli across all 8 agent compositions for one or more papers with --parallel support, error continuation, and optional JSON output
• Makefile: app_batch_run recipe as drop-in alternative to app_sweep using subprocess calls

Fixed

• evaluation_pipeline.py: skip Tier 1 when review is empty or no reference reviews available (prevents false 1.0 and uninformative 0.1 scores)
• evaluation_pipeline.py: cap T1-only composite at composite_weak_reject_threshold (0.4) — incomplete evaluation no longer scores as “perfect”
• evaluation_pipeline.py: route T1-skipped composite through T2+T3 when available instead of raising ValueError
• cc_teams.py: fix stale event type checks (type=="TeamCreate"/"Task" → subtype=="task_started"/"task_completed")
• run_sweep.py: include exception type in sweep error message for actionable diagnostics

Changed

• evaluation_runner.py: DRY paper content loading — renamed _load_cc_paper_content → _load_paper_content and replaced duplicated inline PDF→abstract fallback in _extract_paper_and_review_content with a single call
• Makefile: quiet mode for make validate — per-tool quiet flags (RUFF_QUIET, PYTEST_QUIET, COV_QUIET, CPLX_QUIET) enabled by default; override with make validate VERBOSE=1 for full output. Each recipe prints step indicator with [quiet] when active.

Fixed

• cc_engine.py: normalise CC task_started events to from/to format expected by graph analysis (Gap 5)
• app.py/evaluation_runner.py: thread GraphTraceData from CC artifacts directly to evaluation pipeline, bypassing empty SQLite lookup (Gap 1)
• evaluation_runner.py: load paper content from PeerRead when manager_output is None (CC engine path) so Tier 1/2 see actual paper text (Gap 2)
• app.py: wire cc_model parameter through to evaluation pipeline as chat_model for Tier 2 LLM judge (Gap 4)
• models.py: create_simple_model now routes Anthropic, Gemini, Cerebras, and other providers to their correct backends instead of defaulting to api.openai.com with the wrong API key
• evaluation_runner.py/app.py: chat_model is now threaded from agent execution through the evaluation pipeline, allowing the LLM judge to inherit the correct model when tier2_provider=auto

Added

• graph_export.py: persist agent interaction graph as agent_graph.json (node-link format) and agent_graph.png (static matplotlib render) to per-run output directories; registered with ArtifactRegistry
• run_context.py: graph_json_path and graph_png_path properties on RunContext
• Trace Viewer page (gui/pages/trace_viewer.py): read-only SQLite browser for traces.db
• agent_system.py: log_coordination_event() wiring for single-agent detection
• docs/architecture.md: Entry Points section, expanded Benchmarking Infrastructure (SweepConfig, engine modes, compositions, CLI reference), fixed 12 stale references
• config_app.py: TRACES_DB_FILE constant
• make setup_uv, make setup_bert_model recipes; devcontainer.json HF_HOME persistence
• app.py: up-front RunContext.create() before engine dispatch; execution_id parameter on _run_agent_execution() and run_manager()
• test_app_run_context_wiring.py: Hypothesis fuzz tests for path traversal safety

Changed

• run_cli.py: extract helpers to reduce cognitive complexity (23→9)
• Streamlit theming: native config.toml light/dark sections replace sidebar dropdown; styling.py luminance-based detection

Removed

• judge/agent.py (JudgeAgent), judge/trace_store.py (TraceStore): unused in any production path
• 44 low-value tests: one-time deletion checks, inspect.signature tests, mock-only assertions, stale RunContext wiring tests

Fixed

• traditional_metrics.py: re-enable BERTScore with Levenshtein fallback
• llm_evaluation_managers.py: auto-resolved providers use PROVIDER_REGISTRY.default_model
• MkDocs: pin >=1.6.1,<2.0, fix all build warnings, enable strict: true

Security

• cc_engine.py: _sanitize_cc_query() mitigates CWE-78 argument injection
• test_artifact_registry.py: replace hardcoded /tmp with tmp_path (Bandit B108)

[4.1.0] - 2026-02-22

Added

• make lint_links recipe and lychee link checker npm dependency for dead-link detection in docs
• Agent type indicator (red prefix) in Claude Code status line script (.claude/scripts/statusline.sh)
• WakaTime extension to devcontainer VS Code configurations
• lychee.toml link checker config accepting bot-blocked status codes (403/401/429), excluding archived sprints and unreachable domains
• AGENT_LEARNINGS.md: “Stale Test Fixtures Cause Cross-File Pollution” pattern
• Sprint 10 PRD v3.0.0 — E2E CLI/GUI parity for CC engine, graph visualization, expanded providers, PydanticAI migration
• docs/analysis/Inference-Providers.md: free-tier analysis for 19 providers with recommended models, context windows, rate limits
• docs/reviews/sprint10-prd-review.md: post-task review with 17 findings and resolution status

Changed

• AgentConfig.tools typed as list[Tool[Any]] with type propagated to _create_optional_agent
• Test verification suite (STORY-009): expanded AC1 scope to 14 files, replaced asyncio-marker check with LLM-mock-presence check (AC2), added spec= to collaborator mocks in test_cli_token_limit.py
• Docs: security-advisories.md Related Frameworks section restructured into Threat Modeling and AI Risk & Governance subsections; added NIST AI RMF 1.0, ISO 23894, ISO 42001
• Docs: Sprint 9 PRD expanded to 13 features incorporating review findings
• Commit skill updated for stats ordering and GPG signing
• pytest CI workflow: replaced pip+Python 3.12 with uv+Python 3.13, added project deps install and pull_request trigger
• .claude/scripts/statusline.sh: consolidated 12 separate jq invocations into single tab-delimited call; fixed threshold comments (< → ≤ to match <= operators); added POSIX trailing newline
• Sprint 9 archived: PRD, prd.json, progress.txt moved to docs/sprints/archive/ and ralph/docs/archive/sprint9/
• README badges reordered: docs + llms.txt first, CI/CD second; “DevEx” label removed
• Docs: roadmap.md v4.4.0 — Sprint 9 Delivered, Sprint 10 In Progress
• Docs: architecture.md — Sprint 9 implementation status, development timeline updated

