Two changes, one binary batch:
1. app_install/app_uninstall (task 3): '!ai please install bitcoin knots'
correctly said it can't. Both tools are category-Apps, destructive, and
ride the 13-08 confirm gate (node-authored descriptions added). Install
validates catalog membership BEFORE the dialog (a typo never spends an
approval); uninstall resolves installed ids. Both reach the SAME
package.install/package.uninstall spawns every authenticated caller
uses, via a curated Arc-taking sibling of assistant_dispatch_tool.
2. S6: cloud legs no longer strip prior USER turns from replayed history.
Turn-minimality's allowlist is now the whole conversation's operator
turns (the node's own D-08 transcript, same trust class as this turn),
still mechanically matched, B1 secret scan and 64KB cap unchanged,
fabricated user messages still truncated. The model no longer sees its
own answers without the questions.
Co-Authored-By: Claude <noreply@anthropic.com>
Live evidence, two ways: the 9abc1623 banner never fired because D-16 hides
ungranted tools (model never calls → refused_categories always empty), and
the [[needs:id>]] marker fix failed because a small local model answers with
a workaround narrative instead of emitting structured markers.
The model's reliable, trained behavior is tool CALLING — so disabled tools
are now listed in a DISABLED prompt section and remain in the schema. A call
hits the execution gate, which refuses and records the category → the
trusted chrome offers Settings → AI Data Access. Deterministic and
model-independent. The prompt split is UX/attack-surface shaping; the
security boundary remains the server-side grant re-check in execute_tool
(loop_.rs), unchanged and now the single enforcement layer by design.
Tests: ungranted_tool_only_ever_in_disabled_section (section-aware),
disabled_tools_are_listed_as_callable_but_refused, marker extraction kept
as a harmless safety net. 127/127 assistant suite green.
Co-Authored-By: Claude <noreply@anthropic.com>
The 9abc1623 banner waited on refused_categories, but refused_categories
only fills when the model CALLS a gated tool — and D-16 hides ungranted
tools from the prompt, so the model never calls: it answered 'I can't do
that' in prose and the banner never fired. Live-verified: revoke media,
ask for content, no banner.
- build_system_prompt takes the disabled categories and teaches a marker:
'say it can be switched on in Settings → AI Data Access and end with
[[needs:<id>]]' — category names only, never tool names (D-16 holds)
- extract_needs_markers strips the markers from the reply and folds them
into refused_categories; unknown ids pass through as text (an offer is
the worst a bad marker can cause — never a grant)
- egress's seed-screen test now covers the new paragraph too
Co-Authored-By: Claude <noreply@anthropic.com>
S4: /aiui/api/web-search proxied straight to SearXNG with no auth — anyone
reaching the web port ran searches attributed to the node's IP. Now routed
through the daemon's session-gated model proxy like the claude/ollama legs
(both nginx server blocks), forcing format=json upstream (the client never
sent it — search could 200 with HTML that parsed as nothing).
S3: the forwarder also serves the STANDALONE frontend, whose bodies carry
full history/images with no assistant loop behind them — a pasted seed
phrase went to Anthropic unscreened. The forwarder now runs the egress
secret-shape scan (G-B1) with the node's own secrets dir as deny corpus on
Claude bodies and search queries; blocked requests get a plain-language 400.
Also fixes a REAL gap in the egress tokenizer found by these tests: a JSON
key glued to a string value's first word ('content":"abandon...') dropped
that word, so an exactly-12-word seed pasted as a bare message yielded an
11-member run — checksum misses, backstop misses. Non-member words now
rescan within the token. egress 15/15 + model_proxy 10/10 green.
Co-Authored-By: Claude <noreply@anthropic.com>
Second on-device failure in one session: after the wordlist fix, dev3
blocked cloud turns AGAIN mid-session as 13-10's history grew — splitting
on every non-alphabetic character let words from unrelated JSON fields
chain into one run. Both failures took the whole feature down rather than
protecting anything, which is the worse failure for a screen to have.
Shape is the wrong signal. A real mnemonic's last word encodes a checksum
over the rest, so an accidental run of English words parses as a mnemonic
only about one time in sixteen. Candidate runs are now validated with the
same bip39 crate the wallet uses:
- tokenize on whitespace (a seed phrase is space-separated); a token's
leading alphabetic segment counts, and alphanumerics after it end the
phrase, so a seed glued to a closing quote is still caught
- block only if a 12/15/18/21/24 window parses as a real mnemonic
- IMPLAUSIBLE_MEMBER_RUN (20) backstops checksum-invalid material such as
a typo'd 24-word seed, which prose cannot plausibly produce
Documented trade-off: a checksum-invalid run under 20 words no longer
blocks. The rule that did block it also blocked every legitimate turn,
twice, on a live node. 15/15 egress tests green, including the real
system prompt, scattered-JSON prose, and a genuine mnemonic in JSON.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The shape-only heuristic ('any 12 consecutive lowercase 3-8-char words')
matched ordinary prose — including the node's own system prompt — and
blocked 100% of live cloud chat turns (found on dev3, the first real
Claude call through this screen; log: kind=bip39-word-run every turn).
