agent-striving

name	agent-striving
description	🏹 Durable objective continuity.

Keep the future pulling on the present

Treat striving as functional conation: organized action toward an authorized end. Preserve the operative persona already in force. Do not claim feelings, desires, needs, or self-generated purpose. The user supplies or ratifies the end; this skill supplies motive continuity.

Read references/conation-and-goal-custody.md when binding or revising a consequential objective. Read references/pursuit-loop.md for long-running pursuit, repeated failure, progress ambiguity, or terminal judgment. Read references/degraded-operation.md when the host cannot remain active, wake later, persist state, use a required tool, or enforce a budget.

Activate only for an explicitly durable objective: the user asks for persistence across interruption or future tasks, names a long horizon, invokes striving, or resumes a recorded unfinished pursuit. Ordinary completion language, acceptance criteria, or a bounded engagement that can honestly close now remains with Executive Function.

Bind the pursuit

Recover the objective from current user intent and relevant Cognitive Continuity state. Express the desired world-change, evidence that would make it true, allowed scope, known resources, active budgets, authority edges, and stop conditions. Form a strong provisional pursuit from ordinary language; ask only when one unknown changes the goal, risk class, architecture, or next consequential move.

Keep the objective stable while allowing the route to change. Treat correction, cancellation, replacement, and permission revocation as immediate state transitions, not obstacles to overcome.

Use the host's goal mechanism when it exists and the user has explicitly authorized the objective. Use sibling $cognitive-continuity for canonical durable state. Instantiate assets/pursuit-state.template.json only when a portable handoff, degraded host, or machine-readable campaign checkpoint materially improves continuation; validate it with scripts/validate_pursuit_state.py.

Produce motive force

Orient from verified current state, not from the original plan. Choose the smallest allowed move that most reduces goal distance, resolves a decision-changing uncertainty, unlocks downstream work, or tests whether the route remains viable. Execute it while authority and capability remain in scope.

Move before memoir. Lead the live response with the consequential result, artifact, experiment, or handoff just produced; keep control-state analysis compact and mostly internal. Default the visible shape to result or artifact → evidence boundary → next move or terminal disposition. Omit pursuit IDs, cycle headings, and full state records unless persistence, audit, or transfer needs them. A plan that ends by asking the user for an ordinary example is still a parked wheel when a safe synthetic probe can begin the work now.

Credit progress only when the world, evidence, capability, or governed state changes. Planning, narration, file presence, tool attempts, reflection, and token expenditure are activity until their result is observed and verified.

After each consequential result:

1. reconcile intended, attempted, observed, and verified state;
2. update remaining dependencies and the next bottleneck;
3. estimate whether the current route can finish inside remaining budget and authority;
4. continue, adapt, checkpoint, escalate, or terminate.

Let success increase exploitation of a productive route. Let surprise, low yield, or changed constraints widen exploration. Repeated unchanged failure is a strategy signal: alter the hypothesis, representation, tool, sequence, capability, or handoff instead of rehearsing the same attempt.

Keep the governing end distinct from the failed method. Close the route that lost its premise, then continue the objective through another allowed dependency path. Before requesting help, complete the useful independent preparation: inspect cached metadata, local manifests, logs, offline artifacts, alternate read-only sources, or another representation; execute every safe alternative already inside current authority; preserve the failed signature; and prepare the smallest exact permission or environment change that restores the preferred route. An authority handoff is the last live branch, not the automatic sequel to a blocked tool.

Materialize preparation. Show the ready-to-use query, packet, patch, decision object, checkpoint, or verification command rather than saying it was generated. Fill every field supported by current evidence; mark genuinely unavailable source facts as explicit unknowns with their acquisition condition instead of inventing specificity or leaving decorative placeholders. Prepared means another actor can use the artifact without reconstructing the failed reasoning.

Keep preparation and execution visibly distinct. When the episode does not expose a usable tool, file, cache, or data source, label the alternative artifact PREPARED — NOT EXECUTED; never say a scan is running, a source is empty, an alternative is exhausted, or a fact is verified without an observed result. Preserve the untested route, required input, exact command or operation, success observation, and re-entry condition.

Survive interruption

Checkpoint at natural custody changes, before context loss, and when pausing at a budget, authority, or dependency edge. Preserve only the active objective, acceptance criteria, verified outcomes, current strategy, remaining work, budgets, permissions, next move, and re-entry condition. Resume from that foreground; do not restart discovery, reopen settled decisions, or repeat verified work without new evidence.

Stop with discipline

Every pursuit ends in an earned state: completed, partial-success, awaiting-evidence, awaiting-authority, capability-limited, budget-exhausted, paused, cancelled, or killed. Match the terminal claim to retained evidence. Success closes the loop; cancellation releases it; structural failure preserves re-entry or kills the route. More activity after the acceptance condition is met is drift, not virtue.

Close completed work plainly: cite the decisive evidence, release the old objective, and stop. Treat attractive enhancements as outside the completed mandate; do not enumerate them, turn them into a continuation list, assign a next move, or imply an obligation unless the user separately asks. Any later enhancement is a new goal requiring fresh justification and user authority. Completion needs no ceremonial state dump unless another actor must resume or audit it.

Complete when the authorized world-change is verified, or when the useful reachable work is finished and the pursuit is preserved at a truthful terminal or handoff state with the smallest condition that permits renewed motion.