Two threads in this one. rpiv-pi learns to spend model budget unevenly: which model runs each step, and how hard it reasons, is now configurable per skill, per stage, and per agent. And the workflow runner, which learned to resume a dead run in v1.17, now resumes one that died mid-fan-out or mid-iterate, and can be cancelled between stages. Underneath both, the family standardizes on a single model-key form.

Upgrade notes.

One breaking rename (embedders only). rpiv-workflow’s exported WorkflowContext type is now WorkflowHostContext. If you type host handles against it, update the import. No runtime behavior changed.

Model keys moved to slash form. Persisted keys are now provider/modelId (was provider:modelId). The colon form is still read and auto-migrates on the next save. Caveat: if you roll back across this release, re-run /advisor on the old version first, or the advisor silently disables itself (the older parser is colon-strict).

Two new commands: /rpiv-models to configure model/effort overrides, and /wf @<run-id> to resume a failed run.

Right-size the model (rpiv-pi)

One global model is a blunt instrument: research synthesizes, design decides, commit writes a sentence, and the locator subagents barely reason at all. The new /rpiv-models command and its config file (~/.config/rpiv-pi/models.json) let you set both axes, model and reasoning effort, at four granularities:

Per-skill (skills.<name>): follows a skill everywhere it runs, both as a workflow stage and as a standalone /skill:<name> you type yourself. Standalone overrides arm only on an explicit entry, so your session model stays sovereign otherwise.
Per-stage (stages.<name>): keyed on the workflow graph position.
Per-preset stage (presets.<workflow>.stages.<stage>): one stage of one bundled workflow, resolved ahead of the flat stage rule.
Per-agent (agents.<name>): the research subagents, applied globally.

Most specific wins: preset-stage, then stage, then skill, then defaults.

Reasoning effort gained a first-class off. The vocabulary is now off | minimal | low | medium | high | xhigh, and omitting the field (inherit your session baseline) is distinct from off (disable reasoning); the effort picker offers both as separate choices. /rpiv-models ships full CRUD: a ✓ on every configured entry, per-entry reset, a gated reset-all, and a session-start warning when a key matches no real skill, stage, agent, or workflow.

One operational note: agent overrides live in agent frontmatter, written at sync time. A fresh session picks up your latest agents.* edits on its own, but to apply them mid-session (or to force them over an agent file you have hand-edited) run /rpiv-update-agents, which now re-reads models.json before syncing. The new Right-size the model guide walks the whole thing.

Resume and cancellation (rpiv-workflow)

The audited JSONL trail every run leaves behind is now fully resumable. /wf @<run-id> (or resumeWorkflowByRunId programmatically) replays the trail, rebuilds the accumulated state, and re-enters at the first stage that never completed, including a stage that died mid-fan-out or mid-iterate, where it re-pulls only the unfinished units. A looped stage resumes its trailing generation. Resume guards the one boundary it can check: if a FanoutFn/IterateFn recomputes a different unit list than the run recorded, it refuses rather than run the wrong unit, so a non-deterministic generator fails loudly.

Alongside it:

Cooperative cancellation. Every options bag accepts an optional signal: AbortSignal. The runner checks it at each between-stage seam, records an aborted row for the stage about to run, and returns { success: false }. It does not interrupt a stage already streaming; cancellation lands at the next stage boundary.
Two one-shot helpers. runWorkflowByName and resumeWorkflowByRunId fold load, find, and run (or resume) into a single call. Both return a failure envelope and never throw on a bad name or unresolvable run-id.
Typed STOP. The terminal-edge sentinel is re-exported from the package entry, so authors can write edges: { commit: STOP }; the bare "stop" literal still works.

See Run a workflow for the resume command and the programmatic surface.

One model-key form across the family

modelKey and parseModelKey are consolidated into @juicesharp/rpiv-config so every consumer shares one codec. The canonical form is provider/modelId (slash). Reads accept both separators and prefer slash when both are present; writes emit slash only, so persisted colon-form keys migrate the next time any consumer re-serializes. rpiv-advisor configs auto-migrate on the next /advisor save (mind the rollback caveat above).

Fixes worth calling out

Stale extension context after auto-compaction no longer produces spurious warnings or errors. This was a cross-package issue, fixed in rpiv-pi (guidance injection, git-context injection, model-override lifecycle), rpiv-btw, and rpiv-todo.
Startup no longer crashes with a barrel-initialization race when loading rpiv-workflow.
Workflow model-override lifecycle resets its baseline before restoring at onWorkflowEnd, so a failed restore can no longer leave the override armed and poison the next workflow.
blueprint / design slice-overlap detection now partitions deterministically by file and symbol, cutting verification time further on large plans.

On the site

rpiv-site shipped the new Right-size the model guide, and refreshed Run a workflow, Pick your path, and Compose skills as skills for the resume command, the typed STOP sentinel, the IterateFn/FanoutFn determinism contract, and the corrected polish count.

Grab it

npm install @juicesharp/rpiv-pi@1.18.0

Or let your normal upgrade flow pick it up. Run /rpiv-update-agents once afterward if you configure per-agent models, and migrate any embedder that types against WorkflowContext.

See you at v1.19.0.

Release notes: v1.18.0