Review and dispatch contract
Status: Normative, authorized by ADR-063, accepted 2026-10-11. Owner: Farmslot maintainers.
Verdicts, severity, configurable approval policy and inline locations follow the shared review policy, including gate refresh.
Selection and sequence
A review selection has exactly four configurable dimensions: runner, model,
effort, and mode. mode MUST be warm or cold. Runner choices MUST come
from the registered capability catalog; there is no same sentinel or
worker-relative reviewer default. Model and effort may explicitly select the
runner default. No policy, count, gate action or label may depend on runner
diversity. Every review MUST run in a separate review session.
ReviewSequence is an ordered collection of these selections. One shared
serializer emits the first row as the request-review action payload and any
remaining rows in its optional sequence field. Backlog, dispatch and queue
records use the same selection model in reviewSequence. Rechecks update the
existing review result; retry attempts do not increase the qualifying-review count.
When a run carries a reviewSequence, row 1 also selects the runner, model and
effort of its pipeline review and of that review's rechecks after the evidence
or head changes (ADR-063 Amendment 3). When row 1 is COLD, the publication gate
counts a passing pipeline review of the gate head as row 1's result and requests
only the remaining rows. A WARM row 1 continues its own review lineage at the
gate. Runs without a sequence take the pipeline reviewer from the project
review configuration.
Reviewer recipe replays
Status: Accepted under ADR-032 Amendment 2 (approved by the lead 2026-10-24).
The reviewer-specific execution contract MUST direct recipe replays into a fresh
directory under that reviewer's temporary scratch root, using the project's
existing recipe run or call command. It MUST clear both RECIPE_TASK_DIR and
FARMSLOT_TASK_DIR and pass an explicit --artifacts-dir. A replay MUST preserve
the command's failure status and MUST NOT replace the author's selected recipe
attempt, acceptance ledger or evidence. This rule overrides copied worker
instructions that bind proof commands to the parent task.
COLD and WARM reviewers MUST receive this contract. Their checklist markers,
signals, feedback and verdict files retain their existing paths. The reviewed-input
fingerprint MUST continue to cover author recipe attempts, their evidence and
latest-valid-recipe-run.json; an author evidence change still invalidates a pass.
The private gateway contract adds no public API or harness configuration. It uses the existing recipe CLI flag and environment variables named above.
Automatic re-review limits
Status: Accepted under ADR-063 Amendment 8, approved by Arthur 2026-10-26.
After a worker fix, the gateway MUST stop the automatic loop when a captured delta contains zero source files and the finding set is unchanged. It MUST retain the findings as open at the gate and MUST NOT automatically redeliver them after restart.
A completed fix that changes only evidence/artifacts or at most 50 source lines MUST skip the automatic warm re-review and resolve the prior findings for gate eligibility. Source lines are additions plus deletions under the project's existing source filter. The original reviewer attempt MUST remain unchanged. The gateway MUST record acceptance of the worker fix, its current contribution snapshot and reviewed-input fingerprint without launching or charging a reviewer. Evidence-only changes at the same HEAD MUST be distinguished from an empty fix using the existing reviewed-input fingerprint. Failed or blocked fixes, unavailable captures, missing original review baselines and unmeasured source changes MUST retain the existing review path. Normal and recovered fixes MUST use the same limits.
Round budget
Except for the review-pr and update-branch scopes below, the gateway MUST default to at most five reviewer rounds for a repository/PR. Before a PR identity exists, admission is scoped to its canonical run. Initial, retry, WARM/COLD, pipeline, standalone and publish-gate reviewer attempts MUST spend this same budget; only a counted attempt spends a round (see No-verdict attempts). A worker fix is not a reviewer round. Reattaching to an admitted attempt MUST reuse its receipt. An ambiguous qualified GitHub reference alone MUST NOT establish PR scope for a ticket flow. Existing PR linkage, a pull URL or a PR-bound flow supplies that context.
Admission receipts MUST persist in the existing run snapshots before a runner can start. Replay, restart and a different gate or run MUST NOT refund rounds by themselves; only the admission owner's no-verdict settlement does. Copied review histories MUST NOT count the same admitted attempt twice. The round budget is separate from the number of qualifying passing reviewer sessions.
