Story 012: Remote Agents Solve a Problem Through a Shared Chair Terminal¶

Status: Implemented; baseline, vfkit manifest-repair, and additive PowerDNS/Bielik single-host acceptance pass

As a node operator, I want three model-backed participants on separate Orbiplex nodes to deliberate through a shared Room while the chair exposes a bounded, read-only view of its Workbench terminal. The participants should use the same current terminal state as evidence, propose a solution, and leave terminal actuation under the chair node's local authority.

This story composes existing contracts rather than introducing a second agent, terminal, or collaboration runtime:

Corpus supplies the deliberation policy, roles, and inert answer draft;
Agent supplies one bounded controller on each participating node;
Inquirium supplies model inference to each Agent;
Room supplies membership and the network collaboration carrier;
Sensorium Workbench owns the chair-side PTY and command execution;
Sensorium Interfaces publishes its bounded visible viewport;
the destination daemon resolves a generic Agent observation need through Interaction Broker and the Room/Sensorium adapter.

The active execution profile is the VFKit-backed variant from the shared Story 012 vertical: a digest-pinned full-system guest is started and observed as the chair terminal source, with additive vfkit/profile wiring from the Story 012 acceptance trackers (including the critique-gated extension). In practice, this story is implemented and validated as the same deliberation/repair loop over a Workbench runtime backed by that VM environment, with direct control remaining local to the chair node.

The profile is executable. Its process runner reuses the extracted Story 011 three-node federation/bootstrap layer, then composes the Workbench, Sensorium Interface, Room observation, Agent, and story fixture layers without creating a second trust bootstrap or collaboration runtime.

The three node processes use distinct loopback addresses on one host. This is explicitly multi-address single-host acceptance: it strengthens HTTP/WSS process and TLS-identity evidence, but does not claim multi-host deployment, public relay reachability, NAT traversal, or independent host failure domains. Those remain owned by the P070 deployment profiles.

An explicit single-address-single-host fallback allows unattended execution without privileged loopback aliases. It preserves process and port isolation but is intentionally weaker than the default address-distinct profile and is reported as such.

An additive physical profile is planned under P074. It consumes a reusable operator-owned topology instead of embedding hostnames in this story. Story 012 maps the stable slots as follows: node-a owns the query, Room, Chair executor, Workbench VM, and aggregate report; node-b supplies the solver Agent; and node-c supplies the reviewer or facilitator Agent. The topology remains free of these semantic duties so the same three hosts can serve other acceptance stories.

The profile may be orchestrated from any participating host. The invoking host runs its local slot directly and controls the other slots through the P074 SSH executor, while one lease owned by node-a fences the scenario and permits an idempotent resume. SSH is not a Room relay or Sensorium tunnel: Room, Corpus, WSS, Sensorium Interfaces, and model-facing product traffic must cross the real daemon endpoints. Completion requires per-node evidence and a closed aggregate manifest bound to the exact topology digest.

The physical profile also treats host filesystem layout as an admitted input. The harness strips inherited ORBIPLEX_MODEL_ROOT from daemon children and injects one distinct writable model-store root per slot; only verified source model bytes may be shared read-only. The vfkit executor on node-a uses a short canonical workspace that fits every derived Unix socket path and, on macOS, shares an APFS volume with the verified guest base image so disk preparation can use clonefile. Durable ledgers and reports may remain under the descriptive acceptance state/report roots.

Concrete Problem¶

The chair node owns a small fixture whose deterministic manifest test fails because input paths are emitted in filesystem iteration order rather than in canonical lexical order.

The expected collaboration is:

the chair-side operator runs the failing test in a Workbench terminal;
the remote implementer Agent observes the bounded visible viewport and proposes canonical ordering before digest calculation;
the remote reviewer Agent asks for a regression case proving that different input enumeration orders yield the same manifest;
the chair Agent synthesizes the proposals into an inert action plan;
the local operator applies the change and reruns the test;
all admitted Agents observe the passing terminal state; and
Corpus accepts the chair Agent's terminal outcome as an unpublished answer draft with evidence references.

The exact fixture may change without changing the story contract. It must remain small, local, deterministic, network-independent, and incapable of accessing credentials or the operator's ordinary working tree.

Actors¶

Node A / chair node owns the Corpus query, Room authority, chair Agent, isolated Workbench workspace, terminal session, and Sensorium Interface publication. Its local operator remains the only terminal actuator in the first profile.
Node B / implementer node runs a Room-attested Corpus participant Agent. It may deliberate and observe the shared terminal only after receiving both Room membership and an exact Sensorium Interface observation grant.
Node C / reviewer node runs a separately admitted participant Agent under the same dual-authority rule. It reviews the diagnosis and the proposed regression evidence but receives no terminal-control authority.

Agents remain node-local. The story federates observations and deliberation; it does not migrate an Agent runtime or its private memory between nodes.

Architectural Profile¶

Node A Workbench PTY
  -> bounded visible-screen projection
  -> Sensorium Interface resource
  -> exact participant-scoped observe grants
  -> active Room relay epoch over WSS
  -> Node B / Node C Room recipients
  -> host-owned observation admission
  -> bounded Agent passage context
  -> Inquirium
  -> inert Corpus reasoning turns
  -> Node A chair Agent
  -> unpublished Corpus answer draft

The terminal screen is not a Room message and is not appended to the Corpus transcript. Room carries a cursor-free, coalesced sensorium-interface-read-result.v1 containing one inline latest-state snapshot. The source cursor, PTY handle, source credentials, and Workbench lease remain local to node A.

