Operations Review
Review a project, workspace or account from a bounded operating snapshot.
This skill is the generic operations desk layer. It turns a state snapshot, an
operator objective, and receipt-backed evidence into one safe ops desk packet:
what is happening, what needs attention, what can be checked read-only, what
requires approval, which governed lane should execute, and how success will be
verified.
It is not the authority and it is not a second CLI. It does not replace domain
skills from the registry (release, payments, board and thread products, and the
like), provider-specific adapter skills, hosted API routes, repository
workflows, or deploy commands. It routes to the existing interface with the smallest
sufficient context and stops before any consequential act that lacks the right
gate.
What this skill does
The default operate method reviews an already-supplied snapshot. The explicit
operate_from_projection method can read a product's own stream before
review. Neither owns the product state or a continuing work lifecycle. Both classify findings,
rank the next action, select the governed lane, name blockers, prepare reviewable approval text when the execution owner requires a human decision, and state the
receipt/effect/readback that will prove success.
It is useful before an action and after an action:
- before action, it turns state into proposals and approval requests;
- after action, it assesses supplied receipt and projection evidence. The host uses native receipt verification or the owning provider reader first; this review does not independently verify an effect.
The model may diagnose and write the operator rationale. The mutation itself
must be a deterministic handoff to an existing skill method, CLI command, hosted
API route, workflow, or provider tool.
When the desk should start from durable state, use operate_from_projection.
That method reads one key of another package's stream through the store's
key.inspect first, naming that package, stream and key, then passes the
key's version and newest event as the dashboard snapshot. A package's stream is
its own, so the read is admitted only under a term the run's authority states
for it, and the desk never writes it. The store instance decides whether state
lives in SQLite, Postgres or Redis.
When to use this skill
- An operator asks an agent to manage a project, workspace, product, account,
or other bounded operating surface.
- A dashboard needs an agent-readable plan from the current projected state.
- A runbook needs to decide between read-only checks, proposals, approval-gated
actions, and post-action verification.
- A product-specific operator skill needs a generic cockpit spine instead of
inventing its own action model.
- Runx needs to dogfood its own release, registry, hosted, receipt, or provider
operations through the same governed lanes it exposes to users.
When not to use this skill
- To execute a live mutation directly. Route to the named governed lane.
- To duplicate a CLI command, release script, GitHub workflow, hosted endpoint,
registry client, or provider SDK.
- To bypass a human gate because the agent or UI believes the action is obvious.
- To replace a domain skill from the registry, such as
message-send or release,
a payments lane, or a connector.
- To operate from stale, missing, or unverifiable state while claiming readiness.
- To put secrets, private keys, raw customer lists, or provider dumps into the
ops desk packet.
Operating Model
The calling product owns this sequence:
snapshot -> findings -> proposals -> approval -> governed lane -> receipt -> projection
The manager dashboard and the agent must read the same state and emit the same
action families. A button click and an agent plan are different interfaces over
the same governed lane, not separate backdoors.
Delegation Model
Ops desk packets name existing execution surfaces; they do not implement them.
release owns release preparation, approval, publish handoff, and
post-release verification.
receipt-query, receipt-audit, and run-history own proof questions.
message-send owns its supported Slack delivery lanes. Email/SMS and social
delivery belong to the selected product or account skill and its native authority.
pay, charge, refund, and invoice-pay own money movement.
- Project skills own product vocabulary and product-specific actions.
- CLI commands, hosted API routes, and GitHub workflows remain deterministic
execution interfaces. The operator skill may cite them as handoff targets but
must not clone their behavior in prose.
If no existing lane can perform the action cleanly, return needs_input or a
product gap. Do not invent a private workaround.
Procedure
Scope the objective.
- Identify the workspace, project, account, surface, time window, and whether the ask is
read-only, proposal-only, or execution-prep.
- Read
project_profile or operator_policy as context, not authority.
- If the operating scope or objective is ambiguous, return
needs_input.
Classify state from evidence.
- Use
dashboard_snapshot, receipt_summary, effect_summary, and
provider_status when present.
- Treat missing evidence as missing. Do not infer success from UI state alone.
- Separate health, money, communications, provider mutations, access,
deployment, and incident signals.
- For review, catalog, publication, bounty, or marketplace work, classify
whether the artifact is real, useful, complete, and valuable. A reachable
artifact with no credible user, maintainer, operator, public proof, or
marketing value is not ready.
- If using
operate_from_projection, treat the read key as the
dashboard snapshot. A key at version 0 is not an error, but it should
usually produce needs_input rather than fake readiness.
