/simulate-meeting

/persona-studio:simulate-meeting

$ARGUMENTS

Preamble — session grounding flag check

if [ "$(.venv/bin/python -m persona_studio.grounding.session status)" = "disabled" ]; then
    GROUNDING_ENABLED=false
    echo "[grounding] disabled for this session — skipping the grounding pipeline"
else
    GROUNDING_ENABLED=true
fi

When GROUNDING_ENABLED=false: SKIP the [EVIDENCE BANK] retrieval, verify_claims calls, Tier-2 external-verification fallback, Tier-3 CoVe pass, and post-meeting audit.py invocation. The meeting still runs and produces a plain transcript. Users re-enable via /persona-studio:studio → Toggle factual grounding.

Step 0 — Normalize inputs

Same participant validation as /persona-studio:simulate-debate (including the dual-scope Glob over ./personas/*.md + $HOME/.persona-studio/personas/*.md with project-priority dedup, and project-priority agent-file lookup). 2-6 participants allowed.

Agent invocation strategy: Default to dynamic persona pattern — subagent_type="general-purpose" with the persona file body inlined into the prompt. See /persona-studio:simulate-debate Step 0.5 for the full recipe and rationale. This works for both pre-existing and session-fresh personas, unlike native persona-<slug> subagent types which only work if the file existed at session start.

Ask how the agenda is composed (AskUserQuestion):

• Auto-generate — Main Claude drafts an agenda from the topic.
• Enter manually — user dictates 3-7 agenda items.

Record the final agenda as an ordered list.

Step 1 — Facilitator sets the stage

Main Claude, acting as the facilitator, announces:

• Meeting topic
• Participants
• Agenda list
• Flow: for each agenda item the facilitator nominates one participant to lead, then picks one more for a rebuttal / complementary view, and closes the item with a consensus summary

Step 2 — Agenda loop

For each agenda item i in order: a. Facilitator phrases a concrete opening question (≤ 2 sentences). b. Pick a lead participant most relevant to item i based on their persona's Knowledge Domains. Retrieve the lead's evidence bank for this agenda item: bash .venv/bin/python -c " from persona_studio.grounding.retriever import retrieve_evidence chunks = retrieve_evidence('<lead>', '<topic> <item>', k=8) for c in chunks: print(f'- [{c.source}:{c.line_start}-{c.line_end}] {c.text[:240]}') " Then invoke with the evidence bank appended to the prompt: Agent(subagent_type="persona-<lead>", prompt="[EVIDENCE BANK]\n<retrieved>\n\n[CITATION RULE] For stats/dates/quotes, cite from EVIDENCE BANK or mark [OPINION]/[INFERENCE]. Never fabricate sources.\n\nGive your recommendation for agenda item i in under 300 characters. Topic: <topic>. Item: <item>.") c. Pick one challenger participant with a contrasting Debate Style or overlapping domain. Invoke: Agent(subagent_type="persona-<challenger>", prompt="Read the lead's answer and, in under 300 characters, rebut the weakest point or add what's missing. Lead's answer: <quote>") d. Optional: if either mentioned an absent domain expert among participants, invoke that third persona with a direct question (one exchange max). e. Tier-2 external verification pass for both lead and challenger replies: Same pattern as /persona-studio:simulate-debate Step 3.5 — pipe each reply through python -m persona_studio.grounding.verify_claims with --only-high-risk, call the Tier-2 tool (Perplexity if available, WebSearch otherwise) for any UNVERIFIABLE high-risk claims, and insert [VERIFIED-EXTERNAL: ...] or [UNVERIFIED-EXTERNAL] tags inline. e2. CoVe pass on residual UNVERIFIABLE high-risk claims (same Chain-of-Verification flow as /persona-studio:simulate-debate Step 3.6).

 Before the first agenda item, initialize a per-avatar budget:
 ```bash
 COVE_BUDGET_PER_AVATAR=5
 # for each participant p:
 declare "COVE_USED_<p>=0"
 ```

 For each claim in the lead's and challenger's `CLAIMS_JSON` where
 `status == "unverifiable"` after Tier-2, `is_high_risk == true`, and
 `COVE_USED_<avatar> < COVE_BUDGET_PER_AVATAR`:

 1. Generate the verification question:
    ```bash
    Q_JSON=$(printf '%s' "<claim.text>" \
        | .venv/bin/python -m persona_studio.grounding.cove generate-question)
    ```
 2. Independently answer it via a sub-Agent
    (`subagent_type="general-purpose"`), prompt:
    ```
    Answer the following question strictly from the provided evidence.
    If the evidence does not contain the answer, reply with the single
    word: unknown.