Flow¶

Node A creates the Corpus query and opens a bounded Room under the same policy and invite model used by Story 011.
Nodes B and C accept signed Room invitations and create narrowed collaborative-participant Agent bindings. Node A creates the corresponding collaborative-chair binding.
Node A creates an isolated Workbench terminal session under an allowlisted story workspace, pins sensorium-operational-context.v1 with impact/class = test to that exact environment, and starts the deterministic failing fixture.
Node A publishes only the visible terminal-screen representation as a latest-state Sensorium Interface. Ordered terminal-event replay is refused.
The host issues exact, expiring subscribe grants for that interface to the admitted B and C participant subjects. Room membership by itself does not satisfy this step. The first profile does not add an independent one-shot read grant.
The active Room relay projects the view only to recipients in the intersection of current Room observation rights and current Sensorium Interface grantees.
Each Agent binding fixes a generic need, opaque source ref, payload schema, freshness, and byte bound. The recipient daemon resolves that need through its Room/Sensorium adapter and validates the read result, inline interface frame, Room, relay epoch, Room membership source sequence, recipient, Agent binding, classification ceiling, byte cap, exact source generation, operational context, effective publication, and freshness before producing ephemeral inert context.
The Agent controller uses the accepted latest state in one bounded passage. The durable step trace contains only schema, refs, classification, policy digest, and content digest; it contains no terminal bytes or prompt text.
B and C publish inert Corpus reasoning turns through the existing corpus.room.turn effect and human-in-loop policy. They do not invoke the terminal.
The chair Agent proposes a bounded next action. In the first profile, only local control on node A may enact terminal input, command execution, resize, signal, patch, or file mutation through Workbench.
The host revokes C's exact interface grant and waits until the source-side projection reports the reduced recipient set. C is refused before any repaired terminal state exists.
Node B is dirty-restarted to prove durable Agent and invitation recovery without retaining terminal content in the recipient process.
After the local operator applies the accepted change and reruns the fixture, the latest-state projection shows the passing result to all still-authorized participants. B observes the new source version; C remains refused.
Corpus accepts the chair's agent.outcome.v1 as an inert answer draft. Rendering, publication, settlement, and durable terminal capture remain separate transitions.

Authority Contract¶

The story requires two independent authorities for every remote observer:

current Room membership with the relevant observation right; and
a current exact-resource Sensorium Interface subscribe grant.

Neither authority implies the other. In particular:

Room membership does not grant access to the terminal view;
an interface grant does not admit its holder to the Room;
terminal subscription does not grant sensorium.interface.read, sensorium.interface.invoke, or sensorium.interface.manage;
a participant Agent cannot use terminal input, resize, signal, command, patch, or file-write operations;
the chair Agent does not gain ambient Workbench authority merely because its operator owns the terminal; and
revoking either authority stops future delivery without closing the durable Room.

Observation-To-Agent Boundary¶

The Agent contract is horizontal. agent-core carries only a bounded AgentObservationNeed, a durable AgentObservationBinding, and prompt-free resolution evidence. Their source refs are opaque; the core neither imports nor interprets Sensorium, Workbench, Room, or provider types.

Operator-authored JSON-e Flow configuration may predeclare a bounded mapping from need/ref to source/ref, payload schema, freshness, and byte limits, together with separate grant requests. The configuration is schema-validated and digest-pinned. Rendered flow data may select or narrow a predeclared mapping but must not construct or widen one, and Agent/model/observation data never interpolates an authority-significant wiring field.

The destination daemon is the composition root. For this story it selects the Room/Sensorium resolver, which:

binds the read result and its single inline snapshot to the exact interface, Room, relay epoch, Room membership source sequence, recipient subject, Agent, durable Agent binding, and generic observation need;
rejects unbound, dynamically selected, changed-schema, or widened-bound needs before source I/O;
rechecks both Room and interface authority before and after the broker read;
admits only the declared terminal-screen snapshot schema and latest-state delivery profile;
enforces classification, age, item-count, and byte ceilings before prompt assembly;
coalesces superseded snapshots rather than replaying terminal history;
exposes the accepted observation as inert context, never as an effect request;
records only prompt-free generic metadata, the validated source causal/context, source-version/ref, resolution/ref, policy evidence, and a host-keyed content digest in durable Agent trace; and
discards terminal bytes after the bounded passage unless local control performs a separate classified Workbench capture.

Responsibility for summarizing terminal content belongs to a separately authorized component. The Agent context bridge may select and bound an existing snapshot, but it must not silently become a transcript store or summarizer.

