The shutdown drain send could block past the deadline on a wedged
media thread with a full command channel — same whole-round-trip
principle as readyz's 5ce18bf. Plus session_map doc coherence and a
plan deviation note pointing snippet-copiers at the landed code.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012QndwfhjyTiZcUYp87dwW8
Signed-off-by: Aaron D. Lee <himself@adlee.work>
Add spawn_tap_teardown (tap_engine.rs) — hands the 750ms bounded
close_tx-then-abort sequence to a tokio task instead of running it
inline via block_on on the media thread. The old Delete arm blocked
the 20ms tick loop for EVERY live call on the node whenever one
session's brain was unresponsive; teardown is brain I/O and must
never run on the timing thread. Rewire MediaCmd::Delete to call it.
TDD: new test spawn_tap_teardown_returns_immediately_and_aborts_stuck_engine
asserts (a) the call returns in <100ms even with a permanently-stuck
engine task, and (b) the stuck task is aborted within the 750ms+slack
cap, observed via a Drop-guard channel.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012QndwfhjyTiZcUYp87dwW8
Signed-off-by: Aaron D. Lee <himself@adlee.work>
Task 5 review fixes: the 250ms readiness bound now wraps the whole
Stats round trip (a wedged thread can stall the send, not just the
reply), and the readyz integration test tears down its MediaThread
like the sibling test does.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012QndwfhjyTiZcUYp87dwW8
Signed-off-by: Aaron D. Lee <himself@adlee.work>
Implement task 5: liveness and readiness probes for the HTTP stack.
- GET /healthz: 200 OK always (process + HTTP stack up). Deliberately
does NOT consult the media thread — liveness and readiness have
different restart semantics.
- GET /readyz: 200 + MediaStats JSON iff media thread answers Stats
within 250ms AND node is not draining AND sessions < max_sessions;
else 503 + JSON or "media thread unresponsive" (zombie case:
dead thread while HTTP stack still answers).
Unit test: readyz_503_when_media_thread_gone_but_healthz_200 (routes.rs)
shows the zombie-node failure path from 2026-07-04 review.
Integration test: readyz_200_with_stats_json_when_thread_alive
verifies 200 + valid JSON from a live media thread.
Both handlers follow the brief's signatures; no seam touches.
All existing tests remain green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012QndwfhjyTiZcUYp87dwW8
Signed-off-by: Aaron D. Lee <himself@adlee.work>
MediaCmd::Drain flips a draining flag that rejects new Registers with
the "draining" contract string (routes already map this to 503, Task
3) while existing sessions keep ticking; the reply fires once the
session map empties, checked right after the eviction sweep so a tick
that closes the last session completes the drain in the same
iteration. MediaStats.draining now reports the real flag instead of
the Task 3 placeholder.
config::drain_deadline parses RUTSTER_DRAIN_DEADLINE_SECS, defaulting
to 0 (today's instant dev-loop shutdown). main.rs restructures the
serve/shutdown tail into two phases: on SIGTERM, send Drain while axum
keeps serving (in-flight calls still need DELETE/offer, and Task 5's
readyz will flip so the LB pulls the node), wait for drain-complete or
the deadline, THEN quiesce HTTP and call media_thread.shutdown() for
the hard stop. Calls are non-migratable by design, so drain-then-
terminate is the only graceful scale-in shape.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012QndwfhjyTiZcUYp87dwW8
Signed-off-by: Aaron D. Lee <himself@adlee.work>
Register now sheds the marginal call ("node full") once sessions.len()
>= max_sessions instead of degrading every in-flight call's 20ms
budget. The fixed sleep(META_TICK) tail is replaced by a compensated
sleep that tracks tick_overruns/last_tick_micros — the ADR-0010
benchmark's primary readout and the autoscaling signal. MediaCmd::Stats
exposes {sessions, max_sessions, draining, tick_overruns,
last_tick_micros} for the future /readyz handler. routes.rs maps
"node full" and "draining" to 503 so an LB retries elsewhere.