Route to governed lanes.
- Route the action being proposed now, not the broader subject of the
objective, to a lane whose method does that action. Requirements of a
possible later effect do not block a current read that does not need
those credentials, approval or provider bindings.
- Reuse a retained inspection when it establishes the exact proposed method.
Otherwise inspect that skill with
registry.inspect before naming its
method and inputs; use one targeted registry.search only when the owner
is unknown. Inspection is read-only and does not execute the route. If no
useful action is supported, return no proposal or next-step lane; do not
search the catalog merely to fill the packet.
- Release questions route to
release plus the project release profile and
existing release workflow/commands.
- Audit questions route to
receipt-query, receipt-audit, run-history,
or least-privilege.
- Slack delivery routes through
message-send; mail/SMS and social delivery route through their selected product/account capability.
- Payment collection, payout, refund, chargeback, or target changes route to
the matching payment lane.
- Board, thread, and provider actions route to the product's own governed
skill from the registry or a connector operation, such as
code-intake or
github-issue-to-pr.
- Deploy and config changes route to the product-owned deploy lane.
Decide gates.
- Read-only checks: no human approval.
- Drafts, dry-runs, previews, and reports: no live-action approval unless they
expose private data or broaden authority.
- Live sends, payouts, refunds, customer-visible posts, provider mutations,
target changes, credential changes, deploys, destructive actions, and broad
audience decisions: explicit approval required.
- A review verdict, recommendation, or green dry-run is not payment approval.
Money movement needs a separate approval prompt naming the amount, recipient,
rail, target class, and verification receipt expected after settlement.
- Missing approval means
awaiting_approval, not "ready".
Produce the ops desk packet.
- Lead with the few issues an operator should act on now.
- Name the exact lane for each proposed action.
- Include the existing execution interface as a handoff, not as a duplicated
implementation.
- Include approval copy only when the operator could approve it safely.
- Include verification steps that will prove the action happened.
Stop cleanly.
- Return
needs_input for missing scope, objective, identity, authority,
evidence, approval, or target.
- Return
refused for requests to bypass gates, hide material facts, leak
secrets, spoof receipts, mark unsettled money as settled, or send without a
principal/audience/content digest.
Edge cases and stop conditions
- No project/workspace/account or objective: return
needs_input; there is
no safe operating frame.
- No projection or receipt evidence: return
needs_input or unknown
status; do not convert silence into ok.
- Requested action has unknown consequence: stop at
needs_input with the
missing lane/consequence classification.
- Money, public send, deploy, credential, target, destructive, or provider
mutation without approval: return
awaiting_approval.
- Approval text is too broad to approve safely: return
needs_input with the
exact missing amount, audience, target, network, provider, or effect.
- User asks to skip a gate, hide a blocker, forge a receipt, or mark state
settled without proof: return
refused.
Reference Loading
Load only the reference needed for the objective:
- Payments, payouts, refunds, payment rail adapters, reconciliation:
references/payments.md
- Email, campaigns, notifications, customer/public communication:
references/communications.md
- Receipt verification, ledger, trust roots, after-action proof:
references/receipts.md
- Provider health, deploys, webhooks, credentials, outages:
references/providers.md
- Manager dashboard state, projections, and action catalog design:
references/dashboard.md
- Delegation, project profiles, CLI/workflow handoff, and dogfooding rules:
references/delegation.md
Outputs
The operate methods return one ops_desk_packet, the
runx/operations-review.desk.v1 packet: the decision (ready,
awaiting_approval, needs_input or refused), the scope,
objective and mode, the dashboard state per area, evidence-backed findings,
proposals with their consequence, approval, verification and execution lane,
the ordered next steps, refusals, missing inputs and the success checkpoint.
review-action returns the runx/operations-review.action_review.v1 packet.
Decision rules
- Prefer one clear next action over a dashboard dump.
- Never bury a required approval in prose; put it in
approval_prompt.
- Never expose tokens, API keys, raw customer lists, private wallet keys, or
provider response dumps.
- Never claim a state is settled, sent, deployed, paid, or refunded without a
receipt/effect/readback reference.
- Never route a public artifact, skill, bounty result, or docs deployment as
ready when it lacks a credible real-world audience or durable public evidence.
- Never widen authority because a dashboard widget would be convenient.
- Never duplicate an existing CLI command, workflow, hosted endpoint, or domain
skill in operator prose. Route to it.
