v0.3.0 — harness profiles and the task trace

[0.3.0] — 2026-10-11

Two headlines. Harness profiles: Claude Code, Codex and Pi are each
one declarative HarnessProfile, with conformance tests over every row
and a shared TypeScript shim runtime. A new harness is now a new
profile, not a fork. Task trace: one durable kind: task note
follows a unit of work across in-session subagents, /wrap, and devloop
runs. The trace is verified end to end by the task-360 evaluation suite
on real Claude and Pi sessions.

Upgrade notes

• Run weave doctor --migrate once. It folds stored task asks to
  github:owner/repo#N refs, rejoins split task notes, and upgrades
  pre-#260 devloop rounds.
• The Codex skill bundle moved from skills/ to codex/skills/ (PR
  #276). Re-point hand-made links in ~/.codex/skills/ to
  <repo>/codex/skills/thinkweave-*.
• Pi needs the pi-mcp-adapter extension for MCP
  (pi install npm:pi-mcp-adapter). weave doctor --mcp --harness pi
  checks for it.

Added

• HarnessProfile contract (#191, PR #201): Claude Code, Codex, Pi
  and OpenCode are profile data. A parametrised conformance suite and an
  AST guard against per-harness forks enforce it. docs/HARNESSES.md
  generates its capability matrix from the profiles, and install output
  names every degradation.
• shim-core (#194, PR #201): a shared TypeScript translator runtime
  for harnesses that expose lifecycle events but not Claude Code hooks.
• Pi at E3 (captured) (#114, PRs #207, #211): the Pi shim captures
  prompts, tool calls and stops, and injects SessionStart context.
  weave import pi imports Pi transcripts. MCP routes through
  pi-mcp-adapter, and skills link into ~/.pi/agent/skills.
• Task trace (#184 stack, PRs #222, #227, #229, #240, #260, #261,
  #263, #277, #301): the kind: task note contract; live subagent
  boundary capture; weave task open|close|render; a task note as a
  ledger of rounds; a task pass in the /wrap finalize tail;
  devloop rounds that land on the epic's task; child_of edges; and
  an Open Tasks section in SessionStart. The outcome-judge closes tasks
  whose tracker issue closed, and files PR review comments as
  feedback_for notes.
• task-360 evaluation suite (PRs #261, #264, #267, #272, #301):
  evals/task360/ drives named scenarios through Claude Code, Codex and
  Pi in a throwaway vault. An oracle scores each scenario from the
  index.
• /explain (#298, PR #300): composes a session or diff explainer
  into one vault note. weave explain locate|check finds and validates
  these notes.
• Session notes record their harness and model (#285, PR #301).

Changed

• Leaner Claude Code context (PR #276): Claude Code no longer lists
  each skill twice, and SessionStart drops placeholder sections.
  weave_project_snapshot still serves the full payload.
• Index-backed retrieval (PR #206): timelines, prompts and landings
  query the index instead of scanning the vault. weave timeline drops
  from over 60 s to about 0.04 s.
• Plugin-route launchers bootstrap their dependencies once (#164, PR
  #201).

Fixed

• /wrap fails loudly (PR #278): a dropped weave_create, a no-op
  weave_extract, or an empty finalize now reports an error instead of
  exit 0.
• Codex capture fidelity (PRs #208, #212): PostToolUse events land,
  one session maps to one note, and guardian sidecars are filtered out.
• Wrap and judge resolve session notes by exact id and the harness
  session id (PRs #209, #214, #233).
• shim-core builds and its test suite runs again: npm test passes a
  file glob on Node 24, and the subagent hooks have timeout budgets
  (PR #213).

Internal

• The devloop pin follows funloops main, and loop.toml uses the
  four-gate pipeline (PRs #216, #242, #260).