Jump to text
On this page
Copy and source
Published source confirmed at last check

Source changed 2026-11-18 21:11:39 UTC · snapshot created 2026-11-18 21:12:35 UTC · last check 2026-11-18 21:55:20 UTC

C.24 - Plan Tool or Service Calls for a Fixed Action (C.Agent-Tools-CAL)

Type: Calculus (C) Status: Stable Normativity: Normative

Plain-name. Plan tool or service calls for a fixed action.

Intent. Help a practitioner turn an already fixed action or option into a budgeted call plan, then revise that plan or return a bounded checkpoint without confusing planning with selection or execution.

Depends on.

  • A.15 and its planning and Work patterns for Methods, descriptions, plans, performed Work, and attribution;
  • A.15.7 for a situation-responsive next-action decision;
  • C.11 for a fixed choice among a current OptionSet;
  • C.2.1 when the relied-on decision or checkpoint needs a persistent episteme;
  • C.18 for candidate-generation records;
  • C.19 for live-pool policy;
  • C.19.1 for a scale-based comparison or waiver;
  • C.16 for measured comparison inputs;
  • OPS.11.1:4.2–4.3 for constructing a route from required contributions and reconciling its resource conditions;
  • G.6 when a call trace must remain replayable;
  • B.3 only when a named assurance use needs one bounded assurance result.

Coordinates with. U.PromiseContent for service acceptance conditions, C.28 when a planned call is intended to support a causal use, and E.17/E.24.PUB when an already obtained result is being published or made available.

C.24:0 - Use this when

Use A.15.7 first when ongoing Work still needs the next action to be chosen from current facts within a domain Method. Enter C.24 only after that action is fixed and tool or service calls must be planned. A call plan is neither the situation-responsive decision nor proof that the chosen action was performed.

Use C.24 when an applicable domain prescription or a completed decision has fixed the action or option and the practical question is now:

  • which admitted Methods to call, in what order;
  • which applicable time, compute, cost or risk budget to reserve;
  • what stops or replans the route; and
  • whether the useful output is a CallPlan or a CheckpointReturn.

Do not use it to generate candidates, keep a live pool, choose among unresolved options, execute calls, or score completed Work.

C.24:0.1 - What goes wrong if missed

  • a route is scheduled by an opaque heuristic, so its budget use and stop conditions are hidden;
  • unresolved choice or pool-policy work is hidden in a plan;
  • a route description is mistaken for a Method, a plan for performed Work, or a successful probe for a decision to adopt and execute the route; and
  • replanning loses the basis that fixed the action and its conditions.

C.24:0.2 - What this buys

  • one small, tool-neutral plan that cites the accepted action basis;
  • visible budgets, stop conditions, and replan triggers before calls are made;
  • recoverable call facts and provenance when later use needs an execution account; and
  • one bounded checkpoint stating whether to probe again or seek a decision to adopt the proposed route.

Primary working object. One ATC.CallPlan : U.WorkPlan. Each step selects a U.Method. A route description may help locate or constrain that Method, but remains a separate U.MethodDescription. Actual calls are dated U.Work and remain outside this planning result.

First useful move. Recover the action and completion conditions fixed by the domain prescription, A.15.7 decision or C.11 choose-now result. Reuse an adequate known route.

Otherwise, work backwards from the required result to admitted Methods and their needed inputs, then follow the proposed route forwards from the available inputs as in section 4.1.

Return the supported ordered calls, applicable budget and stop or replan conditions, or the particular missing contribution or condition. Add route-description refs where they are needed to understand or qualify the route.

Not this pattern when. Use C.11 while fixed-option choice is unresolved, C.19 while treatment of a live pool is unresolved, G.5 when the current task is selector-facing result declaration, A.15.5 for work-entry readiness, and A.15.1 when the question is what Work actually occurred or which Method it enacted.

C.24:0.3 - First-minute questions

  1. Which applicable domain prescription, A.15.7 decision or C.11 choose-now result fixes the action or option now being planned?
  2. Does every planned step name an admitted Method, rather than only a vendor route or endpoint label?
  3. Which budget is current: a still-upstream probe budget or an enactment/call budget?
  4. What event stops or replans the route?
  5. Is the useful output a plan, a checkpoint, or a return to a neighbouring pattern?

