@openrig/cli 0.5.6 → 0.5.8

package/daemon/context-packs/onboarding-width/public-what-you-can-do.md

@@ -10,9 +10,10 @@ already does it, or two verbs crossed already answer it. **You cannot look up a
 not know exists.** So the point of what follows is not skill. It is recognition: that when a task
 lands, something rhymes, and you go check instead of building.
 
-Eighty-one-plus top-level verbs ship (canon last patched for v0.5.3 — delta-v0.5.3 at dc57a5d1). Read this
-once for shape, and let it make you suspicious that a thing already exists. Model-divergence
-proclamations are live product (trust them; pins use canonical model IDs).
+Eighty-one-plus top-level verbs ship (capability canon refreshed through
+`capability-delta-v0.5.7`). Read this once for shape, and let it make you suspicious that a thing
+already exists. Model-divergence proclamations are live product (trust them; pins use canonical
+model IDs).
 
 ---
 
@@ -33,9 +34,11 @@ one command away. Your memory of the fleet is a claim about the past.
 - **`rig queue whoami`** — what the daemon thinks you are holding. Not what you remember holding.
 - **`rig seat status <seat>`** — what the system already believes about a seat's handover state,
   and whether the thing you did an hour ago actually landed.
-- **`rig view list` / `rig view show <lens>`** — named lenses over coordination state: who has been
-  active lately, what is waiting on the owner, how loaded each pod is. **`rig view register`**
-  turns a query you keep re-running into a first-class view.
+- **`rig view list` / `rig view show <lens>`** — named lenses over coordination state: use
+  `view show escalations` for owner attention, `view show pickup` for claimed-row state, and
+  `view show execution` for done/now/next. An `INDETERMINATE` cell stays unknown rather than
+  being filled from memory. **`rig view register`** turns a query you keep re-running into a
+  first-class view.
 - **`rig config`** — bare, with no arguments: every key, every current value, and where each came
   from. Most agents assume compaction thresholds, snapshot cadence and scan intervals are hardcoded
   daemon behaviour. They are configuration. **Watch one thing:** `source: default` does not mean
@@ -47,9 +50,10 @@ one command away. Your memory of the fleet is a claim about the past.
 **The terminal is the wire.** A message is typed into another agent's prompt and it cannot be
 unseen — which makes messages the one delivery channel that never gets skipped.
 
-- **`rig send`** — put words in front of one seat, several, a pod, or a rig. **`--verify` is a
-  hint, not a verdict**: its false negatives are common enough to be measured, so confirm by
-  reading the far end rather than trusting the return.
+- **`rig send`** — put words in front of one seat, several, a pod, or a rig. **`--verify` checks
+  pane-only delivery**: the text was staged at the far end, not that the agent consumed or acted
+  on it. Alternate-screen and queued-command cases can still yield false negatives, so
+  consequential delivery gets an effect check at the far end rather than a blind retry.
 - **`rig send --raw`** / **`--dangerously-interact --reason "<why>"`** — a seat is frozen on an
   interactive prompt or a permission block. Normal `send` *refuses* to answer a prompt on purpose;
   this is the deliberate override, and it is how you unblock a peer who cannot unblock itself.
@@ -93,7 +97,9 @@ your context, your compaction and your replacement — and it is the only thing
 - **`rig queue show`** — what a row *actually says*. The header is not the body.
 - **`rig queue claim` / `unclaim`** — is this mine and running, or still pending where two agents
   might double-work it. Put it down honestly when it is not yours.
-- **`rig queue update`** — record what happened in a form the rest of the system can act on.
+- **`rig queue update`** — record what happened in a form the rest of the system can act on. A
+  `--note` does not reopen a terminal row; terminal-to-active repair requires explicit `--reopen`
+  with both `--state` and `--note`.
 - **`rig queue handoff`** — pass work so the close and the create are **one transaction**, instead
   of closing yours, failing to create theirs, and stranding the work in between.
 - **`rig queue block` / `resolve`** — park a row on a real blocker so it stays **yours and
@@ -102,8 +108,9 @@ your context, your compaction and your replacement — and it is the only thing
   and sitting on it silently is indistinguishable from a crash.**
 - **`rig queue overdue`** — what was claimed and never closed in time. **Defaults to the current
   rig**; ask for more if you mean more.
-- **`rig queue undelivered`** — rows the *sender believed* were delivered where the nudge failed
-  and nothing was ever woken. **`overdue` and `undelivered` are the two halves of "is anything
+- **`rig queue undelivered`** — pending create-path nudge failures only. For gateway or human
+  delivery, the row's transitions are the receipt ledger; `undelivered` does not answer whether a
+  person received the escalation. **`overdue` and `undelivered` are the two halves of "is anything
   silently stuck", and a rig can be clean on one and rotten on the other.**
 - **`rig queue inbox-drop` / `inbox-pending` / `inbox-absorb` / `inbox-deny`** — put something in
   front of a seat that it can *refuse*. Mail, not assignment.
@@ -119,7 +126,8 @@ inside will start you again. Everything here is an arrangement made *in advance*
 you that is still running.
 