Without an authorized extension, the gate MUST show N/5 and refuse a sixth round before consuming its request
or starting a reviewer. The refusal MUST tell the operator to stop and diagnose
a design problem or scope creep, then narrow the scope or redesign. Spending all
five rounds MUST NOT prevent publication when all existing approval conditions
are satisfied.
Review-pr budget
Status: Accepted under ADR-063 Amendment 9.
Each review-pr run MUST have its own five-round budget. Its admission lock,
admissions, settlements and no-verdict cap MUST exclude every other run,
including archived review requests and feature runs for the same repository/PR.
New review-pr requests MUST NOT inherit feature extensions. Other runs MUST NOT
inherit new review-pr admissions or settlements. Review-pr remains ineligible
for budget extensions.
Receipts, settlements and grants already copied into a run's snapshot before this amendment MUST keep counting in that snapshot; no history is rewritten. Feature and update-branch scopes MUST stop reading review-pr snapshots. A review-pr receipt not already in one of their snapshots no longer counts toward those scopes.
Replay, restart and reattachment MUST retain that run's admissions and reuse an existing receipt. A sixth counted attempt in the same run MUST still refuse; a new review-pr request starts with its own budget. Feature PR budgets and update-branch accounting remain unchanged. WARM lineage and publication checks MUST NOT change with budget scope.
Update-branch budget
Status: Accepted under ADR-063 Amendment 6 for the lead-approved repair.
Each update-branch run MUST have a two-round budget independent of the PR's
feature budget. The gate MUST show N/2 and refuse a third counted attempt
before launch. Parent extensions MUST NOT raise this limit or appear in its
review instructions. Update-branch remains ineligible for extensions.
New update-branch admissions MUST record the owning run ID in
Run.reviewRoundOwners, keyed by the unchanged admission receipt. The shared
PR lock and durable snapshot copies MUST retain admissions, ownership,
settlements and grants. Feature budgets MUST exclude owned maintenance
receipts; an update-branch budget MUST count only its own receipts. No-verdict
counts, failure reasons, activity rounds and gate labels MUST use the same scope.
The shared reviewRoundBudgetForRun and reviewRoundAdmissionsInScope SDK
helpers own that projection. Activity MUST prefer durable receipts and settlements
over stale gate counts, retaining gate-count fallback for legacy records without
admission metadata.
Legacy unowned receipts MUST retain their PR accounting. Reattaching an existing receipt MUST NOT assign ownership, spend another round or clear a settlement. Restart, replay, archive and retirement MUST retain ownership metadata wherever admission history survives. A later update-branch run gets its own two rounds. Existing review and publication checks still apply.
When pipeline REVIEW exhausts its budget, the run MUST wait at the existing result decision with its slot, checkout and worker retained. A budget refusal is not a reviewer verdict. The operator may extend the budget, accept the captured result with a written reason without publishing it, or close it. Extensions keep their existing eligibility rules. A grant enables review to resume on the same run; it does not start work by itself. Named-slot re-binding remains only a fallback for historical failed runs that already lost ownership.
Scoped extension
Status: Accepted (lead, 2026-10-15) under ADR-063 Amendment 1.
run.extendReviewBudget MUST use existing admin authorization and capture the
authenticated request origin. Caller-supplied approver names or system/worker
signals MUST NOT grant authority. The operator must have lead approval for the
recorded scope; a service credential retains its service identity and cannot
claim to be a person. Existing credential and principal revocation checks apply.
The request names a stored, nonarchived unpublished run, including a failed or interrupted run, its full current head SHA, a finite integer absolute limit greater than its effective limit by one to five rounds, a nonempty scope brief, an idempotency key and the expected prior limit. A new grant MUST refuse an already published, archived or unrelated run, invalid inputs, unknown/live-head drift, unauthorized origin or conflicting prior limit before any state or runner mutation.
