Authorization Contract v2

Normative rules governing which actors may trigger agent dispatch.

This version preserves v1's event-actor and model-bound admission gates while adopting selective entity acquisition and decision-focused audit evidence for the future entity-first path under ADR 0147, with bounded CEL actor-verdict reuse under ADR 0148. Existing dispatch may continue to use v1 until migrated. The historical decision and rationale are recorded in ADR 0054. The NormalizedEvent v1 specification defines the actor.role, actor.kind, actor.role_verified, and actor.bot_role fields consumed by this contract. Bot identity and role resolution are specified by ADR 0107.

ADR 0107 target contract — not yet implemented: The bot identity fields, resolver outcomes, and bot-specific fail-closed rows below describe the behavior to be implemented by the adapters and dispatch path. Until that migration lands, existing compatibility behavior remains authoritative.

Role ordering

Fullsend uses a forge-neutral role hierarchy. Roles are listed from most to least privileged:

admin > maintain > write > triage > read > none > external

Role	Meaning
admin	Full repository control
maintain	Repository settings without destructive admin
write	Push, label, comment, merge
triage	Label and moderate without push access
read	Read-only collaborator
none	Authenticated user without explicit repository permission
external	Actor outside the repository or project (currently Jira-only; GitHub maps non-collaborators to read on public repositories or none; GitLab maps them to none)

Forge-native permission mapping

Forge	Native level	Fullsend role
GitHub	admin	admin
	maintain	maintain
	write	write
	triage	triage
	read	read
	(no collaborator entry)	none
	(fork/non-collaborator)	read (public repository) or none
GitLab	Owner	admin
	Maintainer	maintain
	Developer	write
	Reporter	triage
	Guest	read
Jira	Administrators	admin
	Developers	write
	(other named role)	read
	(no project membership)	external

On GitHub the mapping source is the collaborator permission API (GET /repos/{owner}/{repo}/collaborators/{username}/permission), which returns the user's effective role including inherited org grants regardless of membership visibility. A built-in role_name is used as is. Custom repository roles are an organization-level GitHub Enterprise Cloud feature, so most repositories only ever return built-in names; Fullsend does not interpret a custom role's name or its fine-grained permissions. A custom repository role resolves to the strongest base role set to true in user.permissions (admin, maintain, push → write, triage, pull → read), or none if no flag is set. Only when user.permissions is absent or null does the legacy permission field (admin, write, read) apply; anything else is none. Legacy read stays read, because it cannot distinguish Read from Triage. Fork authors and non-collaborators are mapped to read (public repositories) or none, never external. The external role is currently produced only by the Jira adapter for actors without project membership. The default thresholds deny read, none and external alike. The author_association field is not used because it does not correctly reflect private org membership (see Excluded fields).

A GitHub repository that lists owners_file under authorization in .fullsend/config.yaml consults its repo-root Prow OWNERS file before the collaborator API:

OWNERS list	Role granted
approvers	at least write
reviewers	at least triage
(not listed)	none; the collaborator API decides

OWNERS can only raise a role, never lower one. An entry that names an OWNERS_ALIASES key stands for that alias's members. A login equal to any alias key never matches, and nested aliases are not expanded. The files are read from the base branch for PR-scoped events and from the default branch otherwise, so a PR cannot add its author. See the authorization field.

On Jira the mapping source is the project's role membership roster for the configured --jira-project, resolved once per poll cycle for event-backed dispatch. Entity-source CEL may reuse a completed provider-verified actor verdict across poll cycles and bound run checks only under the bounded cache rules below; event admission does not use that cache. An actor is matched by Jira account ID. Role names are matched case-insensitively against the names above; unrecognized role names fall through to read. Actors with no membership in the configured project — or whose issue belongs to a different project — are mapped to external (fail-closed). See the Jira integration guide for details on the name-based matching limitation.