Substrate Gates¶

Gate	Current state	Required evidence before execution
Story 011 Corpus/Agent deliberation	available	selected participant and chair Agents deliberate over Room with restart-safe bindings and inert final draft
Room Phase 6A relay	available	three-node member-visible WSS relay carries bounded Room and Sensorium Interface payloads with epoch fencing
Workbench terminal source	available	isolated PTY, bounded visible-screen snapshot, exact environment generation/context, local actuation authority, and classified explicit capture
Sensorium Interface Room projection	available	exact grants, complete context-bearing `latest-state`, recipient intersection, host-only terminal status, supersession, revocation, recipient-side restart recovery, and no terminal control; the source-host pump remains process-local and must be recreated after a source-host restart
Agent observation admission	available	substrate-neutral need/binding/evidence in `agent-core`, bounded neutral context qualifiers, preserved P081 source causality, static fail-closed JSON-e wiring, daemon-owned Room/Sensorium resolution, process-local revocable latest-state inbox, resource-bound Interaction Broker source, host-owned pre-inference caution, prompt-free trace, and restart/retention refusal tests
Story 012 process runner	available	the baseline runner extends the shared Story 011 topology; the additive vfkit v2 profile runs the pinned guest fixture and full collaboration through one Workbench runtime without copying trust/bootstrap logic

The acceptance pack must refuse execution while any gate is missing. Marking a documentation row complete is not sufficient evidence; the runner must probe the corresponding runtime surface or execute its refusal vector.

Acceptance Profile¶

The operator-facing pack lives in:

node/tools/acceptance/story-012-shared-chair-terminal/

Its checked-in profile is executable through profile_plan.py smoke. The shared three_node_federation.py layer owns profile rendering, federation-root bootstrap, process lifecycle, and dirty restart for both Stories 011 and 012. The Story 012 runner adds only Workbench, Sensorium Interface, Agent-observation, Room-deliberation, and deterministic fixture behavior.

The smoke activates the local relay epoch through a reserved .invalid endpoint because it is proving composed authority and lifecycle behavior, not network deployment. P070's separate multi-process host-TLS acceptance remains the deployment evidence for the external relay boundary.

The composed profile binds every required refusal to named executable evidence. The process smoke directly proves:

three distinct daemon identities and node-local Agents;
signed Room invite and membership admission for B and C;
no terminal view before the exact interface grant exists;
no view from Room membership alone or from an interface grant alone;
cursor-free bounded latest-state delivery over the active Room relay epoch;
exact generic need/binding plus refusal of a binding belonging to another Agent;
refusal of unbound, dynamically interpolated, changed-schema, or widened source mappings before source I/O;
B and C can deliberate from the shared view but cannot invoke or manage it;
B and C receive the same source-owned test operational-context qualifier before their first feed-dependent turn, while publisher summary text remains unprivileged;
C's revocation converges before repair, stops both the current read and the new passing-state read, and leaves the Room plus B active;
dirty restart of recipient B restores its durable Agent and Room invitation while its process-local observation inbox starts empty and refreshes from current state;
terminal bytes do not enter Room messages, Memarium Agent facts, status, notifications, or prompt-free traces;
A raises the effective context to production through an immutable replacement; B's old statically bound Agent refuses the superseded publication and a new exact binding admits the replacement, while C remains refused;
the passing result is observed after local chair-side actuation; and
the chair outcome remains an unpublished Corpus answer draft.

Lower-stratum suites named in profile.json prove wrong Room, interface and participant binding, conflicting relay-position digests, stale relay epochs, classification ceilings, ordered-event refusal, and remote actuation refusal. This keeps protocol conformance in P070/P082/P083 while making the Story-level evidence ownership closed and machine-validated instead of implying that one process runner duplicates every lower-layer test.

The post-MVP operational-context extension tracked by P082, P064, P069, P071, and P073 is implemented. The composed smoke proves that B and C receive the same source context qualifier before their first feed-dependent Inquirium call, that publisher summary text is not retained as privileged instruction evidence, and that a monotone test -> production host floor is published only through immutable replacement. The old Agent binding refuses the superseded interface/id; the replacement requires a new exact binding and current grant. Missing, malformed, oversized, stripped, downgraded, generation-mismatched, or superseded context fails the collaborative observation passage closed before terminal bytes reach a model, with lower-stratum P064/P071/P082 tests owning the vectors that the composed story does not duplicate. P082 owns this freshness predicate; Story 012 adds no TTL.

The additive profile-vfkit-full-system.json keeps the completed baseline unchanged while proving the stronger vertical in one Workbench runtime. Version 2 boots the pinned full-system guest, verifies the digest-pinned in-guest repair fixture, projects failing and passing PTY state through P082, fences repair through P083, and runs the same three-node collaboration, revoke, dirty restart, export, and inert Corpus-draft lifecycle. Its schema-gated report names the evidence boundary single-runtime-vertical; it does not infer that claim from independent harness results.

The sibling story-012-cloud-hypervisor-full-system profile proves the same single-runtime vertical on a real Linux x86_64 KVM host with the cloud-hypervisor-system.v1 backend. The retained 2026-08-25 report passes all 11 closed checks in 1,503,122 ms, including the three-node Room/Agent flow, failing and passing guest terminal observations, exclusive P083 repair authority, observer revocation, dirty restart, stale-generation refusal, artifact export, and an unpublished Corpus draft. Its file digest is sha256:ea7f5b4a3cff2a784a7d26cad22c0bf0a1ea385e7324ced96e372482376aa920. This is deterministic deployment and mechanism evidence; it does not replace the separate real-model deliberation profiles.

