`initialize_agent_pool` spawned and handshook each slot serially, so a pool of
N paid N x adapter-startup of dead wall-clock before it could serve anything.
With a mock adapter sleeping 3s in `initialize` and 6 workers, serial took
18275 ms; the concurrent version takes 3100 ms (5.9x of a theoretical 6x).
`AcpClient::spawn`/`initialize` hold no statics, locks, `OnceCell`, or
`env::set_var`, so there is no hidden serialization to defeat the win.
The bound is a field on `PoolStartup` (`init_concurrency`) rather than a
constant, so a future scale-from-1 pool can reuse it as its grow-batch size
without changing the signature. It is clamped to `>= 1` because a bound of zero
would deadlock every acquire.
`from_config` defaults it to 4 rather than the pool size. The measured win is
mostly retained at 4 (10 locally-installed adapter slots: ~330ms at 4, ~160ms
unbounded, ~1070ms serial), and the cap limits how many adapters handshake at
once. That matters because neither probed adapter is gateway-backed, so a
full-pool burst against a shared gateway is exactly the case the measurements
do not cover. Promoting the default toward pool size should follow that probe;
the field means doing so needs no signature change.
Three invariants a naive concurrent rewrite silently breaks:
1. Slots are positional. `AgentPool::from_slots` requires `agent.index` to
equal the agent's position in `agents`, because `return_agent` writes
`agents[agent.index]` and `crash_history[idx]` charges the circuit breaker
by the same number. A `JoinSet` yields in completion order, so results are
assigned into a pre-sized `Vec` by index and never pushed. Getting this
wrong is silent corruption, not a crash: given slots holding agents labelled
{1, 0}, a single `try_claim`/`return_agent` cycle turns two live agents into
one via `return_agent`'s "already occupied - overwriting" branch.
2. Shutdown is observed at every unbounded await, including the wait for a
permit. `acquire_init_permit` selects `shutdown.changed()` against
`Semaphore::acquire` with `biased`, so a cancelled batch's released permits
are not inherited by queued tasks that would then spawn adapter children
during teardown. Each in-flight init also reaps the child it owns rather
than being aborted from outside, and a post-loop `has_changed()` check reaps
survivors and returns `Err`, preserving the serial loop's contract that a
pool built during shutdown is never handed back as `Ok`.
3. A partial pool is valid: only an entirely dead pool is an error. Dead slots
stay present as `None`; packing them out would shift every later index.
`initialize_agent_pool` had zero test coverage - a version wrong in the first
three ways above passes 607/607 pre-existing buzz-acp tests. Ten tests now pin
it: positional placement under reversed completion order, positional partial
failure, all-dead is `Err`, wall-clock overlap, `init_concurrency` actually
bounding in-flight inits, prompt cancellation on shutdown, three for
`acquire_init_permit`'s cancellation contract, and one pinning the default
bound below pool size. Pool mocks are `bash -c` scripts, matching the existing
`spawn_script` convention.
Mutation-tested, 10 mutants all killed: push-instead-of-assign, pack-out-dead-
slots, drop the all-dead guard, revert to serial, ignore `init_concurrency`,
drop the post-loop shutdown check, remove the per-task shutdown select, ignore
shutdown while acquiring a permit, remove `biased` from that select, and revert
the default to pool size. Two needed the tests strengthened rather than the
kill claimed: removing the per-task select initially survived because the test
asserted only the error while promptness went unenforced, and the permit-wait
mutant initially hung the suite instead of failing, so that wait is now
explicitly bounded.
The permit-acquisition gap in invariant 2 was found in review by Max, who also
made the case for the conservative default.
Co-authored-by: Dawn (sprout agent) <c6237ef84fa537c78dcee78efd2d4e59f728859c7f194da42ac51ededfa0be05@sprout-oss.stage.blox.sqprod.co>
Signed-off-by: tlongwell-block <109685178+tlongwell-block@users.noreply.github.com>
## Problem
The Databricks model dropdown offers a handful of stale models — and
there's no way to tell that list apart from the real one. The AI Gateway
exposes **66** chat/embedding endpoints on `block-lakehouse-production`,
but the picker was showing a short list that includes models the gateway
no longer serves and embedding endpoints that can't chat at all.
Three independent defects, all on the discovery path:
**1. Live discovery never ran for agents with no saved provider.**
`get_agent_models` gates every in-process discovery attempt on the
provider (`is_openai_compatible_provider` / `is_anthropic_provider` /
`is_databricks_provider`), reading it straight from `record.provider`.