C.24:0.4 - First output

The first useful output is one of these:

CallPlan:
  actionBasis:
    domainPrescribedAction?:
      methodRef
      prescriptionRef
      prescribedAction
      applicabilityAndStop
      intendedWorkPlanRef?
    situationResponsiveDecisionEpistemeRef?  # A.15.7; only when this plan relies on the retained decision
    fixedOptionChoiceResultRef?               # C.11; only a choose-now result
  objective
  plannedCallsInOrder:
    - methodRef
      methodDescriptionRef?          # only when the route description is needed
      dependsOnPlannedStepRefs?      # only when dependency changes the route
      mayRunInParallelWithStepRefs?  # only when safe parallelism matters
  plannedBudgetEnvelope
  stopOrReplan
  nextPlannedAction
CheckpointReturn:
  actionBasis:
    domainPrescribedAction?:
      methodRef
      prescriptionRef
      prescribedAction
      applicabilityAndStop
      intendedWorkPlanRef?
    situationResponsiveDecisionEpistemeRef?  # A.15.7
    fixedOptionChoiceResultRef?               # C.11 choose-now result
  objectiveOrTaskFamily
  testedMethodRefs
  testedMethodDescriptionRefs?
  evidenceRefs
  burnedBudget
  residualBudget
  recommendedNextAction
  commitTrigger

nextPlannedAction and recommendedNextAction are local fields. Exactly one action-basis branch is present.

Add one of the branch-specific refs in C.24:4.4 only when that constraint still affects the plan. A plan with no current policy branch needs no policy placeholder.

If neither output can cite its accepted action basis and state what happens next, the C.24 work is unfinished.

If the domain prescription, its applicability or the action it requires changes, reopen that domain basis before revising the plan. If the A.15.7 decision changes, is withdrawn, or no longer fixes the action, reopen the plan and return to A.15.7. If the C.11 ChoiceResult changes or no longer says choose now, return to C.11. A changed live-pool branch returns separately to C.19. Do not revise the call plan as though its action basis were still settled.

C.24:1 - Problem frame

Tool-using Systems may plan across web services, local programs, instruments, robots, or human-operated routes. The implementation may be an LLM agent, a search system, a conventional planner, or a fixed program. The planning problem is the same: turn a fixed action or option into an ordered and bounded route without hiding route grounding, budget, or stop logic.

A local system-role kind or assignment is recorded only when that separate fact matters.

When planning, revision, or a call is claimed as precise performed Work, recover each performer’s identity through A.13 and let A.15.1 independently admit the dated Work from its performer, Method, interval, and containment facts.

Add the exact A.2.1 assignment reference and F.6 only when the plan or receiving use expressly consumes precise assignment-bound attribution through the same obtaining A.13 assignment; F.6 identifies neither assignment nor performer, and missing or failed F.6 attribution alone leaves the Work intact while its independent A.13/A.15.1 grounds still hold.

C.24:2 - Problem

We need a tool-neutral way to produce or revise one call plan under explicit budgets and policy while keeping Method, route description, plan, performed Work, service promise, trace representation, the action basis that fixed the action, and any assurance result distinct.

C.24:3 - Forces

ForceTension
General method vs local shortcutA scalable approach may improve with data or compute, while a narrow route may be safer or cheaper in the present task.
Exploration vs deliveryA bounded probe may reduce uncertainty, while service and cost limits require a decision to adopt a route or stop.
Assurance vs autonomyA named high-consequence use may need a bounded assurance result, while ordinary planning should not inherit assurance apparatus.
Description vs enactmentA callable route description helps planning, but it is not the Method, plan, call, or evidence of performance.

C.24:4 - Solution

C.24:4.0 - Local objects and boundaries

  • ATC.CallRouteDescription is a U.MethodDescription whose EntityOfConcern is the selected admitted Method and whose claims explain how that Method is carried out through the route, under A.3.2. When it carries vendor-local route data, it states the vendor or source scheme, exact scheme or API edition, intended use, and selected Method ref before any access details, inputs, outputs, or route limits.
  • ATC.CallPlan is a U.WorkPlan for intended calls. Its steps select Methods and may cite route descriptions.
  • ATC.CheckpointReturn is a C.2.1 result episteme stating what was tested, what budget was burned, and what route action is recommended next. It is not the tested Work.
  • ATC.CallGraphRef cites the applicable G.6 trace representation over actual call Work. The representation records or points to facts; it creates none of them.

actionBasis contains exactly one of three branches:

  • domainPrescribedAction identifies the applicable domain Method, the exact prescription and its required action, applicability and stop conditions. An optional intended WorkPlan may carry that intention; it establishes neither authority nor permission. No decision wrapper or ongoing Work is required for this branch.
  • situationResponsiveDecisionEpistemeRef refers to an episteme identified under C.2.1 because this plan relies on an A.15.7 decision; the episteme states the selected action, deciding System, intended performer, action-changing fact, relevant Method limit, and stop or feedback condition.
  • fixedOptionChoiceResultRef refers to a C.11 ChoiceResult whose result is choose now.