Bot identity fields

The adapter/provider MUST classify the verified actor using authoritative source-system metadata before authorization. This may include a provider actor name when that forge gives it bot-specific semantics, such as GitHub's [bot] logins; labels, review types, and arbitrary event-content strings are not bot identity signals. For post-migration producers with actor.kind: bot, actor.role MUST be none, and actor.role_verified is true exactly when the provider completes the bot-role lookup, including a successful no-match result; it is false when resolution fails. For actor.kind: human, actor.bot_role MUST be absent or null, and actor.role contains the forge permission role when actor.role_verified is true. For non-label dispatch, only a non-null, provider-resolved actor.bot_role can pass the bot gate. Once that gate succeeds, the actor.role-keyed observation and mutation thresholds below do not apply to the bot; platform authorization instead requires a valid normalized event, an applicable source and target, and a transition supported by the selected harness. CEL may further restrict routing but cannot create or broaden bot identity authorization.

For humans on non-label paths, role_verified: false denies the event regardless of the role string. A missing role_verified on a trusted pre-migration event is treated as legacy input and retains the current human authorization behavior; new adapters MUST emit the field, and their false value MUST fail closed. Legacy status is determined by the adapter implementation contract, not by guessing from this field's absence. On the permanent label exception, role_verified is not an authorization input: an accepted forge label mutation authorizes a human, and a bot additionally requires positive provider bot classification. For bots, omission means that bot-role lookup is not applicable, not that it failed.

Default thresholds

Category	Minimum role	Rationale
Observation (triage, review)	triage	Read-only analysis; lower barrier to reduce maintainer toil
Mutation (code, fix, retro slash command, prioritize)	write	Agents that push commits or alter state require push access

A role satisfies a threshold when it is at or above the minimum in the role ordering. For example, admin satisfies both triage and write thresholds.

These thresholds apply to human actors and to legacy bot events that still use forge permissions during migration. They do not apply to a bot after successful actor.bot_role recognition, because actor.role remains the compatibility value none for that target representation.

The bash dispatch implementation uses a parameterized has_repo_permission(username, min) helper that encodes this comparison (see Enforcement point).

Implementation note: the Go dispatch path (IsAuthorized() in harnessdispatch) currently uses a coarse write+ filter for all non-exception transitions — it does not yet distinguish observation from mutation thresholds. Poll-based dispatch (GitLab/Jira) uses the Go path, so a triage-role user triggering observation through that path is denied today. The bash path (GitHub webhook dispatch) implements the full parameterized threshold.

Fail-closed behavior

The authorization gate is fail-closed: when a role cannot be determined, the actor is denied.

The entity-evaluation rows below specify the future ADR 0098 path as revised by ADR 0147 and ADR 0148 and do not describe behavior currently implemented by fullsend dispatch or fullsend poll.

The bot-specific rows in the following table are ADR 0107 target behavior; legacy events remain subject to the compatibility rules until implementation migration is complete.