The receipt MUST retain the request identity, run and source/PR identity, approved head, prior/new limits, scope brief, authenticated origin and time. Serialize the grant with review admission, recheck state/head after asynchronous reads, and persist the receipt before using its extra capacity. An exact retry returns the original receipt; reuse of its key for different inputs refuses. Existing admissions MUST remain unchanged. Restart, replay and a copied result MUST NOT refund attempts by themselves or apply the same extension twice.
For extension-eligible feature runs, the effective limit and used count MUST come from the existing admission owner. Before a PR exists, the receipt applies only to its canonical run. Once the run has a verified repository/PR identity, the existing shared scope MUST preserve that receipt and its admissions across active and archived history. A different feature run MUST NOT obtain a fresh default budget for the same PR. The original run remains the execution and publication owner.
The gate shows N/L and the grant's scope/provenance. The current approved
scope brief MUST accompany additional reviewer instructions. A grant does not
start a reviewer, reset worker-fix retries, mark a result passed or authorize
publication. Existing review requests/replay collect real results on the same
run, preserving WARM lineage, immutable ranges, no-verdict accounting and
the number of required qualifying reviewers.
Validation MUST cover unauthorized/stale requests, exact retry and conflicting keys, concurrent grants/admissions, restart/archive scope preservation, refusal without an extension and real sixth/seventh admissions after a bounded grant. An extension MUST NOT make stale reviews or an old package publishable.
No-verdict attempts
Status: Accepted (lead, 2026-10-18) under ADR-063 Amendment 2.
Each admitted reviewer attempt settles once, through the admission owner, when
it ends. An attempt that returned its required result for the reviewed head with
a pass or issues verdict is counted and records kind: verdict. These ends are
no-verdict and spend no round; Run.reviewRoundSettlements records the reason
and time against the receipt:
- a native reviewer that timed out, exited, left a missing or unusable result, failed to launch, or was cancelled or superseded;
- a reviewer a restart could not reattach, or whose recovered signal or result shows a failure;
- a workspace review whose launch failed after admission, or whose monitor ended without recording a result, a cancel or supersession included. A replay that dispatches again at a new engine generation is a new attempt; a restart that re-enters DISPATCH at the same generation keeps its receipt, and a monitor-only replay still settles the attempt its DISPATCH admitted.
A receipt without a recorded verdict spends no round. New admissions record
kind: pending and reserve concurrent launch capacity while the reviewer is in
flight. Uncertain delivery does not count as a completed verdict. Retained
validated contexts and recorded result rows preserve historical verdict charges.
Admitting a receipt again (a reattach) keeps its settlement. When a settled
attempt's pass or issues verdict arrives later, through restart recovery or a
workspace monitor replay that records it, its disposition becomes verdict and the
attempt counts, except for the farm-caused replacement below.
Settlements MUST be merged across the admission scope and copied to every live
run in it. Rounds used count only admissions with a recorded reviewer verdict;
ReviewRoundBudget.noVerdict counts only ended no-verdict outcomes against noVerdictLimit
(default 3).
Below the cap, run behaviour is unchanged. At the cap, these automatic launches stop with a reason naming the count and the last failures:
- the re-review after a fix: the loop stops before its next fix pass and keeps the findings for Continue Fixing;
- the re-review of a fix pass recovered after a restart;
- a publication review plan no operator request started, including configured sequences, reviewed-input reruns and a plan restart recovery resumes: the gate records a launch refusal.
An operator's Request Review still launches, and so does the first fix pass and
re-review of Continue Fixing or Send back. Their no-verdict ends count again.
The gate shows N/L and, once there are any, the no-verdict count against its
cap.
Farm-caused fallback
Status: Accepted (lead F144 direction, 2026-10-25) under ADR-063 Amendment 7.
A retained WARM review that fails before initial command acceptance because its native handoff or resume cannot be established MUST first reconcile the same reserved context, lease and command once. Uncertain delivery MUST NOT launch another reviewer. The existing owner, slot, profile, generation, lease, saved conversation and launch-digest checks MUST still pass.
A confirmed stopped, unaccepted WARM attempt that requires a farm-caused delta-scoped COLD replacement MUST keep its admitted receipt and record a gateway-owned ReviewRoundSettlement with kind: 'infrastructure-fallback', reason and time. That admission MUST spend no review round. A valid verdict from the replacement MUST NOT clear this kind, including during reattach, restart recovery or scope merging. The failed WARM launch still counts once against the existing no-verdict cap. Admission history, grants and findings MUST NOT be removed or reset.