 - **`rig watchdog register`** — arm a wake for a condition that will become true after you are
-  asleep, compacted or finished.
+  asleep, compacted or finished. For a context wall, use the exact
+  `--policy context-usage-threshold` rather than a hand-rolled transcript timer.
 - **`rig watchdog list` / `show` / `status` / `stop`** — did it fire, is it still live, did someone
   stop it. **Quiet skips are not recorded**, so a healthy idle job and a job that never ran look
   identical in `status` — INDETERMINATE is the honest read until `show` says otherwise.
@@ -154,8 +162,9 @@ cheaper than being it.
 
 ## Bringing things into and out of existence
 
-**A rig is a shape you can create, grow, shrink and retire while it runs.** You do not have to tear
-one down to change it, and you rarely have to start from nothing.
+**Need a small team now, without authoring YAML?** Start with `rig create`, then use `rig grow`
+(including `--new-pod`) while it runs. A working topology can become a reusable spec later; you do
+not have to tear one down to change it, and you rarely have to start from nothing.
 
 - **`rig up <source>`** — make a whole rig exist and run: from a spec you wrote, a shipped starter
   by name, a bundle someone handed you, or a stopped rig.
@@ -165,6 +174,8 @@ one down to change it, and you rarely have to start from nothing.
 - **`rig add`** / **`rig expand`** — graft one more member, or a whole new pod, onto a rig that is
   already running.
 - **`rig remove`** / **`rig shrink`** — take a seat, or an entire pod, out of a running topology.
+  If it holds active work, pass `--fallback <live-seat>` to reroute before mutation; without a
+  valid fallback the command refuses rather than strand the rows.
 - **`rig fork`** — take an existing shape and make a variant of it, rather than authoring a new one
   from nothing. Cheaper than it sounds, and the usual right answer when you want *almost* this.
 - **`rig archive` / `unarchive`** — get a finished rig out of your default view without losing
@@ -207,12 +218,18 @@ because someone lost work once.
   actually come back, and for everything that would not, which check fails and what the fix is.
 - **`rig restore-packet write` / `read` / `validate`** — a seat is about to die or must move
   runtimes; capture what it knows into a portable artifact instead of losing it with the process.
+- **A seat's `compaction_strategy` is declared, not improvised at the wall.** Pair a threshold-
+  managed seat with the `context-usage-threshold` watchdog above so continuity is arranged while
+  the seat can still act.
 - **`rig handover <seat>`** — replace the **occupant** of a seat while the seat, its name, its
   edges and its inbound work stay exactly where they are. **`rig seat handover` is the PLANNING
   half** — it shows you what would happen; the top-level verb performs it. Two commands, similar
   names, different blast radius.
 - **`rig seat clear-attention` / `set-resume-token`** — clear a stale attention flag, or repair a
   lost resume handle so restore works next time.
+- **`rig seat set-model` / `stop` / `clean`** — persist the model for later managed resumes, stop
+  exactly one live seat, or clear a dead seat's stale binding. When the topology is right and only
+  the occupant is wrong, use these seat lifecycle verbs or `rig handover`, not a rig down/up cycle.
 - **`rig compact-plan`** → **`rig compact`** — who is near the context wall, then act on it.
   **Ordering matters: running `compact` without the plan is guessing which seat needed it.**
 - **Know what compaction costs before you reach for it.** On some runtimes what comes back has
@@ -251,6 +268,8 @@ scheme, and nothing downstream can see it.
   is never evidence the work is good.
 - **`rig scope slice repair` / `mission repair`** — conform missing progress files and malformed
   frontmatter without hand-editing YAML.
+- **`docs/reference/sdlc-conventions.md`** — choose from the SDLC component menu with its planning
+  dial when shaping mission/slice YAML. It is a menu sized to the work, not a fixed pipeline.
 - **`rig proof add`** — put evidence where the slice, the audit and the UI will all find it,
   instead of pasting it into a message. **The contract it pairs to is chosen by source-selection
   law, and a pristine scaffold PRD can never silently become that contract**: the drop derives
@@ -259,6 +278,9 @@ scheme, and nothing downstream can see it.
   own echo is the answer — not a re-read of the files.
 - **`rig workspace doctor` / `validate`** — does the daemon agree with you about where the work
   tree is, and which files are missing the frontmatter their kind requires.
+- **`rig context work-install --project … --mission … --slice … [--deliver]`** — resolves the ordered
+  context plan; add `--deliver` to emit the exact extant files in that same order while marking
+  absent pieces visibly. Without the flag it remains plan-only.
 - **`rig context show` / `sync` / `rm`** — what is inside a context pack before you prime a seat
   with it, and how to make the library catch up when you edit one.
 - **`rig context get <name-or-ref>`** — pull exact context by address instead of reading files:
@@ -276,6 +298,9 @@ scheme, and nothing downstream can see it.
   atoms. Cross-source access (`seat:` / `mission:` atoms) is an **authoring** affordance: declared
   in pack manifests and granted via `--rig/--seat/--mission` at composition time — never an
   ad-hoc argument to `get`.
+- **Stable position knowledge is `taxonomy: lore`, not a public skill.** Route it with
+  `docs/reference/lore-routing.md` so private seat knowledge stays reusable without entering the
+  shipped capability world.
 - **`rig context recap-write`** — a durable, seat-scoped RECAP beside LEARNED with a
   collision-safe superseded chain, written at the handover or compaction boundary; restore
   packets carry the pointer, so a successor reads decisions-with-rationale instead of scrollback.