The additive profile-powerdns-bielik-vfkit.json is the first full-system service-configuration specialization. The image builder pins pdns-server=4.8.3-4build3, pdns-backend-bind=4.8.3-4build3, the closed PowerDNS fixture, and the final image digest. Build-time vfkit NAT exists only while obtaining those exact Ubuntu packages; runtime vfkit still has no NIC, host share, or SSH service.

Nodes B and C each supervise a separate direct local llama-server over HTTP and select the same stable Bielik alias through their own Inquirium runtime candidate. The acceptance consumer verifies the actual Agent product in the daemon object store by digest and uses its bounded text as an inert Corpus turn. No model output is interpreted as a command. After the existing HIL gate, node A alone may apply bounded host-owned experiments and the checked-in deterministic configuration through P083 exclusive leases. Success requires PowerDNS to listen only on 127.0.0.1:53, answer authoritatively for localdomain, and return exactly a -> 127.0.0.1, b -> 127.0.0.2, and c -> 127.0.0.3. The report consumes the guest's structured PASS evidence and retains the observed transaction ids, expected and actual addresses, and localhost peer rather than reconstructing a declarative success list.

The current additive runner strengthens the deliberation without widening that effect boundary. The requester query is the opening contribution; accepted Corpus instruction overlays frame B as solver and C as reviewer. A controlled Chair Agent maps Corpus baton to Room round-robin, producing the ordered cycle requester context -> solver -> reviewer. The soft deadline is five minutes, the structural guard is 64 cycles, and at least two cycles are required. After the first solver/reviewer pair, node A applies one bounded deliberately incorrect CandidatePlan through a short-lived P083 lease; the DNS verifier rejects it and the failure becomes fresh read-only terminal evidence for the second cycle. That cycle observes the newer terminal state before node A submits a distinct corrected CandidatePlan through a fresh Agent decision, HIL decision, and P083 lease. Both plans are currently constructed by host-owned acceptance code and bound to solver turns; their commands are not derived from the model text. The profile therefore proves feedback and control mechanics, not causal authorship of the plans by the deliberating Agents. Natural-language output is never compiled into terminal input. The checked-in fixture is permitted only after the deadline as explicit diagnostic recovery; a run that needs it cannot emit the promoted 17-check report.

This bounded executor keeps the acceptance effective even when node A has no local model or terminal-capable Agent: it can execute only the host-owned experiment catalog and fixture. It does not solve the general transfer of deliberation intent into actuation. That later path must carry a typed inert inquirium.candidate-plan.v1 or experiment proposal from the deliberation, attribute it to the requester-selected executor, join node A's admission, and compile through the ordinary Sensorium boundary. The requester may select a local host compiler, a local Chair Agent, or a remote Chair/designated participant Agent. The first implemented vertical enables only the local Agent; the other modes fail daemon admission until their passage adapters have evidence. Room role and membership never imply terminal authority: remote execution uses the P083 / Solution 046 Sensorium Interactive Interfaces claim/control/invoke path and still requires a separate resource-scoped Sensorium grant, current generation, bounded lease, review policy, and local refusal path. Free-form Room text is never an executable fallback.

Critique-gated technical deliberation successor¶

The additive ready profile node:tools/acceptance/story-012-shared-chair-terminal/profile-powerdns-bielik-critique-gated-vfkit.json closes that causal gap without turning Room prose into terminal input. Its Room policy permits bounded shell fragments, file paths, configuration fragments, diagnostic and verification commands, and rollback instructions as inert technical evidence. It rejects credentials, binary payloads, unbounded scripts, ambient authority claims, and direct terminal input. The profile narrows the ordinary 64 KiB Room message ceiling to 16 KiB per technical contribution.

The baton becomes requester opening -> solver -> reviewer -> Chair. A solver contribution must state a hypothesis, evidence refs, proposed changes, verification, and rollback, and must publish a content-addressed inert CandidatePlan. The reviewer emits an exactly bound accept|revise|reject judgment over the proposal digest and fresh terminal-state digest. revise must identify a replacement CandidatePlan rather than silently changing prose. The Chair defaults to block and may emit only block, request-revision, or admit-reviewed-candidate, bound to the exact review and reviewed plan digest.

Chair admission still does not execute anything. Node A validates the selected artifact, compiles a pending InquiryFlow, rechecks current grants, classification, budget, generation, idempotency, and Room authority, obtains HIL for the exact effect, and invokes Workbench only through P083 claim -> invoke -> release. A missing review, stale terminal-state binding, digest mismatch, rejected proposal, or unavailable Chair resolves to no effect. No control lease spans inference.

The typed Corpus review and Chair-decision contracts, Agent-product-bound CandidatePlan publication path, durable host admission, critique-gated runner, and closed 26-check report revision are implemented. A retained 2026-07-25 macOS arm64 run passes all 26 checks with two real Bielik runtimes and no deadline fixture. The deterministic fixture cannot satisfy that evidence claim.

Model-authored discovery successor¶

The stronger additive story-012-powerdns-bielik-discovery-vfkit profile removes the ready PowerDNS solution wrapper from the guest and from every model-facing prompt. The solver emits closed structured process or file-write intents; the Agent compiles the exact values into a content-addressed CandidatePlan; the reviewer may publish a replacement through the same requester-owned policy; and the Chair can select only the exact reviewed publication. HIL and P083 remain the only actuation path. The legacy fixture is isolated outside discovery mode and cannot satisfy the discovery report.