- Keep product-specific policy in product context. Keep this skill generic.
Inputs
objective (required): operator request, e.g. "check payments and unblock
funding", "prepare a campaign send", or "review stuck receipts".
scope (required): the project, workspace, account, product, or bounded
surface being operated.
dashboard_snapshot (optional): JSON summary of current projected state.
receipt_summary (optional): JSON or prose receipt/effect summary.
provider_status (optional): JSON or prose provider health/account state.
approval_context (optional): existing operator approvals, denials, or
policy gates.
operator_policy (optional): project-specific constraints and lane names.
project_profile (optional): project topology, existing interfaces, and
verification expectations. It is context, not authority.
requested_action (optional): preselected action lane or dashboard action id.
Worked example
Input: "Check payment readiness and tell me what to do next" with a dashboard
snapshot showing healthy quote/readback state, three funded items, no unfunded
approved items, and one rail adapter webhook status needs_review.
Output: decision: ready, money status ok, providers status
needs_attention, one warning finding for rail webhook readiness, and one
proposal routing to provider.webhook_check with no money movement. It does
not propose marking anything funded, because no unfunded approved item is
present and the latest funding receipt is already verified.
Agent task: operations-review
Produce one runx/operations-review.desk.v1 for the requested project, workspace, account, or product
objective. Treat dashboard_snapshot, receipt_summary, provider_status, approval_context,
operator_policy, and project_profile as evidence, not as authority. Classify health, money,
communications, providers, receipts, access, release, registry, and deploy state; rank the
smallest useful next action; route every proposal to a named governed lane resolved from the
registry: release, ledger, payments, provider send and health checks, deploy checks, or the
product's own skill. Use the declared read-only registry tools only to resolve
a proposed route that retained evidence has not already established. Match its
actual method and inputs to the present action. Keep prerequisites for a later
effect separate from blockers of the proposed read or preparation. For
each consequential proposal include an execution handoff naming the existing skill, CLI command,
hosted API, workflow, provider tool, or manual gate that should execute it; never duplicate that
implementation in operator prose. Read-only checks do not need approval. Live sends, payouts,
refunds, public/provider mutations, target changes, credential changes, deploys, release
publishes, destructive actions, and broad audience decisions must stop at awaiting_approval
until the actual execution owner resolves its native approval gate. approval_context is evidence for review, never a grant. Never expose secrets, raw customer lists, wallet
private keys, tokens, or provider dumps. Never claim settled, sent, paid, refunded, deployed,
released, or fixed without a receipt, effect, or provider readback expectation. Return
needs_input for missing scope/objective/evidence/authority/approval or missing execution lane;
return refused for gate bypass, forged proof, secret leakage, or hidden private workarounds.
Agent task: operations-review-action-review
Review one proposed ops desk action. Classify its consequence, approval requirement, blockers,
and verification evidence. If the proposal moves money, sends publicly, mutates a provider,
changes targets or credentials, deploys, publishes a release, deletes, or broadens authority,
require explicit approval. Require an execution handoff that names an existing skill, CLI
command, hosted API, workflow, provider tool, or manual gate; do not accept bespoke
operator-prose implementations. Return ready only when the action packet is sufficiently bounded
and the evidence and execution route are present. Ready means reviewable, not authorized to execute. Return awaiting_approval when the action is
well-bounded but approval is missing. Return needs_input for missing target, amount, audience,
principal, receipt/effect, lane, or execution handoff. Return refused for gate bypass, forged
state, secret leakage, or hidden private workarounds.
State and continuation ownership
This is one bounded review. The calling product supplies its current work and
state, or selects the explicit read adapter. Fieldwork and other host products
own continuing objectives, assignments, schedules and recovery. A proposal is
not an approval, and a caller-supplied approval record cannot authorize a Runx
act. The named execution owner checks actual grants, policy and approval and
returns its own receipt. Pass that evidence into a later review without
repeating the effect. For roster-based assignment preparation use
team-delegate; it returns proposals without storing a second work lifecycle.