Removed

• 5 stale review documents from docs/reviews/ (evaluation-pipeline-parallel-review, gui-comprehensive-audit, sprint5-code-review, sprint5-test-audit, test-audit)
• .cline/config.json and .gemini/config.json — unused agent configs (project standardized on Claude Code)
• 72 stale files leaked from old main during squash merge: .claude/agents/ (9), opik/ (3), docs/sprints/ (15), src/app/evals/ (8), scripts/citation-styles/ (4), assets/images/ (4), tests/ (8), src/examples/ (2), docker-compose.opik.yaml
• Noisy CHAT_DEFAULT_PROVIDER module-level log line from GUI config

Fixed

• test_no_agentops_commented_code_in_login hardcoded /workspaces/ path broke pytest in GHA; replaced with inspect.getfile() for portability
• Sprint 9 Out of Scope stale link: PRD-Sprint10.md → PRD-Sprint10-Ralph.md after file rename
• 24 broken URLs across landscape docs, further_reading, security-advisories (paper-qa repo, OWASP MAESTRO, StableToolBench org, OpenAI Operator, MatterGen publication)
• lychee.toml: accept 400/415/500 status codes, exclude .venv and .github/workflows, add allenai.org and Springer DOI to excludes
• Broken local file links: architecture.md landscape refs, agent_eval_metrics.md further_reading ref, landscape.md trace_processors path, roadmap.md sprint archive paths, src/examples/README.md relative paths
• Claude Code sandbox phantom files added to .gitignore
• Content diffs from stale merge fixed in .gitignore, Makefile, pyproject.toml, agent_system.py, peerread_models.py, datasets_peerread.py, review_persistence.py
• README/SECURITY formatting: extra blank lines between badge sections, email angle brackets for markdown linking

[4.0.0] - 2026-02-19

Added

• Sprint 9 Feature 1: CC engine GUI wiring — PRD feature for routing “Claude Code” radio selection to cc_engine.run_cc_solo/run_cc_teams in GUI execution path (mirrors existing CLI logic in run_cli.py:126-138)
• resolve_service_url(port) in src/gui/config/config.py: detects GitHub Codespaces, Gitpod, and PHOENIX_ENDPOINT override to build correct service URLs in cloud dev environments; PHOENIX_DEFAULT_ENDPOINT now uses it (STORY-014)
• GUI report generation: “Generate Report” button on App page enabled after evaluation completes; report rendered inline as Markdown with a download button; shares generate_report() logic with CLI (STORY-010)
• App page UX: MAS-specific controls (sub-agents, provider, token limit, config summary) hidden entirely when CC engine selected — not just disabled (STORY-013)
• Evaluation Results page: execution_id displayed as caption below composite score; full ID shown in “Evaluation Details” expander (STORY-013)
• Baseline Comparison Configuration: path validation with st.error for non-existent directories; auto-populate from logs/Agent_evals/traces/ if it exists (STORY-013)
• execution_id included in _prepare_result_dict return and threaded to session state via _execute_query_background (STORY-013)

Removed

• 43 implementation-detail tests across 14 files: deleted test_load_settings.py (4 AST introspection guards), test_sprint1_examples_deleted.py (8 deleted-file guards), test_opik_removal.py (12 removal guards); removed file-exists/string-contains test classes from 3 example test files; removed individual hasattr/callable/isinstance checks from 8 test files. No behavioral coverage lost.

Changed

• Docs: README Status section updated to Sprint 8 Delivered; architecture.md Implementation Status and Development Timeline aligned; Sprint 9 PRD stripped of all 14 solved stories (Features 1-8, Story Breakdown, Ralph Loop notes removed)

Fixed

• graph_builder.py: aligned node attribute key node_type → type to match agent_graph.py reader (Sprint 8 Feature 4 residual); fixed stale node_type= fixtures in test_session_state_wiring.py
• settings.py: replaced text("**Enable Sub-Agents:**") and text("**Token Limit:**") with st.markdown(...) so bold formatting renders correctly; removed unused text import

• render_output() in output.py: renamed type parameter to output_type to avoid shadowing Python built-in type (STORY-013)

• GUI a11y/usability: text-prefix badges [WARN]/[ERR]/[INFO]/[DBG] in log panel (WCAG 1.4.1); [CRIT] for CRITICAL; module text color #999999→#696969 for 5.9:1 contrast (WCAG 1.4.3); "Navigation" radio label with label_visibility="collapsed" (WCAG 1.3.1, 2.4.6); "(opens in new tab)" on Phoenix Traces link (WCAG 3.2.5); CSS radio-circle hiding hack removed (WCAG 1.3.3, 1.4.1); display-only warning on Prompts page; HOME_INFO onboarding corrected to Settings-before-App; RUN_APP_QUERY_PLACEHOLDER made domain-specific; include_researcher/include_analyst default to True; Streamlit primary color #4A90E2 (agent graph blue) (STORY-012)

