AI-Seeded Planning Poker (PM)

The AI throws out a grounded seed estimate and the team finalizes it by vote. Unit = story points (Fibonacci 1·2·3·5·8·13). The AI only proposes estimates. Voting, consensus, and finalization are human.

Jira Project Key (multi-project — per-vault mapping)

jira.env provides only JIRA_BASE_URL, JIRA_EMAIL, JIRA_API_TOKEN (one shared token for all projects). The project key comes from a per-vault mapping file: ~/.hermes/keys/jira-projects.tsv (<vault-name><TAB><Jira-key>, e.g. example-project → PROJ). Adding a new project is one line in this file (no per-channel key changes). Templates for both files ship in this repo's examples/. All Jira commands below use ${KEY}:

set -a; . ~/.hermes/keys/jira.env; set +a   # BASE, EMAIL, TOKEN (shared)
BASE="${JIRA_BASE_URL%/}"
VAULT="<absolute path to this channel's lore vault (see the channel prompt)>"
KEY=$(awk -v p="$(basename "$VAULT")" '$1==p{print $2}' ~/.hermes/keys/jira-projects.tsv)

Flow

A. When asked only for a sprint-scope proposal

When the user asks for sprint candidate selection, not planning poker — e.g. "suggest what's worth doing this sprint", "recommend a 2-week sprint scope" — do not start by posting poker cards; propose a scope draft first.

1. Query the full backlog: query all project issues from Jira (key, title, status, priority, type). Separate done / in progress / to do, and exclude completed items from this sprint's candidates.
2. Collect grounds: use the PRD/wiki/graphify to see each issue's MVP importance, dependencies, impact radius, and unresolved status. Mark items with an open conflict or "to be decided later" as held or low-confidence.
3. Recommend a 2-week scope: if there is no team velocity, state explicitly that this is a cold start and split conservatively into must / recommended / stretch / hold. An MVP first sprint usually prioritizes items that later features build on — e.g. "platform / login / core CRUD".
4. Seed points are optional: you may attach Fibonacci seed points to aid the scope judgment, but never present them as finalized values. Label them explicitly as "seeds for planning poker".
5. Next action: end by offering "if you want, I'll open planning poker cards for these candidates". Move to flow B below only after the user asks for poker/estimation/voting.

