pi-ext bugabingas pi extensions
owned by admin
README Source History Refs Compare Notes Search extensions/pipi/skills/pipi-workflows/SKILL.md Raw Rendered preview
name: pipi-workflows
description: "Use before delegating work to helpers, making PiPi calls, managing reusable agents, or continuing child sessions."
PiPi workflows
Choose
Shortlist catalog roles; read their config.json ∧ optional instructions.md before accepting/rejecting; match purpose/authority, not labels; use agent or state no-fit reason in handoff.
Delegate when independent work/context beats handoff, coordination ∧ rework cost.
Pick least-cost adequate model; copy explicit IDs from current pi --list-models, never guess.
Assign and coordinate
Brief: goal, decisive context, authority, file ownership, canonical refs, bounded result/evidence; inline execution-critical detail, reference rest.
Select role tools/context without stripping applicable instructions; inspect effective config ∧ diagnostics before trusting capabilities.
Unused unregistered preset flags → omit from call-local flags (flags: {} if none needed), not false; never drop required capability/permission flags.
Parallelize independent work; serialize prerequisites/shared mutations.
Continue coherent work; inspect surviving artifacts before respawning; fresh sessions for independent judgments.
Handoff files only when size/reuse justifies creation/reads; preserve durable evidence.
Sustained work → discover ∧ maintain one progress view of observed state, blockers, decisions.
Classifiers advise bounded next decisions, never prove; multi-model judgment → pipi-round-table; evidence > votes.
Receive and improve
Keep native result r privately; in codemode print only this projection, never whole messages or assistantMessage fallback:
store(`pipi:${r.sessionId}`, r);
const message = r.outputMode === "message" ? r.output : r.assistantMessage;
const answer = r.outputMode === "message"
? (message?.content ?? []).filter(b => b.type === "text").map(b => b.text).join("\n")
: r.output;
text({ outcome: r.outcome, diagnostics: r.diagnostics, sessionId: r.sessionId, error: message?.errorMessage, answer });
Add relevant usage, effective config, artifact refs; empty/clipped print → load or native locator, not child regeneration.
Completion ≠ proof; before integration verify claimed boundary with smallest parent-side check; preserve public contracts/regression assertions unless behavior change explicitly approved.
Recheck changed/uncertain boundaries, not routine established behavior; report verified outcomes, limitations, authoritative evidence ∧ superseded results.
Keep task overrides local; within write authority improve existing presets for verified reusable lessons, preserve purpose/permissions, avoid near-duplicates.
Verify preset edits on later calls; preset continuation → reselect agent, not assumed persistence.
---
name: pipi-workflows
description: "Use before delegating work to helpers, making PiPi calls, managing reusable agents, or continuing child sessions."
---
# PiPi workflows
## Choose
- Shortlist catalog roles; `read` their `config.json` ∧ optional `instructions.md` before accepting/rejecting; match purpose/authority, not labels; use `agent` or state no-fit reason in handoff.
- Delegate when independent work/context beats handoff, coordination ∧ rework cost.
- Pick least-cost adequate model; copy explicit IDs from current `pi --list-models` , never guess.
## Assign and coordinate
- Brief: goal, decisive context, authority, file ownership, canonical refs, bounded result/evidence; inline execution-critical detail, reference rest.
- Select role tools/context without stripping applicable instructions; inspect effective config ∧ diagnostics before trusting capabilities.
- Unused unregistered preset flags → omit from call-local `flags` ( `flags: {}` if none needed), not `false` ; never drop required capability/permission flags.
- Parallelize independent work; serialize prerequisites/shared mutations.
- Continue coherent work; inspect surviving artifacts before respawning; fresh sessions for independent judgments.
- Handoff files only when size/reuse justifies creation/reads; preserve durable evidence.
- Sustained work → discover ∧ maintain one progress view of observed state, blockers, decisions.
- Classifiers advise bounded next decisions, never prove; multi-model judgment → pipi-round-table; evidence > votes.
## Receive and improve
- Keep native result `r` privately; in codemode print only this projection, never whole messages or `assistantMessage` fallback:
``` js
store(`pipi:${r.sessionId}`, r);
const message = r.outputMode === "message" ? r.output : r.assistantMessage;
const answer = r.outputMode === "message"
? (message?.content ?? []).filter(b => b.type === "text").map(b => b.text).join("\n")
: r.output;
text({ outcome: r.outcome, diagnostics: r.diagnostics, sessionId: r.sessionId, error: message?.errorMessage, answer });
```
- Add relevant usage, effective config, artifact refs; empty/clipped print → `load` or native locator, not child regeneration.
- Completion ≠ proof; before integration verify claimed boundary with smallest parent-side check; preserve public contracts/regression assertions unless behavior change explicitly approved.
- Recheck changed/uncertain boundaries, not routine established behavior; report verified outcomes, limitations, authoritative evidence ∧ superseded results.
- Keep task overrides local; within write authority improve existing presets for verified reusable lessons, preserve purpose/permissions, avoid near-duplicates.
- Verify preset edits on later calls; preset continuation → reselect `agent` , not assumed persistence.