Condition	Outcome
Collaborator API returns a custom role_name with no effective permission signal	Mapped to none; denied
Collaborator API returns an error or times out	Denied (function returns failure)
Bot-role lookup returns no registered identity on a non-label path	actor.role remains none; actor.role_verified is true; actor.bot_role is absent; denied
Bot-role lookup fails or is unverifiable on a non-label path	actor.role remains none; actor.role_verified is false; actor.bot_role is absent; denied, with the failure retained in resolver/audit diagnostics
Collaborator API response is not valid JSON or does not match the expected shape	Denied
actor.role is empty or missing	Event fails NormalizedEvent validation; never reaches dispatch
actor.role_verified is false for a human	Denied regardless of actor.role
Legacy bot event has an actor.role other than none	Valid in the v1 compatibility schema; authorization follows current compatibility behavior during migration and is not an ADR 0107 bot-role authorization result
actor.role_verified is false but actor.bot_role is non-null for a bot	Event fails NormalizedEvent validation; never reaches dispatch
actor.bot_role is non-null for a human	Event fails NormalizedEvent validation; never reaches dispatch
Username is empty	Denied
OWNERS is missing or malformed, or OWNERS_ALIASES is present but malformed (owners_file enabled)	OWNERS check skipped; the collaborator API decides
Username contains characters outside [A-Za-z0-9-] (owners_file enabled)	OWNERS check skipped; the collaborator API decides
Fullsend invocation provenance for poll discovery or a bound run's eventless check is missing or unverifiable	That eventless evaluation is denied
Harness entity sources are missing or malformed	That harness is skipped for scheduled evaluation
Effective platform eligibility policy is missing, malformed, or unverifiable	Scheduled evaluations governed by that policy are denied
Effective platform eligibility policy uses a wildcard without explicit platform-level justification	Scheduled evaluations governed by that policy are denied
A required entity field or history scope is not yet loaded, or pagination/coverage is incomplete	Its value is unknown, never false or empty; read more or defer to bound run; no sandbox starts from an unresolved result
A poll candidate remains inconclusive, cannot be handed off, or its bound run fails before completion or safe deferral	Retain a durable retry path or prove source rediscovery after checkpoint advancement; no sandbox starts from an unresolved decision
A required entity source is unsupported, unavailable, fails, or cannot establish complete coverage within allowed bounds	Evaluation fails closed; no sandbox starts and the failure is not recorded as a false CEL result
A conservative dependency or provider query cannot be shown to include every authorized fact that could change a definitive CEL result	Expand the plan or leave the result unknown; if it cannot be resolved, evaluation fails closed
Current attribution or a live or eligible cached CEL authorization verdict for a decision-relevant actor-originated fact is missing or unverifiable	The fact cannot be admitted as authority; if it could change the result, evaluation remains unknown or fails closed and cannot start a sandbox
A versioned platform policy with an explicit maximum age no greater than five minutes for cross-check CEL actor-verdict reuse is missing or unverifiable	Reuse is disabled; obtain a live provider verdict or leave dependent CEL unknown, with no sandbox on unknown
A needed CEL actor-verdict cache entry is expired, invalidated, or scoped to another source, actor, repository, policy, or revision	Refresh from the authoritative provider; if that cannot complete, the dependent result remains unknown and cannot start a sandbox
A decision-relevant human actor is below the stage threshold, or a bot has no recognized role	The fact is excluded from the authorized CEL view and cannot trigger execution
Versioned normalized-entity contract is missing or incomplete	Entity-first execution denied

Exceptions

Certain transitions are authorized without requiring a write or triage role from the acting user. Each exception is documented with its rationale.

Label application

Current compatibility behavior: GitHub authorizes an added-label event without applying the ordinary actor.role threshold. This existing behavior preserves bot-to-bot handoffs such as adding ready-to-code after triage completes, including handoffs from provider bots that are not Fullsend- registered.

ADR 0107 target behavior: When transition.kind is label_changed and label.action is added, GitHub authorizes the event regardless of actor.role once the adapter has established an accepted forge label mutation and authoritative actor-to- transition provenance. This is a separate platform authorization grant, not an application of the ordinary stage threshold. At the time of this ADR, GitHub is the only implementation; another adapter MUST provide equivalent forge-authoritative evidence before enabling the same grant, and absent or unverifiable evidence denies the event. For a bot actor, the adapter MUST positively classify it using provider-controlled metadata, but actor.bot_role lookup is not required. The platform gate MUST NOT inspect the label name or maintain an agent-role label allowlist; harness/CEL routing owns that mapping and may further narrow the dispatch.

Bot-submitted reviews (GitHub)

Current compatibility behavior: Until the resolver is implemented, fullsend dispatch continues to authorize a GitHub review_submitted event from a bot without requiring a collaborator permission or actor.bot_role. This existing exception is not evidence that the bot identity was resolved.

ADR 0107 target behavior: After the resolver is implemented, a GitHub review event is authorized only when the provider classifies the actor as a bot and resolves its exact registered actor.bot_role; actor.role remains none. The downstream harness CEL trigger may further constrain which bot role and review state are accepted.

Lifecycle close (pull_request_target.closed)

The pull_request_target.closed event that triggers the retro stage is intentionally ungated: any closer may trigger read-only lifecycle accounting. The retro agent performs only observation work and does not mutate repository state. This exception is currently implemented in the bash dispatch path; the Go IsAuthorized() path has no special handling for closed transitions and applies the standard write+ gate.

Schedule and manual dispatch

When source.system is schedule or manual, the actor is the configured operator or service identity (for example, a GitHub App bot or workflow GITHUB_ACTOR). Adapters MUST classify that identity from authoritative provider metadata. Human identities use verified forge permissions and the human thresholds; bot identities use provider-backed actor.bot_role resolution and retain actor.role as none. An unrecognized or unresolved bot identity is denied.

