feat(prompt): tell the agent what its sandbox can run

The agent had no way to know its container ships ffmpeg, ripgrep or
tesseract, so it either declined work it could do or spent a round finding
out. This adds a command list to the system prompt as a **discovery hint** —
explicitly not an inventory.

Every decision follows from it being a hint:

- The allowlist (~35 entries, `container/commands.rs`) is the curation; a
  full PATH dump is 800 entries of coreutils noise. The probe exists so the
  list cannot *lie*, not so it can discover: `command -v` at login means we
  never announce something a container recreate threw away.
- The rendered prose says the list is partial and names `command -v`, so a
  tool outside the allowlist costs one check rather than a wrong conclusion.
  An empty probe renders as an explicit "could not be read", never as
  silence under a heading promising a list.
- Order is the allowlist's own, grouped by kind of work — the grouping is
  the curation, and the reader is a model, not a grep.
- Staleness is cheap both ways, so there is no invalidation machinery: a
  login-time snapshot on `UserContext`, non-fatal, refreshed at next login.

The gate is the tool, not the sentinel. Every AGENT.md carries
`common/sandbox.md` — the four system agents included — and the section is
emitted iff the turn's model is shown `execute_cmd`, derived from
`allow_tools` plus the security group's visibility filter for a root turn
and from `child_defs` for a sub-agent: always the same definitions the model
will see. `has_execute_cmd` therefore joins the PrefixCache key, since the
group is switchable mid-conversation and that switch already rewrites the
tool payload in the same provider cache.

The fragment holds only the heading and one stable sentence; every
conditional claim lives in the renderer, because prose promising
`sudo apt-get install` is not the renderer's to retract when the tool is
absent. `execute_cmd`'s own description loses `(python + node available)`:
its job is steering away from the shell, and a capability advertisement
diluted it.

CLAUDE.md

@@ -145,7 +145,7 @@ The schema is split into two buckets (§5.1), and the split is the point:
 
 **Memory injection into the prompt**: `AgentSystemContext::load_inject_memory` (`loop_adapters/system.rs`) routes each `meta.inject_memory` entry — `user-memory/…` → owner pool, `shared-memory/…` → the shared (`system.db`) pool, both via `memory_docs::get`; anything else (`data/…`, `$WD/…`) is a disk read. The shared pool is threaded `ChatSessionManager` → `UserLoopRuntime` → `AgentSystemContext`. `assistant` and `project-coordinator` inject `user-memory/index.md` + `shared-memory/index.md`.
 
-**Prompt substitutions**: an `AGENT.md` may carry `<!-- KEY -->` placeholders; `agents::resolve_includes` turns each into a `__KEY__` sentinel, replaced at request time. Several are resolved by the system-context source itself (`loop_adapters/system.rs`) from the session owner (`user_id`) + registry (`shared_pool`) + their `UserFs`, so every source (WS, mobile, cron, sub-agents) gets them with no caller plumbing: `__SKILLS_LIST__` (the generated skills index — see Filesystem & containers), `__SHARED_FOLDERS__` (the user's shared-folders table) and `__USER_PROFILE__` (the owner's directory profile: `Name`, `Date of birth` with age computed at build time, `Sex`, `Preferred language`, admin `Notes` — unset values render as explicit `unknown` / `not specified`, the `Notes` line is omitted when empty). Any other key comes from the per-call `SendMessageOptions::system_substitutions` map.
+**Prompt substitutions**: an `AGENT.md` may carry `<!-- KEY -->` placeholders; `agents::resolve_includes` turns each into a `__KEY__` sentinel, replaced at request time. Several are resolved by the system-context source itself (`loop_adapters/system.rs`) from the session owner (`user_id`) + registry (`shared_pool`) + their `UserFs`, so every source (WS, mobile, cron, sub-agents) gets them with no caller plumbing: `__SKILLS_LIST__` (the generated skills index — see Filesystem & containers), `__SANDBOX_COMMANDS__` (the sandbox command hint — see below), `__SHARED_FOLDERS__` (the user's shared-folders table) and `__USER_PROFILE__` (the owner's directory profile: `Name`, `Date of birth` with age computed at build time, `Sex`, `Preferred language`, admin `Notes` — unset values render as explicit `unknown` / `not specified`, the `Notes` line is omitted when empty). Any other key comes from the per-call `SendMessageOptions::system_substitutions` map.
 