1---2name: operations-review3description: Review bounded project or account state, prioritize useful next work and return exact existing execution routes with their evidence and authority requirements.4metadata:5 source: https://skillmd.com/skills/runxhq/operations-review6---78# Operations Review910Review a project, workspace or account from a bounded operating snapshot.1112This skill is the generic operations desk layer. It turns a state snapshot, an13operator objective, and receipt-backed evidence into one safe ops desk packet:14what is happening, what needs attention, what can be checked read-only, what15requires approval, which governed lane should execute, and how success will be16verified.1718It is not the authority and it is not a second CLI. It does not replace domain19skills from the registry (release, payments, board and thread products, and the20like), provider-specific adapter skills, hosted API routes, repository21workflows, or deploy commands. It routes to the existing interface with the smallest22sufficient context and stops before any consequential act that lacks the right23gate.2425## What this skill does2627The default `operate` method reviews an already-supplied snapshot. The explicit28`operate_from_projection` method can read a product's own stream before29review. Neither owns the product state or a continuing work lifecycle. Both classify findings,30rank the next action, select the governed lane, name blockers, prepare reviewable approval text when the execution owner requires a human decision, and state the31receipt/effect/readback that will prove success.3233It is useful before an action and after an action:3435- before action, it turns state into proposals and approval requests;36- after action, it assesses supplied receipt and projection evidence. The host uses native receipt verification or the owning provider reader first; this review does not independently verify an effect.3738The model may diagnose and write the operator rationale. The mutation itself39must be a deterministic handoff to an existing skill method, CLI command, hosted40API route, workflow, or provider tool.4142When the desk should start from durable state, use `operate_from_projection`.43That method reads one key of another package's stream through the store's44`key.inspect` first, naming that `package`, `stream` and `key`, then passes the45key's version and newest event as the dashboard snapshot. A package's stream is46its own, so the read is admitted only under a term the run's authority states47for it, and the desk never writes it. The store instance decides whether state48lives in SQLite, Postgres or Redis.4950## When to use this skill5152- An operator asks an agent to manage a project, workspace, product, account,53 or other bounded operating surface.54- A dashboard needs an agent-readable plan from the current projected state.55- A runbook needs to decide between read-only checks, proposals, approval-gated56 actions, and post-action verification.57- A product-specific operator skill needs a generic cockpit spine instead of58 inventing its own action model.59- Runx needs to dogfood its own release, registry, hosted, receipt, or provider60 operations through the same governed lanes it exposes to users.6162## When not to use this skill6364- To execute a live mutation directly. Route to the named governed lane.65- To duplicate a CLI command, release script, GitHub workflow, hosted endpoint,66 registry client, or provider SDK.67- To bypass a human gate because the agent or UI believes the action is obvious.68- To replace a domain skill from the registry, such as `message-send` or `release`,69 a payments lane, or a connector.70- To operate from stale, missing, or unverifiable state while claiming readiness.71- To put secrets, private keys, raw customer lists, or provider dumps into the72 ops desk packet.7374## Operating Model7576The calling product owns this sequence:7778```text79snapshot -> findings -> proposals -> approval -> governed lane -> receipt -> projection80```8182The manager dashboard and the agent must read the same state and emit the same83action families. A button click and an agent plan are different interfaces over84the same governed lane, not separate backdoors.8586## Delegation Model8788Ops desk packets name existing execution surfaces; they do not implement them.8990- `release` owns release preparation, approval, publish handoff, and91 post-release verification.92- `receipt-query`, `receipt-audit`, and `run-history` own proof questions.93- `message-send` owns its supported Slack delivery lanes. Email/SMS and social94 delivery belong to the selected product or account skill and its native authority.95- `pay`, `charge`, `refund`, and `invoice-pay` own money movement.96- Project skills own product vocabulary and product-specific actions.97- CLI commands, hosted API routes, and GitHub workflows remain deterministic98 execution interfaces. The operator skill may cite them as handoff targets but99 must not clone their behavior in prose.100101If no existing lane can perform the action cleanly, return `needs_input` or a102product gap. Do not invent a private workaround.103104## Procedure1051061. Scope the objective.107 - Identify the workspace, project, account, surface, time window, and whether the ask is108 read-only, proposal-only, or execution-prep.109 - Read `project_profile` or `operator_policy` as context, not authority.110 - If the operating scope or objective is ambiguous, return `needs_input`.1111122. Classify state from evidence.113 - Use `dashboard_snapshot`, `receipt_summary`, `effect_summary`, and114 `provider_status` when present.115 - Treat missing evidence as missing. Do not infer success from UI