An explicit COLD request, an ordinary completed WARM review, or an incompatible requested model, effort or validation-depth change MUST retain normal accounting. No waiver of the replacement's findings or publication requirements is granted.
A native static-code attempt MUST capture its launch validation depth on AgentContext.reviewValidationDepth and derive launch environment from explicit project settings and pinned sources, excluding unrelated pool tool selections. Full-live attempts MUST retain their pool tool environment. A retained legacy session without this marker MUST preserve its previous launch policy. The already digest-validated handoff launch MUST be carried into its stopped-session resume rather than rebuilt from newer pool values.
The optional fields ship in @farmslot/protocol 0.71.1. The gateway writes and consumes them; existing shared budget projections remain authoritative. No new CLI flag or RPC is added and wire protocol remains 0.63.0.
Session and range
Amended by ADR-063 Amendment 5 (2026-10-20).
- A review's first pass, and a COLD a person asks for, MUST create a fresh review session and review the whole diff at captured base/head SHAs. A reset of an existing session is insufficient. That session becomes the lineage's retained session.
- A re-review MUST continue the retained session of the lineage whose work it
checks, with the same runner, and review only
lastReviewedHead..currentHeadwith the earlier findings in view. A lineage qualifies only with a validated earlier result and a recorded reviewed head; otherwise its next review is a first pass. - The binding key is the run family plus the durable
reviewIdlineage. Once the run has a PR, the PR MUST match too, so a lineage started before the PR continues after it opens. A lineage MUST NOT continue in a run of another family. Request/sequence context identifies the lineage being continued; a recheck updates that same result. Missing or ambiguous lineage in a request MUST be refused. Lineage is provenance, not a fifth selection dimension or a runner-relative default. - A retained session whose process ended MUST be resumed from its saved runner session. A busy session is waited for. A retained session is unavailable only when its record is missing or can't resume on this slot under the same result contract, the requested model or effort differs from its recorded one, another lineage holds it, or the current head no longer descends from the lineage's last reviewed head.
- A re-review whose session is unavailable, or whose resume fails to
start, MUST run COLD in a fresh session in the same attempt and under its
admitted receipt, scoped to the lineage's delta, with the earlier findings in
its prompt. That session becomes the lineage's retained session. The review's
session trace records the reason and the compared ranges (
fallback), and the gate shows both with that review. - The delta is
lastReviewedHead..currentHeadwhile the current head descends from the reviewed head. After a rebase it compares the branch's own change at both heads withgit range-diff: the reviewed head against its merge-base with the base branch, and the current head against its current merge-base. Changes that came only from the base branch are left out. - Operator and person WARM re-reviews MUST use the same delta fallback. An unavailable session MUST NOT turn a re-review into a full-diff review.
- Legacy completed contexts MAY establish the retained baseline through the validated verdict, reviewed head, matching provider session, owner, family and artifact scope. A missing newer validation timestamp or an older runtime review alias alone MUST NOT invalidate that provenance. An attempt that fails before inspection MUST NOT replace the last validated baseline. A binding transferred to a failed attempt MAY continue that provider session only with its gateway-owned no-verdict settlement and the earlier validated provenance.
- Passing review eligibility requires its separate session, current reviewed head and verdict evidence. Choosing a different runner proves nothing.
- A retained reviewer is a resumable session identity, not a running process. Once a native reviewer's result is recorded, the Gateway MUST stop its process and the MCP servers it started, keeping the session, lease and conversation for WARM. A WARM re-review transfers the stopped session and resumes its saved conversation, on native and tmux runs alike. An idle settled reviewer MUST NOT hold a live process. A stop the host does not confirm is logged and leaves the binding held, so the next reuse of the slot stops it.
Mode resolution
@farmslot/review-dispatch.resolveReviewMode owns mode selection. Its precedence
is an explicit selection, Run.reviewMode, Run.prWork.review.options.mode,
the retained Run.repeatReviewContext.mode, a supplied project default, then
COLD. An explicit COLD selection MUST override a configured WARM default.