Membership in the crate's own bip39 English wordlist (already a dep via
seed.rs) distinguishes prose from seed material: glue words break runs,
real seeds are nothing but members. Regression test pins the real system
prompt + a clean wire body to Allow; the 12-word genuine-seed case still
blocks. 13/13 egress tests green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Task 1 decision (proceed-docs-with-probe-first, operator-selected via
AskUserQuestion 2026-08-05): 13-ROUTSTR-FINDINGS.md observed 0 of 9 cited
protocol claims (no live provider was announcing on any of the 3 default
relays in a 30s window on 2026-08-03; relay reachability itself WAS
confirmed). This backend is written against docs.routstr.com's cited shape,
with the first live chat-completions call doubling as the capability probe:
a non-success HTTP status or a response missing the expected
choices[0].message shape fails loudly (bails with the real status/body)
rather than silently degrading.
- assistant/backends/routstr.rs (new): RoutstrBackend implements the
Backend trait — discover_providers subscribes for kind-38421
provider-announcement events over the existing Tor-proxy-aware Nostr
client (nostr_discovery::build_nostr_client, never a second relay
client), process-cached with a 5-minute TTL; select_provider picks the
globally cheapest affordable (provider, model) price across every
discovered provider (Routstr has no fixed target model the way
Ollama/Claude do — CONTEXT.md delegates provider selection strategy to
Claude's discretion), preferring an onion endpoint when Tor is up;
attach_payment calls the existing budget-capped auto_pay_token verbatim
(never hand-rolled); parse_openai_tool_calls parses the one
string-encoded function.arguments shape exactly once at this adapter's
edge; screen_outbound (G-B1/G-B2) runs before any body leaves the node,
exactly as it does for Claude; ROUTSTR_MAX_TOKENS caps every request
explicitly.
- assistant/egress.rs: message_is_turn_own gains "system" and "tool" role
handling plus an OpenAI tool_calls-sibling-field check — the pre-existing
function was written only against Claude's wire shape (system as a
top-level field, tool results wrapped in role:"user") and would have
silently stripped Routstr's system prompt and tool-result context out of
every outbound request via G-B2's fail-closed default arm. Fixed with 4
new regression tests pinning both wire shapes.
- assistant/backends/mod.rs: registers `pub mod routstr;`. select_backend's
actual wiring of the Routstr leg (budget-gated, per D-05) is Task 3's
commit, once AssistantBudget exists — this task's own acceptance criteria
do not require select_backend integration, only the adapter itself.
30/30 assistant::backends:: tests pass (17 new in routstr.rs, 3 new in
egress.rs's OpenAI-shape regression tests were run separately at 12/12).
Zero new packages (nostr-sdk/reqwest already in-tree); dispatcher.rs and
Cargo.toml untouched.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
assistant/egress.rs: screen_outbound(body, ctx) -> EgressVerdict runs on
every request body about to leave this node for a cloud backend. G-B1
scan_secret_shapes checks for macaroon-shaped hex runs, BIP39-length word
runs, ecash/Nostr-key-shaped strings, and the literal contents of files
under data_dir/secrets — a hit fails closed (BlockFallBackLocal), logging
only the match's kind, never the value. G-B2 assert_turn_minimal checks the
outbound body against a mechanical allowlist of this turn's own fields (the
user's turn, this turn's granted tool names, this turn's own tool results);
an unrelated earlier tool result or content wrapped for a different turn is
truncated out rather than eyeballed. An unparsable/ambiguous body also fails
closed. MAX_OUTBOUND_CONTEXT_CHARS caps body size independent of minimality.
Wired into backends/claude.rs's send() before the outbound HTTP request (on
a block, send() errors before anything is sent — Rule 3, outside this
task's originally-declared file list but structurally required to give
screen_outbound a real caller); never wired into ollama.rs — nothing leaves
the node on that leg, so paying the scan cost would be pointless.
mod.rs: AssistantCounters/OwnerNotice — grant refusals, validation
failures, turns-per-request, untrusted-content-present,
cloud-escalation-while-local-up, blocked-egress and MAX_TURNS-reached
counters, each raising an owner_notice() at its own AI-SPEC §7b threshold.
Local and owner-facing only: no exporter, no /metrics, no OTLP anywhere in
assistant/ or rate_limit.rs. backends/mod.rs's select_backend raises a
cloud-escalation-while-local-up notice when Ollama is reachable but its
configured model isn't tool-capable (Rule 3, same file-scope reasoning).
9/9 assistant::egress:: tests pass in this task's own isolated state
(Task 1's 56 plus these 9 — ToolExecCtx's counters field and its loop_.rs
call sites are Task 3's own commit, since nothing in this task's behavior
needs them yet).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