Current compatibility behavior: Until provider-backed bot-role resolution is wired into the adapters and dispatch path, adapters set actor.role to the service identity's effective repository permission (typically write), so the standard permission gate applies.

Target contract, not yet implemented: Existing schedule/manual compatibility handling remains authoritative until provider-backed bot-role resolution is wired into the adapters and dispatch path.

Entity-source CEL evaluation

This future path was adopted by ADR 0098, clarified by ADR 0144, and revised for selective acquisition and audit by ADR 0147, with bounded cross-check actor-verdict reuse under ADR 0148. It applies to an entity-source harness's CEL predicate whether a verified prompting event is present or event == null; legacy event-only dispatch is unchanged. The entity-source path is not yet implemented by fullsend dispatch, fullsend poll, or bound fullsend run. Entity-first execution MUST NOT be enabled until the versioned normalized-entity contract defines CEL-eligible and prompt-eligible fields and their coverage states. That contract is a platform allowlist; a harness may request fewer fields but cannot expand it.

With a prompting event, the event-actor authorization gate above still runs before harness CEL; neither discovery nor an entity-source query replaces it. The event actor authorizes that event, not other actor-originated entity facts. The verified GitHub label-added exception authorizes CEL to use that exact accepted label transition without an ordinary actor-permission threshold. It does not authorize other label applications, the entity's current label state without transition provenance, or other records from the same actor. Current label state may use that grant only when the adapter proves it is the projection of the same accepted transition, with no intervening replacement. For an eventless check, the following Fullsend-origin and platform-eligibility gate applies instead.

fullsend poll may enumerate candidates without synthesizing a NormalizedEvent or event actor. A bound fullsend run may instead recheck only its canonical entity with event == null, including after consuming an initial event. Both paths require eligible entity sources and effective platform policy permitting eventless evaluation. They are authorized only when the caller has trusted Fullsend-controlled invocation provenance: a verified, non-user-assertable platform execution identity bound to the invocation and its target, such as an attested workflow/job identity or installation credential. Neither the actor of an earlier event nor an authenticated wake peer supplies this authority. A CLI flag, request header, or other caller-supplied claim is insufficient. A caller that cannot establish that provenance MUST be denied. Wildcard eligibility (* or all) requires the same explicit platform-level justification as any other wildcard allowlist in this contract.

Each entity-source evaluation's audit record MUST identify its verified initiator (the event actor for an event-bearing check, or the trusted Fullsend invocation identity for an eventless check), target repository, effective policy, harness revision, canonical entity, acquisition-plan identity, and decision outcome, including UNKNOWN, deferred, or denied outcomes. It MUST record which source scopes and bounded queries were needed, their coverage and completeness evidence, and the current attribution and live or eligible cached CEL authorization outcome for individual actor-originated records used as decisive witnesses. Complete negative or aggregate claims require bounded query and source-coverage evidence, bound to the effective authorization policy, permission source, and check time or equivalent authorization epoch. Summaries MAY use counts and digests of admitted, excluded, and unresolved actor checks; they do not require an audit roster of every record and actor examined to establish completeness. When an actor verdict is reused for CEL, the audit MUST record bounded evidence of its authoritative source, lookup start time or age measured from that start, the completed verdict, the maximum age allowed by the effective platform policy, effective authorization policy and source revision, and whether a live, local-cached, or shared-cached verdict was used. This may be aggregated for collection-wide checks rather than listing every actor. If a cached actor verdict contributes to a false CEL result, the audit MUST also identify the bounded recheck or rediscovery guarantee, or show that the candidate remains retryable. Unrelated history is not fetched or listed solely for audit. The audit does not retain raw rejected content. An eventless child cycle attributes initiation to the Fullsend identity, never to an earlier event actor or historical entity actor.