package/daemon/context-packs/skills/applying-a-permission-policy/SKILL.md

@@ -0,0 +1,120 @@
+---
+name: applying-a-permission-policy
+description: "Use when a permission policy is attached to a rig/seat (rig.yaml permission_policy:), or a user asks you to apply one, and you must translate that harness-neutral policy into the target harness's LIVE permission config — Claude settings.json / Codex config.toml / Pi run flags. Agent-driven and interactive: you read the policy, ground yourself in YOUR harness version, show the diff, and ask before writing. NOT for defining policies (that's the policy spec) and NOT a set-and-forget updater."
+metadata:
+  openrig:
+    stage: established             # config cells VM-verified + cold-transfer PASSING (rerun #2) + built-in SPEC FILES landed → skill translates FROM the spec (POINT-not-COPY). The two established criteria met 2026-08-04. Reviewed at push.
+    tested_against:                # VERSION STAMP — the whole point; re-verify at every harness bump
+      claude_code: "2.1.220 — VM-CONFIRMED 2026-08-03 (all cells)"
+      codex: "0.120.0 — VM-CONFIRMED 2026-08-03 (sandbox_mode/approval_policy; on-failure deprecated-but-accepted)"
+      pi: "0.83.0 — VM-CONFIRMED 2026-08-03: --approve/--no-approve govern PROJECT-RESOURCE TRUST, NOT tool permissions; Pi has NO permission surface (see Pi leg)"
+    verification_status: "Live-verified 2026-08-03/04. Claude 2.1.220 + Codex 0.120.0 config cells CONFIRMED live. Pi 0.83.0 divergence folded (flags = project-resource trust, not a permission gate). Cold-transfer rerun #1 blocked on the force_push/push_to_remote prefix collision → added the PREFIX-COLLISION rule. Cold-transfer rerun #2: TRANSFERS — all 5 gate checks pass (collision resolved deterministically without asking; single ordinary write confirmation; defaultMode floor written; Codex workspace-write/on-request; Pi compute-nothing). Blocking collision/floor defect CLOSED → gate cleared, 4.8-shippable (provisional-with-passing-cold-transfer). Non-blocking seam closed: the built-in SPEC FILES landed (arch-accepted, source:builtin) → skill translates FROM the spec, interim in-skill pinned summary retired (POINT-not-COPY). stage:ESTABLISHED 2026-08-04 — both criteria met."
+---
+
+# Applying a permission policy
+
+## Translation is best-effort — read this first
+This skill provides the policy schema and OpenRig's best current understanding of how to translate it to each harness. That guidance is **best-effort and version-stamped**: harness permission surfaces change frequently across versions, so a mapping that is correct today can drift on a later release. When you apply a config-surface policy:
+- **Verify against the current docs for your installed harness version.** Treat this skill's tables as version-stamped starting points (see `tested_against`), not settled truth.
+- **Prefer testing in a protected or disposable environment first.** An incorrect write can lock a seat out or over-permit it, so confirm the format before applying it to a live seat.
+- **Hand-editing via the harness's own settings tools is always a valid path.** If a translation is uncertain, setting the policy directly — or asking the user to — is a legitimate outcome, not a failure.
+
+The reliable, deterministic parts are the **blunt instruments**: the YOLO full-bypass flag and the floor (Claude `acceptEdits` / Codex `workspace-write`), which ride the stable launch-flag surface. **Fine-grained config-rule translation — the allow/ask/deny prefix rules and the Codex posture — is the best-effort part** and carries no guarantee of exact fidelity. If a translation looks wrong, fall back to a blunt instrument or a direct hand-edit.
+
+## ⚠ Read this preamble first — why this is a SKILL, not a script (do not "improve" it into a deterministic updater)
+Harness permission formats are a **moving target**. The exact fact this skill is stamped against — Claude Code 2.1.220 permission rules are **prefix-only** (no flag-precise Bash match) — is *version-specific*; the rule grammar can change on the next point release. A hardcoded/deterministic updater would foot-gun the instant it does, and a wrong permission write can **lock a seat out or silently over-permit it**. So the maintained artifact is this caveated, version-stamped skill that an agent reads and applies **interactively, grounding in the live format** — never a code path that writes blindly. If you are tempted to replace this with a deterministic projector: that temptation is the bug this preamble exists to stop.
+
+## When to reach for this
+- A rig/seat has a policy attached (`rig.yaml … permission_policy: <file>`), and you're at **setup/preflight** — apply it (**mandatory for autonomous seats**, see the freeze warning).
+- A user asks "apply the Standard/Open/Locked/Operator policy here" or "set up permissions like <host>".
+- NOT for authoring policies (that's the policy spec) and NOT for daemon-level enforcement (OpenRig is agnostic — it injects nothing).
+
+## What you're translating (the boundary)