A new review-pr's explicit fresh-chain context also selects COLD when no mode
was supplied, before a configured WARM default.
Admission, continuity, task instructions and recorded review snapshots MUST use
this precedence. Run creation MUST use the same shared default. The resolver does not verify lineage or authorize session reuse.
Under ADR-063 Amendment 5 this precedence selects only a review's first pass. A
re-review defaults to WARM in its lineage: the re-review after a worker fix,
including after a restart; Continue Fixing and Send back; a gate Request Review
that continues a lineage; and the gate's re-review of changed inputs, which
continues the lineage of the row the pipeline review ran, named by the pipeline
result's session.reviewId. A re-review runs COLD only when a person selects
COLD for it, in the gate's Request Review picker or with --review-mode cold on
the gate request. A sequence row's stored mode, Run.reviewMode, PR options and
the project review.mode select only first passes.
For each requested row, the gate's Request Review preselects the run family's
most recent lineage reviewed by that row's runner, in WARM, with that lineage's
model and effort, including when its session is unavailable, and only for one
row per lineage. A later row with the same runner takes that runner's next most
recent lineage, or is new and defaults to COLD. A row whose lineage is
unavailable stays WARM and shows why it will review the delta in a fresh session. The gateway lists the gate's lineages
and their availability in ReadyGatePayload.reviewLineageOptions.
Automatic reviews that create a separate new lineage select the shared COLD default. A re-review whose session is unavailable runs the delta-scoped COLD fallback above, including operator and person WARM requests. Eligibility filtering for authorized alternatives remains separate from mode selection and cannot downgrade an explicit WARM request.
Shared picker
@farmslot/review-dispatch owns one ReviewDispatchViewModel, catalog/default
resolution, row operations, serialization and validation. The backend and SDK
use its validators. Lit and React Native each have one thin renderer. Call
context supplies allowed actions, defaults and sequence availability; callers
MUST NOT duplicate selection state or normalization.
Gate requests, dispatch wizard, backlog options, PR execution and Companion equivalents consume this model. Effort MUST appear next to model. Every review row MUST expose WARM/COLD. Unsupported adjustable effort shows the supported default explicitly. Resource, publication and QA controls remain host concerns. The duplicate units listed in ADR-063 MUST be removed after their callers move.
Public surface
- CLI consumers are
run create, review/dispatch entry points and gate requests:--runner,--model,--effort,--review-mode warm|cold, and--review-sequence <json>. Existing run--moderetains its workflow meaning. @farmslot/protocolownsReviewMode,ReviewSelection,ReviewSequence,ReviewStatus,ReviewAttemptandReviewPolicy. Run creation, PR options, gate resolution, backlog and dispatch queue RPCs consume these types.@farmslot/protocolalso ownsReviewRoundBudget,MAX_REVIEW_ROUNDS,REVIEW_NO_VERDICT_LIMIT,reviewRoundBudgetand thereviewRoundBudgetLabelformatter. Gateway admission, the CLI and both gate renderers consume the shared projection.Run.reviewRoundAdmissionscontains gateway-owned opaque receipts andRun.reviewRoundSettlementstheir no-verdict settlements;ReadyGatePayload.reviewRoundBudgetcarries the display snapshot. No CLI flag or client-written receipt is added.- The shared package exports the view-model and selection/sequence validators for Command Center, Companion, CLI and gateway consumers.
resolveReviewModeandResolveReviewModeOptionsare additive exports from that same package. Pickers and gateway admission, execution and evidence readers consume them. They add no protocol field, CLI flag or harness policy; harnesses inherit the gateway's selected mode and existing lineage checks.- Gate action
request-reviewreplaces both old review-request action IDs. Old fields, action IDs and session/scope enums have no aliases or shims. - This is a coordinated breaking contract update. Persisted configuration and queued requests MUST be converted before deployment; old records lacking proof remain historical provenance, not current passing reviews.