That field is `null` for every agent record created before provider
persistence — and for any agent that inherits its provider from the
build. So all three gates saw `None`, no HTTP discovery ran, and the
request fell through to the `buzz-acp models` subprocess. On the
Databricks path that subprocess returns `discovery_failure_fallback` —
the small hardcoded `DATABRICKS_V2_KNOWN_MODELS` catalog — which the
frontend renders exactly like a live catalog. An internal DMG that bakes
`BUZZ_AGENT_PROVIDER=databricks_v2` and a `DATABRICKS_HOST` still got
the fallback.
**2. The fallback list couldn't represent the running model.**
When discovery genuinely fails, the picker should at minimum be able to
show what the agent is actually configured with. For `DatabricksV2` it
couldn't: the fallback returned only the hardcoded slate, so a model
like `databricks-gpt-5-5` wasn't selectable in its own picker.
**3. Embedding endpoints were offered as chat models.**
`databricks-bge-large-en` was selectable (visible in the dialog today).
The v2 endpoints payload carries no `task` or `state` field, so there is
nothing to filter on but the name.
## Changes
- **`effective_discovery_provider`** (new,
`desktop/src-tauri/src/commands/agent_models_env.rs`) — an explicit
provider (saved record value, or the create/edit dialog's current form
value) still always wins; when there is none, discovery falls back to
the runtime's own provider env var (`GOOSE_PROVIDER`,
`BUZZ_AGENT_PROVIDER`, …) read off the merged env, which by that point
already carries the baked build floor and the process env. Wired into
both `get_agent_models` and `discover_agent_models`.
`SavedAgentModelDiscoveryConfig` now carries `provider_env_var` from
`known_acp_runtime`, so each runtime reads *its own* key rather than a
shared guess.
- The relay-mesh branches in `discover_agent_models` deliberately keep
using `input.provider`: those key off a deliberate provider selection,
never a baked default.
- **Asserted vs inferred matters for missing credentials.** The OpenAI
and Anthropic gates error on a missing API key, while the Databricks
gate falls through; an inferred provider hitting the first two would
have replaced a working subprocess catalog with `config:
ANTHROPIC_API_KEY required` (`export GOOSE_PROVIDER=anthropic` is
goose's documented way to pick a provider, and it keeps the key in its
own keyring). So `effective_discovery_provider` returns a
`DiscoveryProvider` that remembers how the value was resolved, and
`required_env` only reports a missing credential for an asserted
provider. A wrong guess declines and lets the subprocess answer.
- **`is_chat_capable_endpoint`** (new,
`crates/buzz-agent/src/catalog.rs`) — applied in
`parse_v2_endpoints_page`. Drops `*embedding*` and segment-matched `bge`
/ `gte` endpoints; keeps everything unrecognised (fail-open, so a new
model family is never hidden). Segment matching is why it's `split('-')`
and not `contains`: a substring check would swallow legitimate names.
- **`discovery_failure_fallback`** for `Provider::DatabricksV2` now
leads with the configured model (deduped against the known slate,
blank-tolerant), so a failed discovery still yields a picker that can
show the running model. The configured model is trimmed once up front —
`resolve_model` doesn't trim, so a padded `DATABRICKS_MODEL` used to
slip past the dedupe and appear twice.
- **`sort_v2_endpoints_newest_first`** (new, second commit) — the
catalog is now ordered newest-first on each endpoint's
`created_timestamp`, ties broken by name. Previously Buzz sorted
nothing, so the gateway's own order reached the picker: it pages in two
phases (Databricks-managed, then workspace-created — the page token
decodes to `{"phase":"user"}`), each alphabetical, which buried
`databricks-claude-opus-5` 8th behind five older Claude endpoints and
`goose-claude-opus-5` — the newest endpoint in the catalog — 55th of 63.
Sorting in `fetch_v2_models` means both discovery paths inherit it with
no wire or type changes, and the combobox filter preserves incoming
order. Endpoints with an absent or unparseable timestamp sort last
rather than first, so a wire-shape change degrades to "unordered at the
bottom" instead of "shuffled to the top".