    Question: <Q_JSON.question>
    Evidence: <EVIDENCE BANK for this avatar> + <Tier-2 search result, if any>
    ```
    Capture as `$COVE_ANSWER`.
 3. Compare:
    ```bash
    CMP_JSON=$(.venv/bin/python -m persona_studio.grounding.cove compare \
        --claim "<claim.text>" --answer "$COVE_ANSWER")
    ```
 4. Branch on `$CMP_JSON.verdict`:
    - `"discrepancy"`: insert `[FACT-CHECKER CHALLENGE: <reason>]` into
      the transcript line after the flagged claim AND re-invoke the
      avatar with a retract/defend prompt:
      "Your previous statement '<claim.text>' could not be verified.
      Evidence suggests <reason>. Please retract, cite a source, or
      restate with uncertainty. Under 200 characters." Append the
      response as a `[retract/defend]` blockquote line, marking the
      original portion `(original, flagged)`.
    - `"inconclusive"`: leave the Tier-2 annotation; no challenge.
    - `"consistent"`: no action.
 5. Increment `COVE_USED_<avatar>` by 1.

 When a CoVe budget is exhausted for an avatar, annotate remaining
 eligible claims `[UNVERIFIABLE — CoVe budget exceeded]` and stop
 calling CoVe for that avatar.

f. Facilitator synthesizes a 2-3 sentence decision summary + action item candidate.

Step 3 — Meeting close

Main Claude writes:

• Consensus summary (bullet list)
• Action items (owner best-guess, due date TBD)
• Follow-up questions

Step 4 — Save transcript

Write simulations/<UTC-timestamp>_meeting_<topic-slug>.md:

---
kind: meeting
topic: <topic>
participants: [...]
agenda: [...]
generated: <ISO8601>
---

# Minutes

## Agenda 1: <title>
- **Lead (<lead>)**: ...
- **Rebuttal / complement (<challenger>)**: ...
- **Additional remarks**: ...
- **Decision summary**: ...

## Agenda 2: ...

## Conclusion
Every simulation MUST include this section. Write the single clear conclusion the meeting reached in 1-3 sentences. If no consensus was reached, still conclude with "Consensus failed — reason X".

## Satisfaction
- Score: N / 10 (against the criteria the user set at the start)
- Per-criterion score:
  - Criterion 1 (user input): X/10
  - Criterion 2: X/10
  - Criterion 3: X/10
- Retry on miss (Ralph mode only): yes/no

## System Feedback
3-5 improvement points grouped by category:
- **Persona side**: which persona sections should be refined
- **Process side**: agenda design / facilitation improvements
- **Platform side**: tooling / environment / UX improvements

### Consensus · Actions · Follow-ups
Action items table:
| # | Owner | Content | Due |
|---|-------|---------|-----|

Open questions:
- ...

Schema rule (mandatory):

• The three H2 sections ## Conclusion, ## Satisfaction, ## System Feedback MUST be present. If missing, the docs pipeline warns and asks the user to rewrite.
• ## Closing Summary is an optional parent section that may wrap the three sections above together with the Consensus · Actions · Follow-ups subsection, OR each of the three may stand as an independent H2 — both patterns are valid.

Step 5 — Generate Word (.docx) and PowerPoint (.pptx)

Auto-invoke the docs pipeline right after the markdown transcript is saved:

.venv/bin/python scripts/simulation_to_docs.py simulations/<stem>.md

Expected output: same directory, <stem>.docx and <stem>.pptx.

• .docx: full-fidelity archive — every utterance verbatim, facilitator questions and summaries, full agenda and closing summary. For internal minutes or archival distribution.
• .pptx: executive-summary deck — cover + agenda overview + one-line highlight per agenda item + closing consensus + action items. For internal sharing / follow-up presentations.

If the script fails, surface the last 20 lines of its output and ask the user (AskUserQuestion) whether to retry, skip, or open the markdown directly.

Step 6 — Report to user

Show three paths:

• simulations/<stem>.md (markdown source)
• simulations/<stem>.docx (full archive)
• simulations/<stem>.pptx (summary deck)

Then return control to the caller (usually /persona-studio:studio).