docs(roadmap): add #688 — sandbox help json lacks safety schema

ROADMAP.md

@@ -6422,3 +6422,54 @@ Original filing (2026-04-18): the session emitted `SessionStart hook (completed)
 
 448. **`sandbox --output-format json` has contradictory state flags — `enabled:true, supported:false, active:false, filesystem_active:true, allowed_mounts:[]`: claim that sandbox is "enabled" while OS doesn't support namespace isolation and `allowed_mounts:[]` is empty contradicts `filesystem_active:true filesystem_mode:"workspace-only"`** — dogfooded 2026-05-11 by Jobdori on `7244a82b` in response to Clawhip pinpoint nudge at `1503403842920779917` (using fresh-current-main runner at `/tmp/claw-dog-1430` per gajae's 14:00 protocol switch). Reproduction: `claw sandbox --output-format json` on macOS (where `unshare` is unavailable) returns `{"active":false,"active_namespace":false,"active_network":false,"allowed_mounts":[],"enabled":true,"fallback_reason":"namespace isolation unavailable (requires Linux with \`unshare\`)","filesystem_active":true,"filesystem_mode":"workspace-only","in_container":false,"kind":"sandbox","markers":[],"requested_namespace":true,"requested_network":false,"supported":false}`. **Three contradictions in the same envelope:** (a) `enabled:true` AND `supported:false`: what does "enabled" mean if the OS doesn't support sandboxing? Read literally, sandbox is *enabled but unsupported* — semantic nonsense. The likely intent is "user requested sandbox in config" but the field name `enabled` says "is ON". A better name would be `requested:true` or `config_intent:true`, with `enabled` reserved for the actually-active state. (b) `filesystem_active:true, filesystem_mode:"workspace-only"` AND `allowed_mounts:[]`: if the filesystem fence is active in workspace-only mode, the workspace directory itself MUST be an allowed mount. An empty `allowed_mounts:[]` array combined with `filesystem_active:true` means either (i) the fence is being misreported (it's not really active), (ii) the workspace is implicit and `allowed_mounts` only lists *additional* mounts, or (iii) the fence has no allowed paths and nothing is readable — all three are inconsistent with the user-facing summary. (c) `active:false` AND `filesystem_active:true`: the top-level `active` field is a single boolean summary, but it disagrees with `filesystem_active:true` (one component is active). Either `active` is "all components active" (then it should be `false` when any component is off) or "any component active" (then it should be `true` when filesystem is). The current value is `false` despite filesystem being active. **Sibling: no `claw sandbox --help`**: `claw sandbox status` and `claw sandbox --help` go to LLM-prompt fallback or hang (gajae confirmed at 13:00 that `sandbox status` returns typed `cli_parse` but `sandbox --help` is bounded — schema is non-uniform across help paths). **Required fix shape:** (a) rename `enabled` to `requested` or `config_intent` to disambiguate from "currently active"; (b) make `allowed_mounts` explicitly include the workspace when filesystem_mode is "workspace-only" (`allowed_mounts:[{path:"<cwd>",writable:true,reason:"workspace_root"}]`); (c) document the `active` aggregate semantics: pick either "all" or "any" composition rule and document the choice; (d) add `active_components:["filesystem"]` array as a richer alternative to the single boolean — surfaces exactly which sandbox subsystems are live; (e) regression test: when `filesystem_mode == "workspace-only"`, `allowed_mounts` MUST contain the cwd and `active` must agree with the documented composition rule. **Why this matters:** sandbox is the trust surface — automation that checks `sandbox.active == true` before running a risky LLM prompt sees `false` (no namespace, no network) and assumes no isolation, but `filesystem_active:true` means there IS partial isolation. The mixed signal forces consumers to OR all `*_active` fields together. Cross-references #428 (default permission_mode=danger-full-access — paired with sandbox-not-active means zero isolation), #444 (no broad-cwd guard — sandbox is the only safety net and its status is unclear). Source: Jobdori live dogfood, `7244a82b`, 2026-05-11.