Consuming either decision result here preserves its existing result kind.

There is no catch-all ATC.PolicyRef. When a constraint branch is current, cite its object: C.19 PoolPolicyResult or EmitterPolicy, a C.19.1 probe, comparison, local-policy, or waiver result, or a domain constraint whose kind and defining pattern are named. The plan retains applicable time, compute, cost and risk ceilings and its stop or replan conditions.

C.24:4.1 - Owned planning operations

C.24 owns only planning and replanning:

planCalls(
  actionBasis,
  objective,
  admittedMethodRefs,
  routeDescriptionRefs?,
  budget
) -> CallPlan

revisePlan(
  currentCallPlanRef,
  checkpointOrSignalRefs,
  residualBudget
) -> CallPlan | CheckpointReturn | neighborExit

A.3.1 supplies Method admission. The action basis fixes the action or option being planned; it does not admit the Methods chosen for plan steps.

A domain-prescribed basis carries its Method, prescription, required action, applicability and stop conditions. Permission and authority remain independently governed. An A.15.7 basis keeps the selected action, deciding System, intended performer, action-changing fact, relevant domain-Method limit, and stop or feedback condition. A C.11 basis is a ChoiceResult whose lawful result is choose now; probe again, reject current set, and reroute do not fix an action for C.24.

C.18 may supply generated candidate or front material, and C.19 may supply a live-pool treatment that informed the decision; neither record admits a Method. Comparison comes from the selected evaluation Method and, when scale preference is claimed, C.19.1.

A.15.1 governs the actual call or observation Work; G.6 provides a trace representation when those independently established facts, results and provenance relations must remain replayable. C.24 only constrains what the plan or checkpoint must retain for those later uses.

To construct the route, use the backward construction and forward check in OPS.11.1:4.2–4.3 for this fixed action:

  1. Recover usable contributions. Read the available Methods and any needed route descriptions for the results they supply, required inputs, preconditions, relevant effects and limits. Resolve missing interface meaning from the maintained description or provider. Begin with the action’s receiving completion conditions.
  2. Work backwards from the required result. Choose an admitted Method that can supply it. For each input not already available, find a qualified preceding Method or an available source that supplies it. Continue until the route reaches available inputs. Keep required joins and enabling conditions. If a contribution cannot be supplied, return that particular gap rather than adding an ungrounded call.
  3. Check forwards from the available inputs. Follow the proposed outputs into the next inputs. Check consequential units, scope, access, timing and other use conditions. Reconcile shared resources and applicable budget. Elapsed time is not the sum of resource expenditure when calls overlap. Reuse established interface and performance bounds where adequate. Compare the final output with the result promised to the recipient. If a call only prepares the input for delivery or acceptance, include the work still needed and its time and entry conditions in the plan.
  4. Return the route or the unresolved question. A supported route states the calls, intermediate results, order or safe parallelism, resource assumptions and stop or replan conditions needed by the use. If an unknown condition could change feasibility or route choice, identify it and what information could settle it. Section 4.2 selects a bounded scout only when such a scout can usefully answer that question. A changed action returns to its selecting Method.

C.24:4.2 - Bounded scout or probe cycle

When the action remains fixed but an unresolved route condition could change the continuation, first use adequate interface documentation, existing observations or provider information. Select a bounded scout only when its attainable answer can change the warranted continuation enough to justify its burden, within the applicable budget and permission. State what observation would retain, revise or reject the route before planning that scout. Its actual performance can support a CheckpointReturn; an unfamiliar route alone does not require probing.

  • If another probe could still change which option survives the OptionSet, the budget remains a C.11 probe budget and planning returns there.
  • If changed live facts or domain-Method limits could change an A.15.7 action, return there instead.
  • If a prescribed action loses its applicable prescription, return to the domain Method before further route planning.
  • If the action or option remains fixed and only route shape or rollout order is uncertain, the probe uses enactment budget and its checkpoint belongs here.

commitTrigger states the condition for considering a checkpoint decision to end route probing and adopt an identified CallPlan for the fixed action. Name the proposed plan and the work it schedules in recommendedNextAction or in the accompanying explanation. When the trigger holds, the responsible planner considers that transition using the probe evidence, residual budget, and any required safety or assurance result; the planner may adopt the plan, require a further bounded probe, or stop. Satisfying the trigger alone does not make that decision.

