Team Creator

Team Creator

Intro

team-creator derives a provider-neutral AI team by scoring all accessible models on four weighted dimensions (Capability, Cost, Latency, Governance) and mapping the 8 processkit role archetypes onto the resulting heavy / medium / light tiers. It composes 5 existing skills — no new primitives or MCP tools are introduced.

Overview

When to use

Inputs (key parameters)

Tiering formula (summary)

Defaults rebalanced 2026-04-15 after Phase 3 dogfood (see ART-20260415_1545-TeamWeaver-team-creator-dogfood-diff §7). Capability weight raised from 0.40 to 0.60 to prevent tier inversion on same-provider candidate sets where cost and capability are anti-correlated.

TierScore ≥ 0.70 → heavy | 0.40–0.69 → medium | < 0.40 → light

See references/tiering-formula.md for normalisation, thresholds, and a full worked example.

Role archetypes → tier mapping

See references/role-archetypes.md for full override-when rules.

Project-level overrides: if context/team/role-archetypes.yaml exists, it is loaded before archetype mapping and validated eagerly. See references/role-archetypes-override.md for schema and invariants.

Commands

pk-team-create runs the full derivation: queries models, applies the tiering formula, maps archetypes, writes 8 Role + 8 Actor + 8 Binding entities, rewrites context/team/roster.md, and emits a chartering DecisionRecord that explicitly supersedes the prior team DecisionRecord. Use --dry-run to print the plan without writing.

pk-team-review is read-only. Re-scores current assignments against the latest landscape snapshot, diffs tier scores against those stored in the governing DecisionRecord, and surfaces tier-shifts, unavailable models, and new outperformers. No entities are written.

pk-team-rebalance applies a review recommendation. Requires --confirm. For --roles all, the full create logic re-runs and a new DecisionRecord is written. For named roles, the governing DecisionRecord's progress_notes are amended instead.

Outputs

Gotchas

• Missing landscape-summary artifact. This skill ingests a pre-structured markdown artifact tagged landscape-summary, NOT raw HTML. If none exists, error: "No landscape-summary artifact found. Run the Haiku landscape-ingest task to produce ART-*-LandscapeSnapshot before pk-team-create." Do not parse HTML.

• Entity deactivation sequence on re-create. When pk-team-create is re-run: (a) resolve prior team from roster.md; (b) for each prior Binding call binding-management.end_binding with reason="superseded by pk-team-create run <timestamp>"; (c) for each prior Actor whose model is NOT in the new team call actor-profile.deactivate_actor; (d) for prior Actors whose model IS re-assigned to the same role REUSE the existing Actor — do not create a duplicate; (e) then create new Bindings.

• Formula weight persistence. Weights live in the chartering DecisionRecord's spec.inputs_snapshot block. pk-team-rebalance reads them via decision-record.get_decision. No skill-local config file for weights.

• Tier-collapse (< 3 tiers accessible). Promote the two highest- scoring light models to medium. Never fail; degrade gracefully.

• PM clone cap. Parallelism cap never applies to project-manager; PM is always exactly 1 (per DEC-20260414_0900).

• Snapshot staleness. Artifact older than 90 days → warn and proceed. Surface the artifact date in the DecisionRecord.

• Provider-neutrality. Never hardcode any model name, provider name, or tier label anywhere in this skill. All identifiers flow in from model-recommender.

• Override audit trails (OpenWeave layers 1–4). Each override layer produces a distinct audit entry in the chartering team DecisionRecord's inputs_snapshot:

  • Layer 1 (landscape): landscape_artifact_source records "explicit" | "project-tag" | "kit-default" and the resolved ART-id. The project-tagged artifact (ART-*) is itself the audit record for who created the custom landscape.
  • Layer 2 (weights): weights_source + weights block records which DEC--TeamWeights (if any) was applied, or "cli" / "skill-default". The DEC--TeamWeights record carries its own rationale and deciders.
  • Layer 3 (thresholds): tier_thresholds_source + tier_thresholds block; co-located with Layer 2 in the same DEC-*-TeamWeights record (causally coupled — cannot drift).
  • Layer 4 (role pins): archetype_override_file, archetype_override_semantics, and archetype_overrides list each overridden role with its new pin. The context/team/role-archetypes.yaml file itself is the persistent record. Any run with overrides: query the single chartering team DecisionRecord to fully reconstruct the run's configuration.

