# Repair + Continue: keep completed agent work This is an existing runtime capability, not a new graph primitive. A resume reuses recorded answers for a matching completed prefix and executes the repaired source from the first fresh call onward. It is not deterministic LLM generation, a JavaScript stack snapshot, or a general per-node cache. ## Author's procedure Read the stopped run diagnostic and the evidence it names: child result, transcript or answer; use the run journal when no child pointer is proven. Find the project-owned source and the original goal. Preserve the same workflow target and workspace contract. Change the failing prompt or contract and the required suffix; keep unaffected labels, prompts, order and phases unchanged. Do not rename every node while fixing one of them. Use stable business keys where the graph already has them, without claiming that keys bypass the prefix rule. Validate the exact repaired source with the supported checker and Node syntax/source checks, without importing unchecked source. Report what changed and which prefix is expected to remain reusable. Hand the source path, diagnostics and original run id to the [run skill](../../locus-pi-workflow-run/SKILL.md). Repair-only stops at this handoff; authorized Repair + Continue resumes now through that skill, without a new approval ritual. Preserve original attempts and check the new terminal evidence; checking source alone does not complete continuation. ## Current limitations, by name ### Legitimate quality refusal A correct quality refusal can identify a product defect, not a broken workflow. Preserve its diagnostics and acceptance criteria; a generic "fix the workflow" hint is not a reason to weaken the gate. Existing authorization for scoped implementation covers repairing that feature; a refusal does not create a new owner-approval requirement. New product scope or separately unauthorized external effects retain their actual authorization boundary. Give an explicit, bounded correction stage to a named tool-capable owner with the full findings, diagnostic/probe evidence, current tree and original baseline. Preserve only the valid completed prefix before that correction. Follow it with fresh independent review and the original read-only gate, passing the correction report and fresh review as current evidence. Do not replay an old blocked choice or treat pre-correction review as acceptance of changed bytes. If prerequisites drifted, move the fresh boundary earlier instead of claiming invalid reuse. Validate the repaired graph, then resume through the run skill under the existing authorization. A further refusal stays a refusal with fresh diagnostics. This explicit continuation is not a hidden retry or an unbounded automatic loop. The read-only gate records findings; it does not become the correction owner. ### Missing tools in an external child Inspect the actual child launch arguments and tool schema before changing the workflow. Pi `tools: ["*"]` cannot undo an adapter's restricted CLI tools. Repair the owning adapter or profile under existing authorization, then prove shell/git, read and temporary report writing through that same transport before retrying the review. Do not spend more review rounds on an unchanged technical obstacle. A reviewer may repair technical prerequisites without editing reviewed product source; product corrections still need fresh independent review. ### Provider admission failures Read the provider error and actual executed route before changing the workflow. HTTP 402 with `limit_source: openrouter_key_limit` is a provider key allowance failure, not a handoff-length error, an implementation verdict or missing repeat owner approval. Raising `maxTurns` cannot repair it, and there is no item-length option left to raise. A requested `max_tokens` may come from SDK model metadata rather than the workflow prompt. First check whether that provider/transport was intended. Apply an already authorized route correction only to the required unfinished work and verify the completed prefix remains reusable. If the API route is still intended, report the real key-limit/reset prerequisite; do not automatically change financial limits, switch keys, invent a token cap or retry the same rejected request. Neither a valid return shape nor a zero process exit converts admission failure into acceptance. Preserve the failed result and require fresh terminal evidence after the cause is resolved. ### Unconfirmed call after process loss An absent process with no child completion or terminal result is an unconfirmed call. Preserve its journal, replay, transcript and partial artifacts. Read the canonical [reconciliation path](../../../docs/workflows/recovery-and-continuation.md#reconcile-an-unconfirmed-call) before editing source: it requires assessing current effects and any reusable prerequisites. Direct interrupted recovery does not admit an unfinished child. Review the whole stage from its recorded base through committed and uncommitted changes. Commit only under the task's existing authorization. Do not create a baseline archive or content manifest merely to reconstruct the diff. When a verified terminal ancestor remains usable, keep its completed prefix and give the first fresh stage the accepted goal, current tree, preserved work and reconciliation findings. Its job is to retain valid work, finish or repair what is missing and produce fresh verification before ordinary review and QA. Do not blindly rerun an effect, reset the tree to the ancestor, or accept a