Adopting the plan settles which route to use. Permission to perform its calls remains subject to the applicable work-entry conditions, and actual performance is separate U.Work. Where an autonomy budget governs the route, use the checkpoint decision and guard or override path in E.16-S1.A; the trigger grants no wider budget or scope.

C.24:4.3 - Planning laws

ATC-1 — Plan the call, not the app. A plan step selects a Method. A route description, endpoint, service promise, trace row, or response does not become that Method or an actual call.

ATC-2 — Use the C.19.1 branch that applies. A bounded call plan with no scale/generalization claim or declared generality policy can proceed under its task, budget and guards. When that claim or policy is current, start with C.19.1’s scale-claim probe. Consume its first result: no scale claim yet, local analogy or policy, bounded scale comparison, or full Scale-Audit selected. Only the latter two open comparison or audit work.

A completed comparison of usable responses over the receiving budget range, with uncertainty and applicable floors, may then warrant a bounded preference or no scale-based preference. A larger slope alone is insufficient; a generality tie-break needs its separate policy basis.

Keep a BLP-waiver separate: it is used only when a declared generality preference would otherwise decide the use, and it records rationale, the admitted review System, the direct waiver-review responsibility or missing governor, and expiry or review.

If comparable evidence is absent, stop the empirical preference; do not invent a slope vector or treat a waiver as evidence.

ATC-3 — Make budgets and harm limits visible. A CallPlan states its applicable planned ceilings. A CheckpointReturn or Work-side record states actual burn. The admitted System stops or replans when a named ceiling or safety condition is breached.

ATC-4 — Keep live-pool exploration declared. Cite a C.19 PoolPolicyResult only while treatment of that still-live pool constrains this plan. Cite its exact EmitterPolicy only when the plan uses that profile; then record explore_share, including 0 when the current profile explicitly plans none. Do not fabricate either ref after the fixed action or option has made pool treatment irrelevant, and do not silently turn illumination or novelty telemetry into a decision criterion.

ATC-5 — Preserve the execution basis needed by later use. An actual call claim retains its dated Work, performer, enacted Method, interval and containing-System relation under A.15.1. When later use needs a replayable account, retain the relevant plan ref, actual budget delta, inputs and outputs subject to privacy, route-description edition used, and G.6 trace reference. The plan and trace do not establish the represented facts by themselves.

ATC-6 — Add assurance only for a named use. When a planning or rollout decision depends on assurance, name the target claim and use, then cite the B.3 result with its basis, disposition, limits, and reopen condition. No policy label or confidence level substitutes for that result.

ATC-7 — Bind vendor routes to the selected Method. Vendor-specific tokens belong in an edition-pinned ATC.CallRouteDescription that recovers the vendor or source scheme, exact scheme or API edition, intended use, and selected Method ref. Access details, inputs, outputs, and limits may follow. An arbitrary profile, executable adapter, or F.9 Bridge does not satisfy this binding. When executable adaptation is current, identify the Method, MethodDescription, System, and performed Work through their direct patterns. Cite an F.9 Bridge only when its relation independently obtains between two local meanings.

C.24:4.4 - Policy and comparison branches

Add only the branch that still constrains this plan:

CallPlan optional branch fields:
  poolPolicyResultRef?             # C.19; only while live-pool treatment still matters
  emitterPolicyRef?                # C.19; only when this exact versioned profile is used
  scaleClaimProbeResultRef?        # C.19.1 first result
  scaleComparisonResultRef?        # only when the probe selected a bounded comparison
  scaleAuditResultRef?             # only when the probe selected a full Scale-Audit
  blpLocalPolicyRef?               # C.19.1 local policy or analogy, not empirical evidence
  blpWaiverRef?                    # separate from the comparison result
  explore_share?                   # only with the applicable C.19 branch
  risk_bound
  cost_ceiling
  time_ceiling
  stop_conditions
  tie_breakers?
  comparison_tolerances?
  assuranceResultRef?              # only for a named assurance use

Each comparison tolerance names its characteristic, bearer, scale, evidence basis, and window. Graduation, rollout, or widening uses a condition defined by the cited result or direct domain pattern.