• Canonical schema fields (processkit v0.16.0). This skill emits five fields introduced in v0.16.0: Role fields primary_contact (bool), clone_cap (int), cap_escalation (string); Actor fields is_template (bool), templated_from (string, nullable). Seed Actors are always is_template: true; rebalance-spawned clones are is_template: false with templated_from pointing at the seed.

Agent-driven discovery

Agents can resolve user intent into the correct override layer using the trigger phrases below. These are additive to the existing team-creator triggers above — they extend the skill-finder entries, not replace them.

Agents MUST consult skill-finder before routing any of these requests. The trigger phrases are designed to match natural-language owner requests without the owner naming the override layer explicitly.

No-skill-inflation rationale

Option C chosen over extending agent-management (heaviest skill; scope bleed) or model-recommender (individual model profiling vs. team orchestration). Pure orchestration consumer: 5 existing skills composed, 3 commands added, 0 new primitives or MCP tools.

Full reference

CLI contract — pk-team-create

pk-team-create
  --subscription <provider>:<tier>     # e.g. anthropic:max-5x
  --providers <list|"any">             # comma-separated; "any" = all accessible
  --parallelism-cap <int>              # max clones per role, default 5
  --governance-floor <0-5>             # G-score floor, default 3
  [--weight-overrides <json>]          # CLI > DEC-*-TeamWeights > defaults
  [--threshold-overrides <json>]       # CLI > DEC-*-TeamWeights > defaults
  [--landscape-artifact <ART-id>]      # explicit landscape; skips discovery
  [--landscape <ART-id>]               # alias for --landscape-artifact
  [--allow-degraded]                   # include models whose availability
                                       # is not "operational"
  [--security-critical]                # promotes developer archetype to heavy
  [--dry-run]                          # print plan; write nothing

Full step-by-step process (8 steps: resolve landscape → query models → score+classify → map archetypes → deactivate prior team → write entities → write roster → write chartering DecisionRecord) lives in commands/pk-team-create.md. Read that file before changing the sequence; the order of (a) eager role-archetypes.yaml validation before mapping, (b) binding-management.end_binding before actor-profile.deactivate_actor, and (c) DecisionRecord written LAST so its ID can be embedded in roster.md is load-bearing.

CLI contract — pk-team-review

pk-team-review
  [--landscape-artifact <ART-id>]      # default: same 3-level resolution
                                       # as pk-team-create
  [--threshold <float>]                # tier-shift sensitivity, default 0.05

Read-only. Re-scores the current team's models against the latest landscape snapshot, diffs each TierScore against the value stored in the governing chartering DecisionRecord's spec.inputs_snapshot.tier_scores block, and prints a tier-shift table to stdout. No entities written; no DecisionRecord amended.

CLI contract — pk-team-rebalance

pk-team-rebalance
  --roles <list|"all">                 # comma-separated archetype names
  --confirm                            # required; no-op without it
  --reason <string>                    # free-text rationale; persisted
  [--landscape-artifact <ART-id>]
  [--weight-overrides <json>]
  [--threshold-overrides <json>]

For --roles all, full pk-team-create logic re-runs and a new chartering DecisionRecord supersedes the current one. For named roles, only the affected Bindings/Actors are rotated and the governing DecisionRecord's progress_notes are amended with the --reason string and the new tier-score for each rotated role.

Skill composition map

team-creator itself ships no MCP server; every write goes via the upstream skills above. This is what keeps provides.primitives: [] and provides.mcp_tools: [] in the frontmatter honest.

OpenWeave 4-layer override model (data sources)

CLI flags ALWAYS win over DEC-stored values, which always win over skill defaults. Eager validation for layer 4 fires BEFORE archetype mapping; layers 2/3 validation is invoked from references/team-weights-decision-schema.md. Validation failures at any layer are hard errors — the run aborts before any entity write.