- Project
review.modesupplies a configured mode default. Removed keys such asminimum_independent_reviews,require_cross_runnerandself_review.session_policyMUST be rejected with a conversion hint, not ignored into a less restrictive default. - Project harnesses consume the typed gateway contract. They inherit session/mode validation; farm-owned QA remains separate under ADR-058. This adds no harness-specific review policy or extension flag.
Required evidence
Tests and live receipts MUST prove fresh COLD identity/full range, retained WARM identity/delta range (including before a PR exists), the delta-scoped COLD fallback of automatic and person re-reviews with its recorded reason and compared ranges, including a rebased delta, legacy baseline adoption, runner-choice-independent eligibility, effort preservation and sequence round trips. Web/mobile proof MUST exercise the shared picker through user input, not injected UI state.
Diff scope
The shared minimal-diff policy is owned by @farmslot/agent-runtime/scripts/review-diff-scope.cjs. Worker task documents, independent review prompts and the cross-review orchestrator consume that definition. Every verdict includes its Diff scope table. Warm review, and the delta-scoped COLD fallback of a re-review whose session can't continue, narrow correctness review to the saved delta; their scope inventory still covers the whole contribution against the publication base.
The public read-only script prints the policy with instructions, or the policy plus changed paths and sensitive-path candidates with packet <repository> <publication-base> <head>. Candidates require review against the frozen ticket; the scanner neither grants an exception nor detects semantic refactors or formatting churn. It does not change source, review results or publication authority.
Grant-time head records provenance. Later fix heads on the same owned branch use current-head review rules and the recorded scope. An extension survives retirement of its granting run within the same PR scope. The command requires --request-id and accepts --expected-limit for a compare-and-set grant.
WARM delta packets
Under ADR-063 Amendment 14, the gateway freezes inputs/warm-delta.json
with lineage and validated baseline identity, prior/current heads, compared
ranges, unresolved findings, changed audit rows and required completion actions.
Existing delta/range-diff rules apply; merge-resolution changes remain visible.
The packet is data, not task authority. Full baseline material remains available
by frozen artifact reference outside the default prompt.
Reviewer coverage reuse
artifacts/review-ledger.md is reviewer-authored checklist coverage, including
registered child checklists. Scope rows are the ## Diff scope table in
artifacts/review.md. The shared runtime's review-report composer retains their
version-1 derived row index at artifacts/review-coverage.json. It runs in the
existing mark completion path before artifact verification/hashing; the existing
accepted-review-result path validates and retains the index with its report.
This index is not another verdict authority. The engine's
artifacts/acceptance-status.json is neither a reuse source nor a reuse target.
A checklist row key is its frozen checklist/unit identity, checkbox step index and rule-source digest; a scope row key is the canonical contribution-relative path. The index records row content and its reviewer outcome/gap, source report hash, validated baseline/head, authorization and rule/template hashes, plus cited source/evidence hashes. Missing or ambiguous row bindings are not reusable. Current row updates travel with the existing structured review result; no new operator command is required.
Rows may be reused only while all those inputs match. Missing, changed or unverified inputs require current evaluation. Gaps and unresolved findings survive reuse. The reviewer rechecks each rule's applicability against the current delta and invalidates assumptions it changes, including newly added paths. Absent legacy indexes mean no row reuse; they never widen the correctness range or change session fallback. Reuse cannot supply engine acceptance verdicts or a current-head review result, or bypass checklist/terminal validation.
The shared composer combines validated reused rows with current updates and renders complete current scope and criteria tables with provenance. The reviewer authors current findings and changed rows rather than regenerating unchanged tables. Existing scope requirements remain in force on every contribution file.
Compaction and validation
ADR-023's runner capability compacts only inside a fully fenced continuation of
a validated review, before incoming task instructions bind. The run snapshot's
ReviewSessionTrace.compaction persists pending intent before send and the exact
session/packet receipt before delivery. Restart reconciles the same operation.
Unsupported compaction delivers uncompacted in that session, with an explicit
reason, without COLD or budget changes. Unconfirmed compaction follows the
uncertain-delivery reconciliation path, not an automatic replacement. Unfinished
same-attempt recovery skips compaction and preserves its in-progress work.
Validation covers stale hashes, missing evidence, new scope paths, rule changes,
merge resolutions, pending/confirmed restart recovery and controlled WARM cost.