 `system.db` still gets **both** bucket functions — but no longer because the migration is unstarted. It gets the owner schema because it *is* the owner of **shared** memory (`memory_docs`) plus, for now, the globally-scoped `secrets` (`SecretsStore` is built on the system pool and shared by reference into every `UserContext`; the global runtime's *config* now lives in the registry table `mcp_global_servers`, and per-user connector config in each user's owner `mcp_user_servers`). The global runtime no longer writes `mcp_events` there: notification persistence is an explicit `McpManager::new` argument (`EventLog::{Persist,Discard}`), `Discard` for the ownerless global runtime and `Persist` for each per-user one, because an event belongs to whoever it happened to and its only reader (event triage) is per-user. Every *other* owner table is created there but never written to anymore — the global owner-bound managers that would write them (chat/jobs/etc.) are inert (see "Current state"). Fully dropping `create_owner_tables` from `system.db` is blocked on the §4 scope decision for secrets, not on call-site migration.
 
@@ -180,6 +180,8 @@ Two views, **one storage**: for the mounted subtree the fs-tools run **host-side
 
 **The skills index is generated, and the sentinel is the knob.** What reaches the model is not a file anyone maintains but a **function of the two trees** (`crates/skald-core/src/skills/`, pure functions in the shape of `LlmCommandManager`): each skill's `SKILL.md` **path** plus its frontmatter `description`, truncated to 200 chars, under an imperative header ("you MUST read its SKILL.md") — the countermeasure to the real failure mode, which is the model *under*-triggering. Printing the full path rather than an id plus a composition rule is what makes a read tool unnecessary: `read_file` on the printed path is one call, and there is no step left for the model to get wrong. Injection is the placeholder `<!-- SKILLS_LIST -->` (normally `<!-- INCLUDE: common/skills.md -->`, a fragment that holds **only** the sentinel), substituted in `AgentSystemContext::build_base` beside `__MCP_LIST__`; `resolve_includes` needs no branch, its generic `<!-- KEY -->` → `__KEY__` arm already covers it. There is **no `meta.json` flag** — the sentinel *is* the switch, so the four `type: system` agents opt out by not including the fragment (an imperative "read it with read_file" is exactly wrong in an unattended turn, and some of those run with `allow_tools: false`). All eleven `chat`/`task` agents carry the include, sub-agents included: in a delegation the one doing the work is the child. Three rendering rules are load-bearing and each closes a specific failure: a **stable order** (scope, then id) because the index sits inside the provider's cache key; a **deterministic tail cut** at an 8 KB budget, announced by a `[N more skills omitted]` line, because a silently truncated index has the model conclude in good faith that a skill does not exist; and **empty in, empty out** — every word of prose lives inside the render, so an instance with no skills spends zero tokens and leaves no orphan sentence (the MCP list is the counter-example: its prose sits *around* the placeholder, and the empty state once had the model inventing a discovery tool). A colliding id is marked `[name collision]` on **both** lines, never shadowed. A malformed skill is skipped with a `warn!`, never fatal — the index is built while assembling a prompt. Freshness has two doors, one per writer. The in-process tools invalidate directly (`Skald::invalidate_prompt_prefix`, called by `skill_register`/`skill_delete`); a hand edit on the box is caught by the **skills watcher** (`skills/watch.rs`, spawned from `spawn_background`): a recursive `notify` on `{WD}/skills` + `{WD}/skills-users`, debounced ~800 ms, that re-digests each touched tree (`skills::tree_digest` — the (id, description) pairs the index is made of) and emits `SystemEvent::SkillsChanged { scope }` only when the digest moved. The subscriber `spawn_skills_freshness` (next to `spawn_user_lifecycle`, same `Weak` shape) maps the scope and calls the same invalidate accessor. Editing a script leaves the digest byte-identical and announces nothing — which is exactly the §6 rule, so an invisible change costs nobody a cache miss. Two gotchas the code carries comments for: the watcher **canonicalizes `{WD}`** (FSEvents reports real paths, and `/var` is a symlink on macOS), and it creates the two trees if absent (a box before its first user has neither).
 