• GUI judge settings: tier2_provider, tier2_model, tier2_fallback_provider, tier2_fallback_model replaced with selectbox dropdowns populated from PROVIDER_REGISTRY and config_chat.json; fallback_strategy exposed as selectbox with “tier1_only”; judge settings expanders set to expanded=False; “Advanced Settings” header added (STORY-011)
• report_generator.py in src/app/reports/: generate_report(result, suggestions) → Markdown report with executive summary, tier breakdown, and weakness/suggestion sections; save_report(md, path) with auto-created parent dirs (STORY-009)
• --generate-report CLI flag (mutually exclusive with --skip-eval) writes report to results/reports/<timestamp>.md after evaluation (STORY-009)
• --no-llm-suggestions CLI flag to disable LLM-assisted suggestions in generated reports (STORY-009)
• SuggestionEngine with rule-based + optional LLM-assisted paths in src/app/reports/suggestion_engine.py (STORY-008)
• Suggestion Pydantic model and SuggestionSeverity enum (critical/warning/info) in src/app/data_models/report_models.py (STORY-008)
• METRIC_LABELS dict and format_metric_label() in evaluation.py for human-readable metric names (STORY-007)
• ARIA live regions (role="status", role="alert") in _display_execution_result for screen reader accessibility (STORY-007)
• Post-run navigation guidance in completed state (“Evaluation Results”, “Agent Graph”) (STORY-007)
• Sidebar execution-in-progress indicator (⏳) when execution_state="running" (STORY-007)
• st.dataframe() alt text below bar charts in _render_metrics_comparison (WCAG 1.1.1) (STORY-007)
• Baseline comparison inputs wrapped in collapsed expander in render_evaluation (STORY-007)
• Delta indicators in _render_overall_results from BaselineComparison.tier_deltas (STORY-007)
• GUI engine selector: radio toggle between MAS (PydanticAI) and Claude Code engines with CC availability check (STORY-014)
• GUI paper selection: dropdown with ID/title, abstract preview, free-form/paper mode toggle (STORY-009)
• GUI editable common settings: log level, max content length with tooltips; logfire consolidated to JudgeSettings (STORY-010)
• GUI real-time debug log: streaming log panel via st.fragment + LogCapture thread-safe polling (STORY-008)
• --engine=mas|cc flag for CLI and sweep, replacing --cc-baseline (STORY-013)
• Sweep rate-limit resilience: exponential backoff on 429 (max 3 retries), incremental results.json persistence (STORY-013b)
• --judge-provider and --judge-model CLI/sweep args for Tier 2 judge override (STORY-012)
• New examples: basic_evaluation.py, judge_settings_customization.py, engine_comparison.py with README (STORY-002)
• --cc-teams boolean flag for CLI (run_cli.py), sweep (run_sweep.py), and SweepConfig model; enables CC Agent Teams mode with --engine=cc (STORY-006)
• CC baseline Makefile recipes (cc_run_solo, cc_run_teams, cc_collect_teams)
• Test coverage improvements: datasets_peerread 27→60%, models 24→76%, agent_factories 39→75% (STORY-014)
• Security test suite: 135 tests across 5 modules (SSRF, prompt injection, data filtering, input limits, tool registration) (STORY-013)
• MAS composition sweep: SweepRunner for N×M×P benchmarking with CC headless baseline, 33 tests (STORY-007)
• CC artifact collection scripts with docs and tests (STORY-004)
• Spec-constrained mock tests for trace collection, review persistence, log config, logfire instrumentation
• UserStory.md: report generation user story and success criterion (Sprint 8 Feature 6 traceability)
• PRD-Sprint7-Ralph.md and PRD-Sprint8-Ralph.md
• Ralph: baseline-aware validation, process management, story-scoped lint, per-story baseline persistence, timeout protection

Changed

• Makefile: renamed recipes for clarity (ruff→lint_src, ruff_tests→lint_tests, test_all→test, test_quick→test_rerun, sweep→run_sweep, quick_start→quickstart, dataset_get_smallest→dataset_smallest, run_puml_*→plantuml_*, run_markdownlint→lint_md); backward-compat aliases preserved
• Makefile: collapsed 3 Ollama setup variants into OLLAMA=1 flag on setup_prod/setup_dev/setup_devc
• Makefile: grouped make help output with section headers from # MARK: comments
• Makefile: renamed MARK sections for consistency (Sanity→quality, run ollama→ollama, etc.)
• --paper-number renamed to --paper-id (string, supports arxiv IDs); --provider renamed to --chat-provider across CLI, sweep, config (STORY-012)
• SweepConfig.paper_numbers: list[int] → paper_ids: list[str]; added judge_provider, judge_model, engine, cc_teams fields (STORY-012, STORY-013, STORY-006)
• render_sidebar() accepts execution_state parameter; shows in-progress indicator when running (STORY-007)
• Engine selector st.radio now includes help= text explaining MAS vs Claude Code (STORY-007)
• Paper selectbox st.selectbox now includes help= kwarg (STORY-007)
• _render_paper_selection_input no-papers message changed from “Downloads page” to make setup_dataset_sample CLI instruction (STORY-007)
• run_cli.py CC branch now delegates to cc_engine.run_cc_solo / run_cc_teams (removes inline subprocess logic) (STORY-006)
• sweep_runner._invoke_cc_comparison delegates to cc_engine; _run_cc_baselines wires through CCTraceAdapter (STORY-006)
• app.main() now accepts engine parameter; run_app._execute_query_background passes it through (STORY-006)
• Makefile cc_run_solo / cc_run_teams recipes use Python CLI entry point instead of shell scripts (STORY-006)
• JudgeSettings.tier2_provider default changed from "openai" to "auto" — judge inherits MAS chat provider at runtime (STORY-011)
• 429 errors in agent_system.py now re-raise ModelHTTPError instead of SystemExit(1), enabling caller retry logic (STORY-013b)
• PeerRead review score fields coerce int→str via BeforeValidator(str) to handle numeric JSON values (STORY-009)
• PeerRead _create_review_from_dict aggregates missing optional fields into single debug log line (STORY-008)
• Composite scoring tests consolidated: 3 files → 1 (test_composite_scorer.py) with BDD structure template in conftest (STORY-007)
• Removed 3 FIXME dead code blocks from agent_system.py and orchestration.py (STORY-007)
• PlantUML diagrams updated: metrics-eval-sweep (sweep workflow + CC path), MAS-Review-Workflow (MAESTRO security boundaries) (STORY-006)
• Docs: architecture.md v3.7.0 — Sprint 8 scope, report generation section, CC stream-json integration, ADR-008 status, researcher tool update
• Docs: roadmap.md v4.3.0 — Sprint 8 description aligned with PRD scope
• Docs updated: README v4.0.0, architecture.md (benchmarking + security sections), roadmap, CC OTel analysis corrected (STORY-003/004/005)
• CC baseline scripts renamed: collect-cc-solo.sh → run-cc.sh, collect-cc-teams.sh → collect-team-artifacts.sh
• ADR-008: CC baseline engine subprocess vs SDK decision
• Testing best practices: added mock safety rules (spec=RealClass) and unspec’d mock anti-pattern
• Ralph: staleness detection, story-scoped commit scanning, complexity checks, sandbox compatibility (sed -n '1p')

