v0.6.11

pi-team v0.6.11

This patch release makes model fallback channel-aware without changing the public tool API or requiring configuration migration.

Highlights

• Same model, another channel first — pi-team now treats model.id as the model identity and provider as the access channel. If a worker fails, an untried provider offering the same model ID is preferred before switching to a different model.
• Channel-aware candidate selection — automatic fallback candidates reserve available slots for the primary model ID through other configured providers, so equivalent channels are not displaced by ordinary metadata ranking.
• Bounded fallback remains intact — each fallback key is attempted at most once. After all channels for the requested model are exhausted, retry continues in the original fallback order and terminates when the candidate list is exhausted.
• No provider-wide guess from one failure — a failure from one provider/model pair does not mark every model exposed by that provider as unavailable.

Compatibility

• No tool schema or configuration changes
• No migration required
• Built-in roles and playbooks remain model-neutral
• Compatible with the existing Pi and Oh My Pi extension entry points

Validation

• 318 automated tests across 9 test files
• File-size, model-neutral default-resource, TypeScript, and test gates pass
• Post-reload DeepSeek worker smoke succeeded on deepseek/deepseek-v4-flash
• Runtime fallback smoke completed in two bounded attempts: openai-codex/gpt-5.3-codex-spark returned no assistant output, then openai-codex/gpt-5.4 succeeded
• Published tarball contents remain limited to src/, the bilingual READMEs, LICENSE, and package metadata

Upgrade

Pi

pi install npm:pi-team@0.6.11

Then run /reload.

Oh My Pi

omp install pi-team@0.6.11

Restart or reload Oh My Pi after installation.

v0.6.10 — truthful attention and model-neutral teams

pi-team v0.6.10

This release tightens captain-led teamwork around truthful evidence, explicit communication, and model-neutral routing.

Highlights

• Model-neutral built-ins — the shipped role and playbook catalog now describes capabilities instead of naming preferred vendors or models. A release gate prevents model-specific recommendations, private resources, and non-English runtime prompts from leaking into public defaults.
• Evidence-aware routing — explicit captain choices still win, while real worker failures, hard provider errors, cached health evidence, strict policy, and lazy fallback now interact predictably without silent substitution.
• Communication-based attention — background attention is driven by missing RADIO/ACK communication rather than runtime, token growth, or tool activity. Each episode alerts once; captain observation or intervention opens one new window only for affected workers.
• Queued → delivered → ACK request tracking — captain requests carry stable IDs and per-worker ledgers, including broadcast delivery/ACK aggregation and terminal-without-ACK evidence.
• Truthful active-first status — attention, running, and pending workers appear before terminal workers in compact, expanded, and team_status views. Degraded workers use a distinct glyph/count, unacknowledged delivery is labeled AWAITING_ACK, and no recorded progress delta is reported factually as live:no-delta.
• Less status noise — default team_status output is a short captain control surface. Detailed telemetry remains in structured result data and appears in text only when it is relevant to an anomaly or control decision.
• Stale notification suppression — attention generated while the captain is busy is revalidated against the exact communication/request/cancel episode before delivery. Terminal, observed, superseded, or invalidated notifications are dropped without blocking monitor rearm.
• Lifecycle hardening — attention persistence is serialized and Windows-safe, unfinished queued notifications are released across shutdown/reload, and Pi 0.80-compatible agent_start/agent_end hooks avoid relying on newer-only lifecycle APIs.

Validation

• 316 automated tests across 9 test files
• File-size, model-neutral default-resource, TypeScript, and whitespace gates pass
• Post-reload multi-model acceptance review confirmed attention concurrency, concise status behavior, Pi 0.80 compatibility, and clean source/deployment parity
• Pi and Oh My Pi deployed extension trees matched all 70 source files by SHA-256

Upgrade

Pi

pi install npm:pi-team@0.6.10

Then run /reload.

Oh My Pi

omp install pi-team@0.6.10

Restart or reload Oh My Pi after installation.

v0.6.9 — push-first teamwork and evidence-aware routing

pi-team v0.6.9

This release makes captain-led teams more reliable, observable, and easier to supervise without babysitting them.