Report delta calls/output and compaction separately, under normal accounting.
Use scripts/runner-validation/warm-delta-packet.recipe.json for the opt-in
coverage-integrity and controlled gateway proof.
The measured target is 8–12 delta calls and under 10,000 output tokens, not a quota
or a reason to omit required inspection. Billing changes remain separate.
Slotless retention
static_review.workspace.retention_days is a positive integer, default 7.
static_review.workspace.root is an optional absolute directory on the selected
execution machine. Omit it to keep the existing node-owned workspace location.
The allocator creates owner-scoped managed directories below that root, checks
ownership and mount availability, and persists exact paths. Later configuration
changes apply to new allocations; they do not move an existing binding.
A completed slotless review records its exact conversation id/path, reviewId,
validated reviewed head and retention deadline. Its retained conversation stays
under the same project, PR, principal generation, machine, runner, model and
effort. A re-review before expiry reuses that conversation and reviews only
lastReviewedHead..newHead. Foreign or ambiguous identity refuses WARM. Unsupported adapter continuation follows the delta fallback below.
After expiry, a fresh full-contribution COLD session is allowed with an explicit expiry reason;
the prior lineage, findings and reviewed head remain recorded.
Existing maintenance reclaims expired owned workspace/session files. It preserves run records, retained result artifacts and lineage receipts. Active runs, unresolved review decisions and queued review requests for the same PR prevent reclamation. Result retention does not confirm process shutdown: a separate durable stop receipt gates released capacity and warm reuse, and terminal reconciliation retries an interrupted stop. A validated external intake replacement waiting behind an earlier retirement remains unadmitted until that retirement settles; cleanup continues to protect every last-accepted intent. The source is re-read before publication. Retirement stays fenced until the final cleanup receipt reaches disk; a failed receipt write retries metadata without deleting the workspace twice. Restored queue rows behind an earlier durable retirement stay visible but unadmitted until settlement; all other pending reviews continue to protect storage. An admitted continuation stays protected until it finishes, even when its predecessor's deadline passes. Deletion and review admission share a storage fence and revalidate PR activity before deleting. Unknown identity, unavailable storage or unconfirmed process closure leaves the files intact.
Verdict-based charging
ADR-063 Amendment 11 separates admission from budget consumption. Before dispatch, the serialized admission owner persists kind: pending in the existing round-disposition map. This reserves capacity but increases neither rounds used nor the no-verdict count. An accepted prompt, a live session or a recovered startup acknowledgement MUST NOT charge a round.
Only validated reviewer-result ingestion records kind: verdict, in the same durable snapshot as the result or after its verified persistence. That transition spends one round and is idempotent. Recovery reconciles a persisted validated context before skipping an already-ingested result. A failed, cancelled, missing-signal or unstarted reviewer spends none, including when its failure settlement is missing. Ended failures retain the existing no-verdict disposition and automatic-launch cap. The established infrastructure-fallback refund remains free.
Existing recorded verdicts seed retained histories through their validated reviewer contexts or recorded result rows. A historical failed/cancelled attempt or an admission alone is not verdict evidence. Receipt identities, PR/update-branch scopes, extensions and history deduplication are unchanged. Dispositions survive retirement of an earlier run. Pending live launches still reserve the last available capacity. A context-free startup reservation belongs to its current gateway and engine generation; an orphan from an earlier gateway does not pin capacity. Saved live reviewer contexts retain their reservations across restart. A confirmed terminal reviewer without a verdict releases capacity.
Fresh review chains
A new review-pr request MAY set reviewContext.freshChain: true, with a null
lineage entry for each execution row, to start a fresh chain. It MUST use COLD.
When mode is omitted, this explicit fresh request selects COLD before a
configured WARM default. An explicit WARM choice remains invalid.
The gateway MUST skip predecessor lookup for both slotless and slot-backed
review-pr launch paths. No prior review findings, verdict, artifact references,
generation history or retained session may enter the task. The writer MUST omit
inputs/prior-review.json and its previous-review instruction. The new run is
generation one of its own chain. Existing run records and artifacts remain
unchanged.