 
+
+449. **`claw session list --output-format json` routes through `CliAction::ResumeSession` and hits the auth gate, returning `kind:"missing_credentials"` — but `session list` is a pure local filesystem read that requires no API credentials; by contrast, `claw session` (without `list`) correctly short-circuits with `kind:"unknown"` + "is a slash command" message without touching the auth gate** — dogfooded 2026-05-12 by Jobdori on `8f55870d` in response to Clawhip pinpoint nudge at `1503638404842131456`. Reproduction (no creds, isolated env): `env -i HOME=$HOME PATH=$PATH claw session list --output-format json` → `{"error":"missing Anthropic credentials...","kind":"missing_credentials"}` exit 1. `env -i HOME=$HOME PATH=$PATH claw session --output-format json` → `{"error":"`claw session` is a slash command...","kind":"unknown"}` exit 1 (no auth check). Root cause: the parser routes `session list` via `parse_resume_session_args` treating `list` as a session-path token, producing `CliAction::ResumeSession { session_path: "list", commands: [] }`. `resume_session()` then calls `LiveCli::new()` which instantiates the Anthropic client and fires the credentials guard. The `SlashCommand::Session { action: Some("list") }` special-case path in `run_resume_command()` (line 3654 comment: "`/session list` can be served from the sessions directory without a live session") is only reachable after auth passes — the no-creds guard fires before the slash-command dispatch loop. **Asymmetry:** the internal code already knows `session list` is credential-free (the comment at line 3654 says so), but the CLI entrypoint forces creds before the command ever reaches that branch. **Sibling: `session list` with no sessions returns `kind:"session_load_failed"` (from `--resume latest` fallback) rather than `{"kind":"session_list","sessions":[],"session_details":[]}` — the empty-sessions case is misrouted to the resume-failure path instead of a list-success with zero entries.** **Required fix shape:** (a) add a dedicated `CliAction::SessionList { output_format }` variant dispatched when `claw session list` is parsed — do not route through `ResumeSession`; (b) implement `run_session_list(output_format)` as a credentials-free function that calls `list_managed_sessions()` directly (same logic as the slash-command special-case at line 3659); (c) ensure empty sessions returns `{"kind":"session_list","sessions":[],"session_details":[],"active":null}` with exit 0, not a `session_load_failed` error; (d) add the same fix for sibling local-only commands that currently hit the auth gate: `session delete <id>`, `session export <id>`; (e) regression test: `claw session list --output-format json` with no credentials returns `kind:"session_list"` exit 0. **Why this matters:** session list is the canonical inventory surface for automation pipelines — `claw session list --output-format json | jq '.session_details[] | .id'` is the idiomatic way to enumerate sessions for replay, export, or resume. Requiring API credentials to read a local directory listing breaks offline use, CI environments with no API key configured, and any scripting that runs before credential setup. Cross-references #357 (session list requires creds — this is the same bug surfaced by that entry; #449 provides the root-cause path trace), #369 (session help/fork require creds), #427 (resume --help hits auth gate), #431 (skills uninstall requires creds). Source: Jobdori live dogfood, `8f55870d`, 2026-05-12.