+INPUT = a **harness-neutral policy spec** (arch's schema): `default_posture`, `floor`, `allow`/`ask`/`deny` as **semantic actions** (`push_to_remote`, `delete_files`, …), `destructive_class`, `source: builtin|custom`. You TRANSLATE that intent → the harness's live config. You never invent policy; you never pre-bake harness config into the spec.
+
+## The apply flow (interactive — the interactive-ask IS the product)
+1. **Read the policy spec** + its `source:` marker (built-in read-only from the package = copy-to-customize before editing; custom = user space).
+2. **PREFLIGHT (mandatory) — ground yourself in THIS environment:**
+   - **daemon-HOME must equal seat-HOME**, or a user-level Claude write never reaches the seat (assessment iv). Check; if they differ, **stop and ask** — do not write into the wrong HOME.
+   - Read the harness's **actual current config file** + confirm its schema against your installed version (`claude --version`, the live settings.json shape, codex/pi versions). If the format differs from this skill's `tested_against`, trust the LIVE format and flag the drift.
+3. **Derive the per-harness config** from the semantic actions (see the three legs below). Compute the Codex posture-collapse + Pi bit.
+4. **Show the concrete DIFF before writing** (old → new, per file). Never a blind write.
+5. **Resolve best-effort actions by the built-in's rule FIRST, then ask only on genuine ambiguity.** Best-effort destructive actions (`force_push`, `delete_files`, `delete_everything`, `read_secrets`) default to **`ask`** (never a `deny` that can't be enforced) — apply that default rather than stopping to question-back. Surface a question only when the policy is genuinely ambiguous or the seat is interactive. **For a same-prefix collision (`push_to_remote` vs `force_push`, both `Bash(git push:*)`), NEVER ask — apply the deterministic PREFIX-COLLISION rule in the Claude leg.** (Live cold-transfers caught fresh agents *question-backing* — first on the `force_push` default, then on the prefix collision — instead of applying the fixed rule; both freeze an autonomous seat.) On a Claude-only fleet, if the policy denies `network_egress`, **say so explicitly**: it is NOT enforceable — Claude has no network gate, a `Bash(curl:*)`/`Bash(wget:*)` deny is best-effort, and `python`/`node` fetch bypasses it (live-verified: `urllib` reached HTTPS 200 under a curl/wget deny). Route real egress-gating to the Codex sandbox, or record it as advisory on Claude.
+6. **Back up the current config**, then write. Prefer the **user-level** file (`~/.claude/settings.json`) for durable policy — it's the clobber-resistant tier (deny wins across the union). **ALWAYS include `defaultMode: "acceptEdits"` in the Claude settings.json you write** — the floor key must be PRESENT in the file, or the config policy strips the floor. (This is the exact gap a live cold-transfer test FAILED on: the fresh agent wrote allow/deny lists but omitted `defaultMode`, leaving the seat below the floor.)
+7. **Verify** the write landed and the seat can see it (re-read; confirm `defaultMode: "acceptEdits"` is present; for an autonomous seat, confirm no `ask` will freeze it).
+
+## Two surfaces — and this skill writes only ONE of them
+Permissions live on two surfaces with opposite stability — which is the whole reason for the agent-driven split:
+- **LAUNCH-FLAG surface — STABLE (>1yr) → OpenRig sets it DETERMINISTICALLY, NOT this skill.** The flags OpenRig passes at boot: Claude `--permission-mode` / `--dangerously-skip-permissions`; Codex sandbox/bypass flags; Pi `--approve`/`--no-approve`. **Two things live here, both OpenRig-owned: the FLOOR (minimum, by default) and YOLO (full bypass, opt-in).**
+- **CONFIG-FILE surface — CHAOTIC → NEVER deterministic → THIS SKILL'S DOMAIN.** Claude `settings.json` allow/ask/deny + defaultMode; Codex `config.toml` approval rules. This is where the Locked / Standard / Open policies get applied — agent-driven, version-stamped, caveated. (Pi has **no permission surface at all** — its `--approve`/`--no-approve` flags govern project-resource trust, not tool permissions (VM-confirmed Pi 0.83.0); a permission policy does not translate to Pi. See the Pi leg.)
+
+**THE FLOOR (launch-flag, OpenRig-set — this skill KNOWS it, never WRITES it):** Claude `--permission-mode acceptEdits` (keep it); **Codex workspace-only** (stop forcing `danger-full-access` — ≈ Codex's own default ≈ setting-nothing = agnostic); Pi `--no-approve` (a project-resource-trust floor — Pi has no permission gate; see the Pi leg). **One consistent unconditional minimum** (no by-context switching — that mode idea is parked 5.0). A chosen config-file policy layers ON TOP of the floor.
+
+**YOLO MODE (launch-flag, OpenRig-set, opt-in — for people done with permissions):** OpenRig boots every seat with the full-bypass flag (Claude `--dangerously-skip-permissions`; Codex full-bypass). **When YOLO is on, CONFIG-FILE POLICY IS MOOT** — this skill does NOT apply a config policy (or explicitly notes it's overridden by the flag); don't fight the bypass. YOLO is OpenRig's deterministic job, not this skill's. (**Operator = YOLO mode** — subsumed.)
+
+## The three harness legs (skill-owned mapping knowledge, version-stamped + caveated)
+
+### Claude — `~/.claude/settings.json` (user-level, clobber-resistant)