RUTSTER_MAX_SESSIONS (default 64) wires the cap from env.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012QndwfhjyTiZcUYp87dwW8
Signed-off-by: Aaron D. Lee <himself@adlee.work>
Multi-agent scalability & infra-fit review (docs/reviews/) and the
slice-5 seams plan it produced (docs/superpowers/plans/). Slice 5
plants the horizontal-scaling seams — MediaAddressConfig (B1),
admission/Stats (M2), drain (M1), EventSink (M3), non-blocking
teardown (M7), ADR-0009 amendment (M5), tap v2 reservations (M4) —
before further slices calcify around their absence.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012QndwfhjyTiZcUYp87dwW8
Signed-off-by: Aaron D. Lee <himself@adlee.work>
Single-agent (no PM/relay) kickoff to close out slice-4: the
secondary-path advisory e2e and the CI seam gate. API facts verified
against main @ 60b65ab (mock schedule API, ReflexMetrics fields,
lib.rs exports); seam gate specced as pinned blob hashes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012QndwfhjyTiZcUYp87dwW8
Signed-off-by: Aaron D. Lee <himself@adlee.work>
These four reviews were filed on slice-1-review-fixes and
strategic-reviews-post-pivot-rescore but never reached main; both
branches are being retired in the post-slice-4 cleanup. Kept for
lineage: the 06-29 post-pivot re-score is where ADR-0007/0008 era
strategy was re-scored, and the 07-03 reviews build on these.
Sources: slice-1-review-fixes @ d2ef53b (gtm-path, vision-sanity-check,
slice-1-claude-adversarial-assessment), strategic-reviews-post-pivot-
rescore @ ea27167 (post-pivot re-score).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012QndwfhjyTiZcUYp87dwW8
Signed-off-by: Aaron D. Lee <himself@adlee.work>
Task 1 review flagged the new reflex module was missing from the
module-map //! catalogue. AGENTS.md learner-facing convention requires
the catalogue entry describing what the module does + why it exists.
Addresses the Minor finding from the Task 1 task review.
Signed-off-by: Aaron D. Lee <himself@adlee.work>
The critical-path foundation for the barge-in reflex. AdvisoryEvent is
the enum carried over a tokio mpsc from TapEngine to Reflex (brain →
FOB). ReflexMetrics is the observable surface. barge_in_flush is the
new AudioPipe trait method (default delegates to clear_playout_ring) —
the kill-now path that clears the ring AND drains rx_audio_out.
Task 1 of the slice-4 plan. Everything else depends on this landing.
Signed-off-by: Aaron D. Lee <himself@adlee.work>
Every commit MUST be signed off (git commit -s). Preserves the
copyright-holder's option to dual-license later — without a DCO/CLA
instrument, every external contribution erodes that option (ADR-0004).
Particularly load-bearing under multi-agent parallel authorship, where
multiple sessions commit in parallel and the DCO is the only per-commit
aggregation instrument.
Adds DCO.md with the full Linux-Foundation DCO text + the agents-sign-off-
with-the-human's-identity rule. AGENTS.md Git workflow gains the DCO bullet.
AGENTS.md +292 lines: the relay (model-agnostic MCP message-bus at
localhost:7110), PM session launch checklist (poller, watch.sh, kitty/tmux
spawner), session handoff (inbox.log / poller.log / relay-log.jsonl), PM
turn-start discipline (drain inbox, list_pending, git log, surface
proactively), multi-dev parallelism (5-rule checklist with the slice-3
anti-pattern), when-the-PM-is-blocked protocol. Plus the slice-4 PM
kickoff prompt under docs/superpowers/kickoffs/.
Load-bearing for the multi-agent workflow executing slice-4 (and beyond):
session discipline (turn-start polling, no standby mode, proactive
surfacing) was the prior failure mode; the prompt + these sections bake
the correction into AGENTS.md so a fresh PM agent session inherits it.
Design record for the operator console — the downstream non-technical
operator's surface. **Not scheduled for current build;** lands after the
spearhead (steps 1-6) and the ACD/escalation rungs (capability ladder
rungs 1-2). The current build target is slice-2; the next build target
after slice-2 is slice-3 (the OpenAI Realtime adapter). An agent that
proposes building the console now is wrong — point it at spec §0.