Plan-local ceilings and stop conditions need no policy object. If another domain constraint is current, give its field the result-kind name and cite its defining pattern; do not put it in a catch-all constraint ref.

When a condition relies on assurance, cite the exact B.3 result and its supported scope; no universal assurance level is inherited from C.19.

C.24:4.5 - Causal action-use field

Add the causal field only for a named causal use in which the planned calls are intended to observe, intervene, collect counterfactual-rung evidence, simulate for a causal claim, condition a counterfactual policy, or evaluate a policy causally:

CallPlan.causalActionUseSpec?:
  causalUseQuestionRef: CausalUseQuestionRef
  targetCausalityLadderRung: CausalityLadderRung
  causalUseClaimKind: CausalUseClaimKind
  causalActionPolicyClass?: CausalActionPolicyClass
  causalEvidenceDesignRef?
  causalSupportComponentRefs?
  causalUseSupportResultRef?: CausalUseSupportResultRef
  supportedUse
  unsupportedUse

The field states the planned causal use and any support already consumed. It does not estimate an effect, prove identification, certify fairness, or turn simulation output into realized counterfactual evidence. Use C.28 for those support questions.

C.24:4.6 - Public quick card

An ordinary written plan can supply this content; the schemas above are optional forms. Retain an existing applicable plan when it already states:

  • exactly one action-basis branch: the grounded domain-prescribed action, an A.15.7 decision episteme, or a C.11 choose now ChoiceResult; add the objective and ordered Method refs;
  • route-description refs only when needed, with their source scheme, exact edition, intended use, and selected Method binding;
  • dependencies or safe parallelism only when they change the route;
  • applicable time, compute, cost or risk budgets, with stop and replan conditions;
  • next planned action; and
  • an exact C.19 or C.19.1 result, B.3 assurance result, causal-use result, provenance ref, or named domain-constraint result only when that branch is current.

This is enough for an ordinary plan. Do not fill the heavier branches merely to make the record look complete.

C.24:4.7 - Closure and worked cases

Close as a CallPlan when route order and budgeted enactment are the current question. Close as a CheckpointReturn after a bounded route probe, stating whether to probe again, seek the checkpoint decision to adopt the proposed plan, or stop. Return to the domain prescription, A.15.7 or C.11 when the corresponding action basis reopens; return to the applicable neighboring pattern when pool treatment, selector declaration, readiness, execution, or publication becomes the current question.

Domain-prescribed action before Work.

In a constructed catalogue-maintenance case, CatalogueMethod_3 requires rebuilding a digest when the catalogue’s source checksum changes. That condition is established; no choice of action remains and the rebuild has not begun. Plan the two admitted Methods directly:

CallPlan:
  actionBasis:
    domainPrescribedAction:
      methodRef = CatalogueMethod_3
      prescriptionRef = catalogue_method_v3#changed-source-digest
      prescribedAction = rebuild_digest
      applicabilityAndStop = changed_source_checksum; stop_if_input_version_changes
  objective = digest_for_the_fixed_catalogue_version
  plannedCallsInOrder = [ReadCatalogueMethod_2, BuildDigestMethod_4]
  plannedBudgetEnvelope = {time<=10_minutes, cost<=catalogue_budget, risk<=read_and_derived_file_only}
  stopOrReplan = input_version_changes_or_budget_is_exceeded
  nextPlannedAction = confirm_required_entry_permission_then_run_read

The prescription fixes the action, while work-entry permission remains separate. If a permitted route probe is needed, its CheckpointReturn retains this same domainPrescribedAction branch and states the tested Methods, findings, spent/residual budget and next action. A changed prescription or input condition reopens the domain basis. Neither form invents an A.15.7 decision or completed Work.

A.15.7 decision into a known route.

During ongoing repository-repair Work, changed source facts make produce_patch_and_verify the next action under the current repair Method. Using the steering Method in A.15.7, the responsible maintainer makes that decision. The retained decision names the repair agent as intended performer, the changed-source fact, and test failure as the stop and feedback condition. Because the call plan relies on the decision later, the team retains it in one episteme identified under C.2.1. The episteme describes the situation-responsive decision; it is not a C.11 ChoiceResult.