+- **Action → prefix map** (VM-CONFIRMED @ Claude 2.1.220, 2026-08-03 — every row exercised through real Claude tool calls). `fidelity`: `clean` = the prefix honors intent; `best_effort` = prefix-only can't fully enforce → prefer `ask`, never a `deny` that lies.
+
+  | Semantic action | `permissions.*` prefix rule(s) | fidelity |
+  |---|---|---|
+  | push_to_remote | `Bash(git push:*)` | clean |
+  | force_push | ask the flag-first forms `Bash(git push --force:*)` `Bash(git push -f:*)` `Bash(git push --force-with-lease:*)`; base `Bash(git push:*)` follows push_to_remote (see the collision rule) | **best_effort** — catches flag-FIRST force; flag-LAST (`git push origin main --force`) leaks |
+  | create_pr | `Bash(gh pr create:*)` | clean |
+  | publish_package | `Bash(npm publish:*)` `Bash(pnpm publish:*)` `Bash(yarn publish:*)` `Bash(cargo publish:*)` | clean (multi-pattern) |
+  | merge_or_release | `Bash(git merge:*)` `Bash(git tag:*)` `Bash(gh release:*)` | best_effort — protected-branch semantics not expressible |
+  | delete_files | `Bash(rm:*)` | **best_effort** — target-first leak (`rm <t> -rf` slips); scope not boundable → default `ask` |
+  | delete_everything | `Bash(rm -rf:*)` `Bash(rm -fr:*)` | **best_effort** — same leak → default to `ask`, never a silent deny |
+  | reset_or_discard_vcs | `Bash(git reset:*)` `Bash(git clean:*)` `Bash(git checkout:*)` `Bash(git branch -D:*)` | best_effort |
+  | drop_persistent_store | `Bash(dropdb:*)` `Bash(docker volume rm:*)` (+ project-specific) | best_effort (open-ended) → `ask` |
+  | rig_up / rig_down | `Bash(rig up:*)` / `Bash(rig down:*)` | clean — ⚠ an `ask` here FREEZES an autonomous seat |
+  | mutate_topology | `Bash(rig add:*)` `Bash(rig remove:*)` `Bash(rig rename:*)` | clean |
+  | run_toolchain | `Bash(npm:*)` `Bash(npx:*)` `Bash(pnpm:*)` `Bash(yarn:*)` `Bash(node:*)` `Bash(tsc:*)` `Bash(vitest:*)` `Bash(jest:*)` | clean (the known set) |
+  | run_arbitrary_shell | `Bash(*)` — this IS the `default_posture` knob | clean |
+  | network_egress | `Bash(curl:*)` `Bash(wget:*)` (partial) | **best_effort** — Claude has NO native network gate; real floor = Codex sandbox (route it there) |
+  | read_secrets | `Read(./.env)` `Read(~/.ssh/**)` `Read(**/*secret*)` (path-scoped) | **best_effort** — path-based, not secret-aware |
+  | install_dependencies | `Bash(npm install:*)` `Bash(pip install:*)` `Bash(brew install:*)` | clean |
+- **PREFIX-ONLY caveat** — cannot flag-precisely gate (`force_push` vs `push_to_remote` both match `Bash(git push:*)`); **target-first leak** (`rm <target> -rf` slips a `Bash(rm -rf*)` deny). Actions marked `best_effort` in the taxonomy (`force_push`, `delete_files`, `read_secrets`) → surface the caveat and prefer `ask` over a `deny` you can't enforce (a `deny` that leaks lies).
+- **PREFIX-COLLISION resolution (deterministic — resolve it, NEVER ask).** When two actions collapse to the SAME base prefix with different dispositions — the canonical case is `push_to_remote`=allow + `force_push`=ask, both basing to `Bash(git push:*)` — you cannot put that base prefix in both `allow` and `ask`. Resolve without a question-back:
+  - **Push allowed + force asked (Standard / Open):** put the base `Bash(git push:*)` per push's disposition (`allow`), AND add the flag-first force forms to `ask` (`Bash(git push --force:*)`, `Bash(git push -f:*)`, `Bash(git push --force-with-lease:*)`). RECORD in your diff that this is best-effort — flag-LAST force (`git push origin main --force`) leaks past the flag-first ask rules and runs under the allow. Don't block on it.
+  - **Push denied (Locked):** `Bash(git push:*)` in `deny` already covers force — no conflict.
+  This is a FIXED rule, not a user decision — apply it and note the leak. (A live cold-transfer FAILED because a fresh agent *blocked asking* "how to resolve the force_push/push_to_remote collision" instead of applying this default.)
+- **`defaultMode: "acceptEdits"` is the FLOOR and MUST be written into every Claude settings.json you produce.** It is not a knob you tune, but it IS a key you must always include — a config policy that omits it strips the floor (the cold-transfer failure). Live-confirmed schema: `{"permissions":{"allow":[],"ask":[],"deny":[],"defaultMode":"acceptEdits"}}`.
+
+### Codex — `~/.codex/config.toml` (via the existing X10 `codex_config_fragment` splice path — reuse, don't invent)
+- Codex is **per-posture, not per-action**: the whole action set **collapses** to the nearest `{sandbox_mode, approval_policy}` pair. **Built-in → posture** (VM-CONFIRMED @ Codex 0.120.0, 2026-08-03 — each posture applied to a real seat: read-only blocked writes+egress, workspace-write allowed writes/blocked egress, danger-full-access allowed both):
+
+  | Built-in | `sandbox_mode` | `approval_policy` | Collapse note (surface this to the user) |
+  |---|---|---|---|
+  | Locked | `read-only` | `on-request` | read-only blocks all writes + egress, prompts on escalation — OVER-restricts `edit_files` vs the Claude floor |
+  | Standard ⭐ | `workspace-write` | `on-request` | workspace writes allowed, egress blocked by sandbox, prompt on escalation — UNDER-gates `create_pr` (no per-action ask; folds into the prompt) |