Authorization of an event actor or Fullsend invocation origin does not authorize other entity content. For each evaluated harness, Fullsend derives a conservative acquisition plan from its CEL predicate and the normalized-entity contract. The platform retains a closed classification of action-indicating categories; it need not fetch every category for every predicate. The plan MUST cover every field, history range, or derived fact that could change the predicate's result, including actor-originated issue or change-proposal bodies, comments, reviews, and label applications or current label state when used as action-indicating evidence. Counts, existence tests, and other aggregates over actor-originated records have the same dependency. The evaluator MAY defer planned reads and stop once partial evaluation yields a definitive result. A source field is actor-neutral for a particular CEL use only when an actor-originated change cannot make the predicate newly actionable. For a service-clock predicate, raw authoritative updated_at means time since any forge update; it is not recomputed from actor-filtered history. Where the provider guarantees that it only advances, updated_at < moving_cutoff can become true as time passes. An update, including one by an unauthorized actor, can defer that staleness run but cannot make it newly eligible. A harness that needs time since the last authorized update must acquire and authorize the relevant history. By contrast, updated_at >= cutoff can become true because of an update and requires provenance and authorization for that causal occurrence, or remains unknown. Raw comments_count, current labels, or newest-comment metadata cannot enter the authorized CEL view if an unauthorized record or change could affect the result. Fullsend MUST recompute such a fact from admitted records, prove the relevant causal occurrence or platform exception, or leave it unknown. Presenting an action signal as current state never bypasses the actor gate. A provider query MAY narrow reads only when it is proven to include every authorized fact that could affect the result; otherwise Fullsend expands the query or leaves the dependent result unknown. A provider aggregate may replace individual reads only if it preserves the actor-filtered meaning under the eligible CEL verdicts for the current check and proves the coverage needed for that result. Unanalyzable dynamic access requires a conservative broader plan or a fail-closed unresolved outcome. Cached fields or history may satisfy a planned read only after their source revision and query coverage are validated for the current check. Accepted events and ordinary entity wakeups invalidate affected entity facts under ADR 0144. They do not by themselves invalidate a separately scoped actor-verdict cache entry; a trusted authorization change does.

Fullsend distinguishes not requested or not loaded, partially loaded, successfully empty, absent or inapplicable, unsupported, and failed data. Unloaded or incomplete data MUST NOT masquerade as an empty collection or a false condition. For every decision-relevant fact that remains actor-originated under the causal classification above, Fullsend MUST re-establish its current source-revision actor and obtain an eligible provider-verified authorization verdict before admitting the fact into CEL: a human must meet the harness's observation or mutation permission threshold. A bot needs a provider-backed exact registered role under ADR 0107. Only the exact transition covered by the verified label-added event exception is exempt from these checks. That exception supplies authority for the transition only, not for unrelated entity history. An unauthorized fact is excluded from the authorized view. Missing or unverifiable attribution or authorization of a potentially decision-changing fact remains unresolved and cannot prove absence. Historical or inferred actor relationships are insufficient; only the bounded provider-verified CEL verdict cache defined below may replace a live actor-permission lookup. CEL may further restrict the authorized view, but cannot create authority or replace this platform gate.

Bounded CEL actor-verdict cache

This exception to a fresh actor-permission lookup for each entity check is adopted by ADR 0148. For entity-source CEL only, Fullsend MAY reuse a completed, authoritative human permission or exact registered bot-role verdict across entity checks and harnesses, poll cycles, separate poll and bound run loops, and the bound run's final CEL decision. The cached value MUST be the resolved role or verified absence, not a harness-specific threshold decision; every check reapplies its own stage policy. Both positive and verified negative results may be cached; errors, timeouts, unresolved lookups, caller claims, and partially verified roles MUST NOT be cached. The key MUST bind the provider, tenant, canonical repository or project, immutable actor ID, verified actor kind, effective authorization policy and its revision or digest, and permission source. When applicable, it also binds the selected ref and the presence or absence and content or revision of both OWNERS and OWNERS_ALIASES, or the bot-registry revision. A login or entity ID is not an actor cache key. Where an authority does not expose a revision token, the authoritative lookup's start time defines the observed epoch; known changes to that source still invalidate its entries.