Removed

• scripts/collect-cc-traces/ shell scripts directory — replaced by cc_engine.py Python module (STORY-006)
• Makefile: dead recipes setup_devc_full, setup_devc_ollama_full, output_unset_app_env_sh
• Legacy config keys paper_numbers and provider in sweep JSON — use paper_ids and chat_provider
• "not-required" API key sentinel in create_simple_model — None lets SDK fall back to env vars
• 3 composite scoring test files merged into test_composite_scorer.py (STORY-007)
• Deprecated examples: run_evaluation_example*.py, run_simple_agent_*.py, utils/, config.json (STORY-001)

Fixed

• Judge 401 auth failures: validated API key now forwarded through _resolve_provider_key → select_available_provider → create_judge_agent instead of being discarded

• run_cli.py CLI parser: space-separated args (--paper-id 1105.1072) now parsed correctly instead of treating value flags as booleans

• CCTraceAdapter._extract_coordination_events() stub now parses inboxes/*.json messages (STORY-014)

• test_download_success_mocked AttributeError — patched correct import path (STORY-007)
• agent_system.py: result.output instead of deprecated result.data
• trace_processors.py: end_execution() now idempotent
• logfire_instrumentation.py: correct “Phoenix”/”Logfire” init log messages
• review_persistence.py: reviews saved under project root, not src/app/
• Log paths aligned to logs/Agent_evals/ (config_app.py, judge/settings.py)
• Ralph: story-scoped lint (no pre-existing violations), restart baseline isolation, validate scope

Changed (Sprint 6)

• STORY-015: Executed Sprint 5 test audit refactoring plan — deleted ~61 implementation-detail tests from test_trace_store.py (basic CRUD and metadata tests) while preserving all behavioral coverage (thread-safety, context manager tests retained)
• STORY-014: Fixed failing test expectations to match actual behavior — removed tests for non-existent error propagation, empty string validation, corrupted PDF errors
• STORY-009: Review tools enabled by default; --no-review-tools to opt out
• STORY-008: Review tools routed to researcher agent (was manager-only); single-agent fallback preserved
• STORY-001: Complete Opik removal (~140 lines) — replaced 13 references with Phoenix/Logfire equivalents
• Logfire scrubbing: get_logfire_scrubbing_patterns() returns only 7 extra patterns not covered by Logfire defaults

Fixed (Sprint 6)

• STORY-013: Log scrubbing edge cases — natural language patterns (“password to”), broadened OpenAI key regex to sk-\S+
• STORY-003: CCTraceAdapter supports sibling and legacy directory layouts with auto-discovery; --cc-teams-tasks-dir flag
• STORY-002: Phoenix Docker recipe (persistent volume, gRPC 4317, auto-restart); graph renders when execution_id exists
• SSRF allowlist: added api.github.com, removed 3 unused LLM provider domains

Removed (Sprint 6)

• STORY-006: Orphaned cc_otel module and tests — wrong abstraction for CC tracing

Security (Sprint 6)

• STORY-010 (CRITICAL): CVE-2026-25580 SSRF mitigation — URL validation with HTTPS-only + domain allowlist in datasets_peerread.py (49 tests). CVE advisories in SECURITY.md
• STORY-011 (HIGH): Prompt injection mitigation — input sanitization with length limits + XML delimiter wrapping in LLM judge prompts (25 tests)
• STORY-012 (HIGH): Log/trace data scrubbing — pattern-based redaction for API keys, passwords, tokens in Loguru sinks and Logfire OTLP exports (13 tests)

Added (Sprint 5)

• STORY-001: Tier 2 judge provider fallback chain (tier2_provider=auto)
• STORY-002: Token limit override (CLI --token-limit, GUI, env var)
• STORY-003: Single-agent mode detection + weight redistribution
• STORY-004: PeerRead optional field handling (“UNKNOWN” defaults)
• STORY-006: Streamlit background execution with session state persistence
• STORY-007: Debug log panel (LogCapture, loguru sink, HTML output)
• STORY-008: Graph visualization wiring (GraphTraceData→NetworkX→GUI)
• STORY-009: Editable judge settings page with pipeline plumbing
• STORY-010: MAESTRO 7-layer security review (2 critical CVEs found)
• STORY-014: wandb import guard tests

Changed (Sprint 5)

• STORY-016: PeerRead tools moved from manager to researcher in multi-agent mode
• STORY-008: main() returns dict; complexity 19→10 via extracted helpers
• STORY-011: Deleted 31 implementation-detail tests (no behavioral coverage loss)
• STORY-005: README, roadmap, architecture docs updated

Removed (Sprint 5)

• STORY-017: Duplicate AppEnv class from load_settings.py
• STORY-011: test_opik_removal.py and test_migration_cleanup.py
• STORY-014: Dead agentops code from login.py

Fixed (Sprint 5)

• STORY-012: OTLP endpoint double-path bug (HTTP 405)
• STORY-013: Tool success_rate overwrite; agent-tool edge weight averaging; dead communication_overhead metric
• STORY-014: wandb/weave import guards; WANDB_ERROR_REPORTING defaults to “false”
• STORY-015: Debug log in get_api_key() for empty .env strings
• STORY-004: PeerRead papers 304-308, 330 — .get() for optional review fields

Security (Sprint 5)

• STORY-010: CRITICAL — CVE-2026-25580 (PydanticAI SSRF), CVE-2026-25640 (Stored XSS)
• STORY-010: MEDIUM — CVE-2024-5206 (scikit-learn data leakage)
• STORY-010: 31 findings across 7 MAESTRO layers — mitigations in docs/reviews/sprint5-code-review.md

Added (Sprint 2)

• STORY-001: JudgeSettings pydantic-settings class replacing config_eval.json
• STORY-002: Post-run evaluation wiring with --skip-eval CLI flag
• STORY-003: TraceCollector integration into agent orchestration with GraphTraceData support
• STORY-004: Graph vs text metric comparison logging in evaluation pipeline
• STORY-005: Logfire + Phoenix tracing infrastructure (replacing Opik)
• STORY-006: Streamlit evaluation dashboard with agent graph visualization

Added (Sprint 3)