- The name tiebreak is load-bearing: eleven managed endpoints share one
placeholder timestamp (`1699610000000`), so without it their relative
order would vary between runs. That placeholder is also not always
accurate — a few genuinely recent endpoints
(`databricks-kimi-k2-7-code`, `databricks-llama-4-maverick`) land at the
bottom with the 2023 batch. The gateway offers nothing better to sort
on.
- Env/provider lookup helpers moved out of `agent_models.rs` into
`agent_models_env.rs`. This keeps the command module under the file-size
limit **without ratcheting the override up** — the existing 1079 entry
is untouched (file is now 1066 lines).
## Verification
Live against `block-lakehouse-production`, release build:
```
BUZZ_ACP_AGENT_COMMAND=$PWD/target/release/buzz-agent \
BUZZ_AGENT_PROVIDER=databricks_v2 \
DATABRICKS_HOST=https://block-lakehouse-production.cloud.databricks.com \
DATABRICKS_MODEL=databricks-gpt-5-5 \
./target/release/buzz-acp models --json
```
- before: 66 endpoints, including `databricks-bge-large-en`,
`databricks-gte-large-en`, `databricks-qwen3-embedding-0-6b`
- after: **63** endpoints, `[.models[] | select(.id |
test("embedding|-bge-|-gte-"))]` → `[]`
Top of the list after the sort commit:
```
goose-claude-opus-5 2026-07-24
databricks-claude-opus-5 2026-07-23
databricks-gemini-3-6-flash 2026-07-20
databricks-gemini-3-5-flash-lite 2026-07-20
databricks-inkling 2026-07-14
```
Tests: 15 new (8 in `catalog.rs` — including the two-wire-shape
timestamp parse, the sort's tiebreak/no-timestamp cases, and the
padded-model dedupe — and 7 plus one assertion in
`agent_models_tests.rs`, 3 of them covering the asserted/inferred
credential split), two existing tests updated. `just check`, `just
test-unit`, and `just desktop-tauri-test` all pass (1636 desktop-tauri
tests, 274 buzz-agent lib tests).
Not run locally: the Docker-backed integration suite (`just test`) —
this diff touches neither `buzz-relay`, `buzz-db`, nor `buzz-auth`.
## Follow-ups (deliberately out of scope)
Two inference-path defects found while investigating, both reproduced
live against the gateway and both independent of discovery:
1. **Gemini thought signatures are dropped.** The gateway returns a bare
`thoughtSignature` on tool calls; the external-model serving endpoints
return it nested as `extra_content.google.thought_signature`. Neither
shape is round-tripped, so multi-turn tool use on `databricks-gemini-*`
fails with a 400 on the second turn.
2. **Array-shaped `content` is silently discarded.** Some models return
OpenAI `content` as a block array rather than a string; `parse_openai`'s
`str_field` returns `None` and the text is dropped.
The legacy `serving-endpoints` path does not work around either one, and
costs reasoning support on the GPT-5 family.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Closes#2600
## Summary
Self-hosted multi-tenant deployments (one relay serving many communities
via host-based tenancy) routinely need more than three communities owned
by the same operator identity. `MAX_COMMUNITIES_PER_OWNER` is currently
a hardcoded const, and hitting it surfaces as a `limit_reached` 409 from
`POST /operator/communities` — which provisioning UIs tend to mislabel
(mine reported it as "subdomain already taken").
This makes the limit configurable per deployment:
- New env var `BUZZ_MAX_COMMUNITIES_PER_OWNER` — read once per process,
must parse as a positive integer; missing/invalid/non-positive values
fall back to the existing default of **3**, so current deployments are
unaffected.
- Enforcement locations are unchanged and stay in the authoritative
relay-layer checks: community provisioning
(`create_community_with_owner`) and ownership transfer (inside the
advisory-lock transaction).
- Parse/fallback rules are extracted into a pure helper
(`effective_owner_limit`) with unit tests, keeping the cached getter
trivial.
## Test plan
- `cargo test -p buzz-db --lib` — new `owner_limit_*` tests cover
default, invalid, non-positive, and positive-override cases.
(Pre-existing unrelated failure on clean main:
`replica_fence::tests::fence_starts_closed_and_opens_on_advance`,
tracked in #2369.)
- `cargo clippy -p buzz-db --all-targets` and `cargo fmt` clean.
