Lil-Task-X is now a fully agentic Product Management pipeline powered by Gemini 2.5 Flash and LangChain. The system runs in two deterministic phases:

• Phase 1 – Product Strategy: a ReAct agent ingests the product brief, performs live competitor research, and returns a market-ready PRD in Markdown plus structured JSON.
• Phase 2 – Delivery Plan: the same agent family converts the PRD into features, stories, tasks, staffing assignments, delivery scenarios (on-track / need-more / reduce-scope), email drafts, and repo watchlist reminders alongside budget projections. A Gemini-backed PDF generator produces visual artefacts.

All reasoning happens through LangChain’s initialize_agent API with AgentType.ZERO_SHOT_REACT_DESCRIPTION. Every LLM call uses temperature=0 and top_p=0.1 to minimise hallucinations. Competitive claims are grounded by an automatic web-search tool (SerpAPI or Exa).

1. LLM: ChatGoogleGenerativeAI (Gemini 2.5 Flash by default, Pro tiers optional when enabled).
2. Tools (src/agents/pm_agent.py):
   • competitor_report: SerpAPI/Exa-powered search that returns structured competitor data.
   • load_csv: exposes repo CSVs (employees, budgets, stakeholders) as JSON.
   • budget_calculator: aggregates numeric spend categories for validation.
   • task_splitter: supplies baseline lifecycle tasks per feature.
   • jira_uploader: mock integration that confirms export readiness.
   • pdf_generator: bridges to the Matplotlib + ReportLab PDF renderer.
3. Agent: ReAct loop plans tool usage, reasons through evidence, then emits the final JSON payloads consumed by each phase.
4. Decision Support: Phase 2 applies budget and capacity checks, surfaces three human-readable delivery options, drafts Jira issues/emails, and flags repo areas to monitor for shareholder alignment.

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

cp .env.example .env

Fill in the .env file with your credentials:

• GOOGLE_API_KEY – required for Gemini APIs (defaults target gemini-2.5-flash).
• SERPAPI_KEY – optional, enables Google web search (preferred).
• EXA_API_KEY – optional, enables Exa search (used if present).
• GEMINI_MODEL – optional override for the primary Gemini model (defaults to gemini-2.5-flash).
• GEMINI_FAST_MODEL – optional override for lightweight requests; defaults to the same value.
• GEMINI_PRD_MODEL – optional override for the phase-one PRD synthesis call (defaults to gemini-1.5-pro-latest).

If both SERPAPI and EXA keys are absent the agent warns that competitor data is unavailable, but the run will continue with local knowledge only.

python -m src.pipeline.main --base-dir "$(pwd)" --outputs-dir "$(pwd)/outputs"

After a successful run you will find three canonical artefacts inside outputs/:

1. phase1_product_spec.md – the polished PRD with market sizing, personas, feature roadmap, risks, and KPIs.
2. phase2_tasks.json – structured data containing features, stories, Jira-ready tasks, staffing assignments, budget reconciliation, delivery options, email drafts, and repo watchlist guidance.
3. phase2_budget_analysis.pdf – charts, tables, and Gemini-generated narrative summarising budget usage and scheduling insights.

final_output.json aggregates the key paths plus structured outputs for downstream automation.

• Phase 1: run PhaseOneProductStrategy(config).run() from a Python shell after initialising PipelineConfig. Inspect the saved prompt (phase1_prompt.txt), raw agent trace (phase1_raw.json), and PRD Markdown.
• Phase 2: feed the Phase 1 PRD into PhaseTwoFeasibility(config).run(prd_markdown). Examine phase2_raw.json for all intermediate ReAct steps, and open phase2_budget_analysis.pdf to verify charts.

For automated regression, stub ChatGoogleGenerativeAI.invoke with canned responses. The agent tools accept JSON strings, so you can simulate tool outputs during tests without hitting external services.

• Place new CSVs under data/ and reference them via relative paths when invoking the load_csv tool (e.g., data/new_employees.csv).
• Ensure column headers are descriptive; they will be preserved in JSON for the agent to interpret.
• Budget files should expose numeric values in a single column. Non-numeric entries are coerced to 0.0 by the budget_calculator tool.

• Hourly costs derive from the employee roster CSV (Hourly_Rate_USD). The agent keeps individuals under 160 hours across the 12-week plan.
• The budget_calculator tool aggregates CSV rows so the agent can cross-check totals before returning budget_report.
• The PDF generator plots two charts: budget allocation (pie) and engineer/tester workload (bar). A Gemini summary validates that every claim aligns with the JSON payload.

jira_uploader currently mocks the API call and returns a confirmation string. To enable a real integration:

1. Replace the tool implementation in src/agents/pm_agent.py with a call to your Jira Cloud/Server REST endpoint.
2. Inject authentication (OAuth or PAT) and handle error responses.
3. Continue returning a concise status message so the agent’s reasoning remains traceable.

• All Gemini calls fix temperature=0 and top_p=0.1 to eliminate creative drift.
• Prompts instruct the agent to cite sources for competitor statements and budget maths. Check phase*_raw.json to review the reasoning chain.
• If competitor scraping fails (missing API key, rate limit), rerun after setting the relevant env var. The agent degrades gracefully but flags the absence of live intel.

• Keep new tools deterministic and JSON-oriented so the ReAct agent can parse outputs reliably.
• When adding files under data/, document them in this README and ensure relative paths work on macOS/Linux.
• Pull requests should regenerate final_output.json to validate that all three deliverables compile successfully.