The requester also owns a staged admission policy. After the read-only bootstrap, each passage may mutate exactly one path already present in active terminal configuration. A later stage may follow a file path authored by an earlier model plan only after that path appears in the active configuration. The stage selector encodes solution-aware listener, declaration, and zone-data sequencing, and its selected scope is disclosed to solver and reviewer. It does not author bytes or decide whether the DNS goal is met. The model authors review findings and an optional replacement, while host policy constrains the admissible verdict for a phase; this profile does not claim an epistemically independent reviewer verdict. Full CandidatePlan digests, terminal completion markers, host-derived bounded and non-truncated effect projection, retained provisioning inventory, reviewer replacement, and restart-safe append-only facts remain checked.

The diagnostic ladder is intentionally cheaper than the deployment story. Pure contract tests run first; deterministic replay exercises policy, review, Chair, and inert-effect transitions over recorded values; a real-model bench measures one planning stage without daemons, Room, or VM; only then may the vfkit profile run as deployment evidence. Every real-model sample records prompt, evidence, response, usage, duration, verdict, and a digest of all policy/model/runtime inputs.

The original 2026-07-29 Bielik calibration did not justify deployment smoke. Its path-aware active-directive oracle accepted two of five retained scoped zone-declaration samples, while both measured zone-data stages produced no goal-ready correction. That result remains useful negative evidence rather than a passing profile claim.

The bounded 2026-07-30 role matrix then compared pinned local candidates over 130 samples without daemons, Room, or VM. Qwen2.5-Coder 7B produced 26 goal-ready and 46 policy-admitted plans from 50 solver attempts, including 9/10 ready plans on the failed-plan correction case. Qwen3.5 4B produced 2 ready and 30 admitted plans from 50 attempts and no ready correction. Qwen2.5-Coder 7B therefore remains the solver default. None of the reviewer candidates crossed the deployment gate: Qwen2.5-Coder 3B, Qwen2.5-Coder 7B, and Phi-4 Mini had 0/10, 1/10, and 2/10 contract-admitted replacements respectively, and all had 0/10 ready replacements. The matrix records no reviewer deployment selection; Phi-4 Mini is only the best observed next challenger. The current-profile model-only four-pair p95 sum is 528,400 ms, but a full authority budget still requires measured HIL, P083, terminal, and verifier latency.

The selected deployment profile now carries both guest-attested challenge descriptors. Its passing retained run selects powerdns-bind-missing-zone-data.v1: PowerDNS already has a valid loopback listener, BIND backend, and authoritative declaration, but the referenced zone-data file is absent and no final localdomain records exist in the image. The prepared-system manifest, completion record, fixture digest, and image digest bind that initial state before vfkit starts. This proves the second variant only; it does not turn one retained run into a general claim over all prepared systems.

The role boundary now has an additive v2 correction path. A reviewer can emit one bounded request-regeneration correction without plan bytes; the daemon binds it to the source proposal and shared correction-state digest, the solver authors a fresh proposal, and a fresh review must bind that successor before Chair admission. The join is append-only, restart-safe, and inert until the existing host, HIL, and P083 gates run. A 2026-08-01 ten-pair Qwen2.5-Coder 7B calibration produced ten valid requests but only one domain-ready staged successor. The mechanism is therefore implemented, while model deployment readiness remains explicitly open.

The stronger story-012-powerdns-qwen25-coder-discovery-vfkit profile now has a fresh retained passing macOS arm64 full-system run. Two separately supervised Qwen2.5-Coder 7B runtimes completed two round-robin cycles against the second challenge. The first model-authored plan was admitted through HIL and P083 but did not satisfy the verifier; after fresh terminal evidence, a different model-authored plan reached the exact DNS goal. The closed report passes 31 checks, records 15,863 prompt-plus-completion tokens, 344,259 ms of model time, 369,053 ms of deliberation, and 561,972 ms end to end, and reports the host-owned structural umask 022 framing separately from unchanged CandidatePlan bytes. It uses no deadline fixture and derives no effect directly from Room prose. The same run also produced complete P086 recordings with zero gaps or drops: 564 observations on requester node A, 19 on solver node B, and 22 on reviewer node C. A separate fresh unseeded full-system success was observed on the same date; fixed-seed repeatability remains open. This is deployment evidence, not evidence that the critique-to-regeneration v2 path has crossed its separate 0.6 promotion gate.

The same retained run closes the package-bound P085 acceptance boundary. Node A activates the operator experiment package before deliberation, uses its exact four-passage inference Flow and package-owned turn-order producer, and records the offer, policy, producer, decision, activation, passage, and terminal-product lineage. The Flow evidence repeats the package ceiling of eight passages and the binding's monotone narrowing to the four passages exercised by this profile. After restart, stale Flow and turn-order replay attempts refuse under the retired process authority; terminal revocation then makes a fresh stale invocation refuse without falling back to distribution policy or a local equivalent.

