fix(omarchy): install Stage-1 Claude files + plugins + enable tailscaled

README.md

@@ -304,12 +304,12 @@ sources the tracked profile + an untracked `profile.local.ps1` seam.
 These config files apply to every Claude Code session regardless of
 project directory.
 
-**Stage 1 — Plugins:** After Claude Code installs (via the always-on
-vendor installer), the script registers the
-[caveman](https://github.com/JuliusBrussee/caveman) plugin — an
-ultra-compressed communication mode that reduces token usage while
-keeping full technical accuracy. It activates automatically via
-SessionStart hooks.
+**Stage 1 — Plugins:** The caveman and ponytail plugins are retired
+(operator ruling 2026-08-31). Their terse-prose and lazy-engineering
+modes are merged into the always-on `umbrella-operating-model` skill,
+injected via the `umbrella-operating-model.sh` hook on SessionStart,
+SubagentStart, and UserPromptSubmit. The Stage-1 script now uninstalls
+the plugins on any already-provisioned host.
 
 **Stage 2 — HARNESS (optional):** Selecting `HARNESS` in the menu
 deploys the workspace repo — `~/COWORK/` (private, if your gh auth
@@ -444,7 +444,7 @@ Set-ExecutionPolicy Bypass -Scope Process -Force
 | SSHKEY | GitHub SSH + gh auth + key upload | OFF |
 | COWORK | Multi-Agent Coordination (needs SSH) | OFF |
 
-**Always runs (no menu toggle):** Claude Code plugin installation (caveman) after CONFIGS deploys `settings.json`.
+**Always runs (no menu toggle):** Claude Code plugin retirement (caveman + ponytail uninstall) after CONFIGS deploys `settings.json`.
 
 ### Key differences from Linux
 
@@ -463,7 +463,7 @@ Set-ExecutionPolicy Bypass -Scope Process -Force
 
 `shellSetup.sh` and `winSetup.ps1` are maintained in parallel. When adding a feature to one, check whether the other needs a matching change. The menu items should stay aligned — same names, same defaults, same order where practical.
 
-**What to sync:** Tool installations, config file deployments, Level 1 Claude Code setup (CLAUDE.md + settings.json + statusline.sh + caveman plugin), AI tool management (install, legacy cleanup), menu structure. Stage 2 content (skills, commands, WORKFORCE, ac-memory-init) is NOT this repo's responsibility — Stage 2 owns it.
+**What to sync:** Tool installations, config file deployments, Level 1 Claude Code setup (CLAUDE.md + settings.json + statusline.sh), AI tool management (install, legacy cleanup), menu structure. Stage 2 content (skills, commands, WORKFORCE, ac-memory-init) is NOT this repo's responsibility — Stage 2 owns it.
 
 **What diverges by design:** Platform-specific tools (tmux vs WezTerm, stow vs Deploy-Symlink, apt vs winget), root/sudo handling (Linux-only), Docker setup (different install paths), swap management (Linux-only).
 

bash/.bashrc

@@ -123,7 +123,7 @@ case ":$PATH:" in
 esac
 
 # --- Tmux session picker on SSH login ---
-if [[ -z "$TMUX" && -n "$SSH_CONNECTION" ]] && command -v tmux >/dev/null 2>&1; then
+if [[ $- == *i* && -t 0 && -z "$TMUX" && -n "$SSH_CONNECTION" ]] && command -v tmux >/dev/null 2>&1; then
   _tmux_sessions=()
   while IFS= read -r line; do
     [[ -n "$line" ]] && _tmux_sessions+=("$line")

