Exploratory Data Analysis - Delivery Constraint & Opportunity Analysis

Description

Use the catalog to reconstruct the delivery cycle, score impact (people×effort plus downstream wait), look until the ranking can be trusted, and triage eliminate / automate / delegate. Streamlit, Jupyter, and MCP are skills, not a separate workflow.

Activities in this Phase

Order	Name	Actions
Constraint and Flow Analysis
#1	Map the BizDevTestSecOps Value Stream	Read the 'CAS-1 Map the BizDevTestSecOps Value Stream' activity from playbook 'Delivery Constraint & Opportunity Analysis' (workflow 'Constraint and Flow Analysis'). Follow its guidance when assisting me. ## Purpose Draw the BizDevTestSecOps map from catalog events that exist, not from a generic GitLab/Jira cartoon. ## Inputs - Queryable catalog views (pipeline done, including closed-issue backfill). - Lake Scope Contract: `docs/lake-scope-contract.yaml` key `views`. - Starter file `streamlit/data/value_stream.yaml` (human notes: `docs/analysis/value-stream-map.md`). - Skill Value Stream and Constraint Analysis → Recipe steps 2 and 4. - Skill Athena SQL Querying → Recipe (workgroup, LIMIT 10, `$.payload.`). ## Steps 1. Open `docs/lake-scope-contract.yaml` → `views`. Those names are* the Glue/Athena tables. 2. Prove each view is queryable: run `scripts/create_athena_views.py` (populate-lake, not a laptop against shared env) or `sql/proof_views.sql` in workgroup `athena_workgroup` (default `ops-lake-analysts`). `SELECT COUNT()` then `LIMIT 10`. Paste `QueryExecutionId` onto the matching row in `streamlit/data/value_stream.yaml`. If n=0 or the table is missing, stop — do not invent a stage. 3. Open `streamlit/data/value_stream.yaml`. Default five stages (skill Value Stream → Recipe step 2): - Biz* — `github_issues` opened (`$.payload.created_at`); join issue number - Dev — `github_pull_requests` opened → merged; `github_commits` in window; join `Closes #N` - Test — `github_workflow_runs` start → first `conclusion=success` - Sec — security-gate job in that run, or a dedicated scan workflow - Ops — deploy workflow run start → complete Delete any row whose `view` is not in `views`. Uncomment gitlab/jira only if those sources exist. Do not use heartbeat `github_project_items` as Biz. 4. Feature-type cohort: skill Value Stream → Recipe step 4 (classifier). Do not map every issue. 5. Commit the filled YAML. That is artifact BizDevTestSecOps Value Stream Map. ## Done when - Map names real view+column pairs for each stage boundary. - Every remaining row has a QueryExecutionId (or an explicit empty-table stop). - No stage points at a table that is not in `views`. ## Feeds Score Cycle Impact.
#2	Score Cycle Impact	Read the 'CAS-2 Score Cycle Impact' activity from playbook 'Delivery Constraint & Opportunity Analysis' (workflow 'Constraint and Flow Analysis'). Follow its guidance when assisting me. ## Purpose Reconstruct touch vs wait, score impact, and mark where AI could help. Do not design agents. ## Formula - wall-clock = touch_time + wait_time - effort = people_in_role × touch_time - impact = effort + (people_downstream × wait_time) ## Inputs - Filled `streamlit/data/value_stream.yaml`. - `sql/stage_touch_wait.sql` and skill Cycle Impact Scoring → Recipe. - Skill AI Leverage Mapping → Recipe. - Role capacity file `streamlit/data/role_capacity.yaml`. ## Steps 1. Open `streamlit/data/cost_of_delay.yaml`. For each stage on the map, set `touch_time` and `wait_time` from catalog events (issue → PR → CI → merge → deploy). Never treat wall-clock as effort. Label measured vs assumed. Pin `query_execution_id` from `sql/stage_touch_wait.sql` (workgroup from the contract; extract `$.payload.`). 2. Fill `streamlit/data/role_capacity.yaml`: people_in_role, people_downstream, source of the assumption. Do not* infer headcount from unique GitHub authors. 3. Join issue↔PR on `Closes #N` in payload, never `repo` or timestamps alone. 4. Compute effort, wall-clock, impact (worked sample in the YAML: Sec wait 48h × 8 downstream + 2h touch × 1 person = 386). Keep ToC #1 visible (`toc_rank_1: true`). 5. Fill `streamlit/data/ai_opportunities.yaml`: opportunity names, `ai_can_help` yes/no/unknown, effort S–XL. Low hanging fruit = high impact AND AI-yes AND low effort. No agent products. 6. Those three YAML files are artifacts Delivery Cycle Reconstruction, Cost of Delay Ranking, and AI Opportunity Prioritization Matrix. ## Done when - Ranking has impact scores with measured/assumed labels and one named #1 constraint. - Matrix has at least the high-impact boxes scored; no agent names. ## Feeds Look until trusted, then Triage: Eliminate / Automate / Delegate. ## Skip Do not optimize a non-constraint. Do not put AI on a low-impact stage because it is easy.