• STORY-007: Plugin architecture with EvaluatorPlugin base class and PluginRegistry
• STORY-007: JudgeAgent orchestrator replacing EvaluationPipeline
• STORY-007: TraceStore for thread-safe trace storage
• STORY-007: Plugin wrappers for all three evaluation tiers
• STORY-009: cc_otel module for Claude Code OpenTelemetry instrumentation with Phoenix OTLP backend
• STORY-010: GUI settings page displays actual values from CommonSettings and JudgeSettings pydantic-settings classes
• STORY-010: GUI prompts page loads directly from ChatConfig.prompts without hardcoded fallback
• STORY-011: Property-based tests using Hypothesis for score bounds, input validation, and math invariants
• STORY-011: Snapshot tests using inline-snapshot for Pydantic model dumps and structure regression
• STORY-012: Optional weave dependency group in pyproject.toml (only loaded when WANDB_API_KEY is set)
• STORY-013: Trace logging to all 6 PeerRead manager tools (get_peerread_paper, query_peerread_papers, read_paper_pdf_tool, generate_paper_review_content_from_template, save_paper_review, save_structured_review) with time.perf_counter() timing
• STORY-013: Property-based tests for trace event schema invariants (agent_id always present in tool_call dicts)
• STORY-013: Snapshot tests for GraphTraceData transformation output structure
• STORY-014: Session state initialization for provider and sub-agent configuration in run_gui.py
• STORY-014: Provider selectbox with all PROVIDER_REGISTRY options on Settings page
• STORY-014: Checkboxes for include_researcher/analyst/synthesiser on Settings page
• STORY-014: Run App page reads configuration from session state and displays current settings
• STORY-014: Session state persistence across page navigation for provider and agent selection

Added (Sprint 4)

• Sprint 4 PRD v2 with Features 1-7 (standalone numbering, operational resilience + Claude Code baseline comparison)
• Feature 1 (STORY-001): Graceful Logfire trace export failure handling (suppress connection error stack traces for both span and metrics exports, affects CLI and GUI)
• Feature 2 (STORY-002): Thread-safe timeout handling in graph analysis using ThreadPoolExecutor (replaces signal-based timeouts)
• Feature 3 (STORY-003): Tier 2 judge provider fallback integration tests and troubleshooting documentation
• Feature 3 (STORY-003): Auth failure detection in all Tier 2 assessments (technical_accuracy, constructiveness, planning_rationality) with neutral fallback scores (0.5)
• Feature 7 (STORY-007): CLI --cc-solo-dir and --cc-teams-dir flags for baseline comparison against Claude Code artifacts
• Feature 7 (STORY-007): _run_baseline_comparisons() function in app.py for evaluating Claude Code solo and teams baselines
• Feature 7 (STORY-007): render_baseline_comparison() GUI section for side-by-side metrics display and three-way comparison tables
• Feature 7 (STORY-007): CLI baseline comparison logging with summary output for each pairwise comparison
• Feature 4 (STORY-004): Complete test suite alignment with hypothesis and inline-snapshot (no BDD/Gherkin)
• Feature 5 (STORY-005): Claude Code trace adapter for solo and teams modes parsing Claude Code artifacts into GraphTraceData format
• Feature 4 (STORY-004): Hypothesis property-based tests for data validation invariants, score bounds, URL construction, execution traces, and metrics output
• Feature 4 (STORY-004): Inline-snapshot regression tests for Pydantic model dumps, configuration outputs, benchmark results, and GUI state structures
• Feature 4 (STORY-004): Test coverage for integration tests (PeerRead dataset compatibility), benchmarks (performance baselines), GUI pages (evaluation/graph/sidebar), and data utilities (datasets_peerread)
• Feature 3 (STORY-003): docs/best-practices/troubleshooting.md with provider fallback chain guidance
• Feature 4: Complete test suite alignment (hypothesis property tests + inline-snapshot regression tests for remaining 12 test files)

Fixed (Sprint 4)

• Feature 2 (STORY-002): Thread-safe graph analysis timeout handling
• Replace signal-based timeout with concurrent.futures.ThreadPoolExecutor
• Fix “signal only works in main thread” error in Streamlit GUI
• path_convergence calculation now works in non-main threads
• Graceful fallback to 0.3 when timeout occurs (maintains existing behavior)
• Added 5 comprehensive tests for thread-safe timeout behavior

Fixed (Sprint 4)

• STORY-001: Noisy ConnectionRefusedError stack traces when Logfire/OTLP endpoint unreachable (both CLI and GUI)
• Feature 5 (STORY-005): Claude Code trace adapter – parse Claude Code artifacts into GraphTraceData in two modes: solo (single Claude Code instance, no orchestration) and teams (Claude Code Agent Teams with delegation), both with full tool/plugin/MCP access
• Feature 6 (STORY-006): Baseline comparison engine – BaselineComparison Pydantic model + compare()/compare_all() for three-way CompositeResult diffing (PydanticAI vs Claude Code solo vs Claude Code teams)
• compare() function for pairwise diffing of any two CompositeResult instances
• compare_all() convenience function for all three pairwise comparisons
• Metric-level deltas for all 6 composite metrics (time_taken, task_success, coordination_quality, tool_efficiency, planning_rationality, output_similarity)
• Tier-level deltas (Tier 1, Tier 2, Tier 3) with graceful handling of missing Tier 2
• Human-readable comparison summaries with average delta and largest metric difference
• Property-based tests for delta symmetry (swap inputs → negated deltas)
• Snapshot tests for model structure and comparison output validation
• Feature 7: CLI & GUI baseline integration – --cc-solo-dir and --cc-teams-dir CLI flags, three-way comparison in GUI

Added

• inline-snapshot as supplementary testing tool
• common module with shared utilities extracted from scattered helpers
• MAS design principles, security, and plugin design skills
• Sprint 3 PRD with Features 5-15 (plugin architecture, judge fallback, GUI wiring, test alignment)