Records the visual-brainstorming session's resolved decisions so they
don't get re-litigated at implementation time:
§1 Principles (load-bearing):
- GUI is a pure API client — **green-zone, never an insider** (ADR-0008);
holds no privilege a third party couldn't get from the public API.
- Config-as-code is the single source of truth; the UI is a reconciled
*view* (the k8s model — declarative desired-state in git, the engine
reconciles; console + CLI + git are views over it).
- Operator persona, not builder. The builder authors flows in **code +
an AI pair** (ARCHITECTURE 'DX spine'); this console is NOT the
authoring surface for the technical builder. **No drag-drop authoring
canvas** — a deliberately-closed decision (vision-revision §8).
- Ships in the batteries-included distro and the all-in-one image
(the reference GUI is bundled — 'boom' includes a usable console).
§2 IA — Ops Console shell: left nav + center panel + persistent live rail.
Dense single-screen; for an operator who lives in it. Live rail always
visible regardless of the selected section.
§3 Dashboards — Live Ops (default landing) + Quality (sibling view). Maps
the k8s split: Live Ops = live operational state; Quality = the data-owned
eval loop (capability ladder rungs 3→4).
§4 Flows editor — master/detail + a generated, navigable preview map.
The preview map is a *view* rendered from the config, not a drag-drop
authoring canvas. Earns its place by showing branching (the escalation
split) the linear stage-list flattens.
§5 Declarative flow schema (illustrative YAML; format is open). The config
the console edits via the API; the builder edits in git; the engine
reconciles. The map is generated from this structure.
§6 Other sections (conventional): Agents · Numbers · Queues · People ·
Recordings · Settings.
§7 Architecture: separate app, public API only; never reaches into the FOB.
Config writes go through the reconcile path. Live data rides the event
stream off Valkey (ADR-0005); never the 20 ms media loop.
§8 v1 vs. later scoping. §9 open decisions (flow config format, map
rendering depth, quality metrics, live transport, framework).
Cross-refs ADR-0002 (fused vertical + composable platform), ADR-0006
(WebRTC ingress / escalation UX), ADR-0007 (rented trunk / Numbers
config), ADR-0008 (FOB / green-zone — the UI is green-zone),
ARCHITECTURE.md 'DX spine' + 'GUI & extension architecture', PORT_PLAN §5
(contact-center domain).
- *.kate-swp: Kate editor's swap files have a different extension pattern
than vim's *.swp; without this, kate-swp files like
docs/reviews/.2026-06-28-slice-1-claude-adversarial-assessment.md.kate-swp
leak into the working tree.
- .superpowers/: the superpowers companion scratches its visual-brainstorming
server state (HTML mockups, server.log, server.pid, server-info) under
.superpowers/brainstorm/ in the project root. The companion explicitly
advises gitignoring this directory; scratch artifacts are not source.
Mechanical propagation of the crate rename ( rutster-signaling-sip →
rutster-trunk ) and the repo URL fix ( github.com/anomalyco →
git.adlee.work/alee ) through the documents that name them:
- docs/DEVELOPMENT.md: crate-layout sketch + stub-crate description.
- docs/superpowers/plans/2026-06-28-slice-1-webrtc-loopback.md: the
workspace members list in the plan's binding-values section —
repository URL updated.
- docs/superpowers/plans/2026-06-28-slice-2-agent-tap.md: workspace
members list in binding-values + crate-layout sketch in the file-
structure section (rutster-signaling-sip → rutster-trunk).
- docs/superpowers/specs/2026-06-28-slice-2-agent-tap-design.md: §2.1
workspace-layout sketch's STUB crate-row.
No content changes beyond the substitutions; the slice-2 spec/plan
body (protocol, TapAudioPipe, TapClient, TapEngine, lifecycle, done-
criteria) is untouched.
ADR-0007 (rent the trunk) + ADR-0008 (FOB/green-zone doctrine) propagation
into the narrative docs that orient contributors + readers. No code changes.
README:
- Wedge bullet 2: 'one secure auditable boundary' now lists media + local
reflexes + spend + tap + audit (was: trunk termination + media + spend).
Adds the honest caveat that PSTN media inside the boundary is the on-prem
*graduation* (ADR-0007), not a day-one claim.