Within-check memoization under ADR 0147 avoids repeated lookups while the verdict remains eligible under an absolute five-minute hard cap. Cross-check reuse is disabled until a versioned platform policy defines and enforces an explicit maximum age no greater than five minutes. The platform MAY choose a shorter age; repository or harness settings MUST NOT widen the hard cap or the effective platform bound. Only a completed verdict may enter the cache. Its absolute expiry is measured from the authoritative lookup's start at the earlier of the hard cap and any shorter platform bound. An entry expires at that fixed time even if it is read, handed from poll to run, or retained by a shared host cache; no hit extends its lifetime. An expired verdict must be refreshed before it can contribute to a definitive CEL decision, including during one long progressive check. Known role, membership, OWNERS, OWNERS_ALIASES, bot-registry, or effective-policy changes MUST invalidate affected entries without waiting for TTL expiry. An ordinary entity wakeup invalidates affected entity facts, not an otherwise valid actor verdict, unless it carries trusted evidence of an authorization change. Shared cache readers MUST enforce expiry with a trusted clock or a conservative skew allowance; uncertain age makes an entry unusable. An expired, invalidated, or mismatched entry requires a new authoritative lookup. If that lookup cannot complete, the dependent CEL result remains UNKNOWN or is denied under the fail-closed rules above; the stale verdict is never used. A shared host cache MUST remain inaccessible to the sandbox and agent.

A false CEL result dependent on any cached actor verdict MUST cause a recheck after that verdict expires or is invalidated, and no later than the first of those events plus a bounded, platform-controlled discovery or reconciliation interval, even when the entity has not changed. The recheck obtains a new verdict and may be scheduled directly or guaranteed by source rediscovery. The platform MUST define that interval and either retain a durable candidate through checkpoint or watermark advancement or prove rediscovery after it. If timely re-evaluation cannot be guaranteed, the cached verdict MUST NOT close the candidate: Fullsend uses a live verdict for a final false decision or keeps the candidate inconclusive and durably retryable under a bounded retry schedule. A cached actor result alone cannot permanently close the candidate.

This deliberately accepts CEL authorization evidence up to the platform maximum age. A still-valid positive cache entry may admit a newly observed actor-originated record after an unobserved revocation, and a negative entry may exclude a newly promoted actor until expiry and the required bounded recheck. A positive entry may likewise keep a negative predicate false after an unobserved revocation. The bound run still checks the current entity facts, source-revision actor, and coverage before its final CEL decision; only the actor authorization verdict may be reused within TTL. The CEL cache MUST NOT satisfy event-actor admission or ADR 0133's model-bound record admission and exposure checkpoints. Those gates continue to resolve fresh authority under their separate rules. Fresh model-bound checks may redact records or deny further execution, but they do not undo a sandbox launch admitted using a positive CEL cache entry. The launch may already have occurred after an unobserved revocation within the permitted age.

Fullsend MAY evaluate the predicate over a partially materialized authorized view. It may return a definitive true or false only when that result is the same for every possible completion of still-unloaded inputs admitted under the actor-verdict rules above and consistent with established source coverage. This quantifier includes still-valid cached CEL verdicts within the bounded platform age, even if actor permission changed without a known invalidation. Coverage MUST account for pagination and concurrent source changes; a required collection remains incomplete unless a stable view or sound reconciliation establishes its relevant range. For example, an authorized witness may settle an existence test as true; false or a negative condition requires complete coverage of its relevant authorized search space. An UNKNOWN result requires further relevant field or history reads and re-evaluation in poll or bound run, or retention as retryable inconclusive work. A provider failure, unsupported source, unresolved authorization, or inability to prove completeness in a remaining decision dependency within allowed bounds fails closed: it cannot be converted to false, empty, or a sandbox run.

fullsend poll MAY hand an unresolved candidate and its acquisition hints to a bound fullsend run without claiming a CEL match. A definitive true result in poll also selects only a candidate, not a sandbox launch. The bound run MUST independently re-resolve current entity state and actor provenance, use a live or eligible cached CEL actor verdict, and reach a definitive true result before starting a sandbox. A definitive false result does not start one. Any deciding verdict that expires or is invalidated before sandbox admission MUST be refreshed and CEL re-evaluated; otherwise the decision remains inconclusive and the sandbox does not start. Any handoff is provisional. Until the bound run reports a definitive false decision with any cache-dependent recheck durably arranged, completes actionable work, or safely defers it under ADR 0144, discovery MUST retain a durable retry record or prove that its source query will rediscover the candidate after advancing a checkpoint or watermark. Scheduling a run, delivering a wakeup, or receiving a process-start receipt is not completion. If the bound run fails or remains inconclusive, the candidate MUST remain retryable and cannot be recorded as false, empty, or handled.