Changed

• STORY-013: Logfire instrumentation now initialized at app startup (app.py:207-209) when JudgeSettings.logfire_enabled=True
• STORY-013: _store_trace() logging enhanced to include full storage path for JSONL + SQLite (trace_processors.py:357)
• STORY-012: login.py conditionally imports weave only when WANDB_API_KEY is configured
• STORY-012: app.py provides no-op @op() decorator fallback when weave unavailable
• STORY-008: Consolidated all evaluation code from app.evals.* to app.judge.*
• STORY-010: GUI settings page refactored from provider selection to read-only settings display with Streamlit expanders
• STORY-010: GUI prompts page updated to load prompts from ChatConfig without PROMPTS_DEFAULT fallback
• Makefile: DRY refactor of quick_start/dataset_get_smallest via _find_smallest_papers helper; removed redundant \ from .ONESHELL blocks
• Pandoc: configurable LoF/LoT generation and unnumbered title support
• Ralph PRD parser rewritten with typed models and safe updates
• All dependencies updated to latest 2026 versions

Removed

• STORY-010: PROMPTS_DEFAULT hardcoded constant from gui/config/config.py (DRY principle - single source of truth in config_chat.json)

Fixed

• STORY-013: agent_id now included in tool_call dicts during trace processing (_process_events at trace_processors.py:269, _parse_trace_events at trace_processors.py:377)
• STORY-013: GraphTraceData transformation succeeds with researcher traces (no “missing agent_id” error)
• STORY-013: Manager-only runs now produce non-empty trace data (tools log trace events)
• Cerebras provider 422 error from mixed strict tool definitions — added OpenAIModelProfile with strict disabled
• Statusline ctx(left) accuracy — compute true usable space by subtracting 16.5% autocompact buffer (33k tokens) from remaining percentage; added CLAUDE_AUTOCOMPACT_PCT_OVERRIDE environment variable support for configurable buffer threshold

Removed

• STORY-008: Entire app.evals module (migrated to app.judge)
• STORY-008: Duplicate src/app/agents/peerread_tools.py (canonical at src/app/tools/)
• Deprecated config/config_eval.json (superseded by JudgeSettings pydantic-settings)
• Opik tracing from evaluation pipeline (replaced by Logfire + Phoenix)
• Unused pytest-bdd and agentops dependencies

[3.3.0] - 2026-02-09

Added

• generating-writeup skill: academic/technical writeup generation with pandoc citation support, IEEE [1] default style, and template.md for document structure
• Pandoc --citeproc integration: BIBLIOGRAPHY and CSL variables in make run_pandoc, auto-resolves IEEE CSL from scripts/writeup/citation-styles/ieee.csl
• Claude Code Skills infrastructure (5 skills): core-principles (KISS, DRY, YAGNI), designing-backend, implementing-python, reviewing-code, generating-prd
• Ralph Loop autonomous execution system (.claude/scripts/ralph/): ralph.sh orchestrator, prompt.md, init.sh
• Template-based state file management (.claude/templates/ralph/): prd.json.template, progress.txt.template
• Makefile recipes for Ralph: ralph_init, ralph, ralph_status, ralph_clean
• AGENTS.md Claude Code Infrastructure section with Skills and Ralph Loop references
• Comprehensive benchmark catalog expansion: 40+ benchmarks from further_reading.md integrated into landscape-evaluation-data-resources.md across 10 categories (General Agent, Web Agents, Code/SE, Tool Use, Scientific, Enterprise, Multi-Agent, Safety/Security, Planning, Specialized Domains). Key additions: CORE-Bench (computational reproducibility - highly relevant for PeerRead), MultiAgentBench (multi-agent coordination evaluation), WebArena/VisualWebArena/BrowserGym (web interaction), ToolLLM/MetaTool/StableToolBench (tool usage), CLEAR framework (enterprise metrics with ρ=0.83 production correlation), safety benchmarks (SALAD-Bench, Agent-SafetyBench, SafeAgentBench, AgentHarm, WASP, CyberGym)
• Research paper expansion: 232+ papers covering 2020-2026 (from 154+)
• Practitioner resources section in further_reading.md with Anthropic engineering insights
• Evaluation frameworks: Inspect AI (UK AISI, PydanticAI support), Bloom (behavioral evals), Petri (alignment auditing), DeepEval (three-layer evaluation model)
• DeepEval integration analysis: Three-layer model (Reasoning/Action/Execution), component-level metrics, GEval custom criteria
• Failure mode taxonomy from Anthropic harness pattern mapped to evaluation metrics
• Key benchmarks from survey papers: GAIA, API-Bank, SALAD-Bench, Agent-SafetyBench, SafeAgentBench, AgentHarm
• AgentBeats Competition benchmarks (20 added): Scientific (SciCode, CORE-Bench), Web (BrowserGym, Online-Mind2Web, WebShop), Enterprise (Spider 2.0, CRMArena, CRMArena-Pro, Finance, MedAgentBench), Code (AppWorld, USACO, VERINA), Safety (DoomArena, WASP, CyberGym, Smart Contract), Gaming/Embodied (Werewolf, Minecraft, ALFWorld, PersonaGym), Legal (LegalAgentBench)
• AgentBeats Competition participation guide (AgentBeats_basics.md): OUTSTANDING opportunities for Research Agent Track ($16k), Multi-Agent Track (category-defining), and AAA Track with dual-track submission strategy, implementation roadmap, and immediate next steps
• Tool evaluation expansion: τ²-bench (2506.07982) dual-control user-agent evaluation
• OS/Web benchmarks: OSWorld (2404.07972) comprehensive task evaluation
• Memory systems research: MAGMA, MACLA, comprehensive memory surveys
• Enterprise evaluation frameworks: CLEAR (ρ=0.83), AgentArch, TheAgentCompany, MultiAgentBench
• Pydantic Evals to Practitioner Resources: Span-based evaluation with OpenTelemetry, Logfire integration, philosophy validates post-execution behavioral analysis
• Arize Phoenix Multi-Agent to Practitioner Resources: Three evaluation strategies (Agent Handoff, System-Level, Coordination), multi-level metrics, five coordination patterns
• Production Framework Integration Matrix: Added Pydantic Evals and Arize Phoenix rows (now 10 frameworks)
• New candidate metrics from production frameworks: path_convergence (Arize Phoenix), handoff_quality (Arize Multi-Agent), semantic_outcome (LangSmith), evaluator_alignment (Self-Improving Evals)
• Research papers: Rubric Rewards AI Co-Scientists [2512.23707], SWE-EVO long-horizon benchmark [2512.18470], Confucius Code Agent [2512.10398], SciSciGPT [2504.05559]
• New Metrics for Implementation section in architecture.md: fix_rate, rubric_alignment, path_convergence, handoff_quality
• SWE-EVO benchmark to landscape-evaluation-data-resources.md with Fix Rate metric for partial progress evaluation
• SciSciGPT to landscape-research-agents.md with LLM Agent capability maturity model
• Research Plan Evaluation methodology to research_integration_analysis.md: rubric-based self-grading, long-horizon partial progress, hierarchical memory assessment
• Claude Evaluation Framework to Practitioner Resources: SMART criteria, grading hierarchy (Code→LLM→Human), Bloom correlation 0.86
• Claude Eval Framework to Integration Matrix (now 11 frameworks): validates three-tier approach
• Clear audience targeting for all main documentation files
• PyPI verification requirement for new package introductions in AGENTS.md
• architecture.md: Comprehensive system architecture documentation
• landscape analysis: AI agent ecosystem split into focused documents (landscape.md overview, landscape-agent-frameworks-infrastructure.md, landscape-evaluation-data-resources.md)
• agent_eval_metrics.md: Evaluation metrics catalog
• Comprehensive three-tier evaluation pipeline test with realistic scientific paper data
• Full pipeline workflow demonstration with PeerRead-compatible data models
• Performance monitoring and observability testing with trace collection
• Error handling and fallback strategy validation across all evaluation tiers
• Task 4.3: Complete PeerRead integration validation framework with 7 comprehensive test files
• Real dataset validation testing for PeerRead data compatibility and performance
• Composite scoring validation with 5 performance scenarios and edge case testing
• Performance baseline documentation with empirical analysis and optimization recommendations
• Enhanced integration testing with multi-paper scenarios and production readiness validation

