v2.0.5

Patch Changes

• 612c38f: Update the bundled local-duo profile to Qwen3.6 4-bit as the default brain, fold profile-level coms gates into agent-hub/herdr, and show a running strip with messageCount on the Fleet Dashboard. Doctor/setup tests now advertise every model collectModelTargets finds so version bumps do not fail CI.

v2.0.4

Patch Changes

• fc8be3e: Check the .pi/extensions, .pi/harnesses, and scripts npm roots in doctor, and stop Fleet Pi launchers with actionable dependency remediation before extension import failures can surface as misleading unknown flags.
• e4f1957: Make agent-fleet setup --allow-exec include and run the consented runtime dependency steps instead of leaving them as skipped plan actions.

v2.0.3

Patch Changes

• 3cf0e2f: Make native research helpers live-only: they appear in the Fleet Dashboard (Alt+A) while running, never as main-dispatcher cards, and are removed immediately on every terminal outcome. Session, transcript, findings, and /af-agents-history records remain; finished rN restart and research-keep are retired.
• 3cf0e2f: Fix installed workspaces on Node.js 22+ by keeping the doctor model-visibility runtime in plain JavaScript and installing deterministic workflow dependencies under scripts/node_modules.

v2.0.2

Patch Changes

• 7ebeb61: Add multi-model poll, debate, and merge (/af-poll, /af-debate, just flow poll / just flow debate) with .pi/agents/voices.yaml panels, model-visibility doctor findings, and optional integrator --apply behind a cross-process writer lease. Planner drops the risk delegate role in favor of voice-1/voice-2/voice-3 for unresolved architectural forks at tier feature/project.
• ebb740e: Stop copying Agent Fleet product runbooks into target repositories. Keep the workflow guide and Codex conductor contract under the managed .pi/agent-fleet/ runtime tree, and safely retire unchanged files from their former repository-root locations during setup.

v2.0.1

Patch Changes

• 8f36484: Add the headless deterministic workflow runtime, workflow drafting skill and generator, scoped TypeScript authoring gate, executable workflow documentation, and installer/package wiring.
• e007380: Add numbered just flow cleanup and just flow merge branch maintenance, persisted source/result metadata, safe Worktrunk removal, and accepted-run squash merging back to the originating branch.

v2.0.0

Major Changes

• 3492f8d: Remove the manual research lifecycle slash commands /af-research, /af-research-cont, /af-research-rm, /af-research-clear, and /af-agents-cont. Read-only reconnaissance now enters through the dispatcher's bounded spawn_research tool or the automatic specialist NEEDS_RESEARCH: pipe; the existing add/drop, kill, and restart controls remain available.
• 1fae5d7: Consolidate Fleet execution controls under Work Mode. Keep /af-work-mode, Alt+M, and --work-mode; remove /af-posture, --posture, and the deprecated fast/standard/strict mode aliases. Rename the runtime types, state, UI, tests, and documentation while retaining read-only compatibility for legacy persisted posture entries.
• 6793d7e: Remove /af-persona and the persona-gate override. The hub dispatcher is the generated Fleet prompt using the Work Mode selected through /af-work-mode; leftover persona-gate keys are ignored and flagged by doctor. agents/orchestrator.md remains as the catalog description of that role.

Patch Changes

• 196b2ec: Extract Agent Hub configuration and full session lifecycle orchestration so index.ts is a bounded composition root while preserving startup and shutdown order.

  Closeout measurements:

  • wc -l .pi/harnesses/agent-hub/index.ts → 1,850 lines.
  • Unicode character count (len(Path(...).read_text())) → 99,182 characters; plan estimate / 4 → 24,795.5 tokens.
  • Phase 6 production modules remain below 600 lines and extracted functions below 200 lines; registration remains 16 tools / 21 commands / 8 flags.

• d3d6cd1: Extract the 21 agent-hub slash command registrars into dedicated command modules while preserving the existing command surface and behavior.

• 196b2ec: Extract Agent Hub fleet dashboard, detail panel, and context-budget projections into bounded UI modules.

• 65dfe30: Extract Agent Hub dispatch orchestration, grid and overlay UI, and execution-history state into explicit typed modules while preserving behavior and the registered surface.