CallPlan:
  actionBasis:
    situationResponsiveDecisionEpistemeRef = patch_action_decision_17
  objective = produce_patch_and_verify
  plannedCallsInOrder =
    - methodRef = InspectRepositoryMethod_4
      methodDescriptionRef = inspect_repo_route_v3
    - methodRef = EditCandidateMethod_2
      methodDescriptionRef = edit_candidate_route_v2
    - methodRef = TargetedTestMethod_7
      methodDescriptionRef = targeted_tests_route_v5
  plannedBudgetEnvelope = {time<=45_minutes, compute<=x2, cost<=y2, risk<=r2}
  stopOrReplan = targeted_tests_fail_twice
  nextPlannedAction = enact_now

An actual call needs its independent A.15.1 grounds; add assignment-bound attribution and a trace representation when the receiving use consumes them.

Constructing a route from a fixed result.

The fixed action is to deliver a weekly report and obtain the recipient’s acceptance. The maintained interface specifies the timestamp unit, the recipient’s week boundary and time zone, and the required report format. Four admitted Methods extract the rows, convert their timestamps, aggregate by week, and render a conforming report. For the specified input size and one sequential performer, they take at most 2, 3, 4 and 1 minutes respectively. The team has 12 minutes.

Work backwards from acceptance. The four calls prepare the report in at most 10 minutes. Their outputs establish that it can be ready to send. The recipient also provides an admitted submission Method. Under its stated conditions it delivers a conforming report and returns the recipient’s acceptance within 2 minutes. The intended performer has access, the receiver is available for the full interval, and the prepared report meets its entry conditions. The five-call plan therefore fits the 12-minute budget. The time bound warrants adopting the plan under the stated conditions. To close the actual job by its deadline, establish when this recipient accepted this report. Acceptance after the deadline means the commissioned job finished late. Delivery itself may have been on time. An undated acceptance can establish receipt and acceptance while leaving timeliness open.

If the export already supplies compatible timestamps, the remaining calls fit in 9 minutes. If the submission service promises only to queue the file within 2 minutes, the team can promise timely submission on that basis. A deadline for recipient acceptance still depends on what happens after queuing.

If the timestamp unit or week boundary is missing, recover that interface or receiving condition from its authoritative source. The magnitude of sample numbers does not establish a unit. A changed report requirement returns to the action basis before replanning.

Unfamiliar route.

In this constructed repair case, ci_route_choice_09 fixes the action: produce a candidate patch and run the targeted tests. The proposed ci_patch_plan_09 orders EditCandidateMethod_2 and TargetedTestMethod_7, has an upper duration bound of 25 minutes, and stops for replanning if the reproduction changes or the tests fail twice. Adoption here requires a supported guarantee that the repair fits the enactment time left after route probing.

Before adopting that plan, the maintainer probes whether the failure can be reproduced reliably enough to guide the patch. For this case, MinimalReproductionMethod_5 requires the same failure in two consecutive runs on the fixed repository revision, with the trace and reproduction evidence collected on that revision. One such run has completed. The enactment budget covers both this route probing and the planned repair.

CheckpointReturn:
  actionBasis:
    fixedOptionChoiceResultRef = ci_route_choice_09
  objectiveOrTaskFamily = unfamiliar_ci_failure
  testedMethodRefs = [LogTraceMethod_2, MinimalReproductionMethod_5]
  evidenceRefs = [trace_result_1, reproduction_result_1]
  burnedBudget = 15_minutes
  residualBudget = 30_minutes
  recommendedNextAction = repeat_minimal_reproduction_then_return_for_checkpoint_decision_on_ci_patch_plan_09
  commitTrigger = reproduction_is_stable_and_required_evidence_is_current

If the repeat establishes the trigger, the maintainer decides whether to end route probing and adopt ci_patch_plan_09: produce the patch, then run its targeted tests. The maintainer compares the time remaining after that repeat with the plan’s 25-minute upper duration bound and checks the applicable safety or assurance conditions.

A 3-minute repeat leaves 27 minutes, which covers that bound. A 6-minute repeat leaves 24 minutes, so the current bound does not establish the required guarantee. A tighter justified bound for the same route, a qualified alternative route or an independently permitted budget revision may establish fit; otherwise stop.

If the plan is adopted, the repair agent may start its first call only with the required work-entry permission. The checkpoint itself records neither that permission nor a performed call.

Two vendor routes with one token.

Vendor A and Vendor B both publish a route called search. vendor_a_search_v2 states scheme VendorA API, edition 2026-07, intended use repository text search, and selected Method RepositoryTextSearchMethod_3. vendor_b_search_v5 states scheme VendorB agent tools, edition 2026-08, intended use web source retrieval, and selected Method WebSourceRetrievalMethod_8. The shared token identifies neither binding; the description fields do. An executable adapter, if used, remains distinct from the Method it implements, and its execution remains separate Work.