- Deployed on my multi-tenant relay (vibecode.casa) with
`BUZZ_MAX_COMMUNITIES_PER_OWNER=100`: provisioning a 4th community for
the same owner succeeds; without the var the stock limit of 3 still
applies.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Signed-off-by: Lee Salminen <leesalminen@gmail.com>
Kind 30175 persona sync events carry plaintext `system_prompt` and
`respond_to_allowlist`. This PR adds **author-only-unless-shared read
semantics**: events without `["shared","true"]` are visible only to the
author; events with that tag are community-readable.
## What changed
### New read class (kind 30175)
Kind 30175 gets per-event gating at every relay read surface. The
`shared` marker is a **tag**, not a content field, so content bytes
(which double as the `source_version` drift basis) are not affected when
toggling share state.
### `event_visible_to_reader` helper (`handlers/req.rs`)
Centralizes the three per-event access predicates —
`is_author_only_event`, `is_unshared_persona_event`,
`reader_authorized_for_event` — into one `pub(crate)` fn callable from
both WS and HTTP adapters. All result-visibility sites now call this
single helper.
### NIP-98 HTTP bridge (`api/bridge.rs`)
- `POST /query` catchall: replaced the two-step author-only +
result-gated checks with `event_visible_to_reader` (now also covers the
persona shared-gate).
- `POST /count`: added `needs_persona_filtering` to the fast-path guard
(forces per-event fallback when filter can match `kind:30175`) and
replaced both fallback loops' individual checks with
`event_visible_to_reader`.
- FTS `/search` bridge helper: replaced `is_author_only_event` with
`event_visible_to_reader` as defense-in-depth (30175 is not in the FTS
allowlist today; comment at site explains the future-proofing intent).
### Ingest validation (`handlers/ingest.rs`)
`validate_persona_envelope` rejects malformed `shared` tags: wrong
value, missing value, duplicates. Accepts exactly `["shared","true"]`
and tag-absent.
### Kind helpers (`buzz-core/src/kind.rs`)
`is_persona_shared_kind`, `is_unshared_persona_event`,
`filter_can_match_persona_shared_kinds`.
### Tests (`e2e_persona.rs`)
8 unit tests in `kind.rs`, 6 in `ingest.rs`, 8 e2e tests total:
- AC-1–6 covering the gated surfaces
- `test_persona_live_fanout_shared_gate`: reworked with explicit
monotonic `created_at` timestamps (t0 < t1 < t2) and per-step head
assertions, eliminating the NIP-33 event-id tie-break race. Also asserts
foreign live subscription receives nothing on shared→unshared
transition.
- `test_persona_ingest_shared_tag_validation`: added `shared=x` and
missing-value wire-level rejection cases.
- `test_persona_mixed_kind_filter_does_not_leak`: publishes a kind-9
event and asserts it IS returned; absence-only assertion no longer
sufficient.
- `test_persona_http_query_cross_author_gate`: NIP-98 `/query`
cross-author gate (authors filter, kindless `ids` — both blocked; shared
`ids` — passes).
- `test_persona_http_count_cross_author_gate`: NIP-98 `/count`
cross-author gate (foreign sees 1/shared, author sees all, wildcard
checked).
### NIP-AP.md
Replaced aspirational "every relay read chokepoint" wording with an
enumerated list of gated surfaces including NIP-98 `/query`, `/count`,
and FTS/search with their enforcement mechanism named. Added
**Non-goal** note for side-band existence oracles
(reaction/report/deletion target resolution).
## Existing tests
All pre-existing `e2e_persona` tests use `{ids:[event_id]}` or
`{authors:[self]}` filters — author self-reads bypass the gate and are
unaffected.
## Gates
`just check` ✅ | `just test-unit` ✅ | `cargo test -p buzz-relay` ✅ (749
passed, 1 pre-existing failure in
`demo_join_forwarded_arm_round_trips_echo` — flaky on `main`, unrelated
to this PR, verified red at `origin/main` before this branch)
---------
Signed-off-by: Will Pfleger <pfleger.will@gmail.com>
Co-authored-by: npub1mn7jgtj4w2pd0g0zeuhxsa6jy6p0rewxz4kujt98my82ahfmp72sxjexk7 <dcfd242e557282d7a1e2cf2e6877522682f1e5c6156dc92ca7d90eaedd3b0f95@buzz.block.builderlab.xyz>
## Summary
- classify IPv4-compatible, IPv4-mapped, and SIIT IPv4-translated IPv6
addresses using the existing IPv4 SSRF policy
- decode IPv4 destinations under the RFC 6052 well-known NAT64 prefix
- conservatively block local-use NAT64, Teredo, and 6to4 ranges
- add boundary coverage for every newly handled transition prefix
## Why
The workflow webhook SSRF guard previously recognized IPv4-mapped IPv6
addresses but not other standardized IPv6 forms that can embed or route
to IPv4 destinations. Private, loopback, or link-local IPv4 targets
represented through those forms could therefore pass address
classification.