- 'Is / isn't': 'isn't a TDM/PSTN-hardware PBX — *and not a SIP stack.*'
No first-party SIP (ADR-0007). PSTN reach is rented transport.
- Memory-safety pillar: 'rutster parses no SIP at all' — entire first-party
wire surface is WebRTC/RTP/SRTP + WebSocket tap/ingress, all memory-safe
Rust. The carrier-SIP interop tail lives outside the trust boundary.
- Spearhead step 5: 'Add a real phone number via rented transport' (was:
'Replace WebRTC ingress with a real PSTN trunk call'). Re-aims at the
AI-telephony frontier; no first-party SIP stack.
- References: 0007 supersedes 0003 in the highlighted ADR list.
ARCHITECTURE:
- New 'FOB and the green zone' section after the fused-vertical framing —
names the build-vs-reuse doctrine with the FOB member list + green-zone
member list, and restates memory-safety precisely ('FOB is 100%
memory-safe Rust; the green zone is trusted OSS kept outside the
boundary — not an over-claim that every byte is Rust').
- 'Inside the boundary': 'Carrier trunk — rented transport, not first-party'
replaces 'Trunk SIP termination — Rust-native.' PSTN audio arrives as
a media-leg ingress from a rented CPaaS raw-media fork or an out-of-tree
SBC for on-prem sovereignty.
- Biggest technical risk: 'No longer the SIP stack — because rutster no
longer builds one.' Retires ADR-0003's named schedule risk; redirects
to 'the reflex loop itself' (turn-taking, VAD-driven barge-in, jitter,
pacing) — which is also the differentiator.
PORT_PLAN:
- Design rule 4: 'rent it, don't own it' replaces ADR-0003's 'Rust-native
trunk SIP, no SBC shield.' ADR-0007 restores the rule's original instinct.
- SIP signaling (trunk) row: disposition flipped to 🔌 Rented / out-of-tree.
- Outbound registration row: disposition flipped to 🔌 Rented / out-of-tree
(handled by the rented transport or out-of-tree SBC; rutster parses no SIP).
- Spend/abuse engine row: 'co-located with call origination + the tap inside
the boundary' (was: 'co-located with trunk termination'). rutster mediates
both the provider call-control API and the brain tap — the brain never
holds the wire.
- Spearhead step 5: 'Add a real phone number via rented transport.'
- Open decisions: SIP line updated to 'Re-decided — ADR-0007.'
AGENTS.md:
- ADR list: adds 0007 and 0008 with annotations; marks 0003 as superseded
by 0007.
- Key decisions to respect: prepends the FOB/green-zone doctrine as the
*the* build-vs-reuse rule. 'When in doubt, default to green zone — the
FOB earns its members, it doesn't collect them. This is why the trunk is
rented (ADR-0007) and Valkey reused (ADR-0005), not rebuilt. Don't pull
green-zone plumbing into the core.'
ADR-0007 lands in code:
- Rename crates/rutster-signaling-sip → crates/rutster-trunk. The crate was
a stub anyway (lib.rs with a doc-comment + crate_compiles() test); the
rename locks the new boundary shape — the future rented-transport
ingress (CPaaS media-leg fork / out-of-tree SBC glue, **no SIP stack**)
lands here at spearhead step 5.
- rutster-trunk/src/lib.rs doc-comment now describes the ADR-0007 split:
CPaaS media-leg adapter as primary, out-of-tree SBC for on-prem
graduation. Cross-refs ADR-0007 + ADR-0008 (FOB vs. green zone: trunk
is green zone; SIP lives outside the trust boundary).
- rutster-spend/src/lib.rs doc-comment updated to quote ADR-0007's
'rutster mediates both the provider call-control API and the brain tap'
framing — the spend gate sits in that boundary, structurally preventing
a runaway brain from exceeding spend/pacing. Pulling spend out into a
service re-introduces the 3-vendor structural hole.
Cargo bumps:
- edition = '2024' (slice-1's pinned Rust 1.85 + edition-2024 floor
already requires this; rutster-media's let-else pattern in
OpusDecoder::decode · Slice-2's let-else in AcceptOffer's Uuid
parse_str · all rely on edition 2024's stabilized let-chains / let-else).