Scale comparison, when current.

The cheap C.19.1 probe for BatchSearchMethod_3 and IndexedSearchMethod_6 returns bounded scale comparison for the same repository-search task and 10k–100k files window. The comparison then uses elapsed time and missed-match rate from repo_search_benchmark_12, including uncertainty and cost limits, and warrants a preference for IndexedSearchMethod_6 only inside that window.

If the evidence for one Method covers only small text files and the evidence for the other covers only large mixed repositories, the task difference leaves their relative performance on the receiving workload open. Return no scale-based preference. A tie-break written for established equivalent performance does not apply to that finding. The project can use a local policy that covers the actual case or seek a waiver when overriding a preference that would otherwise decide the use.

Near misses. A route label with no recovered Method remains probe material. A plan with no Work is still intent. A trace row does not prove performer, assignment, Method, or service acceptance. A satisfied commitTrigger with no checkpoint decision leaves the proposed route unadopted.

Transfer examples. The same result shape works for research assistance, program repair, and lab automation. The Methods and safety conditions differ; the plan/checkpoint boundary does not.

C.24:6 - Bias-Annotation

Keep notation and vendors out of the conceptual contract. Do not average unlike scales. Do not let a route description, plan, trace, response, or confidence label stand in for a Method, performed Work, evidence result, or assurance result.

C.24:7 - Conformance Checklist

  1. Every result states exactly one grounded action basis: an applicable domain-prescribed action with its Method, prescription and conditions; an A.15.7 decision episteme; or a C.11 choose-now ChoiceResult. A WorkPlan alone supplies no permission or authority, and a prescribed action needs no decision wrapper or ongoing Work.
  2. Every planned step names an A.3.1-admitted Method that realizes or supports the fixed action. The action basis fixes the action or option; selecting it does not establish Method identity. Route-description refs remain separate and optional.
  3. The plan records its applicable time, compute, cost or risk ceilings and stop or replan conditions.
  4. C.18 candidates or front material and C.19 live-pool treatment may inform the action basis but do not admit Methods; C.24 owns only planning and replanning results.
  5. A scale branch first cites one C.19.1 probe result, then any selected comparison or Scale-Audit result; a BLP-waiver remains separate from evidence.
  6. Every vendor-bound ATC.CallRouteDescription identifies source scheme, exact edition, intended use, and selected Method; an arbitrary profile, adapter, or Bridge cannot substitute.
  7. Each current policy or constraint ref resolves to its C.19, C.19.1, B.3, or domain-defined object; a plan with no such branch remains valid.
  8. A CheckpointReturn states tested Methods, evidence, burned and residual budget, next action, and commit trigger.
  9. Actual call claims retain their A.15.1 Work grounds. A receiving use that needs replay also retains the relevant plan, budget delta, source editions and G.6 trace refs.
  10. A causal action-use branch uses the current C.28 question, support-component, and support-result contract and grants no downstream authority.
  11. The ordinary quick path remains readable without ontology or assurance apparatus.

C.24:8 - Common Anti-Patterns and How to Avoid Them

  • Planning the whole tool lifecycle. Keep candidate generation, selection, execution, scoring, and publication outside C.24.
  • Route description as Method. Recover the Method or keep the route in probe state.
  • Plan as execution. Put actual burn and call facts in Work-side results and the trace.
  • BLP slogan as comparison. Use C.19.1’s probe and any selected comparison; keep a waiver separate or return no scale claim or no scale-based preference.
  • Catch-all policy or profile ref. Cite the PoolPolicyResult, EmitterPolicy, C.19.1 result, B.3 result, or domain-defined constraint, or omit the branch.
  • Confidence threshold as assurance. Use a direct condition and cite B.3 only for a named assurance use.
  • Executable adaptation by implication. Store the binding in a route description; identify any executable adaptation independently.
  • Successful probe as commitment. Keep the trigger, decision to adopt the proposed plan, and permission to perform its calls distinct; satisfying the trigger supplies no missing decision or permission.

C.24:9 - Consequences

Tool use becomes inspectable before execution: the result shows which applicable prescription or accepted decision fixed the action or option, which Methods are planned, what budget is reserved, and what changes the route. Identical vendor tokens remain distinguishable by source scheme, edition, intended use, and selected Method. The cost is explicit Method grounding and branch-specific constraint refs. Heavy assurance, causal, or scale-comparison records appear only when their use justifies them.