For reproducibility, the story policy supplies a closed allowlisted command-plan template to the solver. The real model must carry the exact CandidatePlan marker in its verified Agent product, and the acceptance runner publishes those retained bytes with the product ref and digest; there is no prose fallback. This proves typed product provenance and critique authority, not unconstrained plan synthesis.

The host bounds approval pressure independently from the story fixture: it admits two distinct experiment proposals per round by default, permits an operator override only up to eight, and does not charge exact proposal or idempotency replay as a new proposal. Each plan remains independently limited to one effect proposal by default and eight at the hard ceiling.

Its evidence boundary is deliberately single-host-full-system: the three Node processes, two local model processes, and one full-system VM cross real HTTP/WSS, PTY, and vsock boundaries on one macOS host. This is not evidence for separate host failure domains. Structural validation is complete; a passed report was retained from the 2026-07-24 expensive run with real llama-server, Bielik GGUF bytes, and the prepared PowerDNS image. The run completed in 155,637 ms, including 45,496 ms of deliberation, and records distinct host-issued inference refs for B and C. Equal product digests remain valid for deterministic responses and are not used as execution identity. The profile therefore proves two separately supervised invocations whose returned bytes converged; it does not claim epistemically independent first judgments. Such a stronger claim would require an explicit isolation and diversity policy rather than an output-digest inequality check. The duration is an observed wall-clock measurement, not a cold-cache benchmark: model and operating-system cache state was not controlled by this profile. The retained report also proves the role-aware multi-cycle extension: distinct solver and reviewer roles, two round-robin cycles, a verifier-rejected first experiment followed by a host-submitted corrected CandidatePlan after fresh terminal feedback, Agent propose decisions, admitted HIL, and separate P083 claim -> invoke -> release lifecycles. Its closed evidence records zero effects derived directly from Room prose; it does not claim that either plan was derived from that prose. The retained report is node:tools/acceptance/story-012-shared-chair-terminal/reports/2026-07-24.story-012-powerdns-full-system.macos-arm64.json with file digest sha256:706e1c78f069a2277cb86dde79919880e6450a1505d2aa8df08d26ab0a739de4. This additive post-MVP profile is manual or dedicated-runner evidence and does not replace, reopen, or gate the completed baseline Story 012 acceptance. The report is registered in Schema Gate with positive and fail-closed duplicate- check fixtures; structural schema passage still does not substitute for a real run.

The Story consumer validates bytes/base64, bytes/count, and bytes/sha256 before interpreting a terminal marker. Chair admission also uses the current Room membership high-water sequence rather than a creation-time or fixture constant.

Failure Modes and Mitigations¶

Failure mode	Risk	Mitigation
Room membership is treated as terminal authority	unauthorized observation	require the exact current interface grant independently on every delivery and Agent admission
Remote Agent receives an actuation grant	participant controls the chair terminal	close the observer capability set to `read` and `subscribe`; refuse `invoke` and `manage` in profile validation
Old screen snapshots are replayed as a transcript	stale or excessive context	use cursor-free coalesced `latest-state`; reject ordered-event interfaces
Terminal bytes enter durable Agent memory	credential or source leakage	retain only refs, classification, policy digest, and host-keyed content digest
A technical Room fragment is treated as executable input	prompt injection crosses into Workbench	keep prose inert; require a content-addressed CandidatePlan, bound review, Chair decision, host admission, HIL, and P083
Chair admits a plan other than the one reviewed	critique is bypassed after approval	bind Chair decision to review ref and reviewed plan digest; mismatch resolves to `block`
Revocation closes the Room	collaboration state is lost with one view	close only the projection/subscription and preserve the durable Room
Restart silently widens authority	stale grants or relay epochs revive	rebuild from durable facts, recheck revocation, and require a fresh current-state delivery
Agent or observed data changes source wiring	confused deputy selects an authority-bearing source	accept only operator-authored, digest-pinned static mappings; rendered flow data may select or narrow but never create or widen them
Story runner duplicates Story 011 trust logic	two drifting federation bootstraps	compose or extract the existing topology/bootstrap helper before implementing the runner
Chair Agent becomes terminal operator by implication	observation and effect authority are complected	keep first-profile actuation local-control-only and require a later explicit effect profile for Agent-driven commands

Non-Goals¶

Giving every participant interactive terminal control.
Sending raw PTY input or ordered terminal events as ordinary Room messages.
Persisting a full terminal transcript automatically.
Migrating Agent state between nodes.
Requiring a non-member federation relay or a public Matrix homeserver.
Treating terminal observation as evidence that a proposed diagnosis is true.
Letting an Agent publish the final Corpus answer or enact changes by itself.