Fullsend MUST omit or minimize unneeded untrusted fields where the normalized entity contract permits. Further trust or injection filtering MAY run after CEL routing and before the harness pre-script. Retained content remains untrusted for dispatch and CEL evaluation throughout. ADR 0133 defines the separate model-bound admission gate: only filtered content whose source record and current actor authorization satisfy that ADR may be exposed to the model under bounded harness scope. Any resulting agent run uses the harness's configured identity and permissions.

Model-bound actor-attributed data admission

The model-bound gate applies to every source-system record that Fullsend may expose to a model, regardless of whether the record arrived in an event, was found by polling or snapshotting, or was read through a proactive or Fullsend-controlled API. Fullsend MUST establish the record's current actor attribution through the source system and resolve that actor's current authorization. A human MUST meet the applicable observation or mutation permission threshold. A bot MUST have fresh provider-backed recognition of its exact registered actor.bot_role under ADR 0107; its actor.role: none is not compared with a human threshold. Missing or unverifiable attribution, unrecognized or unresolved bot identity, and a human below threshold MUST redact the record; there is no data-only fallback. A label-event grant applies only to its exact verified transition and does not admit unrelated content.

Source-record attribution MAY come from a provider-verified event or be resolved from the source system. Model-bound actor authorization MUST be resolved from the current authoritative permission or registered-role source under the effective policy after event or retrieval admission; it MAY be co-fetched with a later record read or lazily resolved when first needed. A resolved result MAY be cached only for the logical lifetime of the containing event or entity structure and agent invocation, subject to the fresh checkpoints below. Re-reading or re-fetching the data, creating a new structure for retry or requeue, or terminating and re-invoking the agent ends that lifetime and requires fresh resolution, even if the same in-memory structure is reused. The cache lifetime is not extended by object identity or by retaining the structure.

Event-carried authorization evidence MAY support event admission and routing, but MUST NOT satisfy model-bound checks. Event-triggered runs MUST resolve record-actor authorization after event admission and before model admission, again immediately before sandbox initialization, and again immediately before first model exposure. If sandbox initialization and first exposure are contiguous, one fresh check MAY serve both. Non-event retrieval paths MUST perform the equivalent checks after retrieval admission. A Fullsend-controlled capability invoked after sandbox initialization MUST resolve authorization immediately before releasing each result to the model. That release check is the mid-run equivalent of the sandbox checkpoints.

A cache entry MUST NOT cross any checkpoint. Implementations MAY coalesce contiguous checks and MAY batch or cache lookups for multiple records from the same actor, provided the cache remains within this logical lifetime. Failure, unavailability, or loss of authorization at any checkpoint MUST redact the record and, if no authorized records remain, deny the run. Revocation after first exposure cannot retract content already present in the model context, but every later retrieval result requires its own release check.

Excluded fields

The following fields are not authorization evidence and must not be used for dispatch gating:

Field	Why excluded
author_association	Does not reflect private org membership; an org admin with private membership gets CONTRIBUTOR instead of MEMBER (github/gh-aw-mcpg#2862). Also reflects contribution history, not current authority.
Contribution history	Past contributions do not confer repository permissions. A former maintainer's contributions cannot substitute for a provider-verified role verdict; a valid CEL cache entry may still reflect a role from before revocation within its bounded age.
actor.is_entity_author	Being the author of an issue or PR does not grant repository permissions. This field supports routing decisions in CEL triggers, not authorization.

Principle: relationship and contribution-history fields are not evidence of current authority. Event-backed authorization must be derived from the forge's permission model at event time, not from cached or inferred relationships. Fullsend-originated eventless evaluation must instead be authorized from trusted Fullsend-controlled invocation provenance for poll enumeration or bound rechecks. Decision-relevant actor-originated entity facts are separately gated by current source-revision attribution and either a live or eligible bounded CEL actor verdict before they can influence CEL; historical actors and relationship fields must not serve as authorization evidence.

Enforcement point

Authorization is a platform-level gate that CEL cannot override. Event-backed dispatch checks the normalized event actor before routing; Fullsend-originated eventless evaluation checks trusted invocation provenance and platform eligibility before resolving the entity. An entity-source predicate may then be evaluated progressively, but each actor-originated fact passes its own actor gate, using live or eligible bounded CEL authorization, before it can contribute to a definitive CEL result. Its current source-revision attribution and coverage are checked separately. Not-yet-loaded facts remain unknown rather than being treated as unauthorized, absent, or empty.