C.24:10 - Rationale and current practice

Qualification window. This comparison uses the source set as of 2026-08-21. Reopen it when a later result changes the relative value of explicit planning, route grounding, active information gathering, checkpoint use and replanning, multidimensional evaluation, or long-horizon budget and dependency handling for the declared use.

ContributionAdopted, adapted, or rejected moveBoundary and trade-off
ToolPlanner, EMNLP 2024, ToolPlanner: A Tool Augmented LLM for Multi Granularity Instructions with Path Planning and FeedbackAdopt: keep path planning, feedback, and replanning explicit instead of hiding them inside one call loop.The gain is replayable route revision; the cost is a plan object. The LLM benchmark does not define universal FPF objects.
PlanningArena, ACL 2025, PlanningArena: A Modular Benchmark for Multidimensional Evaluation of Planning and Tool LearningAdapt: check tool selection, reasoning, user-input interpretation, and execution-relevant constraints separately instead of treating one aggregate score as plan quality.Its scenarios do not set universal weights or safety limits. C.24 keeps only the dimensions that change this plan or checkpoint.
IBM Research, ECAI 2025, From Grounding to Planning: Benchmarking Bottlenecks in Web AgentsRetain: keep route grounding distinct from plan quality, but reject the claim that planning is always the dominant bottleneck.This preserves a cheap diagnostic split without hard-coding a web-agent bottleneck order.
Aghzal et al., 2026 preprint, Why Do LLM-based Web Agents Fail? A Hierarchical Planning PerspectiveAdopt: separate high-level planning, low-level execution, and replanning; a sound plan does not excuse failed grounding or adaptive control.The result makes the IBM split conditional. It does not make every web-agent layer mandatory in a known fixed route.
DeepPlanning, ACL 2026, DeepPlanning: Benchmarking Long-Horizon Agentic Planning with Verifiable ConstraintsAdapt: retain global budgets, dependencies, or safe parallelism when material, and use the bounded scout and checkpoint cycle for information needed before commitment.Long-horizon benchmarks expose degradation and efficiency trade-offs, but their task schemas do not belong in every ordinary call plan.
C.19.1 current scale-comparison sources and methodAdopt conditionally: use Bitter-Lesson pressure only with the probe and a named bearer, scale window, evidence, cost, safety, and uncertainty.Generality is not a winner by label; local policy and waiver stay separate from empirical comparison.

These contributions distinguish route grounding, plan quality, execution and replanning. Keep only fields that change the route, stop, reliance or replay. Reconsider the plan when new knowledge changes one of those decisions; another benchmark alone need not change an adequate plan.

C.24:12 - Relations

  • A.3.1 supplies admitted Method identity; neither decision branch admits a Method merely by selecting an action.
  • The steering Method in A.15.7 is used to reach the situation-responsive decision cited through situationResponsiveDecisionEpistemeRef; when this plan relies on the decision, its episteme has the identity conditions defined in C.2.1.
  • A C.11 ChoiceResult whose result is choose now is cited through fixedOptionChoiceResultRef.
  • C.18 supplies generated candidate or front material, and C.19 supplies PoolPolicyResult or EmitterPolicy only when live-pool treatment still constrains the plan. Neither admits a Method.
  • C.19.1 supplies the scale-claim probe, any selected comparison or Scale-Audit result, and any separate local policy or BLP-waiver.
  • A.15, A.15.1, A.15.2, A.2.1, and F.6 keep Method, description, plan, Work, performer, and attribution distinct.
  • G.6 supplies the trace representation cited by ATC.CallGraphRef.
  • B.3 supplies one bounded assurance result only when a named assurance use is current.
  • C.28 supplies causal-use support when the plan is used for causal evidence, intervention, policy, fairness, or counterfactual work.
  • C.27 evaluates temporal claims about speed, narrowing, recovery, or stop/replan rate. More calls or faster narrowing is not success by itself.
  • E.23 may use C.24 plans and checkpoints inside improvement Work; C.24 does not restate the improvement loop.
  • E.10.MOVE, E.11.PUR, and A.15.5 recover project moves, pattern-use recommendations, and work-entry readiness when those questions are not plan-local.

C.24:End

Referenced in the corpus

81 literal mentions in other sections. Read their context to establish the relation.