Highlights

• Push-first background runs — normal runs now return await_completion_push. Pi follows up when the team finishes, while a runtime attention monitor reports only sustained recorded silence. It never cancels or reroutes workers automatically.
• Evidence-aware model health — synthetic timeout, rate-limit, and provider-error probes are advisory. Recent real worker outcomes carry more routing authority, while confirmed auth and model-not-found failures still block invalid choices.
• Safer lazy fallback — only selected primaries are probed eagerly; fallback models remain lazy and are tried after real execution failure. Cached hard failures are respected before fallback dispatch.
• Model-compatible thinking — requested thinking levels are mapped through model capability metadata instead of model-name rules. Probe sessions omit synthetic thinking defaults.
• Actionable captain decisions — role-specific roleId=provider/model overrides now reject malformed, unknown, or unconfigured choices immediately, accumulate across messages, and preserve strict/task-first timeout behavior.
• Better status evidence — team_status now exposes tool-call and tool-error totals, model-health evidence source, terminal delegation-lane state, and clearer liveness information.
• Runtime hardening — improved cancel target resolution, SOP id validation, worker-attempt accounting, fallback event aggregation, and current-run model-health persistence.

Validation

• 281 automated tests across 8 test files
• TypeScript, file-size, and whitespace gates pass
• Post-reload runtime smoke confirmed: background dispatch returned await_completion_push, completed without polling, delivered a completion follow-up, and closed its delegation lane successfully

Upgrade

Pi

pi install npm:pi-team@0.6.9

Then run /reload.

Oh My Pi

omp install pi-team@0.6.9

Restart or reload Oh My Pi after installation.

v0.6.8

v0.6.8

This release improves Pi and Oh My Pi compatibility, captain observability, and worker control for multi-model team runs.

Highlights

• Added clearer Pi and Oh My Pi installation metadata and README guidance.
• Improved team status and handoff evidence with factual worker previews and compact routing information.
• Persisted worker routing/fallback details into status, handoff, and artifacts.
• Added durable observation tracking to avoid duplicate background completion notifications after a captain has already observed a terminal run.
• Made worker cancellation more tolerant while failing closed on ambiguous keys.
• Clarified that team_message is cooperative steering, not an interrupt.
• Added bounded-evidence worker guidance and search exclusions for .pi/team/ and .omp/team/ run artifacts.
• Improved model routing order: explicit model preferences first, fresh role recommendations before configured defaults, no hardcoded model-name strength rules.
• Added degraded worker/run status handling and tests.

Validation

• npm run check passed.
• 232 tests passed.
• npm pack --dry-run includes 80 files with no dev-only/private artifacts.

v0.6.7 — captain experience & reliability fixes

Captain rules now visible in team() tool description. Workers get bash by default for file discovery. Review/code tasks default to high thinking. Budget-exceeded workers marked degraded and unblock downstream.

Highlights

• Review/code modes now default to thinking:"high" — no more review tasks running at off
• Workers get bash + read + write by default — ls/find/grep before reading, never guess paths
• team() tool description includes mandatory captain rules summary
• New "degraded" worker status: budget-exceeded workers keep partial output and don't block downstream rounds
• Per-model thinking mapping (taskThinkingLevel) for coding and review modes
• 214 tests passing (213 deterministic + 1 flaky B6)

Install

pi install npm:pi-team

Or pin from git:

pi install git:github.com/Kurptas/pi-team@v0.6.7

Then run /reload in Pi.

v0.6.6 — omp compatibility & reliability fixes

Compatibility and reliability patch for pi-team.

pi-team now runs correctly on both Pi and omp (OnMyPi) CLIs, with fixes for 6 bugs found in a cross-model audit.

Highlights

• omp compatibility: team state now writes to the correct config directory (.omp/team/) on omp
• B1: worker artifact write failures no longer kill the whole run
• B2: double-dispose guard prevents cleanup crashes on abort
• B3: degraded models properly included in dead model sets
• B4/B6: runId collision prevention + stale run state TTL cleanup
• 206 tests passing

Install

pi install npm:pi-team

Or pin from git:

pi install git:github.com/Kurptas/pi-team@v0.6.6

Then run /reload in Pi.

v0.6.5 — first public release

First public release of pi-team.

pi-team lets Pi assemble a small squad of AI agents and put them to work on a single task — different models covering each other's blind spots, each one doing what it does best.

Highlights

• Run any task across several models at once: code review, research, financial analysis, debugging, decisions
• The right model on each job — strong models judge, lean ones scan
• Stay in the loop: watch each member live, steer or stop mid-run
• Ready-made team templates, or let Pi design one on the spot
• Per-role tool whitelists keep each worker confined to what it needs

Install

pi install npm:pi-team

Or pin from git: pi install git:github.com/Kurptas/pi-team@v0.6.5

Then run /reload in Pi.