Done When¶

[x] The story document and executable acceptance profile agree on topology, authority, data lifetime, and refusal behavior.
[x] The additive physical multi-host profile runs the Story on three separate machines through P074's reusable topology and any-host orchestrator, retains one topology-digest-bound aggregate report, and proves that product traffic used direct daemon transports rather than the SSH control channel. Its preflight also proves distinct slot-owned writable model stores, bounded vfkit socket paths, and same-volume APFS image cloning on the macOS executor.
[x] P069 and P073 track the substrate-neutral Agent observation port, daemon-owned Room/Sensorium resolver, and composed process evidence.
[x] The profile validator rejects terminal actuation grants, membership-as-authority, ordered-event delivery, and durable terminal content.
[x] Every substrate gate has executable evidence and is marked available.
[x] The composed three-node smoke completes the concrete problem from failing test through independent B/C deliberation, C revocation, dirty B restart, local repair, passing-state observation by B plus continued refusal for C, and an unpublished answer draft.
[x] The additive vfkit v2 profile proves full-system, interface/fencing, and collaborative behavior through one Workbench runtime without weakening or relabeling the completed baseline.
[x] Terminal marker checks consume verified byte evidence and Chair authority binds the current Room membership high-water sequence.
[x] One-runtime vfkit Story 012 evidence uses a digest-pinned guest fixture rather than a host-path copy and emits the closed story-012-vfkit-full-system-report.v1 artifact.
[x] Retain one passing story-012-powerdns-full-system-report.v1 from the additive post-MVP single-host profile using two real local Bielik runtimes and the exact PowerDNS guest image. The profile, runner, digest-verified Agent-product bridge, deterministic host-owned actuation fixture, localhost/DNS assertions, structural tests, and schema-gated closed report validator are implemented. The 2026-07-24 retained report proves the real run independently of those structural gates.
[x] Implement the additive role-aware PowerDNS deliberation runner: requester opening, solver/reviewer overlays, Corpus baton to Room round-robin, two required and at most 64 cycles under a five-minute deadline, one bounded verifier-rejected guest experiment, a separately admitted corrected CandidatePlan from fresh terminal evidence, append-only proposal/flow/operation/receipt execution projections, restart quarantine without blind retry, and fresh P083 leases. Fixture fallback is forbidden before the deadline and cannot produce a passing 17-check report.
[x] Run the expensive role-aware profile with real Bielik runtimes and vfkit, and replace the retained report only after its closed deliberation evidence passes the exact 17-check profile validator and Schema Gate. The retained 2026-07-24 report proves separate roles, two round-robin cycles, failed and corrected experiments, Agent propose, HIL, P083 lease release, and no effect derived directly from Room prose.
[x] Add the first adaptive follow-up profile in which a typed inert inquirium.candidate-plan.v1 or experiment proposal crosses from Corpus deliberation to the requester-selected local Chair Agent. The portable policy also names deterministic local compiler, remote Chair Agent, and designated participant Agent modes, but daemon admission rejects those modes until their passage adapters have executable evidence. Every plan remains inert until node A verifies attribution and a separate resource-scoped Sensorium grant, review, classification, generation, lease, budget, and idempotency policy. The implemented local vertical binds the signed proposal to a retained solver turn and the author's Room invite, compiles a pending InquiryFlowV1, prepares fresh latest-state evidence, performs an Agent passage, verifies the content-addressed Inquirium product under a closed propose|no-effect decision contract, and prevents no-effect from claiming or invoking. For propose, node A rechecks the exact flow node, interface, grant, generation, operational context, method, input schema, payload digest, classification, lease ceiling, and proposal expiry, requires an admitted operator-question decision, and lets P083 acquire and release the lease only around invoke. Execute future admitted remote control through P083 / Solution 046 rather than a Story-local proxy; the remote executor carrier remains additive work and is not claimed by this local evidence. Free-form prose is rejected as effect authority.
[x] Define the additive critique-gated technical deliberation profile with bounded shell/configuration guidance as inert Room evidence, ordered solver -> reviewer -> Chair control, fail-closed Chair decisions, exact digest lineage, and unchanged HIL/P083 effect authority. The checked-in profile is structurally validated and marked ready.
[x] Freeze typed corpus-reasoning-experiment-review.v1 and corpus-reasoning-chair-experiment-decision.v1 contracts, plus positive and fail-closed fixtures for missing review, stale terminal evidence, unbound replacement plans, digest substitution, and Chair bypass.
[x] Implement the critique-gated runner and Agent-authored CandidatePlan publication path. It must prove an unsafe technical proposal is blocked with no lease or effect, then prove a reviewer-accepted or reviewer-revised exact plan crosses host admission. Host-constructed fixtures cannot satisfy this gate.
[x] Add a closed 26-check report revision proving technical proposal lineage, reviewer verdict, Chair decision, exact admitted plan digest, HIL, P083 lifecycle, and zero direct effects from Room prose.
[x] Retain one real vfkit/Bielik run that passes the closed critique-gated report validator without using the deadline-only deterministic fixture. The retained 2026-07-25 macOS arm64 run passes all 26 checks with two round-robin cycles, one failed experiment, one reviewed correction, HIL, and P083 lease release.
[x] Implement the model-authored discovery contracts, solution-free guest fixture, structured intent compiler, reviewer replacement, full-plan terminal digest binding, staged active-configuration-derived mutation scope, real-model bench, deterministic replay, append-only calibration ledger, closed report schema, and fail-closed profile validator.
[x] Measure the pinned Bielik 4.5B Q8 planning ceiling before deployment smoke. The scoped declaration stage reaches two ready plans in five re-evaluated samples; both measured zone-data stages reach zero ready plans in five samples.
