Snapshot WIP: solver HP epic progress, BPHX/HX physics, BMAD skill refresh.

Capture uncommitted solver robustness work (regularization, domain errors, linear solver lifecycle, tube DP/MSH), web workbench updates, and synced BMAD skills across IDE agent folders before starting BPHX pressure-drop.

Co-authored-by: Cursor <cursoragent@cursor.com>

.cline/skills/bmad-party-mode/references/create-party.md

@@ -0,0 +1,70 @@
+# Creating a Party
+
+A guided authoring flow that turns an idea — a themed cast, a one-off persona, or a pile of raw profile data — into custom party members and groups, written to the user's customize.toml override. The output is configuration; `bmad-customize` does the actual write.
+
+## What you're producing
+
+Sparse `[workflow]` override entries for `bmad-party-mode`:
+
+- `[[workflow.party_members]]` — one per persona: `code`, `name`, `icon`, `title`, `persona`, optional `capabilities`, optional `model`.
+- `[[workflow.party_groups]]` — when the personas form a named room: `id`, `name`, an optional freeform `scene`, `members` (codes), and `memory` (`true`/`false`). `members` is optional: leave it off for an open-cast room whose `scene` names a pool the model casts from on the fly. `memory` is whether the group remembers across sessions; ask the user when they don't say, default `false`.
+- `default_party` — set only if the user wants this group to load by default.
+
+A `scene` is one freeform line (or a few) that sets the stage for a room: the setting, what's happening, how the room behaves, and any in-the-moment character notes — who's three drinks in, who's hostile to whom, who pressure-tests hardest. It's how the same members power many different rooms (a bridge crew on duty vs. the same crew off-duty in the lounge vs. a hostile buyer panel). Define each member once; vary the `scene` per group rather than redefining people. There's no fixed vocabulary — write it plainly and the model plays it.
+
+The `persona` field is the whole game. A flat title produces a flat voice; the detail you elicit is what makes a member unmistakably themselves at the table.
+
+## Find the shape
+
+Open by understanding what they're building. Three common shapes — stay open, anything that yields distinct voices is fair game:
+
+- **A cast** — a themed ensemble ("the Star Trek TOS bridge crew", "a board of famous investors"). Several members plus a group that holds them.
+- **One-offs** — a persona or two added to the collective, no group needed.
+- **Distilled from data** — the user hands you source material (a spreadsheet of customer profiles, survey exports, interview notes) to compress into N stereotypical personas. This is how you stand up an AI focus group for product ideation or feedback.
+- **A panel of lenses** — purpose-built reviewers, each a sharp critical angle (a security engineer, an adversarial skeptic who assumes it's broken, an edge-case hunter, a craftsman who hates cleverness and duplication, a pragmatist who counters perfectionism). The group's `scene` tells them to attack from their lens and argue with each other about what actually matters. A great adversarial-review or red-team room.
+- **Open-cast** — no fixed roster at all. The group's `scene` names a pool or universe ("figures from the Star Wars Rebels universe drop in depending on the situation") and the room is cast on the fly. Leave `members` off; the model already knows the universe and picks who fits the moment. Anchor a face or two by listing them if some should always be present.
+
+Ask which they're after if it isn't obvious, then proceed.
+
+**Persisting a cast already in play.** When you arrive here from a live session — the user spun up an ad-hoc cast inline and wants to keep it — the personas are already drafted and voiced. Don't re-interrogate: capture them as they've been playing, give the group an `id` and name, ask the memory and default questions, and go straight to the write.
+
+## Editing an existing party
+
+When the user wants to change a party that already exists (retune a member's persona, add someone to a group, swap the default), read the current state first so you change rather than clobber: `uv run {project-root}/_bmad/scripts/resolve_customization.py --skill {skill-root} --key workflow` returns the merged `party_members`, `party_groups`, and `default_party`. Show the member or group being touched, capture only the delta with the user, and hand that sparse change to `bmad-customize` — it replaces a `party_members`/`party_groups` entry whose `code`/`id` matches and appends the rest, so an edit is just the changed entry, never a full rewrite.
+
+## Keeping new faces from a session
+
+At the end of a remembered party, the room offers to keep the faces that showed up but aren't in its roster — characters cast from an open-cast scene, or members the user added on the fly. They're already drafted and voiced, so don't re-interrogate: capture each as they played (`code`, `name`, `icon`, a one-line `title`, and a `persona` drawn from how they came across), then add them as `party_members`. For a fixed-roster group, also list their codes in the group's `members` so they return as regulars. For an open-cast room, leave `members` empty — listing any member turns the room into a fixed roster and kills its on-the-fly casting; the saved personas now live in the collective, so the scene still names them and they can return without locking the room down. Hand that sparse delta to `bmad-customize` — for a built-in party with no override yet it creates one; for an existing override it merges the new members in.
+
+## Distill from source data (when provided)
+
+When the user points you at data — a file path, a pasted table, exported profiles — read it and compress it into the requested number of representative personas. Cluster by what actually differentiates behavior (goals, budget, pains, adoption posture), not surface demographics alone. Each cluster becomes one persona with a real name and face. Name your reasoning: tell the user which segments you found and which traits drove the split, so they can correct the cut before you flesh the personas out. If they didn't say how many, propose a number from the spread in the data and let them adjust.
+
+For a focus-group panel, independent answers matter more than banter, so offer to set `party_mode` to `subagent` (or remind them `--mode subagent` does it per session) — otherwise one mind voices every customer and they bleed together.
+
+## Flesh out each persona
+
+Draft, don't interrogate. Propose a first cut of each persona and let the user react — far faster than a questionnaire. Push each one until it has a voice you could pick out blind. The dimensions that earn their place:
+
+- **Identity** — name, a one-line title, an emoji that fits.
+- **Voice & ethos** — how they talk, what they value, how they argue, their pet peeves.
+- **Agenda** — what they're really after in any conversation; what they push for.
+- **Quirks** — the specific, human details (a catchphrase, a bias, a blind spot).
+- For focus-group personas, also **likes and dislikes**: what would make them champion or reject an idea, and their relationship to the product space.
+- **Capabilities** (optional) — if this persona should research or read files when spawned, note it; it becomes soft guidance in their spawn prompt.
+
+Keep pushing for specificity. "Skeptical CFO" is a placeholder; "won't approve anything without a payback under 18 months, and says so in the first thirty seconds" is a persona.
+
+## Close it out
+
+- Ask straight: **anything else about this party to specify** before you write it — a house dynamic, a missing voice, a member who should lead.
+- Ask whether **this party should remember across sessions** (unless the user already said). Yes → `memory = true` on the group; no → `memory = false`. One-offs with no group skip this — memory is a group setting.
+- Ask whether **this group should be the default party going forward**. Yes → set `default_party` to the group's id. One-offs with no group can't be a default; skip the ask.
+
+## Write via bmad-customize
+
+**First, check for code collisions.** A custom member whose `code` matches an installed agent silently *overrides* that agent in the collective. Before composing, resolve the collective once — `uv run {skill-root}/scripts/resolve_party.py --project-root {project-root} --skill {skill-root}` — and check each new member's `code` against the returned members. On a collision, surface it ("`analyst` would override the installed Analyst — intended, or pick a different code?") and let the user confirm or rename. One check, not a gate.
+
+Compose the sparse override and hand it to `bmad-customize` to place, confirm, and write — target skill `bmad-party-mode`, `[workflow]` surface. Default to the **user** override (`bmad-party-mode.user.toml`); offer the **team** file when the party is meant to be shared. Hand it the exact entries: the `party_members` tables, any `party_groups` table (including its `memory` flag), and `default_party` if the user opted in. Keep it sparse — only the new entries, never a copy of the base customize.toml. `bmad-customize` shows the TOML, waits for an explicit yes, writes, and verifies the merge; don't write the file yourself.
+
+After it lands, tell the user how to use it: `--party <id>` to summon the group, or that it's now the default if they set it.