- repository = 'https://git.adlee.work/alee/rutster' (the self-hosted
Gitea remote — matches 'git remote -v' origin). The github.com/anomalyco
URL was stale from the pre-pivot copy.
- Cargo.lock regenerated by cargo for the rename.
No behavioral code changes — the trunk crate's body is still the stub
crate_compiles() test. FOB membership (per ADR-0008) is unchanged:
rutster-trunk will be FOB-internal at step 5 because it's the media-leg
ingress (hot path); spend-spend stays FOB because spend is
security-constitutive.
ADR-0007 — Trunk/PSTN strategy: rent the transport, no first-party SIP stack.
Supersedes ADR-0003 (Rust-native trunk SIP). The 2026-06 strategic-relevance
review stress-tested ADR-0003 against the 2026 market + solo-build realities
and it did not survive: a trunk SIP/media core is the highest-cost,
lowest-differentiation square on the board (the perceived-quality battle and
the buying criteria both live above the transport), and no production-hardened
pure-Rust SIP stack exists to stand on (ezk is bus-factor-1 / pre-1.0;
str0m self-describes as 'not for production' for SIP).
rutster owns no SIP stack. PSTN reach is rented transport in three layers:
1. Primary (demo, most users) — CPaaS raw-media fork (Twilio Media Streams,
Telnyx). Media-leg ingress, core-as-server (parallel to WebRTC ingress
per ADR-0006). Use the raw-audio fork, *not* managed Voice-AI products
that would consume the reflex loop.
2. Graduation (on-prem) — out-of-tree SBC (Kamailio/FreeSWITCH/drachtio +
rtpengine) B2BUAs carrier SIP into rutster as clean RTP/tap media,
outside the trust boundary.
3. Never — a first-party Rust trunk SIP stack. rsip/ezk stay off the
critical path.
The only things rutster owns in Rust: the call model, the reflex loop, the
agent tap. Everything that touches a carrier is rented (layer 1) or
out-of-tree (layer 2).
ADR-0008 — The FOB and the green zone: the build-vs-reuse doctrine. Names
the boundary criterion implicit across ADRs 0002-0007 as one mechanical rule
every contributor (human or agent) applies the same way.
- FOB (build in Rust) — admitted only if it passes one of: hot path,
security-constitutive, differentiating. Current FOB: media termination
(rutster-media on str0m) · real-time reflexes (VAD/barge-in/jitter/pacing)
· call model (rutster-call-model) · agent tap (rutster-tap) · spend/abuse
gate (rutster-spend) · control API + state/bus trait (rutster).
- Green zone (reuse at arm's length) — its own process/container/trust
domain, never in the FOB's address space. Admitted when it fails all
three FOB tests AND a trusted, actively-maintained project already does
it well. Current green zone: Valkey (ADR-0005) · carrier trunk (ADR-0007)
· agent brain · reference GUI · object storage · KMS · OTel collector ·
container supervisor.
- 'Actively maintained' is the load-bearing gate: excludes a first-party
Rust SIP stack and webrtc-rs; admits Valkey and str0m-for-WebRTC.
- When in doubt, default to green zone. The FOB earns its members.
Restates the memory-safety pillar precisely: the FOB is 100% memory-safe
Rust; the green zone is trusted battle-tested OSS kept outside the boundary
— not an over-claim that every byte is Rust.
ADR-0003's status is flipped to Superseded with a forward-pointer to ADR-0007.
Kept as the historical record of the 'own trunk SIP + media termination in
Rust, no SBC shield' plan; reversed under the strategic-relevance review.
- session_map.rs: AppState::close outer timeout bumped 500ms → 750ms
so the inner close-arm bound (500ms in tap_client.rs) has room to
finish `ws.close(None).await` cleanly before the abort fallback
fires. Outer > inner per final-fixes re-review Important #1.