#3	Look until Trusted	Read the 'CAS-3 Look until Trusted' activity from playbook 'Delivery Constraint & Opportunity Analysis' (workflow 'Constraint and Flow Analysis'). Follow its guidance when assisting me. ## Purpose See the delivery cycle as people live it until the impact ranking can be trusted. This is not a query-console step. ## Inputs - `streamlit/data/cost_of_delay.yaml`, `ai_opportunities.yaml`, `role_capacity.yaml`. - Mart URL: `docs/deploy-manifest.yaml` → `streamlit_mart_url`. - Notebook `notebooks/look-until-trusted.ipynb`. - Skills EDA with Streamlit Data Mart, EDA with Pandas and Jupyter, MCP-Driven Conversational Analytics, Metric Render Catalog → Render pairings / Starter mart tabs. ## Steps 1. Open the mart URL. Home is Delivery cycle, then Opportunity quadrant. Feature Cycle Time (`pages/11_Feature_Cycle_Time.py`) is a check, not the ranking. 2. Look at the whole flow first. Use Metric Render Catalog pairings to challenge the #1 constraint (lead/cycle, throughput, MR hours, CI failure, load). Skip any family whose contract views are missing. 3. Challenge assumed numbers in `role_capacity.yaml`; relabel or remeasure. Jira `status_segments`: touch/wait may be measured (business days). Do not mix with PR hours. 4. Chat/MCP findings are hypotheses until reproduced in `notebooks/look-until-trusted.ipynb` (Restart & Run All). 5. Copy survivors into `streamlit/data/promoted_eda.yaml` (metric id, pairing, unit, n, window, agrees_with_ranking, query_execution_id, tsy_section). Empty file (`rows: []`) is valid if nothing earned TSY — that is not a Present Findings bug. ## Pitfalls - If `LoginRefreshRequired` / `CreateOAuth2Token` expired → `aws login`, restart kernel, not a Glue/Athena ACL bug. Skill Athena SQL Querying / EDA with Pandas and Jupyter. ## Done when - Impact ranking survives a second look (or was corrected in YAML). - Any finding that will inform triage has query text + QueryExecutionId. - Catalog checks that disagree with the ranking are explained or the ranking is updated. ## Feeds Triage: Eliminate / Automate / Delegate.
#4	Triage: Eliminate / Automate / Delegate	Read the 'CAS-4 Triage: Eliminate / Automate / Delegate' activity from playbook 'Delivery Constraint & Opportunity Analysis' (workflow 'Constraint and Flow Analysis'). Follow its guidance when assisting me. ## Purpose Decide eliminate vs automate vs delegate for ranked impact constraints, in that order, in writing. Tag cause layers (dE / dK / dI) and route each tag. ## Inputs - `streamlit/data/cost_of_delay.yaml` and `ai_opportunities.yaml`. - Skill Eliminate/Automate/Delegate Triage → Cause tags / Routes. - Template `streamlit/data/action_register.yaml`. ## Steps 1. Open `streamlit/data/action_register.yaml`. One row per ranked constraint, impact order. Fill eliminate first; automate only if eliminate is written no; delegate only if automate cannot encode the judgment. 2. The AI matrix may inform where automate or delegate could apply. It does not skip eliminate and it is not an agent backlog. 3. Automate shape: one Lambda + one Guardrail (policy / allow-list / human approval — name the check, not a product). 4. Tag cause layers on each row (multi-tag allowed). Skill → Cause tags: - dE environment (tool, process, incentive, data, time, authority) - dK knowledge/skill (trainable / selectable) - dI individual (selection, capacity, motivation) "Knows the right move, environment punishes it" stays dE, not dK. 5. Route per tag (skill → Routes): dE+eliminate → drop; dE+process/tooling → process/tooling (not AI SDLC); dE+automate/delegate → AI SDLC candidate; dK → Team Upskill Engine input (stage, clock, unit, n, role); dI → staffing. Do not author TUE modules. 6. Escalate #1 if neither eliminate nor automate works — flow redesign. Do not implement the Lambda in this activity. ## Done when - `action_register.yaml` has decision, cause tags, route, and evidence (clock, unit, n, role) for #1 and every ranked row you touched. - No dK row claims a finished curriculum. No dE process/tooling row is labeled AI SDLC unless automate or delegate was chosen. ## Feeds Tell the Data Story (publish this register — do not re-classify).