[x] Complete the discovery correction loop with a bounded three-attempt causal lineage and terminal-delta projection. Preserve exact plans host-side, expose normalized process calls and digest-only historical file-write identity to the models, carry bounded reviewer findings and verifier outcomes, and exclude historical payload bytes and hidden reasoning from prompts. A file value later observed in the active configuration remains fresh terminal evidence. Solver and reviewer now receive the same closed story-012.powerdns-correction-state.v1 value.
[x] Add a deterministic novelty gate and structured reviewer findings. The implemented first slice rejects an exact failed read-only experiment, or a subset of the same failed probes plus service activation, emits a typed inert refusal, and requires any reviewer replacement to pass the same gate. Replay records the redundant original decision and novel selected decision separately.
[x] Add a relevant-precondition digest and extend deterministic novelty rejection to exact failed file mutations only while challenge, VM generation, resolved admission scope, literal configuration, normalized service state, and verifier state remain unchanged. The projection consumes only the last complete nonce-marked, host-owned read-only snapshot from the cumulative PTY feed; a changed digest keeps the retry novel. The gate checks all three retained experiments rather than only the immediately preceding attempt.
[x] Benchmark solver and reviewer candidates independently with at least ten samples per discovery stage. Compare the pinned Qwen2.5-Coder 7B solver baseline with a runtime-conformant Qwen3.5 4B challenger, and evaluate Qwen2.5-Coder 3B plus Phi-4-mini for the lower-latency reviewer role. Select by domain correction success and measured latency, not a general leaderboard. The 130-sample matrix retains Qwen2.5-Coder 7B as solver and deliberately leaves the reviewer deployment default unset because none of the candidates produced a ready replacement.
[x] Replace accumulated experiment prose with one bounded state capsule carrying at most three exact host-side lineages, prompt-safe effect summaries, learned facts, refuted assumptions, next tests, terminal delta, and effective scope. Process summaries include canonical argv digests. The closed acceptance report retains bounded review analysis, the relevant-precondition digest, and original plus selected typed novelty decisions, while leaving raw precondition state and historical file bytes in runtime/replay evidence rather than retained metadata.
[x] Add an additive critique-to-regeneration path in which the reviewer may emit bounded typed findings without authoring a replacement CandidatePlan, the solver regenerates a candidate from those findings and the shared correction-state capsule, and the regenerated candidate receives a fresh review before the Chair may select it. Bind every transition to the original proposal, terminal evidence, correction-state digest, and regenerated plan digest; apply the same novelty, policy, budget, and host-admission gates as for any other candidate. Keep the current reviewer-authored revise replacement contract valid as a compatibility path. The v2 review and regeneration schemas, negative fixtures, signature golden vectors, append-only daemon join, restart recovery, requester-owned API, live baton integration, strict role grammars, host-side current-role and effective-membership checks for exact implementer and reviewer assignments, and prompt-free replay are implemented.
[ ] Promote critique-to-regeneration from opt-in mechanism to a deployment-ready Story 012 model path only after a role-specific bench of at least ten correction pairs reaches the declared thresholds without policy relaxation. The pinned Qwen2.5-Coder 7B run on 2026-08-01 produced 10/10 typed requests but only 1/10 domain-ready solver successors in the staged zone-declaration case, below the 0.6 ready-rate threshold. This measured model-capability gap does not reopen the completed baseline Story 012 acceptance.
[ ] Derive budgets from the append-only run ledger. Fit at least four complete feedback cycles inside one authority epoch or renew authority between cycles without widening an individual grant or holding a P083 lease during inference. The matrix yields a provisional model-only envelope of 720,000 ms and 24,576 tokens after a 25 percent margin. The retained full-system run adds four model turns, 15,863 total tokens, 344,259 ms of model time, 369,053 ms of deliberation, and a 120,687 ms maximum turn. Keep this task open for explicit between-cycle authority renewal: four cycles at the measured effect-feedback cost do not fit safely under one canonical Room authority ceiling.
[ ] Require deterministic replay before vfkit and retain repeatable seeded E2E evidence with prompt-free per-role latency, token, cycle, rejection, experiment, and correction measurements. Two fresh unseeded full-system successes now exist, and the later closed 30-check report is retained. An explicit fixed-seed repeat is still required before claiming repeatability.
[x] Add the resolved guest-attested powerdns-bind-missing-zone-data.v1 prepared-system challenge: valid loopback listener, BIND backend, and authoritative declaration referencing an absent zone-data file, without final localdomain records. Select challenge/id and verifier contract ref/digest from pinned guest/image evidence rather than a runner literal. The exact prepared-system source, manifest, image completion record, and retained deployment report now bind the selected second variant.
[x] Retain a passing story-012-powerdns-discovery-full-system-report.v1 from a real vfkit, llama-server, and pinned local-model run that reaches the exact DNS goal through model-authored trial and error without a solution fixture. Do not promote a report until the staged bench demonstrates a plausible correction path and the closed validator accepts the exact deployment bytes. The retained 2026-08-23 Qwen2.5-Coder 7B run passes 31 checks after one nonpassing and one successful admitted experiment, with fresh terminal feedback, HIL, P083 lease release, reported host structural framing, zero direct Room-prose effects, and three complete zero-gap P086 recordings. The same report proves activation and use of the P085 package Flow and turn-order producer, post-restart stale-replay refusal, terminal revocation, and stale-authority refusal without fallback.