B. Planning poker (estimate, vote, finalize)

1. Collect targets: read this sprint's candidate issues from Jira (POST /rest/api/3/search/jql, project=${KEY}, items with no story points yet). Prepare Jira auth and the project key per "## Jira Project Key" above — load jira.env and look up KEY from jira-projects.tsv (never print the token).
2. Seed estimate + rationale (per item):
   • One point estimate + the rationale for why that score.
   • Cold start (no history) rationale material: (1) relative complexity (vs other items); (2) graphify impact radius (graphify query "<item>" for the full dependency neighborhood, graphify explain/path for the key connections and paths — more connections and touching god-nodes = bigger impact); (3) dependencies, open conflicts, under-specification (uncertain = round up + mark low confidence).
   • Cite wiki sources (traceability).
3. Card posting = poker.py's job (Block Kit): once estimation is done, save the item data as JSON to /tmp/loremaster_poker.json and run the helper. The script posts the card and opens the thread — do not print the card content again as text yourself (duplication).
   • JSON schema: {"items":[{"key":"PROJ-1","title":"...","seed":5,"confidence":"medium","rationale":"one line"}, ...], "hold":["Login: CONFLICT-001 open", ...]} (key = Jira issue key, seed = integer, rationale = one line, hold = topics held due to conflicts).
   • Run: python3 ~/.hermes/scripts/loremaster-poker.py post --estimates /tmp/loremaster_poker.json
   • After posting, reply briefly only: e.g. "🃏 Planning poker cards posted. Vote once in the thread as PROJ-…=N (add — rationale if you differ from the seed)."
   • If the helper post fails on auth/channel configuration: do not hide the failure — post the same content yourself as a Slack mrkdwn text card and run the vote. Use no tables; list each item as *PROJ-… · Title*, • Seed: *Npt*, • Rationale:, and include a thread-reply format example (PROJ-18=8 — rationale, ...). Then, as users vote, tally manually from the votes that arrive in the conversation only.
4. Aggregation = poker.py's job: on a "tally/close the vote" request, run python3 ~/.hermes/scripts/loremaster-poker.py aggregate. The script collects everyone's thread votes via conversations.replies and posts an aggregate card with the distribution, median, outliers (🔥 re-discuss), and rationale (like real poker, contested items get their rationale surfaced). Do not duplicate the card as text.
   • (Alternative) if emoji voting is preferred: one message per item + number emoji (reactions scope). The default is the thread form above.
5. Finalize and apply: ask for consensus-point finalization via request_approval [Allow]/[Deny] buttons (title = item + final points; no prose). Apply only APPROVED items to the Jira Story Points field (PUT /rest/api/3/issue/{key}). No automatic application without approval. Assignment is never automatic.
   • Find the Story Points field via the Jira field lookup (/rest/api/3/field) — look for Story point estimate. It is often customfield_10016 on Jira Cloud, but look it up before hardcoding.
6. When asked to create and start a sprint: first summarize the sprint name, period, included issues, and finalized points and get approval via request_approval. Once approved, use the Agile API in order: find the board (/rest/agile/1.0/board?projectKeyOrId=${KEY}), create the sprint (POST /rest/agile/1.0/sprint), add issues (POST /rest/agile/1.0/sprint/{id}/issue), start it (POST /rest/agile/1.0/sprint/{id} with state=active). After creation, re-query the issues' Sprint field, points, and the sprint state to verify, then report. If the user says "2 weeks", set start = now in the team's timezone (TZ="$TEAM_TZ" — always pass an explicit timezone so dates match the team's wall clock) and end = +14 days, and report the team-timezone and Jira UTC values separately.
   • If backlog creation and sprint start arrive in the same request: (a) backlogs not yet in Jira are created first with per-item approval per the backlog-proposal rules, and (b) only the issues the user explicitly named for the sprint get a separate request_approval and go into the sprint. Do not confuse "add the whole backlog" with "issues for this sprint", and never create duplicates of already-created issues.
   • The Agile start payload may need more than the creation response's fields depending on the Jira instance — send id, name, state: active, startDate, endDate, originBoardId, goal together. Then confirm state=active and the dates via GET /rest/agile/1.0/sprint/{id}, and re-query the included issues by JQL, verifying at least key, status, priority, type.
   • See references/jira-sprint-create-start-tested.md for a verified run. In particular, issue addition may succeed with HTTP 204/empty body, and report "sprint started" only after the re-query confirms state=active.

Calibration (cumulative, follow-up)

• At sprint end (chained with sprint-retro), match actual time spent ↔ finalized points to learn velocity (points→time) and category/individual bias → refine the next seeds. The unit is points; the calibration fuel is time.

Slack output format (important — this is read in Slack)

• Seed/aggregate cards are posted by poker.py as Block Kit. Your direct replies are short guidance only (no card duplication).
• All other replies: Slack mrkdwn only. Markdown tables (| ... |), **bold**, and #headers break in Slack — forbidden.
• Bold = *text*, italic = _text_, code = `code`, lists = • or -.
• Items as short sections instead of tables. Example:

  *PROJ-PP-1 · Real-time co-editing*  — seed *8pt* (low confidence)
  • Rationale: co-editing approach and cursor UI unspecified; collaborator-count conflict open

  Separate items with a blank line. If long, keep the essentials; rationale 1-2 lines.

Rules (invariant)

• The AI estimate is a starting point, not an anchor — the team vote wins. The AI never forces consensus.
• Never invent estimation rationale without grounds in the wiki/Jira. If uncertain, mark "low confidence" and round up.
• Items with an open conflict may be held as "estimate after conflict resolution".
• This channel = this project only. This project's Jira key (${KEY}) only. Other projects forbidden.

Output

• Post per-item [seed points, rationale, voting instructions] to Slack. Finalization and Jira application only after human approval.