Changed

• AGENTS.md: Added Skills and Ralph infrastructure references, updated Agent Role Boundaries section
• Ralph scripts location: Moved from scripts/ralph/ to .claude/scripts/ralph/
• Research integration analysis: Updated to 208+ papers with 2022-2026 coverage (version 3.0.0)
• Further reading document: Enhanced with 54 new papers including safety benchmarks and memory systems
• Research validation references: Added CLEAR framework and evaluation taxonomy citations
• Enhanced error handling in evaluation pipeline with detailed context logging and specific guidance for different error types
• Improved performance monitoring with bottleneck detection and comprehensive metrics collection
• Enhanced fallback strategy reliability with better status reporting and detailed failure tracking
• Added configuration validation to prevent invalid pipeline configurations
• Improved logging throughout pipeline execution with performance insights and failure analysis

Fixed

• Evaluation pipeline integration: Fixed data model field name mismatches in composite scorer tests
• Test integration issues: Corrected field mappings between Tier3Result model and composite scorer expectations
• trace_observe_methods.md: Observability analysis
• Modular architecture with functional separation: agents/, evals/, llms/, tools/
• LLM provider abstraction layer with multi-provider support
• Agent factory patterns for creating specialized agents
• Evaluation engine orchestration and management systems
• Configuration-based prompt management for evaluation agents
• Comprehensive docstrings for all major modules (agents/, data_models/, config/)
• GitHub models integration with OpenAI-compatible API
• Three-tier evaluation system with traditional metrics, LLM-as-judge, and trace collection
• Typed Pydantic models for LLM assessment results (TechnicalAccuracyAssessment, ConstructivenessAssessment, PlanningRationalityAssessment)
• Composite scoring system integrating all three evaluation tiers into unified assessment
• MetricNormalizer with six normalization functions for consistent scoring ranges
• CompositeScorer with weighted aggregation of normalized metrics
• RecommendationEngine with threshold-based paper acceptance recommendations
• TierIntegrationManager orchestrating fallback handling for missing evaluation tiers
• Performance-optimized composite scoring achieving <100ms latency target
• Comprehensive test suite with >95% coverage for composite scoring system
• Enhanced type safety with explicit type annotations for trace processors and evaluation pipelines
• Modern datetime handling with timezone-aware UTC timestamps
• Comprehensive exception handling patterns for evaluation fallback mechanisms

Changed

• Documentation structure: clarified purpose statements for README.md (humans), AGENTS.md (agents), CONTRIBUTING.md (shared)
• AGENTS.md: streamlined content, removed duplicated architecture information
• CHANGELOG.md: reduced boilerplate, consolidated change type descriptions
• README.md: major reorganization, moved detailed content to dedicated docs
• Restructured codebase from monolithic to modular architecture
• Migrated LLM functionality from single file to focused modules
• Updated GUI import paths to use new modular structure
• Refactored agent creation to use configuration-based prompts
• Shortened evaluation prompt strings for improved readability

Fixed

• Import errors after architectural restructuring
• Line length violations in evaluation modules
• CLI argument parsing for proper provider selection
• GUI import paths to work with new module structure
• Time score calculation in traditional metrics to prevent negative values
• PydanticAI deprecation warnings where feasible
• Datetime deprecation warnings by migrating to datetime.now(datetime.UTC)
• Type safety issues in trace processors with explicit type annotations
• BaseException type issues in evaluation pipelines with proper exception handling
• LLM assessment model definitions with comprehensive Pydantic validation
• Data model imports using direct module references instead of __init__.py exports
• Graph analysis engine configuration validation now allows partial weight specifications for improved usability
• Test data structures in graph analysis tests to include required fields
• NetworkX error handling test expectations to match actual fallback behavior

Removed

• Obsolete context/ directory and .claude/commands framework references
• FRP (Feature Requirements Prompt) command references from documentation
• Redundant architecture details from AGENTS.md
• Monolithic llm_model_funs.py file after successful migration

[3.2.0] - 2025-08-19

Added

• Evaluation engine for PeerRead dataset
• Documentation updates: separation of human and agent files, clear CONTRIBUTING.md

Fixed

• Cleaned up obsolete documentation and logs, removed outdated markdown files and datasets

[3.1.0] - 2025-08-10

Added

• Inspected paper visualization
• PlantUML local generation with Docker

Changed

• Updated project documentation
• Sprint plans
• PlantUML diagrams with CSS for better clarity and consistency