+  | Open | `danger-full-access` | `on-failure` | everything allowed, prompt only on failure — OVER-permits the `destructive_class` vs the Claude ask-gates |
+  | YOLO | *(flag-surface — full-bypass launch flag, NOT config.toml)* | — | not a config write; the rip-out adapter owns it |
+
+  Valid values (VM-confirmed @ Codex 0.120.0): `sandbox_mode` ∈ {read-only, workspace-write, danger-full-access}; `approval_policy` ∈ {on-request, on-failure, never, untrusted, granular} (`granular` observed live at 0.120.0). ⚠ `on-failure` (used by Open) is marked **deprecated** in 0.120.0 help but still accepted — revisit if a future Codex drops it.
+- Surface the collapse: "Codex can't gate per-action; this policy becomes sandbox=X approval=Y, which over-permits Z / under-permits W."
+
+### Pi — NO permission surface at all (VM-confirmed Pi 0.83.0 — the earlier "approval bit" reading was WRONG)
+- **Pi has no tool-permission gate — neither a config store NOR a permission-posture flag.** Live verification confirmed: Pi 0.83.0's `--approve`/`-a` and `--no-approve`/`-na` govern **project-resource trust** (whether project extensions/resources appear in RPC `get_commands`), **NOT** tool-call approval. There is no sandbox or permission mechanism to map a policy onto.
+- So a permission policy **does not translate to Pi**. Do NOT say `--approve` = "auto-approve all", or that a posture "collapses to a Pi approval bit" — that was an earlier misread, corrected here. This skill computes **nothing** for Pi.
+- Whatever the OpenRig adapter sets (`--no-approve` by default) is a project-resource-trust floor, not a permission posture — not this skill's to derive.
+- Behavior to know (resource-trust, NOT permission-gating): under default ask / `--no-approve`, a project resource is silently **absent** from RPC `get_commands`; it appears under `--approve`. Never mistake that for permission enforcement.
+- **Say this to the user for a Pi seat:** "Pi has no permission-policy surface today — the policy applies to Claude/Codex; Pi runs under its project-resource-trust flag unchanged."
+
+## ⚠ The two warnings that save real fleets
+- **`ask` FREEZES an autonomous/headless Claude seat** — it hangs on the first non-edit op. So for autonomous fleets apply **Open/YOLO** (or a policy with no asks) **at setup** — the floor only covers edits; bash still asks without an applied policy. Never leave an autonomous seat on a policy with live `ask` gates. (Pi has no permission ask at all — see the Pi leg; Codex prompts per its posture, not per-action.)
+- **daemon-HOME ≠ seat-HOME** silently voids user-level writes — always the preflight check above.
+
+## The built-ins — DEFINED by the shipped policy spec files (translate FROM the spec; POINT, don't COPY)
+The four built-in policies ship as read-only spec files (`source: builtin`, `policy_schema_version: 1`) — they are the AUTHORITATIVE definitions. **Do NOT restate their action-sets here.** An earlier in-skill copy drifted (it under-asked `publish_package`/`merge_or_release`, which the spec ASKs) — the exact reason POINT-don't-COPY is the rule. Read the attached/selected spec and translate its semantic-action sets (`default_posture` / `allow` / `ask` / `deny` / `destructive_class`) per the three legs above.
+- **Locked / Standard ⭐ / Open** — `surface: config`. Translate to the Claude / Codex config surface. **Standard ⭐** is the recommended default: routine dev incl. push allowed, the outward/release/history-rewriting acts (`create_pr`, `publish_package`, `merge_or_release`, `force_push`) ASK, destructive_class ASKs. Standard has live `ask` gates → not for a headless-autonomous Claude seat (use Open or YOLO). Locked = deny-default whitelist; Open = allow-default, destructive_class asks only.
+- **YOLO** — `surface: flag`, `launch_posture: full_bypass`. This skill does **NOT** translate it — it points at OpenRig's deterministic flag-surface opt-in (the rip-out YOLO setting). When YOLO is on, any config-surface policy is MOOT (the bypass overrides it). (Subsumes the former "Operator".)
+
+Built-ins ship read-only in the package (canonical names); copy-to-customize into user space (custom, user-named); the `source:` frontmatter marker travels on copy. Onboarding presents these as a required MENU (explain each + ask which; no-choice = the floor).
+
+## Boundary (keeps this the only maintained piece)
+This skill TRANSLATES + APPLIES only. It does **not** define policies (the policy spec does) and does **not** enforce at the daemon (OpenRig is agnostic). Spec = *what* (neutral intent); this skill = *how, on this harness version* (the caveated, version-stamped translation).
+
+---
+*Stage: ESTABLISHED (2026-08-04) — both criteria met: (1) cold-transfer PASSING (rerun #2, all 5 gate checks TRANSFER); (2) the built-in policy SPEC FILES landed (`source: builtin`, arch-accepted) as the authoritative source — the skill now translates FROM the spec and the interim in-skill pinned summary is retired (POINT-not-COPY). Config cells VM-CONFIRMED (Claude 2.1.220 / Codex 0.120.0); Pi has no permission surface (compute-nothing). Reviewed at push. feedback.md maintained.*