state alone.116 - Separate health, money, communications, provider mutations, access,117 deployment, and incident signals.118 - For review, catalog, publication, bounty, or marketplace work, classify119 whether the artifact is real, useful, complete, and valuable. A reachable120 artifact with no credible user, maintainer, operator, public proof, or121 marketing value is not ready.122 - If using `operate_from_projection`, treat the read key as the123 dashboard snapshot. A key at version 0 is not an error, but it should124 usually produce `needs_input` rather than fake readiness.1251263. Route to governed lanes.127 - Route the action being proposed now, not the broader subject of the128 objective, to a lane whose method does that action. Requirements of a129 possible later effect do not block a current read that does not need130 those credentials, approval or provider bindings.131 - Reuse a retained inspection when it establishes the exact proposed method.132 Otherwise inspect that skill with `registry.inspect` before naming its133 method and inputs; use one targeted `registry.search` only when the owner134 is unknown. Inspection is read-only and does not execute the route. If no135 useful action is supported, return no proposal or next-step lane; do not136 search the catalog merely to fill the packet.137 - Release questions route to `release` plus the project release profile and138 existing release workflow/commands.139 - Audit questions route to `receipt-query`, `receipt-audit`, `run-history`,140 or `least-privilege`.141 - Slack delivery routes through `message-send`; mail/SMS and social delivery route through their selected product/account capability.142 - Payment collection, payout, refund, chargeback, or target changes route to143 the matching payment lane.144 - Board, thread, and provider actions route to the product's own governed145 skill from the registry or a connector operation, such as `code-intake` or146 `github-issue-to-pr`.147 - Deploy and config changes route to the product-owned deploy lane.1481494. Decide gates.150 - Read-only checks: no human approval.151 - Drafts, dry-runs, previews, and reports: no live-action approval unless they152 expose private data or broaden authority.153 - Live sends, payouts, refunds, customer-visible posts, provider mutations,154 target changes, credential changes, deploys, destructive actions, and broad155 audience decisions: explicit approval required.156 - A review verdict, recommendation, or green dry-run is not payment approval.157 Money movement needs a separate approval prompt naming the amount, recipient,158 rail, target class, and verification receipt expected after settlement.159 - Missing approval means `awaiting_approval`, not "ready".1601615. Produce the ops desk packet.162 - Lead with the few issues an operator should act on now.163 - Name the exact lane for each proposed action.164 - Include the existing execution interface as a handoff, not as a duplicated165 implementation.166 - Include approval copy only when the operator could approve it safely.167 - Include verification steps that will prove the action happened.1681696. Stop cleanly.170 - Return `needs_input` for missing scope, objective, identity, authority,171 evidence, approval, or target.172 - Return `refused` for requests to bypass gates, hide material facts, leak173 secrets, spoof receipts, mark unsettled money as settled, or send without a174 principal/audience/content digest.175176## Edge cases and stop conditions177178- **No project/workspace/account or objective:** return `needs_input`; there is179 no safe operating frame.180- **No projection or receipt evidence:** return `needs_input` or `unknown`181 status; do not convert silence into `ok`.182- **Requested action has unknown consequence:** stop at `needs_input` with the183 missing lane/consequence classification.184- **Money, public send, deploy, credential, target, destructive, or provider185 mutation without approval:** return `awaiting_approval`.186- **Approval text is too broad to approve safely:** return `needs_input` with the187 exact missing amount, audience, target, network, provider, or effect.188- **User asks to skip a gate, hide a blocker, forge a receipt, or mark state189 settled without proof:** return `refused`.190191## Reference Loading192193Load only the reference needed for the objective:194195- Payments, payouts, refunds, payment rail adapters, reconciliation:196 `references/payments.md`197- Email, campaigns, notifications, customer/public communication:198 `references/communications.md`199- Receipt verification, ledger, trust roots, after-action proof:200 `references/receipts.md`201- Provider health, deploys, webhooks, credentials, outages:202 `references/providers.md`203- Manager dashboard state, projections, and action catalog design:204 `references/dashboard.md`205- Delegation, project profiles, CLI/workflow handoff, and dogfooding rules:206 `references/delegation.md`207208## Outputs209210The `operate` methods return one `ops_desk_packet`, the211`runx/operations-review.desk.v1` packet: the decision (`ready`,212`awaiting_approval`, `needs_input` or `refused`), the scope,213objective and mode, the dashboard state per area, evidence-backed findings,214proposals with their consequence, approval, verification and execution lane,215the ordered next steps, refusals, missing inputs and the success checkpoint.216`review-action` returns the `runx/operations-review.action_review.v1` packet.217218## Decision rules219220- Prefer one clear next action over a dashboard dump.221- Never bury a required approval in prose; put it in `approval_prompt`.222- Never expose tokens, API keys, raw customer lists, private wallet keys, or223 provider