+**The sandbox command list is a discovery hint, and the tool — not the sentinel — is the knob.** `container/commands.rs` probes the user's container at login (`UserContextFactory::build`, right after `ensure()`, **non-fatal**) with one `docker exec` running `command -v` over a curated ~35-entry `PROBE_ALLOWLIST`, and the result rides `LoopConfig.sandbox_commands` → `AgentSystemContext` → `__SANDBOX_COMMANDS__`. Three decisions carry it and each is the answer to an obvious-looking alternative. **The allowlist is the curation, and the probe is there so the list cannot lie** — not the other way round: a full `PATH` dump is 800 entries of coreutils noise, so what is worth tokens is decided by hand, and `command -v` exists only so we never announce something a container recreate threw away. A tool outside the list therefore never appears, which is fine because **the rendered prose says the list is partial and names `command -v`** — an inventory the model reads as exhaustive is the failure this shape avoids, the same one the skills index's `[N more skills omitted]` line closes. Order is the allowlist's own (grouped by kind of work), never sorted: the grouping *is* the curation, and the reader is a model, not a `grep`. **Staleness is cheap in both directions**, which is why there is no refresh machinery at all: a mid-session install is known to the agent that ran it, and a container recreate costs one `not found` plus the `apt-get install` the agent was already able to do. Gating is the one part that is not the skills pattern: every `AGENT.md` carries `<!-- INCLUDE: common/sandbox.md -->`, **including the four `type: system` ones**, and the section is emitted iff the turn's model is shown `execute_cmd` — computed from `allow_tools` plus the security group's visibility filter (`session/handler/config.rs`) for a root turn, and from `child_defs` for a sub-agent, i.e. always from *the same definitions the model will see*. Hence `has_execute_cmd` is in the `PrefixCache` key: the group is switchable mid-conversation from the chat's shield pill, and keying on it costs nothing because that switch already rewrites the tool payload sitting in the same provider cache. The fragment holds only the heading and one stable sentence; **every conditional claim lives in the renderer** (a departure from the `__MCP_LIST__` shape it otherwise follows), because prose promising `sudo apt-get install` is not the renderer's to retract when the tool is absent. Three rendered cases, and the middle one is why this is not a one-liner: the list, the *unreadable-probe* line (empty ≠ bare sandbox — rendering nothing under a heading that promises a list is how the MCP section once had a model invent a discovery tool), and the no-`execute_cmd` line. `execute_cmd`'s own description deliberately carries **no** capability advertisement — its `(python + node available)` was removed when this landed, since its job is steering the model *away* from the shell for work a file tool does better, and the two messages dilute each other.
+
 **Containment** (`resolve_host_path`) is unchanged and still guards **the host branch**: every path that lands on a mount is canonicalized (following symlinks) and prefix-checked against its mount base, **fail-closed**. That check is what it always was — the defence against a symlink planted from inside the container pointing at the **host's** `/etc`, which the host-side tool would otherwise follow off the box. Opening the container branch does not weaken it: that branch never touches the host filesystem, so there is no host to escape from, and the check keeps applying to everything mounted. `grep_files` stays disk-only (regex ≠ FTS; memory → `memory_search`) but resolves its root the same way. `execute_cmd`'s `workdir` is an agent path mapped to its container path via `UserFs::to_container`.
 
 The threading: `UserContext.fs` (built by `container::build_user_fs` at login, snapshotting shared memberships) → `ChatSessionManager` → `ChatSessionHandler.fs` → `ToolContext.fs`. **Admin CRUD is wired** (`src/frontend/api/shared_folders.rs` — `GET/POST /api/shared-folders`, `PATCH/DELETE /api/shared-folders/{id}`, `POST`/`DELETE .../members[/{user_id}]`; UI `shared-folders.js`): a create/describe/delete + per-member `can_write` surface, and each mutation emits `SystemEvent::UserMountsChanged`, on which the lifecycle reconciler runs `Skald::refresh_user_mounts` — rebuilding the affected user's fs + container mounts **in place**, so a membership change lands without a re-login (blueprint §6's "admin CRUD" + "membership refresh without re-login" TODOs, now closed; it still settles at next login/boot if the live remount fails). `execute_cmd` /stop is robust: the command runs under `setsid -w` in its own process-group (leader pid recorded in a container pidfile), and a `KillReaper` drop-guard reaps that group on /stop **or** timeout via a detached `docker exec` that walks `/proc` and kills members by **positive pid** (the container's dash mishandles `kill -<pgid>`); the pidfile is passed positionally (`$1`), and the container's `--init` (tini) reaps the killed processes so no zombies accumulate. **Per-user MCP connectors now run inside this container** (§7) — the container infra enabled it; see the MCP connectors section.