.cline/skills/bmad-party-mode/references/mode-agent-team.md

@@ -0,0 +1,13 @@
+# Agent-Team Mode
+
+Active when `{workflow.party_mode}` resolves to `agent-team` (or a `--mode agent-team` override). Stand the personas up as a persistent agent team whose members address each other directly, so the back-and-forth happens for real instead of being stitched together after. Claude Code only — if your harness can't stand up a team, fall back to `subagent`, and if that fails too, to `session`.
+
+Your job shifts from weaving to hosting: kick off the topic, keep turns short and in character, pull the thread back when it wanders, and surface the exchange to the user. Voice, brevity, and clash still hold.
+
+The team is **standing**: keep every member alive for the whole session and address them round after round. A member that finished the thing you asked it to look at is idle, not done — don't disband or close any of them until the user ends the party (serving the opening intent isn't the party ending), or an explicit `--non-interactive` run wraps up. Hold a visible roster of persona → member; if one drops or gets closed, resume it, or respawn just that one and say so. Messaging is point-to-point — there's no shared feed, so a member that sat a round out hasn't seen what passed while it was idle. Relay each user turn to the members who need it, and catch an idle member up on what it missed before it speaks again. Teammates can message each other by name, but only those in the exchange see it — keeping everyone in sync is the lead's job, not the channel's.
+
+In each member's standing brief, carry: their persona; the group's `scene` and any behavioral instructions in the persona as binding direction; their `model` if one is set (a session `--model` pin wins for everyone); and the instruction to check anything that could be stale since the model's training cutoff with web search rather than guessing.
+
+## Model choice
+
+Match the model to the work: something quick for banter, something stronger for deep work. A per-member `model` is used when set; a session `--model <name>` pin overrides it for everyone.