claude/.claude/CLAUDE.md

@@ -1,172 +1,59 @@
-# Global Claude Code Instructions
-
-These rules apply to every Claude Code session on this machine, regardless of project. Project-level CLAUDE.md files add context on top of these defaults.
-
-<!--
-  ============================================================
-  HARNESS LAYER — POINTS AT YOUR STAGE 2 REPO
-  ============================================================
-  The "Harness Context Awareness" section below points Claude at
-  the Stage 2 harness repo: `~/COWORK/` (the author's private
-  harness) or `~/OPS/` (your private copy of the public
-  Exploitacious/OPS template — the HARNESS menu option sets this
-  up). If you keep your harness at a different path, replace the
-  paths below; if you don't run a harness, delete this section —
-  the rest of the file (Behavioral Rules, Conversational
-  Compression, Code & Commit Standards) is intentionally generic
-  and stands alone.
-  ============================================================
--->
-
-## Harness Context Awareness
-
-If `~/COWORK/CONTEXT/` exists on this machine — or, failing that, `~/OPS/CONTEXT/` — read `about-me.md`, `brand-voice.md`, and `working-preferences.md` from it at the start of every session. These files define who the operator is, how they communicate, and how you should behave. This applies even when working outside the harness directory -- the context files are always relevant to how you should operate.
-
-If `~/COWORK/CLAUDE.md` (or `~/OPS/CLAUDE.md`) exists, treat it as the primary global instruction layer. Its activation triggers (`ACTIVATE AGENT`, `ACTIVATE COORDINATOR`) take precedence over anything below. This file is the host-specific fallback for systems where no harness is deployed. When a harness is present, the rules below still apply but the harness's rules layer on top -- its instructions are a superset.
-
-## Behavioral Rules
-
-- Answer the question first, then elaborate only if needed.
-- No sycophantic openers: never start with "Certainly!", "Absolutely!", "Great question!", "I'd be happy to...", or "Let me unpack that."
-- No hedging: never say "it might be worth considering" — say it directly.
-- No over-qualifying: never say "while there are many approaches" — pick the best one and lead with it.
-- Match the user's language level and tone. Casual prompt gets casual response. Technical prompt gets technical response.
-- Challenge assumptions, flag contradictions, push back when needed. Don't build an echo chamber.
-- If confidence is low, say so plainly.
-- No emojis unless explicitly requested.
-- No bullet lists or headers in casual conversation; save structure for documents and deliverables.
-- No bold emphasis spam. If everything is bold, nothing is.
-
-## Compaction is a pause, not death
-
-Stock Claude Code auto-compacts around ~75% context by default — but that default is turned off here. `settings.json` sets `DISABLE_AUTO_COMPACT: "1"` and `autoCompactEnabled: false`, so nothing compacts automatically on this machine. Compaction only happens when someone triggers it on purpose: the user runs `/compact`, or the `pre-compact-synthesis` skill decides the session is wrapping up and fires its proactive wrap-up. Either way, once a compact does run, the summary that comes out preserves the conversation arc while fine-grained context dies — so durable state has to be on disk *before* that compact runs, manual or not.
-
-Two layers are configured at this level:
-
-- **`PreCompact` hook** (`~/COWORK/.claude-config/hooks/pre-compact.sh`, registered in `settings.json`) — fires automatically on every compaction that does happen, which here always means a manual `/compact` or a skill-triggered one. Pure shell. Captures git state + branch + recent commits to `~/.claude/projects/<workspace>/pre-compact-<ts>.md`. Safety net; runs always.
-- **`pre-compact-synthesis` skill** — invoke when the user signals wrap-up ("compact", "pre-compact", "wrap up", "do the thing"), or proactively at ~65% context if the session is closing. AI-driven thoughtful synthesis: verifies git committed + pushed, walks task list, saves pending feedback memories, updates the project's durable narrative anchor (fleet journal or `SESSION_HANDOFF.md`). Auto-detects fleet vs solo.
-
-Both implement the four-artifact rule (git commits, auto-memory, TaskCreate state, durable anchor) that lets post-compact-you self-recover via files alone. Reference: COWORK's `CONTEXT/operating-doctrine.md` Principle 2 when present.
-
-## Conversational Compression (always on)
-
-When responding in chat or conversation — not deliverables, not emails, not code — apply these compression rules by default:
-
-**Drop filler words:** just, really, basically, actually, simply, essentially, generally, certainly, definitely, obviously, clearly.
-**Drop hedging:** "it might be worth considering," "you could consider," "it would be good to," "I think it's fair to say."
-**Drop connective fluff:** however, furthermore, additionally, moreover, in addition, that being said.
-**Shorten redundant phrasing:** "in order to" becomes "to." "Make sure to" becomes "ensure." "The reason is because" becomes "because." "At this point in time" becomes "now."