Tiering formula — canonical definition

norm(x)        = clamp((x - min) / (max - min), 0.01, 1.00)
TierScore(m)   = C·norm(C_raw(m))
               + K·norm(K_raw(m))
               + L·norm(L_raw(m))
               + G·norm(G_raw(m))

C_raw(m)       = (swe_bench_verified + swe_bench_pro + community_rating) / 3
K_raw(m)       = 1 / (output_price_per_1M × quota_burn_vs_flagship)
L_raw(m)       = output_tokens_per_sec
G_raw(m)       = G_score / 5             # normalise 0-5 → 0-1

Defaults       C=0.60  K=0.20  L=0.10  G=0.10   (sum = 1.0 ± 0.001)
Tier cuts      heavy ≥ 0.70   medium ≥ 0.40   light < 0.40

Worked example and edge cases in references/tiering-formula.md.

Role archetype pin table (kit defaults)

Schema fields (primary_contact, clone_cap, cap_escalation, is_template, templated_from) were introduced in processkit v0.16.0 — older entity files predate them. See references/role-archetypes.md for the full responsibilities lists that flow into create_role.responsibilities=[…].

chartering DecisionRecord — inputs_snapshot schema

spec:
  inputs_snapshot:
    subscription: <provider>:<tier>
    governance_floor: <0-5>
    parallelism_cap: <int>
    weights:                  {C: <float>, K: <float>, L: <float>, G: <float>}
    weights_source:           cli | dec-team-weights | skill-default
    tier_thresholds:          {heavy_min: <float>, medium_min: <float>}
    tier_thresholds_source:   cli | dec-team-weights | skill-default
    landscape_artifact:       ART-<id>
    landscape_artifact_date:  <ISO date>
    landscape_artifact_source: explicit | project-tag | kit-default
    tier_scores:              {<model-id>: <float>, ...}
    weight_overrides_applied: <bool>
    archetype_override_file:  present | absent
    archetype_override_semantics: delta | replace | null
    archetype_overrides:      [{role, kit_default_pin, override_pin, rationale}]

pk-team-rebalance and pk-team-review both read this block via decision-record.get_decision. It is the single source of truth for a team's configuration — there is no skill-local config file for weights, thresholds, or pins. inputs_snapshot.tier_scores is what pk-team-review's diff is computed against.

Referenced source files

Extension points

• New role archetype. Add to references/role-archetypes.md, extend the override schema in references/role-archetypes-override.md (PM-immutability invariant must continue to apply), and add a row to the pin table above. The 8-archetype assumption is hard-coded in pk-team-create step 6 — bump it carefully.
• New scoring dimension. Extend the weight set (currently {C, K, L, G}); update weights schema in inputs_snapshot, the worked example in references/tiering-formula.md, and the validation rule that weights sum to 1.0 ± 0.001.
• New override layer. Add a layer-N row to the OpenWeave table, define an audit field in inputs_snapshot, and document the precedence rule (CLI > DEC > defaults) for the new layer.

Corner cases

• Tier collapse (< 3 distinct tiers). Promote the two highest-scoring light models to medium. Never abort — degrade gracefully. Recorded in DEC inputs_snapshot.notes.
• Snapshot staleness > 90 days. Warn and continue; surface the artifact date in the DecisionRecord rather than hiding it.
• PM clone-cap. --parallelism-cap is silently overridden to 1 for project-manager regardless of CLI / archetype overrides (DEC-20260414_0900). Layer-4 overrides cannot raise it.
• Reused Actor identification. The seed Actor for an unchanged role is identified by is_template: true, NOT by name or ID prefix — this is the authoritative test (template clones drift but the seed never does).
• Empty preferred-providers tie-break. Two models within 0.05 TierScore and no preferred provider supplied: pick the higher raw Capability score, then alphabetical model-id as final tie-break.
• Provider neutrality. No model name, provider name, or tier label may be hardcoded anywhere in this skill — all identifiers flow in from model-recommender. Hardcoding is caught by release-audit's skill_structure check on the next release.