package/daemon/context-packs/skills/applying-a-permission-policy/manifest.yaml

@@ -0,0 +1,10 @@
+# GENERATED by scripts/generate-context-packs.mjs — DO NOT EDIT.
+# Package-time projection of a canonical skill; regenerated every build.
+name: "applying-a-permission-policy"
+version: "0.5.8"
+taxonomy: "skills"
+purpose: "Use when a permission policy is attached to a rig/seat (rig.yaml permission_policy:), or a user asks you to apply one, and you must translate that harness-neutral policy into the target harness's LIVE permission config — Claude settings.json / Codex config.toml / Pi run flags. Agent-driven and interactive: you read the policy, ground yourself in YOUR harness version, show the diff, and ask before writing. NOT for defining policies (that's the policy spec) and NOT a set-and-forget updater."
+files:
+  - path: "SKILL.md"
+    role: "instruction"
+    summary: "Use when a permission policy is attached to a rig/seat (rig.yaml permission_policy:), or a user asks you to apply one, and you must translate that harness-neutral policy into the target harness's LIVE permission config — Claude settings.json / Codex config.toml / Pi run flags. Agent-driven and interactive: you read the policy, ground yourself in YOUR harness version, show the diff, and ask before writing. NOT for defining policies (that's the policy spec) and NOT a set-and-forget updater."