+
+
+
+450. **`prompt` emits `kind:"missing_credentials"` JSON on STDERR (not stdout), leaving stdout at 0 bytes — automation pattern `output=$(claw prompt hello --output-format json)` captures nothing on auth-absent failure; `doctor` correctly surfaces `auth.status:"warn"` with `api_key_present:false` but exposes no `prompt_ready:false` field that automation can check before invoking `prompt`** — dogfooded 2026-05-16 by Jobdori on `a35ee9a0` in response to Clawhip pinpoint nudge at `1505208225321062521`. Exact reproduction (isolated env, no creds, fresh git repo, HEAD `a35ee9a0`): `timeout 5 env -i HOME=$ISOLATED_HOME PATH=$PATH CLAW_CONFIG_HOME=$PROBE/.claw-cfg claw prompt hello --output-format json > stdout.txt 2> stderr.txt` → stdout = **0 bytes**, stderr = 195 bytes containing `{"error":"missing Anthropic credentials…","exit_code":1,"hint":null,"kind":"missing_credentials","type":"error"}`, exit code 1. Confirms Gaebal's `1505208553793781792` pinpoint that `prompt` timeout + zero bytes was the prior state — HEAD `a35ee9a0` now correctly exits 1 with `kind:"missing_credentials"` **but the envelope is still routed to stderr** (issue #447 class, same class as prior entries #422, #435). **Contrast with `doctor`:** `claw doctor --output-format json 2>/dev/null` succeeds to stdout with `checks[auth].status:"warn"`, `api_key_present:false`, `auth_token_present:false` — but the auth check has no `prompt_ready:false` field. Automation that gates on `doctor` before invoking `prompt` must re-derive readiness from `api_key_present && auth_token_present` — there is no single canonical boolean. **Three compound problems:** (a) **stdout-empty on `--output-format json` failure**: same class as #447; `prompt`'s error envelope goes to stderr, not stdout. The canonical automation idiom `if ! result=$(claw prompt "q" --output-format json); then echo "$result" | jq .kind; fi` sees `$result=""` on failure — the jq call gets nothing. All `--output-format json` error paths must route JSON to stdout per #447 contract; (b) **`doctor` missing `prompt_ready` field**: `doctor --output-format json` already knows auth is absent (`api_key_present:false`) but surfaces no derived `prompt_ready:bool` or `prompt_blocked_reason:string` field. Automation must infer readiness from `api_key_present || auth_token_present || legacy_*_present` — a 5-field OR across legacy fields that is fragile as auth mechanisms evolve. A single `prompt_ready:false` (with `prompt_blocked_reason:"auth_missing"`) inside the `auth` check would give downstream a stable contract; (c) **`claw prompt` with no auth does no preflight and fires straight at the API**: the preflight check that `doctor` runs (auth discovery) is not reused by `prompt` to emit a fast typed error before attempting the network call. Both Gaebal's pinpoint (prompt hanging silently on older HEAD) and the current behavior (prompt hitting auth gate after a brief API attempt) stem from the same root: prompt does not short-circuit at the point where `doctor` already knows auth is absent. If `doctor` can emit `kind:"doctor"` with `auth.status:"warn"` in ~20ms without a network call, `prompt` should emit `kind:"missing_credentials"` in the same window and output it to stdout. **Required fix shape:** (a) `prompt --output-format json` must write the `kind:"missing_credentials"` JSON envelope to **stdout**, not stderr — same fix as #447 for all error envelopes; (b) add `prompt_ready:bool` and `prompt_blocked_reason:string|null` to the `auth` check in `doctor --output-format json`; derive it as `api_key_present || auth_token_present || legacy_saved_oauth_present`; (c) `prompt` must run the credential preflight check (same codepath as doctor's auth check) before attempting any API call and emit `{"kind":"missing_credentials","prompt_blocked_reason":"auth_missing"}` on **stdout** with exit 1 if the check fails; (d) `--output-format json` stdout routing fix must cover: `prompt`, `session list` (cross-ref #449), `skills uninstall` (cross-ref #431), `resume` (cross-ref #435), `acp serve` (cross-ref #443) — the full `kind:"missing_credentials"` class; (e) regression test: `claw prompt hello --output-format json` with no creds writes JSON to stdout (0 bytes stderr), exits 1, `kind:"missing_credentials"`, in under 200ms (no network attempt). **Why this matters:** `prompt` is the primary consumer entry point. Auth-absent failure routing to stderr breaks every automation wrapper that captures `$(claw prompt ... --output-format json)`. The `doctor` preflight metadata gap means auth-readiness checks require parsing 5 legacy fields instead of reading one boolean. Cross-references #447 (all JSON error envelopes on stderr), #449 (session list hits auth gate), #431 (skills uninstall hits auth gate), #357 (auth gate on local ops cluster), #422 (exit-code parity). Source: Jobdori live dogfood, `a35ee9a0`, 2026-05-16.