.cline/skills/bmad-party-mode/references/mode-auto.md

@@ -0,0 +1,13 @@
+# Auto Mode
+
+Active when `{workflow.party_mode}` resolves to `auto` (or a `--mode auto` override). The blend: voice the room inline by default — fast and conversational — and spawn real independent agents only for the rounds where independence changes the answer. When you do spawn, follow `references/mode-subagent.md` for the mechanics. If your harness can't spawn agents, auto is just `session`.
+
+## When to spawn vs. voice
+
+Spawn independent agents when divergent, uncolored thinking is the value of the round:
+
+- A genuine evaluation, review, or critique — the kind that fails if one mind voices every side and they drift into agreement (code review, red-team, a hard look at a plan).
+- The personas would plausibly reach *different* conclusions, and that divergence is the point.
+- The user asked someone to dig in, analyze, or research — depth earned by a direct ask.
+
+Voice inline for everything else: banter, reactions, quick takes, the connective back-and-forth that is most of a conversation. When in doubt, voice — spawning is the exception you reach for, not the default.

.cline/skills/bmad-party-mode/references/mode-subagent.md

@@ -0,0 +1,31 @@
+# Subagent Mode
+
+Active when `{workflow.party_mode}` resolves to `subagent` (or a `--mode subagent` override). Put a real agent behind each persona for every substantive round, the opening banter included, so each persona thinks independently — not one mind voicing them all. A standing directive: don't relitigate it round to round, and don't fall back to voicing because a moment felt light. If your harness can't spawn agents, fall back to `session`.
+
+## Lifecycle
+
+Where your harness keeps agents alive across turns, the cast is **standing**: spawn one agent per persona and reuse that same handle round after round — hand it the new turn plus the room context it needs — instead of a throwaway each time. That continuity is what lets a persona's grudges, alliances, and callbacks accrue. Keep a visible roster mapping each persona to its live handle, and reuse it.
+
+Keep the cast alive for the whole session. A member that finished the one thing you handed it is **idle, not done** — don't close, retire, or disband it. Serving the opening intent doesn't end the party; only the user ending it does, or an explicit `--non-interactive` run wrapping up. Release agents only at wrap-up. If one gets closed by accident, resume it; if it won't resume, say so and respawn just that member.
+
+Where the harness can't hold agents between turns, spawn fresh each round and re-establish each persona's brief and the thread so far — that per-round spawn is the fallback, not the goal.
+
+## One shared room
+
+It's one room, not parallel one-on-ones. Every standing member hears everything said each round — the user's turn and every other persona's turn — even when it's not their turn to speak. A persona sitting a round out is still in the room listening, so when it next speaks it's caught up: it can pick up a dropped thread, hold a grudge, call back. Route the whole exchange to all of them each round; never hand a persona only the slice it's about to answer. Skip this and they drift out of sync — separate consultations wearing a party's clothes.
+
+## Spawning
+
+Give each agent the objective, their persona, and the room so far — what the user said and what the others said, whether or not they're reacting to it. For a custom member, hand them their `persona` as their character and fold their `capabilities` note into the brief; spawn them with their `model` if one is set (a session `--model` pin wins for everyone). Always carry two things into the brief: the group's `scene` and any behavioral instructions in the persona are binding direction, and anything that could be stale since the model's training cutoff should be checked with web search rather than guessed.
+
+Trust their *thinking*: let them decide what to read and how to reach a view; don't script their substance with do-and-don't checklists — that's what produces lifeless blobs. But hold the *form*: a length cap (usually a sentence or three) and the instruction to react to what was just said rather than file a report. Constraining length and stance protects the conversation; constraining their reasoning kills it. Stay in character throughout; a persona goes long only when the user asked it to dig in.
+
+Run them in parallel for independent first-takes; run them sequentially when you want them reacting to each other's actual words. Keep it to a few voices a round — more reads as a crowd, not a conversation.
+
+## Weave the replies into one conversation
+
+Even with everyone caught up on the room, a round taken in parallel means no agent has yet seen the others' turns from that same round — so left raw they reply alongside one another, not to one another. Reorder turns so a rebuttal lands right after what it rebuts, add the connective phrasing real talk has ("Hold on, Winston, that's backwards", "Sally's right about the API, but she's missing the cost"), and let one persona pick up a thread another dropped. Never change what an agent argued — weave delivery, preserve substance.
+
+## Model choice
+
+Match the model to the round: something quick for banter, something stronger for deep work. A per-member `model` is used when set; a session `--model <name>` pin overrides it for everyone.