Standing team, repository, trigger-rule and farm workflow policies MUST reject
freshChain: true. Only a one-off request host may opt in to the fresh-chain control;
review capability alone MUST NOT expose it in a persistent policy editor.
The selection MUST survive existing request serialization, queue persistence and replay of that request. It MUST NOT change a previously frozen request's policy or reset its review budget. Omitted or false retains the current COLD or WARM continuation contract outside the comparison default below. Fresh plus WARM or retained lineage context is invalid. QA and publication-gate review requests MUST reject the fresh policy.
A new review-pr request with lane: "comparison" and a nonempty variant
MUST default to a fresh chain unless its primary review context explicitly
supplies a retained reviewId. Omitted mode selects COLD before a configured
WARM default. Explicit lineage continues through the existing mode, identity
and session checks. This default applies through preview, queue and creation;
it adds no CLI flag or SDK field. Production requests keep their existing
continuation contract. Runner or model names MUST NOT imply comparison
classification. Historical run records and frozen inputs remain unchanged.
The shared review-and-dispatch model owns this request choice. New review-pr
hosts opt in; publication-gate hosts do not. Command Center and Companion use
thin renderers of that model. The CLI maps run create --fresh-review-chain to
this same context and COLD, refusing a conflicting explicit mode or lineage.
Manual validation uses
scripts/runner-validation/fresh-review-chain.recipe.json. The request proof
uses the checkout-local CLI and an authenticated disposable gateway; the task
input proof uses the existing task writer. Neither starts a model. Browser
checks select the shared checkbox through normal user interaction and verify
that changing the flow removes the choice. These rendering fixtures do not
claim reviewer-session behavior.
Adapter-owned continuation
Under ADR-063 Amendment 13, WARM MUST use the lineage owner's shared runner
saved-session continuation capability. Adapter support MUST NOT participate in
identity matching. A valid, unexpired lineage with an unsupported adapter MUST
use a fresh COLD session in the same attempt over
lastReviewedHead..currentHead, preserving prior findings and artifacts as input.
Its fallback trace MUST record session-resume-unsupported and the compared
range. The successor becomes the retained session for that lineage.
Selection, subject freezing, task generation and result ingestion MUST preserve
that delta scope even though the effective mode is COLD. Ownership generation,
project, repository/PR, lineage, runner/model/effort, ancestry, busy-session,
capacity and publication fences MUST still pass. A retained checkout is required
only when resuming it; unsupported continuation uses a newly allocated checkout.
Missing or foreign lineage identity MUST NOT use the unsupported-adapter fallback.
A missing saved conversation is distinct from missing lineage identity and keeps
its existing delta fallback. Unsupported capability MUST NOT emit an
unsupported-runner refusal; that reason remains readable in historical traces.
This fallback uses normal verdict accounting and grants no refund. Expiry retains its
existing full-contribution COLD contract.
A stopped transfer that returned no verdict MAY continue the same conversation
through its latest execution owner under the same identity and busy-session fences.
RepeatReviewContext.priorRunId remains the validated baseline;
ReviewSessionTrace.priorRunId identifies the physical source owner. An attempted
source-head advance MUST NOT replace the last validated comparison base.
The successor MUST capture its own support admission even when immutable support
content is reused. A predecessor's support binding is not that receipt.
The gateway records ReviewWorkspaceBinding.transferredTo on the predecessor
before moving execution authority, including archived predecessors. It records
continuedFrom and a separate task path on the successor before node allocation.
The node verifies the former immutable source identity and stop receipt before
advancing that same checkout. Retry reuses this transfer; prior task artifacts
are preserved. A transferred predecessor cannot replay execution or reclaim the
successor's source. The current owner alone starts a new retention period.
Public surface
ReviewWorkspaceBinding in @farmslot/protocol adds optional continuedFrom
and transferredTo receipts for gateway allocation, replay and retention readers.
Existing bindings omit both. These additive fields grant no caller authority.
ReviewSessionRefusalReason in @farmslot/protocol adds
session-resume-unsupported for the existing fallback trace, consumed by the
review result and gate displays. No command, flag, RPC or harness extension is added.