+
+688. **`sandbox --help --output-format json` is JSON-valid on current main but message-only (`{kind, command, topic, message}`), while actual `sandbox --output-format json` exposes the safety/trust schema (`active`, `supported`, `enabled`, namespace/network/filesystem flags, `allowed_mounts`, `fallback_reason`, markers) that help does not describe; automation cannot discover sandbox-state fields or their intended semantics without invoking sandbox and reverse-engineering the payload** — dogfooded 2026-05-24 for the 23:30 Clawhip nudge at message `1508250775162196070`, reproduced on a freshly rebuilt current `origin/main` binary (`git_sha f8e1bb726`) from `/tmp/cc-probe-main-2130`. Active claw-code sessions: none.
+
+    Reproduction:
+
+    ```bash
+    $ env -i HOME=/tmp/iso43/home PATH=/usr/bin:/bin TERM=dumb \
+        claw sandbox --help --output-format json
+    {
+      "command": "sandbox",
+      "kind": "help",
+      "message": "Sandbox\n  Usage            claw sandbox [--output-format <format>]\n  Purpose          inspect the resolved sandbox and isolation state for the current directory\n  Output           namespace, network, filesystem, and fallback details\n  Formats          text (default), json\n  Related          /sandbox · claw status",
+      "topic": "sandbox"
+    }
+    ```
+
+    Actual sandbox JSON on the same binary exposes the trust-state fields:
+
+    ```bash
+    $ claw sandbox --output-format json | jq 'keys'
+    [
+      "active",
+      "active_namespace",
+      "active_network",
+      "allowed_mounts",
+      "enabled",
+      "fallback_reason",
+      "filesystem_active",
+      "filesystem_mode",
+      "in_container",
+      "kind",
+      "markers",
+      "requested_namespace",
+      "requested_network",
+      "supported"
+    ]
+    ```
+
+    Help does not expose any structured `output_fields`, `component_fields`, `mode_fields`, `local_only`, `requires_credentials:false`, `requires_provider_request:false`, `mutates_workspace:false`, or documented `active` aggregation semantics. This matters even more because #448 already shows the sandbox payload can carry confusing combinations such as `active:false` with `filesystem_active:true`; help should be the place where those fields are machine-described.
+
+    **Why distinct from existing items:** #448 is about contradictory sandbox state values and missing/implicit mount semantics in the actual payload. #688 is about help-schema discoverability: even if the payload values are corrected, `sandbox --help --output-format json` still gives automation only a prose `message` and no field contract. #325 is broad top-level help prose-wrapper; #687 is status schema help; #686 is doctor check-schema help. This one is sandbox-specific because sandbox is the safety/trust surface and its component-state vocabulary needs to be discoverable before callers trust a run.
+
+    **Why this matters:** claws and wrappers gate risky prompt execution on sandbox state. They need to know, from help or schema metadata, which fields exist (`active_namespace`, `filesystem_active`, `allowed_mounts`, `fallback_reason`), whether the command is local-only and credential-free, and how to interpret the aggregate `active` versus component-active fields. Today they must invoke sandbox, inspect an example response, and guess the semantics.
+
+    **Required fix shape:** (a) Extend `sandbox --help --output-format json` with structured fields: `usage:"claw sandbox [--output-format <format>]"`, `formats:["text","json"]`, `related:["/sandbox","claw status"]`, `local_only:true`, `requires_credentials:false`, `requires_provider_request:false`, `mutates_workspace:false`, `output_fields:["kind","active","supported","enabled","requested_namespace","active_namespace","requested_network","active_network","filesystem_active","filesystem_mode","allowed_mounts","fallback_reason","in_container","markers"]`, `component_fields:["namespace","network","filesystem"]`, `filesystem_modes:[...]`, and `schema_version`. (b) Add a machine-readable `active_semantics` field, e.g. `"all_requested_components_active"` or `"any_component_active"`, matching the eventual #448 fix. (c) Keep `message` as human summary only. (d) Derive the help schema from the same sandbox report struct/registry used to render the payload so added sandbox fields cannot drift from help. **Acceptance check:** `claw sandbox --help --output-format json | jq -e '.command=="sandbox" and .local_only==true and .requires_credentials==false and ([.output_fields[]] | index("filesystem_active") and index("allowed_mounts") and index("fallback_reason")) and (.active_semantics | type == "string")'` should pass; currently those structured fields are absent. Source: gaebal-gajae dogfood for the 2026-05-24 23:30 Clawhip nudge.