[3.0.0] - 2025-08-03

Added

• MAS review engine using PeerRead dataset

Changed

• Agent Context

[2.1.0] - 2025-07-25

Added

• PeerRead dataset functionality
• PeerRead agent usage documentation to reflect new architecture with data_models instead of datamodels path structure
• Eval functionality in separate system
• Gemini CLI as fallback for Claude Code CLI

[2.0.0] - 2025-07-06

Added

• Claude Code CLI commands and settings

[1.1.0] - 2025-07-05

Added

• Makefile command and devcontainer.json for Claude Code CLI usage

Changed

• Moved streamlit_gui and examples to /src
• Moved app to /src/app

[1.0.0] - 2025-03-18

2025-03-18

• refactor(agent,streamlit): Convert main and run_manager functions again to async for streamli output
• fix(prompts): Update system prompts for manager,researcher and synthesiser roles to remove complexity
• chore(workflows): Update action versions in GitHub workflows for consistency
• chore(workflows): Update action versions for deploy docs to pgh-pages
• docs(deps): Add documentation dependencies for MkDocs and related plugins to pyproject.toml

2025-03-17

• feat(main,agent): refactor entry point to support async execution and enhance login handling
• feat(cli,login,log): refactor entry point to integrate Typer, enhance logging, added login every run
• feat(streamlit): replace load_config with load_app_config, enhance sidebar rendering, and improve output rendering with type support
• feat(streamlit): enhance render_output function with detailed docstring and improve query handling in run_app
• feat(streamlit): enhance render_output function with additional info parameter and improve output handling in run_app
• feat(streamlit,app): add Typer dependency, update main entry point for async execution, add streamlit provider input
• feat(agent): update configuration and improve agent system setup with enhanced error handling and new environment variables
• feat(config,login,catch): add inference settings with usage limits and result retries, enhance login function to initialize environment and handle exceptions, comment out raise in error handling context to prevent unintended crashes
• feat(login,catch): integrate logfire configuration in login function and improve error handling context

2025-03-16

• feat(devcontainer): Refactor devcontainer setup: remove old configurations and add new setup targets for development and Ollama
• feat(devcontainer): Changed from vscode to astral-sh devcontainer
• feat(devcontainer): Changed to vscode container, added postcreatecommand make setup_env
• feat(devcontainer): restructure environment setup with new devcontainer configurations
• feat(devcontainer): update environment names for clarity in devcontainer configurations
• refactor(agent): Added AgentConfig class for better agent configuration management, Refactored main function for streamlined agent initialization.
• feat(config,agents): Update model providers and enhance configuration management, examples: Added new model providers: Gemini and OpenRouter, src: Enabled streaming responses in the agent system
• chore: Remove unused prompt files, update configuration, and enhance logging setup
• refactor(exception,logfire): Enhance error handling and update model configurations in agent system

2025-03-14

• feat(scalene): Add profiling support and update dependencies
• refactor(Makefile): Improve target descriptions and organization

2025-03-13

• refactor(API,except): .env.example, add OpenRouter configuration, enhance error handling in run_simple_agent_system.py, and update ModelConfig to allow optional API key.
• feat(streamlit): add Streamlit app structure with header, footer, sidebar, and main content components
• feat(streamlit): enhance Streamlit app with detailed docstrings, improved header/footer, and refined main content layout
• feat(makefile,streamlit): update Makefile commands for CLI and GUI execution, and modify README for usage instructions, add streamlit config.toml
• feat(streamlit): restructure Streamlit app by removing unused components, adding new header, footer, sidebar, and output components, and updating configuration settings
• chore: replace app entrypoint with main, remove unused tools and tests, and update makefile for linting and type checking
• chore: Enhance makefile with coverage and help commands, update mkdocs.yaml and pyproject.toml for improved project structure and documentation
• test: Update makefile for coverage reporting, modify pyproject.toml to include pytest-cov, and adjust dependency settings
• test: Add coverage support with pytest-cov and update makefile for coverage reporting
• test: makefile for coverage reporting, update dependencies in pyproject.toml for improved testing and coverage support
• chore: Remove redundant help command from makefile
• refactor(agent,async): Refactor agent tests to use async fixtures and update verification methods for async results
• fix(Dockerfile): Remove unnecessary user creation and pip install commands from Dockerfile
• feat(agent): Update dependencies and add new example structures; remove obsolete files
• chore(structure): simplified agents.py
• fix(pyproject): Replace pydantic-ai with pydantic-ai-slim and update dependencies
• feat(examples): add new examples and data models; update configuration structure
• feat(agent): update dependencies, enhance examples, and introduce new data models for research and analysis agents
• feat(examples): enhance prompts structure and refactor research agent integration
• feat(examples): improve documentation and enhance error handling in agent examples
• feat(agent): Added data models and configuration for research and analysis agents, Added System C4 plantuml
• feat(weave,dependencies): update dependencies and integrate Weave for enhanced functionality in the agent system
• feat(agent): initialize agentops with API key and default tags for enhanced agent functionality
• feat(agent): integrate logfire for logging and configure initial logging settings
• feat(agent): adjust usage limits for ollama provider to enhance performance
• feat(agent): refine system prompts and enhance data model structure for improved agent interactions
• feat(agent): update system prompts for improved clarity and accuracy; add example environment configuration
• feat(agent): enhance agent system with synthesiser functionality and update prompts for improved coordination
• feat(agent): add Grok and Gemini API configurations; initialize logging and agent operations
• feat(agent): improve documentation and refactor model configuration handling for agent system
• feat(agent): update environment configuration, enhance logging, and refine agent management functionality
• feat(agent): refactor login handling, update model retrieval, and enhance agent configuration

[0.0.2] - 2025-01-20

Added

• PRD.md
• C4 architecture diagrams: system context, code
• tests: basic agent evals, config.json

Changed

• make recipes

[0.0.1] - 2025-01-20

Added

• Makefile: setup, test, ruff
• devcontainer: python only, w/o Jetbrains clutter from default devcontainer
• ollama: server and model download successful
• agent: tools use full run red
• pytest: e2e run, final result red
• Readme: basic project info
• pyproject.toml