• 196b2ec: Extract Agent Hub pool presentation, shortcuts, completions, and research controls into explicit modules.

• 196b2ec: Extract Agent Hub roster, model substitution, capability, and work-mode policy into typed policy modules.

• 71ce9ef: Extract Agent Hub system-prompt templates and ordered session-start orchestration into explicit typed modules while preserving prompt text, startup behavior, and the registered surface.

• 196b2ec: Extract Agent Hub research creation, retention, streaming, guarded spawn, and completion delivery into a typed research runtime.

• 196b2ec: Extract Agent Hub mutable state, budget, assertion, artifact, and archive APIs behind explicit root-owned context ports.

• 196b2ec: Extract Agent Hub dispatch and research tool execution orchestration behind narrow typed executor ports.

• 27a0b61: Extract all 16 Agent Hub tool registrations into typed registrar modules while preserving the existing tool schemas and behavior.

• 196b2ec: Extract Agent Hub context-pressure, compaction, deferred-input, and ordered turn lifecycle handlers behind explicit ports.

• 74265fa: Add a registration-surface regression fixture and exclude generated version snapshots from code search results before decomposing the agent-hub entry point.

• 5486704: agent-fleet setup no longer overwrites an existing .ai/agent-fleet-overrides.md.

• 626a7bb: Deduplicate the standalone coms harness and Agent Hub by routing registry, transport, presence, discovery, messaging, and shutdown behavior through one shared coms core.

v1.0.4

Patch Changes

• 7730d44: Add Compact/Verbose Fleet Detail logs with immediate event-driven refresh and a two-second recovery ticker. Local specialists, research helpers, and delegate children now expose wrapped assistant output, provider-emitted thinking, tool arguments, complete textual tool results, status, and duration. Per-run JSONL transcripts are owner-only, bounded in live memory, and redact common secret formats before persistence and display.
• 6885455: Make grilling of unspecified design forks mandatory during planning, spec writing, and implementation. Already-stated requirements in chat, prompts, PRDs, and rules are not re-asked; remaining multiple-valid-way, contradiction, and competing-code-pattern choices wait for an explicit decision.

v1.0.3

Patch Changes

• 10e9150: Add a two-step Fleet Dashboard model-substitution picker and make session-wide source-to-target mappings apply dynamically to future persona and delegate spawns. Bare /af-agent-models-substitute now opens the same picker while the two-argument form remains available.

v1.0.2

Patch Changes

• f2007a3: Replace Agent Hub's double-confirmation budget UX with one localized Yes/No ask_user confirmation. Accepted turn continuations renew the current turn in-place; accepted task continuations open an audited tranche while preserving task tier, assertions, capability packs, blockers, label, and progress. Exclude human ask_user wait time from both turn and task active-time clocks. Remove the redundant /af-new-task slash command; genuinely different work still resets through set_task_tier(new_task: true). Add a visible inline m model picker over Fleet Detail logs that exposes every model Pi currently reports as available and applies the choice to a local specialist, research helper, or nested delegate's next run without interrupting current work. Normalize legacy and Kitty keyboard sequences so arrows and paging keys navigate the picker consistently.
• c028354: Add the read-only /af-context Agent Hub context-budget diagnostic with provider reconciliation and separate fleet context planes.
• e1e8bff: Refresh the root, MCP-extension, and harness runtime lockfiles to patched Pi, HTTP, schema-validation, and YAML transitive dependencies. Add an offline lock-floor regression so vulnerable resolutions cannot silently return.
• e1e8bff: Reduce standing Agent Hub context with automatic, intent-gated capability packs and selected managed-child context manifests. Prevent same-turn overflow with 80%/90% pressure diagnostics, single-flight automatic compaction, and exact-once deferred input replay. Restore persisted named rosters against current configuration while keeping stale orchestrator sessions fail-closed, and document recovery without session-file edits.

v1.0.1

Patch Changes

• ae46472: Prevent empty-roster research cards from crashing the Agent Hub, correct task active-time and new-task reset lifecycle accounting, and persist sanitized mode/reset routing diagnostics in Pi session entries.