previous child's self-report. Changed prerequisites may require an earlier fresh stage. Use [external-locus-pi](../../external-locus-pi/SKILL.md) for an inspectable, retained Pi session for long execution; create-only still stops before launch. ### Read-only review evidence The reviewer inspects the full stage diff with its own tools, including any uncommitted changes, and accounts for every in-scope path — additions, deletions, renames, tests, docs and diagrams — before a favorable verdict. Current-source sampling is not complete change review, and a favorable verdict is invalid while required coverage is missing. The producer does not rebuild the change as a readable bundle: no evidence entrypoint, locator list, content manifest or tree hash. If the reviewer cannot read the diff, fix its tools or the stage boundary, not the handoff format. After a repair, rerun the incomplete review and its dependent suffix through ordinary resume, preserving the unaffected completed prefix. ### Unfinished implementation Read the failed child's transcript and verification evidence before classifying the stop. If an environment problem was repaired within the accepted scope, reuse the prepared environment after checking it is still suitable and continue the assigned implementation. Preparation and passing baseline tests do not prove the requested behavior was implemented. Do not terminate a child voluntarily just because work remains while an authorized next step is available. A real blocker names the remaining obstacle, its evidence and why an in-scope repair cannot resolve it. An observed budget, provider or resource stop retains its exact machine reason and completed work; do not invent an owner-approval gate. Preserve real authorization, scope and safety boundaries. For example, missing dependencies fixed and baseline green, but implementation absent means continue implementation; required access still denied means report that obstacle; an actual turn-budget failure follows the section below. Keep acceptance checks: a valid return shape or `[]` is not proof of completion. Repair the unfinished source through the normal continuation route, preserving completed calls. Do not add an unbounded retry or turn incomplete work into success. ### Repeated output-contract failures Inspect the entire unfinished suffix, including reachable saved child workflows, for the same defect before continuing. For each occurrence, check the actual downstream consumer: ordinary reports stay plain text, and their next semantic reviewer receives the complete text. Keep existing checks for completeness, verdict and blockers before subsequent writes. Preserve genuine choices that route the graph. A source that still declares a removed shaped-result option fails before any child with that option's name; migrate it to text, a choice, or an exact caller-assigned file. Do not replace content review with text-length checks. Keep completed calls' prompts and effective options unchanged, and verify their replay identity. Do not redesign the graph or sweep completed stages as part of this repair. A raw-value example is a correction aid, not proof that the next model will submit a declared member; strict validation remains required. A large answer is never the cause. The runtime has no answer-size policy, so a stopped choice call means one of three named things: the value still missed the declared members after its clarification turn (`[workflow:return]` journals how many), the transport could not host `workflow_return` (`output-contract-unavailable`: run it on a host that can, or make the call plain text — not a script bug), or an explicit budget stopped the run. Do not shorten the answer, ask for a character limit in the prompt, or add a size option the runtime refuses by name. ### Turn-budget failures `Child exceeded its cumulative ... assistant-turn budget` counts SDK model cycles (`turn_start`) during the whole child session, including ordinary tool use before its first result. It does not mean the workflow restarted or the agent attempted that many `workflow_return` calls. Read the failing child's transcript and journal to distinguish useful work from format repair or a loop. For legitimate work, raise the explicit `maxTurns` on the unfinished suffix with a one-line reason. Keep other budgets and the original completion criteria. There is no package default: an undeclared `maxTurns` is unbounded. A completed call keeps the value its record was written with (a prefix recorded with `maxTurns: 20` keeps `maxTurns: 20`); changing every call to the new allowance would change their request keys and end reuse. Verify actual prefix reuse before resuming. The [runtime reference](../../../docs/workflows/index.md) owns the axis list and timer constraints. Ordinary narrative still uses plain text; there is no character limit to exceed, and increasing turns does not fix a wrong result mode. Use `choice` for a routing decision and an exact caller-assigned file for discovered work units, including sequential slice queues. ## Runtime handoff The [run recovery procedure](../../locus-pi-workflow-run/references/recovery.md) owns launch admission, original input/workspace binding, operator answers and proof of actual prefix reuse. Do not copy those runtime rules into a repaired workflow. Saved answers restore no files or tool effects; retain suitable prerequisites and require fresh verification before irreversible effects. A larger legitimate work list does not authorize an invented call cap or an automatic budget increase.