This also covers the legacy SIIT IPv4-translated prefix
(`::ffff:0:0:0/96`), which Rust's `Ipv6Addr::to_ipv4()` does not
recognize but an SIIT-enabled network may route to the IPv4 value in the
final 32 bits.
Network-specific NAT64 prefixes remain a deployment concern and should
be restricted through egress policy; they cannot be inferred generically
from an IPv6 address.
## Test plan
- `cargo fmt --all -- --check`
- `cargo test -p buzz-core network` (35 passed)
- `cargo clippy -p buzz-core --all-targets -- -D warnings`
- `git diff --check`
---------
Co-authored-by: Amp <amp@ampcode.com>
## Summary
Disable automatic system-proxy discovery for workflow webhook requests.
## Why
Webhook destinations are resolved, validated, and pinned before the
request to prevent DNS-rebinding SSRF. If reqwest uses a system proxy,
the proxy can resolve the original hostname itself instead of connecting
to the validated address, bypassing that pinning guarantee.
Calling `no_proxy()` keeps these security-sensitive requests on the
directly validated connection path. Redirects remain disabled.
## Test plan
- `cargo fmt --all -- --check`
- `cargo test -p buzz-workflow --features reqwest` (149 passed)
- `cargo clippy -p buzz-workflow --all-targets --features reqwest -- -D
warnings`
- `git diff --check`
Co-authored-by: Amp <amp@ampcode.com>
Fixes#2637 — full analysis and reproduction there.
## Problem
Audit entries are stamped and hashed with `Utc::now()` (nanoseconds),
then stored in a `TIMESTAMPTZ` column (microseconds). `compute_hash`
covers `created_at.to_rfc3339()`, and chrono emits 0/3/6/**9**
fractional digits depending on the value — so the digest written at
`service.rs:103` is computed over `…T12:00:00.123456789+00:00` while
`verify_chain` recomputes over the `…T12:00:00.123456+00:00` that
Postgres hands back.
Every hash chain backed by a real database therefore fails verification
at its first entry, on untampered data. That is not just a broken
feature — it means a genuinely forged row is indistinguishable from the
permanent baseline failure, so `HashMismatch` carries no signal.
It is invisible in CI because all six chain tests are `#[ignore =
"requires Postgres"]`, and the in-process `hash.rs` tests use a fixture
timestamp of `2026-01-01T00:00:00Z` — zero sub-seconds, the one value
where the bug cannot appear.
## Solution
Reduce `created_at` to the stored precision *before* hashing, so the
in-memory entry and the row are byte-identical:
```rust
pub fn to_storage_precision(created_at: DateTime<Utc>) -> DateTime<Utc> {
created_at.trunc_subsecs(6)
}
```
`log_inner` is the only place that assigns `created_at` — every caller
goes through `NewAuditEntry`, which carries no timestamp — so this is a
single choke point. It is wrapped in a `log_timestamp()` helper purely
so the invariant is assertable without a database.
I chose truncation at the write path over the alternative (hashing a
precision-independent encoding such as
`timestamp_micros().to_be_bytes()`). Both fix the mismatch, but
truncating keeps the existing hash preimage format and gives the
stronger invariant: the `AuditEntry` returned from `log()` is now
exactly what a later read returns.
Truncation matches what actually happens on the wire — sqlx encodes
`DateTime<Utc>` as microseconds since the Postgres epoch, truncating —
so the value hashed is the value stored.
## Validation
Toolchain note: built on Windows with the `x86_64-pc-windows-gnu`
toolchain (no MSVC linker locally).
**Before**, against Postgres 17 with `migrations/*` applied:
```
$ cargo test -p buzz-audit --lib -- --ignored --test-threads=1
test service::tests::chain_links_within_one_community ... FAILED
test service::tests::chains_are_independent_per_community ... FAILED
test service::tests::community_chain_starts_at_seq_1_with_null_prev ... ok
test service::tests::cross_community_row_does_not_verify ... ok
test service::tests::verify_detects_tampering_within_a_community ... FAILED
test service::tests::verify_empty_range_is_false ... ok
test result: FAILED. 3 passed; 3 failed
```
with `HashMismatch { seq: 2 }` / `HashMismatch { seq: 1 }` on untampered
chains.