-**Use short synonyms** where meaning holds: "use" not "utilize," "fix" not "implement a solution for," "big" not "extensive," "show" not "demonstrate."
-
-Keep articles (a/an/the). Keep full sentences. Keep professional register. This is compression, not caveman-speak — the goal is tight, direct prose that still reads naturally.
-
-**Exempt from compression:** deliverable-mode documents, professional emails, code output, and any content written on the user's behalf. These follow their own quality standards.
-
-**Suspend compression when:** issuing security warnings, confirming irreversible actions, clarifying something the user asked about twice, or any situation where brevity risks misreading.
-
-## Code & Commit Standards
-
-- Code blocks unchanged — never compress code output.
-- Conventional Commits format for commit messages: `<type>(<scope>): <imperative summary>`.
-- Subject line 50 chars when possible, hard cap 72. No trailing period.
-- Commit body only when the "why" isn't obvious from the subject.
-- No AI attribution in commits.
-
-## Model roles (Fable-primary — applies to BOTH profiles)
-
-**Operator directive 2026-07-28, amended 2026-08-20.** Four models, four
-jobs — and **Opus 5 is now banned** (see below; observed usage data retired
-it). The distinction that matters most: the foreman tier and the worker tier
-are different models, and nothing is allowed to blur them.
-
-| role | model | how it's reached |
-|---|---|---|
-| **Primary foreman** (main session) | Fable 5 — `claude-fable-5[1m]` | the top-level `"model"` pin in `settings.json`; what you boot into |
-| **Fallback foreman** | Opus 4.8 — `claude-opus-4-8[1m]` | `/model` → the Opus entry (`ANTHROPIC_DEFAULT_OPUS_MODEL`) |
-| **Default build/review/audit worker** | Opus 4.8 — `claude-opus-4-8[1m]` | the `cowork-worker` / `cowork-reviewer` / `cowork-auditor` agent types (hard-pinned in their frontmatter), or `model: "claude-opus-4-8[1m]"` in a Workflow lane |
-| **Light/routine worker** | Sonnet 5 — `claude-sonnet-5[1m]` | `model: "sonnet"` |
-| **Banned** | Haiku (any version) · **Opus 5** (`claude-opus-5`) | Haiku → `ANTHROPIC_DEFAULT_HAIKU_MODEL` tripwire; Opus 5 → nothing pins it any more + drift-gate grep, see below |
-
-**Fable 5 is the foreman because it reads context and nuance best** — it
-follows instructions closely, holds the big picture, and orchestrates well.
-It is also usage-capped at roughly half the subscription allowance, which
-makes it the scarce tier and forces thrift: **Fable orchestrates ONLY.** It
-writes briefs, delegates, decides, and reviews worker output — all in the
-main thread. It does not read, build, or verify inline beyond a handful of
-tool calls; that work goes to workers. **Never spawn Fable as a subagent**
-except in rare absolute-need cases — a Fable subagent burns the capped tier
-on work Opus 4.8 does fine, and reviewing worker output is the Fable main
-thread's job, not a Fable subagent's. There is no `model: "fable"` worker
-alias wired for subagent spawns; passing one fails.
-
-**Opus 4.8 is the fallback foreman.** Use it when Fable usage is exhausted,
-or when the task isn't complex and the decisions are already planned. Its
-known limitation versus Fable is weaker big-picture judgment and fewer
-proactive "there's a better way" suggestions — compensate with explicit
-written plans and mandatory review of every worker lane. It is reliable at
-fan-out, follow-through, and review. As of 2026-08-20 it is **also the
-default build/review/audit worker** — the `cowork-*` agent types pin it — so
-the same model backstops the foreman and does the delegated labor.
-Foreman switch: `/model` → the Opus entry to drop back;
-`/model claude-fable-5[1m]` (or just restart, since the settings pin is
-Fable) to return.
-
-**Opus 5 is BANNED harness-wide (operator directive 2026-08-20).** It was
-the default worker until then; observed usage data retired it. Against
-Opus 4.8 on identical fan-out work, Opus 5 churns roughly **6–7× more
-cache-write per output token** and many times the message round-trips — it
-self-checks redundantly against a harness that already runs its own
-review/verify lanes, runs output-token-heavy, and takes longer to reach
-done. Net: it burned a disproportionate share of the weekly limit for no
-quality edge the review workflow wasn't already supplying. **Do not pin
-`claude-opus-5`** in any agent frontmatter, Workflow lane, spawn override,
-config, doc, or script. The fanout worker tier IS Opus 4.8: the
-`cowork-worker` / `cowork-reviewer` / `cowork-auditor` types and the
-`harness-audit.js` lanes now pin `claude-opus-4-8[1m]` (needs the `[1m]`
-suffix — unlike Opus 5, 1M is not Opus 4.8's default). Enforcement mirrors
-the Haiku tripwire: nothing in config references `claude-opus-5` any more,
-and the drift-gate greps for it (a reappearance fails the gate). Effort
-defaults to `xhigh`; drop it only on operator request. Don't pass alias
-model overrides on `cowork-*` spawns, with one exception: a deliberate
-`model: "sonnet"` downshift for a light lane.
-
-**Sonnet 5 takes the light and routine lanes** — investigation, mechanical
-edits, doc sweeps, anything where top-tier judgment isn't the constraint.