- tap_integration.rs: reconnect-path test now asserts the resumed
`next_pcm_frame()` returns the FRESH marker (samples[0] == 9, not
the stale samples[0] == 7), actively witnessing the §5.3 step 4
"no stale bleed-through" contract. Pre-kill seeding strengthened
to ≥2 frames so step 7's silence-after-flush assertion is non-
vacuous. Per final-fixes re-review Important #2.
Spec ref: 2026-06-28-slice-2-agent-tap-design.md §5.1 step 5, §5.3 step 4.
Code-review findings from the slice-2 final whole-branch review:
- tap_client.rs: close arm sends session_end (not bye) per §5.2, awaits
brain bye with bounded 500ms timeout, then ws.close(). Promotes the
Task-7-documented Minor to Important's required spec compliance.
- session_map.rs: AppState::close() now clears channel.tap = None before
Closing->Closed, sends close signal, bounded-awaits engine task for the
teardown handshake (500ms cap), aborts as fallback. Per §5.1 step 5.
- pcm.rs + tap_audio_pipe.rs: AudioPipe trait gains clear_playout_ring()
(default no-op for EchoAudioPipe); TapAudioPipe clears the VecDeque.
RtcSession::clear_playout_ring delegates. Per §5.3 step 4 playout-on-
disconnect-flush contract.
- tap_engine.rs: spawn_tap_engine creates + shares a flush_tx mpsc
channel; run_engine_loop signals flush after each failed pump loop
(before backoff); TapConn.flush_rx is drained by drive_all_sessions
to call RtcSession::clear_playout_ring.
- tests/tap_integration.rs: new reconnect-path test asserts Channel
stays Connected, metrics.reconnect_attempts increments, playout
returns None during outage, playout resumes after EchoServer
restart with fresh frames (no stale bleed-through). Per §8.4.
- static/index.html: one static line acknowledging external-brain
routing per §7.2 (defer real-time tap status to step-3 GUI).
Spec ref: 2026-06-28-slice-2-agent-tap-design.md §5.1 step 5, §5.2, §5.3 step 4, §7.2, §8.4.
- tap_engine.rs: Backoff::next_delay now matches spec §4.3 exactly:
250ms -> 500ms -> 1s -> 2s -> 5s cap (no intermediate 4s step from
naive doubling). Test asserts the spec enumeration including the
"stays at 5s forever" cap-after-step-5 invariant.
- session_map.rs: removed the dead 'Closing && tap.is_some()' branch
from drive_all_sessions. AppState::close already handles teardown
inline (fires close_tx, aborts task, clears field, advances
state, removes entry). The branch was never exercised and
pre-paved a "future peer-initiated close" path -- both AGENTS.md
anti-patterns ("don't pre-pave the wrong pattern").
Spec ref: 2026-06-28-slice-2-agent-tap-design.md §4.3 step 4 + §5.1 step 3.
- pub struct TapHandle(()) — zero-sized marker (Option<TapHandle> compiles
to a bool, no allocation). The live tap connection lives in the binary's
DashMap<ChannelId, TapConn>, keyed by the channel's existing ChannelId
(== wire session_id per spec §5.3).
- Channel grows pub tap: Option<TapHandle>. None until Connected, set on
Connected, cleared before state advances to Closed.
- State invariant: tap and ChannelState are tied (Connected+None=bug).
- Slice-1 backwards compatible: new field is Option<...>;
Multi-tap-per-channel is the future trigger for a TapId newtype (YAGNI).
Spec ref: 2026-06-28-slice-2-agent-tap-design.md §5.2, §6.
- run_tap_client: drives a connected WebSocketStream with a tokio::select!
over rx_pcm_in (inbound PCM → audio_in WS frame) and ws.next() (brain
frame → audio_out mpsc or control handling).
- Handshake: send hello, await brain hello (bounded 2s timeout).
- seq gap detection (log + count, never drop on gap — spec §3.1).
- Hot-path errors (encode/decode failures, send/recv failures) are logged
+ counted; TapClientError is returned only on graceful close, hello
timeout, or WS errors. The TapEngine (Task 7) decides reconnect policy.
- Pure-helper unit test (elapsed_ms); full pump behavior is exercised
against the in-process EchoServer in the Task 8 integration test.
Spec ref: 2026-06-28-slice-2-agent-tap-design.md §4.2.