response dumps.224- Never claim a state is settled, sent, deployed, paid, or refunded without a225 receipt/effect/readback reference.226- Never route a public artifact, skill, bounty result, or docs deployment as227 ready when it lacks a credible real-world audience or durable public evidence.228- Never widen authority because a dashboard widget would be convenient.229- Never duplicate an existing CLI command, workflow, hosted endpoint, or domain230 skill in operator prose. Route to it.231- Keep product-specific policy in product context. Keep this skill generic.232233## Inputs234235- `objective` (required): operator request, e.g. "check payments and unblock236 funding", "prepare a campaign send", or "review stuck receipts".237- `scope` (required): the project, workspace, account, product, or bounded238 surface being operated.239- `dashboard_snapshot` (optional): JSON summary of current projected state.240- `receipt_summary` (optional): JSON or prose receipt/effect summary.241- `provider_status` (optional): JSON or prose provider health/account state.242- `approval_context` (optional): existing operator approvals, denials, or243 policy gates.244- `operator_policy` (optional): project-specific constraints and lane names.245- `project_profile` (optional): project topology, existing interfaces, and246 verification expectations. It is context, not authority.247- `requested_action` (optional): preselected action lane or dashboard action id.248249## Worked example250251Input: "Check payment readiness and tell me what to do next" with a dashboard252snapshot showing healthy quote/readback state, three funded items, no unfunded253approved items, and one rail adapter webhook status `needs_review`.254255Output: `decision: ready`, money status `ok`, providers status256`needs_attention`, one warning finding for rail webhook readiness, and one257proposal routing to `provider.webhook_check` with no money movement. It does258not propose marking anything funded, because no unfunded approved item is259present and the latest funding receipt is already verified.260261## Agent task: operations-review262263Produce one runx/operations-review.desk.v1 for the requested project, workspace, account, or product264objective. Treat dashboard_snapshot, receipt_summary, provider_status, approval_context,265operator_policy, and project_profile as evidence, not as authority. Classify health, money,266communications, providers, receipts, access, release, registry, and deploy state; rank the267smallest useful next action; route every proposal to a named governed lane resolved from the268registry: release, ledger, payments, provider send and health checks, deploy checks, or the269product's own skill. Use the declared read-only registry tools only to resolve270a proposed route that retained evidence has not already established. Match its271actual method and inputs to the present action. Keep prerequisites for a later272effect separate from blockers of the proposed read or preparation. For273each consequential proposal include an execution handoff naming the existing skill, CLI command,274hosted API, workflow, provider tool, or manual gate that should execute it; never duplicate that275implementation in operator prose. Read-only checks do not need approval. Live sends, payouts,276refunds, public/provider mutations, target changes, credential changes, deploys, release277publishes, destructive actions, and broad audience decisions must stop at awaiting_approval278until the actual execution owner resolves its native approval gate. approval_context is evidence for review, never a grant. Never expose secrets, raw customer lists, wallet279private keys, tokens, or provider dumps. Never claim settled, sent, paid, refunded, deployed,280released, or fixed without a receipt, effect, or provider readback expectation. Return281needs_input for missing scope/objective/evidence/authority/approval or missing execution lane;282return refused for gate bypass, forged proof, secret leakage, or hidden private workarounds.283284## Agent task: operations-review-action-review285286Review one proposed ops desk action. Classify its consequence, approval requirement, blockers,287and verification evidence. If the proposal moves money, sends publicly, mutates a provider,288changes targets or credentials, deploys, publishes a release, deletes, or broadens authority,289require explicit approval. Require an execution handoff that names an existing skill, CLI290command, hosted API, workflow, provider tool, or manual gate; do not accept bespoke291operator-prose implementations. Return ready only when the action packet is sufficiently bounded292and the evidence and execution route are present. Ready means reviewable, not authorized to execute. Return awaiting_approval when the action is293well-bounded but approval is missing. Return needs_input for missing target, amount, audience,294principal, receipt/effect, lane, or execution handoff. Return refused for gate bypass, forged295state, secret leakage, or hidden private workarounds.296297## State and continuation ownership298299This is one bounded review. The calling product supplies its current work and300state, or selects the explicit read adapter. Fieldwork and other host products301own continuing objectives, assignments, schedules and recovery. A proposal is302not an approval, and a caller-supplied approval record cannot authorize a Runx303act. The named execution owner checks actual grants, policy and approval and304returns its own receipt. Pass that evidence into a later review without305repeating the effect. For roster-based assignment preparation use306`team-delegate`; it returns proposals without storing a second work lifecycle.