**After**, same database:
```
test result: ok. 6 passed; 0 failed
```
`verify_detects_tampering_within_a_community` is the one to look at: it
asserts `HashMismatch` lands on the *tampered* entry's `seq`. It was
failing because verification already blew up on an earlier untampered
row — so the assertion proving tamper detection works had never actually
been exercised. It passes now.
Also:
- `cargo test -p buzz-audit --lib` (no Postgres) — 12 passed, 0 failed.
- `cargo clippy -p buzz-audit --all-targets -- -D warnings` — clean.
- `cargo fmt -p buzz-audit -- --check` — clean.
## New tests
Three in `hash.rs`, none needing Postgres:
- `storage_precision_drops_sub_microsecond_digits` — the helper's
contract, and that it is idempotent so a re-read value is unchanged.
- `nanosecond_timestamps_cannot_survive_a_database_round_trip` — asserts
the digests **differ**. This is the trap itself, written down so the
next person changing the hash preimage sees why the precision reduction
is load-bearing.
- `storage_precision_timestamps_survive_a_database_round_trip` — the
invariant the write path must hold.
Plus `log_timestamp_carries_no_sub_microsecond_digits` in `service.rs`,
deliberately **not** `#[ignore]`d, so a regression on the write path is
caught by `just test-unit` instead of only by Postgres-gated tests that
normally never run.
## Compatibility
Rows written before this stay unverifiable — they always were — so there
is no migration. An operator relying on an existing chain has to
re-anchor.
## Relationship to #2620#2620 proposes a shared `verify_entries` walk (anchoring, seq
contiguity, tail-truncation detection) plus a `buzz-admin audit verify`
command. Its Postgres-free unit tests build entries in memory and would
pass regardless, but its `#[ignore]` Postgres tests and the operator
command itself would fail on every real chain until this lands. Worth
taking this first so that work has a verifiable baseline — the two
changes don't overlap in code.
---------
Signed-off-by: Shani Singh <teamdeveloperworld@gmail.com>
## Problem
Auth-class errors (expired OAuth token, HTTP 401) are non-retryable: the
token won't self-repair between attempts. Today, `PromptOutcome::Error`
for an application-class error falls into the generic `queue.requeue()`
path, burning up to 10 retry slots over a long backoff window before
dead-lettering. Will's canary run observed the 401 message being retried
repeatedly.
## Solution
Add `is_auth_error()` that classifies `AcpError::AgentError` messages
matching two narrow patterns observed in the field:
- `"Re-authenticate"` — emitted by the Claude CLI for expired OAuth
tokens
- `"API Error: 401"` — present in Claude/Codex HTTP-401 responses
Conservative matching is intentional: a false positive (misclassifying a
transient error as non-retryable) silently drops a user message, which
is worse than a false negative (extra retries on an auth error).
In `handle_prompt_result`, a new branch intercepts the failing batch
before `queue.requeue()` for auth-class errors and dead-letters
immediately, posting a user-visible notice to re-authenticate the CLI
(e.g. `claude /login` or `codex login`).
The transport/application split in `PromptOutcome::Error` is untouched —
this only changes batch fate after an application-class auth error.
## Tests
6 new tests in `error_outcome_emission_tests`:
1. `is_auth_error` matches `Re-authenticate` message
2. `is_auth_error` matches `API Error: 401` message
3. `is_auth_error` rejects other `AgentError` messages (usage credits,
etc.)
4. `is_auth_error` rejects transport errors (I/O, WriteTimeout)
5. Auth error dead-letters immediately — 0 pending channels after
`handle_prompt_result`
6. Non-auth application error still requeued — 1 pending channel after
`handle_prompt_result`
Full `cargo test -p buzz-acp`: 598/598 passing.
Signed-off-by: Will Pfleger <pfleger.will@gmail.com>
Co-authored-by: npub1mn7jgtj4w2pd0g0zeuhxsa6jy6p0rewxz4kujt98my82ahfmp72sxjexk7 <dcfd242e557282d7a1e2cf2e6877522682f1e5c6156dc92ca7d90eaedd3b0f95@buzz.block.builderlab.xyz>