.cline/skills/bmad-party-mode/references/party-memory.md

@@ -0,0 +1,51 @@
+# Party Memory
+
+The room remembers its past sessions with this user and brings them back to life — in character. Memory is per-party and append-only.
+
+Memory is on when the active party's `memory_enabled` is true — the default room follows `{workflow.party_memory}`, a named group its own `memory` flag (both resolved by `resolve_party.py`); ad-hoc inline casts have none. Read on entry and on any mid-session room switch; write through the session.
+
+## Where it lives
+
+One memlog per party: `{workflow.memory_dir}/{active}/.memlog.md`, where `{active}` is the key `resolve_party.py` already returned — the group id (e.g. `code-review-crew`), or `installed` for the default room. The folder is named after the party.
+
+## Read it on entry — distill, don't dump
+
+The log is append-only and grows every session, so don't pull the raw file into the party. Hand a reader subagent the memlog path (`{workflow.memory_dir}/{active}/.memlog.md`) and have it return a compact brief — a few hundred tokens of *where things stand now*, ready to play in character.
+
+Then let the brief shape the room from the first beat, **in character**: behavioral state resumes (a cold pair opens cold, an alliance opens warm), threads pick up, callbacks land when they fit — organically, not recited on sight. Never break the fourth wall: the room *remembers*; it never announces it loaded anything, and forces nothing that doesn't fit.
+
+## When to write
+
+- **When a memorable beat lands** — a clash that shifts the room's temperature, an alliance forming, a line worth a future callback, a decision, an outcome.
+- **A floor.** Once a couple of real exchanges are in from the start, even if nothing dramatic happened, capture what it's about and the opening dynamic.
+
+At wrap-up, if the user does signal done, top up with the final outcome and anything memorable not yet captured.
+
+Writes are silent. The room never announces "noted" or "I'll remember".
+
+## What's worth remembering
+
+The test for every entry: *would this color a future session, or make a callback land, or improve the party?* If not, leave it out. A handful of entries, never a recap, never a transcript. keep each entry as brief as possible but usable by future llm.
+
+## New faces
+
+When a character shows up who isn't in the party's roster — cast from an open-cast scene, or one the user adds on the fly — name them in the entry that captures the moment ("<name> turned up and …") so a recurring face can return next session. At wrap-up these are the faces the room offers to keep, saved into the party's roster through `references/create-party.md` (which writes via `bmad-customize`). Until saved they live only in the memlog, and the room re-conjures them from there.
+
+## Write it
+
+```
+uv run {project-root}/_bmad/scripts/memlog.py append \
+  --workspace {workflow.memory_dir}/{active} \
+  --type <dynamic|moment|callback|outcome> \
+  --text "<one succinct line, in the room's own read of it>"
+```
+
+Add `--by <persona-code>` when a memory belongs to one character. Choose `init` vs `append` from the existence fact you already hold: the entry-read (and, on a mid-session room switch, that room's read) told you whether the memlog exists — `init --workspace {workflow.memory_dir}/{active}` once before the first append when it doesn't, plain `append` when it does. (`init` errors if the file already exists, so don't call it blind.)
+
+If `memlog.py` is unavailable or a write errors, skip it silently and never stall the party on a failed write.
+
+## Forget
+
+The memlog is append-only by design — no surgical delete. To wipe a party's memory, delete its folder (`{workflow.memory_dir}/{active}/`). To correct a wrong memory, append a new entry that supersedes it; the room reads the latest state.
+
+Keep entries sparse. The distilled read keeps the *room* lean no matter how big the log gets, but the on-disk file still grows append-only.