Forge event
  --> NormalizedEvent (adapter)
  --> Event-actor authorization gate        <-- enforced here
  --> Event-only CEL routing (unchanged), or entity-source resolution
      --> Dependency-driven acquisition and per-fact actor gate
      --> CEL true / false / UNKNOWN
      --> Execution only on definitive true

Fullsend poll discovery / bound run eventless recheck
  --> Trusted-origin authorization gate     <-- enforced here
  --> Enumerate and resolve candidates (poll) / resolve canonical entity (run)
  --> Evaluate each candidate harness (poll) or the bound harness (run):
      --> Dependency-driven acquisition and per-fact actor gate
      --> CEL true / false / UNKNOWN (event is null)
      --> Unknown may continue in bound run after current entity re-resolution
      --> Further trust/injection filtering and execution only on true

CEL triggers: routing only

Harness trigger expressions express routing and may tighten input selection, not platform permission policy. A CEL expression may tighten dispatch conditions (e.g., require a specific label, restrict to non-fork PRs, filter by bot identity) but may never weaken the platform authorization gate. On the event-backed path, an event that fails authorization never reaches CEL evaluation.

This separation is enforced architecturally: on the event-backed path, IsAuthorized() runs before MatchHarnesses() in the dispatch core. On the Fullsend-originated eventless path, the trusted-origin gate runs before poll enumeration or bound-entity resolution and CEL evaluation. For each eligible entity-source harness, the platform admits decision-relevant actor-originated facts only after current source-revision attribution and a live or eligible bounded CEL actor verdict under its stage policy. The verified event grant may authorize only its exact label transition, as described above. Eventless checks use event == null. Neither path lets a CEL expression override or relax an authorization denial. Legacy event-only harness routing does not acquire an entity history under this section.

Per-repo configurability

The authorization gate is a platform-level security boundary. Individual repositories cannot disable it. Per-repo configuration (which stages are enabled, which labels trigger automation) operates within the authorization boundary. A repository can disable a stage entirely but cannot make it available to unauthorized users.

A repository can add permission sources through the authorization field (today only owners_file). These sources extend the has_repo_permission helper and the Go IsAuthorized check; they do not bypass the gate or change the thresholds.

Versioning

This is the v2 normative contract under ADR 0015. ADR 0147 supersedes v1's future full-history entity-evaluation and audit requirements; ADR 0148 adds bounded cross-check reuse of provider-verified actor-role verdicts for entity-source CEL only. It does not change v1 event admission or the v2 model-bound release checkpoints. v1 remains available for existing dispatch implementations during migration. Breaking changes to v2 require docs/normative/authorization/v3/.

ADR 0107 migration exception: The migration from legacy bot authorization exceptions to provider-backed bot identity is also permitted within v2. It is a policy tightening, not a new supported actor population: existing compatibility behavior remains authoritative until an adapter implementation is updated; once updated, registered bots use actor.bot_role and unregistered non-label bots are denied. The adapter implementation itself is the migration boundary; no per-event mode marker or separate deprecation window is required. An updated adapter MUST enforce the target behavior and MUST NOT silently fall back to legacy bot authorization when its resolver fails. The normalized-event fields used to carry this distinction are additive, and the permanent label-added exception is preserved. This exception does not alter the general v3 rule for unrelated authorization changes.

Change	v2 impact
Breaking (requires v3): remove a role from the hierarchy, raise a default threshold, remove a documented exception, change fail-closed to fail-open, or weaken the decision-relevant actor and coverage gates beyond the bounded CEL cache adopted here	Dispatch implementations must migrate
Adapter bot-identity migration (allowed in v2): update an adapter to emit and enforce the ADR 0107 target representation	The adapter's existing compatibility behavior applies until that implementation update; afterward, registered bots use actor.bot_role and unregistered non-label bots are denied
Non-breaking (allowed in v2): add a role, lower a default threshold, add a new exception, add forge mappings, clarify documentation	Existing dispatch behavior is preserved or relaxed