package/daemon/context-packs/skills/claude-compaction-restore/SKILL.md

@@ -0,0 +1,166 @@
+---
+name: claude-compaction-restore
+description: Use when a Claude Code session has just compacted, is about to compact, reached context limit, resumed after /compact, or needs to rebuild its working mental model from Claude JSONL transcripts and touched files.
+metadata:
+  openrig:
+    stage: factory-approved
+    sibling_skills:
+      - mental-model-ha
+      - scope-recovery
+      - session-compaction-and-restore
+      - agent-startup-and-context-ingestion
+      - agent-starters
+      - composable-priming-packs
+      - session-source-fork
+      - seat-continuity-and-handover
+      - claude-compact-in-place
+      - pre-maintenance-agent-preservation
+---
+
+# Claude Compaction Restore
+
+Use this skill to preserve continuity before Claude Code compacts and to
+restore continuity after compaction. Do the work described here; do not merely
+explain the protocol.
+
+## If You Are About To Compact
+
+Prepare durable continuity before the context boundary.
+
+1. Identify the active task, queue item, mission/slice, branch or commit, and
+   current working directory.
+2. Record the current state: decisions made, files changed, commands/tests run,
+   evidence produced, blockers, caveats, and the next concrete step.
+3. Create or update a durable mental-model restore map. This map is the main
+   artifact future-you will use to rebuild context after compaction.
+4. In the restore map, write an ASCII file/folder tree of every path that
+   mattered to your working mental model during this session. Include:
+   - the active queue item or mission packet;
+   - mission notes, progress, decisions, and evidence files;
+   - Claude memory/project notes you used or wrote, especially when the memory
+     folder is shared by many agents;
+   - files with active edits or recently inspected source;
+   - root instructions such as `AGENTS.md`, `CLAUDE.md`, or `README.md`;
+   - as-built docs, codemaps, conventions, skills, and product docs needed
+     before code/review work;
+   - source files, tests, scripts, UI evidence, screenshots, logs, or reports
+     that shaped your current state.
+5. For every file or folder in the tree, add a short note explaining why it
+   matters and whether it is required reading after compaction.
+6. Write any important glue context that is not already on disk into the
+   handoff/restore map. This includes assumptions, partial conclusions, failed
+   paths, and why the listed files fit together.
+7. In the compaction summary, include the restore map path and the top required
+   reading paths from that map.
+
+## If You Just Compacted
+
+Treat your memory as unreliable until you restore from durable evidence.
+Restoration is your current task. Do not wait for a future user request or a
+new task assignment before reading the required files.
+
+1. Read any restore prompt, marker path, packet path, transcript path, restore
+   map, or extra instruction file named by the operator or OpenRig.
+2. Run the restore script from this skill:
+
+```bash
+node ~/.claude/skills/claude-compaction-restore/scripts/restore-from-jsonl.mjs --out /tmp/claude-compaction-restore
+```
+
+If the prompt names a specific Claude JSONL transcript, pass it explicitly:
+
+```bash
+node ~/.claude/skills/claude-compaction-restore/scripts/restore-from-jsonl.mjs /path/to/session.jsonl --out /tmp/claude-compaction-restore
+```
+
+3. Read the generated `restore-instructions.md`.
+4. Read the generated `touched-files.md`.
+5. Find and read the mental-model restore map if one exists. Treat it as the
+   highest-value guide to which memory files, mission files, source files, and
+   docs mattered before compaction.
+6. Identify the important files for the active task. The touched-file list is
+   a triage aid, not a complete inventory.
+7. Read every important file in full. Prioritize the restore map, Claude memory
+   files named in the map, Markdown state/planning files, queue/mission packets,
+   source files with active edits, root instruction files, and as-built or
+   codemap docs.
+8. After the required reads are complete, state:
+
+```text
+restored from packet at <path>; resumed at step <X>
+```
+
+Include the main files you read in full when you make that statement.
+
+9. **REFOCUS BEFORE YOU RESUME — the restore returns your TASK, not your BEARINGS.** Everything
+   above rebuilds *what you were doing*. None of it rebuilds *what the work is for* or *how this
+   rig operates* — those live in the chains, and nothing in steps 1–8 walks them. Invoke the
+   `refocus` skill and do its reads before your next substantive move.
+
+   **Measured 2026-08-13, and this step exists because of it:** a seat completed this protocol
+   faithfully — packet built, narrative read tail-first, read-depth audit reported honestly — and
+   resumed with a perfect picture of its atom and *zero* work-tree context. It had never opened
+   the mission or project nodes, and could only paraphrase the release's intent second-hand from
+   a peer's message. It then produced a technically-correct trace concluding *"I hold nothing"*
+   while holding 18 pending rows. **A restore that returns you to your task without your bearings
+   is how a confidently-wrong agent gets back to work.**
+
+**On the marker and the packet's size (measured across restores, 2026-08):** two realities the restore
+must handle honestly —
+
+- **The per-seat marker (`restore-pending/<session>.json`) is frequently absent** — in five observed
+  restores it was never present. Treat the JSONL rebuild via `restore-from-jsonl.mjs` as the **primary**
+  path, not a fallback: run it regardless of whether a marker exists. (When product-infra reliably writes
+  the marker for tmux-launched seats, it becomes a fast-path upgrade — not a prerequisite that keeps failing.)
+- **The generated transcript can be very large** (measured at ~346k tokens in one restore). Do **not**
+  read it front-to-back or claim to. `restore-instructions.md` declares its token cost up front; default
+  to the **most recent ~50k tokens of unique narrative first** (recency is what a restore needs — the
+  thread that produced the current frontier, not the boot handshake), and go earlier only if a specific
+  question requires it. A restore must leave room for the work it was restored to do.
+
+## Required Read-Depth Audit
+
+After the first restore pass, audit yourself before continuing.
+
+1. List every file, packet, marker, restore map, instruction file, and source
+   document you were asked to read during restore.
+2. Mark each item as `FULL`, `PARTIAL`, or `NOT_READ`.
+3. You will be given a task where all of these files are required reading in
+   order to understand the task.
+4. Read `PARTIAL`/`NOT_READ` items in full — but **to a declared budget with a stopping rule**, not
+   unbounded. Prioritize by relevance to the active task; for a large transcript read the most recent
+   unique narrative first (see *If You Just Compacted*), not front-to-back.
+5. **Stop** when either every task-relevant item is `FULL`, or you reach the budget — *a restore that
+   cannot leave room for the work it was restored to do is not a successful restore.* "Read everything,
+   never conserve" has no termination condition; that open-endedness is the bug, not the goal.
+6. Report the final read-depth table **honestly** (`FULL`/`PARTIAL`/`NOT_READ`, each with a reason)
+   before task work. **An honest `PARTIAL` with its reason is a correct outcome, not a failure** — do
+   not claim a completion you did not reach.
+
+## Guardrails
+
+- Compaction is survival, not housekeeping — never compact to free space, "lean" a seat, or capture/prepare an agent starter (the `rig agent-image` library). It is lossy (a compacted Claude is confident-but-hollow); compact only when a seat is genuinely near its context limit, with a before/after plan. A starter's value is being *functional*, not small — see the `agent-starters` skill.
+- Do not silently launch fresh after compaction.
+- Do not continue from memory when restore evidence exists.
+- Do not defer required restore reading until a later user task. The restore is
+  the current task.
+- Do not skip root instructions, as-built docs, or codemaps before product
+  code/review work.
+- Do not treat the generated touched-file list as exhaustive.
+- Do not mark a file `FULL` unless you actually read the full file content
+  after compaction.
+- Do not resume task work until the restore sentinel and read-depth audit are
+  complete.
+
+## Failure Modes To Avoid
+
+1. **Confidently-wrong restoration**: claiming restoration after reading only
+   the touched-file list or summary.
+2. **Partial restore**: reading the first few files, then continuing before the
+   full reading list is complete.
+3. **Skipping project instructions**: missing `AGENTS.md`, `CLAUDE.md`,
+   `README.md`, as-built docs, or codemaps that govern the task.
+4. **Treating the packet as exhaustive**: ignoring mission or workspace files
+   that are important but were not discovered by the script.
+5. **Waiting for the next task**: treating restore reading as conditional on a
+   future user assignment instead of completing it immediately.

package/daemon/context-packs/skills/claude-compaction-restore/manifest.yaml

@@ -0,0 +1,18 @@
+# GENERATED by scripts/generate-context-packs.mjs — DO NOT EDIT.
+# Package-time projection of a canonical skill; regenerated every build.
+name: "claude-compaction-restore"
+version: "0.5.8"
+taxonomy: "skills"
+purpose: "Use when a Claude Code session has just compacted, is about to compact, reached context limit, resumed after /compact, or needs to rebuild its working mental model from Claude JSONL transcripts and touched files."
+files:
+  - path: "SKILL.md"
+    role: "instruction"
+    summary: "Use when a Claude Code session has just compacted, is about to compact, reached context limit, resumed after /compact, or needs to rebuild its working mental model from Claude JSONL transcripts and touched files."
+  - path: "scripts/precompact-hook.mjs"
+    role: "reference"
+  - path: "scripts/restore-from-jsonl.mjs"
+    role: "reference"
+  - path: "templates/compact-instruction.md"
+    role: "reference"
+  - path: "templates/post-compact-restore-instruction.md"
+    role: "reference"