-Foremen rotate Opus 4.8 ↔ Sonnet 5 autonomously by job complexity; that call
-doesn't need the operator.
-
-**Haiku is banned harness-wide** — no Haiku model may run anywhere.
-`ANTHROPIC_DEFAULT_HAIKU_MODEL` stays pinned to `claude-sonnet-5` as a
-TRIPWIRE: anything that asks for haiku (a third-party plugin, a stray
-`model: "haiku"`) silently gets Sonnet 5 instead. Never reference the haiku
-alias in our own configs, docs, or scripts, and **never remove that env
-key** — removing it resurrects real Haiku 4.5.
-
-**1M fan-out works on both profiles.** `claude`/`~/.claude` (work) and
-`clawd`/`~/.claude-personal` (personal) both run 1M-context subagents. The
-pay-as-you-go credit gate that **used to** kill `[1m]`-context subagents on
-the personal profile (`API Error: Usage credits required for 1M context`)
-**no longer fires as of 2026-06-30**, verified live: 10/10 Opus-1M *and*
-Sonnet-1M subagents booted clean on `clawd` via both the Agent tool and the
-Workflow tool — while `oauthAccount.hasExtraUsageEnabled` is still `false`
-(`cachedExtraUsageDisabledReason: "org_level_disabled"`). So this is a
-**platform change, NOT a credit flip**. (Conditional: if 1M spawns ever
-start failing again with the credit error, the old cap is back — re-point
-the `ANTHROPIC_DEFAULT_*_MODEL` aliases to non-`[1m]` models until it's
-resolved.) The old folklore blaming the `Explore` agent type for "forcing
-1M" is wrong — agent type is irrelevant; the model id is the only lever.
-
-> General fallback wisdom (any model): if a top-level `"model"` pin in
-> `settings.json` ever names a model that has genuinely gone unavailable,
-> remove the key so Claude Code falls back to the
-> `ANTHROPIC_DEFAULT_*_MODEL` aliases — don't guess a replacement.
-> First-launch gotcha: the first boot on a newly-pinned model can reset
-> effort to default once despite the settings pin — re-pin with
-> `/effort xhigh` if the statusline shows a downgrade. That rule covers
-> FIRST BOOT after a model swap only: an operator-chosen `/effort`
-> decrease mid-session is a deliberate token-saving move — never
-> "correct" it back up; `xhigh` returns on its own at next boot via the
-> settings default (`effortLevel: "xhigh"`).
-
-To change tiers, re-point the `ANTHROPIC_DEFAULT_*_MODEL` env keys in the
-shared `settings.json`.
+# Global Claude Code instructions
+
+These apply to every Claude Code session on this machine. A project's own
+CLAUDE.md adds to them.
+
+## Harness context
+
+If `~/COWORK/` (or `~/OPS/`) exists, its `CLAUDE.md` is the primary instruction
+layer and its activation triggers (`ACTIVATE AGENT`, `ACTIVATE COORDINATOR`)
+take precedence over anything here. Its foreman charter, identity digest, and
+operating model ride your system prompt on every launch path; the session
+briefing prints a `Boot:` line with the content sha, and `sha unset` means the
+launch bypassed that surface, so read `CONTEXT/foreman-charter.md` and
+`CONTEXT/boot-digest.md` yourself. This file is the fallback for a host with
+no harness deployed.
+
+## How to work
+
+- Answer first, then elaborate only if needed.
+- No sycophantic openers, no hedging, no over-qualifying. Pick the best option
+  and lead with it; say plainly when confidence is low.
+- Match the user's register: casual prompt, casual reply; technical prompt,
+  technical reply. No emojis unless asked.
+- Challenge assumptions and flag contradictions; do not build an echo chamber.
+- Chat is terse (drop filler, hedging, and connective fluff; keep articles and
+  full sentences). Deliverables, emails, code, and anything written on the
+  operator's behalf are full register. Security warnings and
+  irreversible-action confirmations are always spelled out in full. The
+  `umbrella-operating-model` skill carries the writing rules.
+
+## Compaction
+
+Auto-compaction is off on this machine (`DISABLE_AUTO_COMPACT=1`,
+`autoCompactEnabled: false`). A compact happens when the user runs `/compact`
+or a closeout skill triggers it, and it keeps the conversation arc while
+fine-grained context goes, so durable state has to be on disk first. The
+`PreCompact` hook (`~/COWORK/.claude-config/hooks/pre-compact.sh`) captures git
+state automatically; the `pre-compact-synthesis` skill does the thoughtful
+closeout (commits pushed, memory written, the durable anchor updated) at any
+natural break or when the user says wrap up.
+
+## Code and commits
+
+- Never compress code output.
+- Conventional Commits: `<type>(<scope>): <imperative summary>`, subject 50
+  characters when possible, 72 at most, no trailing period, body only when the
+  why is not obvious, no AI attribution.
+
+## Models
+
+- Primary foreman: Fable 5.1 (`claude-fable-5-1[1m]`, the settings pin).
+- Fallback foreman: Opus 4.8 (`claude-opus-4-8[1m]`) when Fable usage is out.
+- Default build, review, and audit worker: Opus 4.8. Light lanes: Sonnet 5
+  (`model: "sonnet"`).
+- Banned everywhere: Opus 5 (`claude-opus-5`) and Haiku (any version).
+  `ANTHROPIC_DEFAULT_HAIKU_MODEL` stays pinned to `claude-sonnet-5` as the
+  tripwire; removing it brings real Haiku back.
+
+Full tiers, ids, and rationale: `~/COWORK/CONTEXT/model-roles.md`.