v0.13.3

Fixed

• A LangChain run inside a generic trace is now its child, not an unrelated root. An app
  can wear both tracers at once — @decimalai.trace / start_trace() around the call and the
  LangChain adapter installed process-wide by instrument(), the composition the README
  documents — but no adapter consulted the generic context, so one logical run shipped as TWO
  root traces. The generic envelope only collects what log_llm_call puts in it, so with the
  model calls going through LangChain it arrived with zero llm_calls and zero spans. Measured
  on DecimalAI's own fleet (2026-09-07): 3,604 such traces in 3 days, 31.9% of the langchain
  path against 0.00% on every other framework, each an empty envelope beside a full sibling.
  Downstream, an empty envelope reads as having used no components and is graded keep at
  score 1.0 against every manifest (100% keep, versus 80.1% for real traces), so those runs
  could never be flagged for replay. The adapter now captures the live generic trace id when a
  root run OPENS and sends it as parent_trace_id; the platform already walks children when
  grading, so the envelope's grade becomes evidence-based with no server change. An explicit
  sub-agent parent_trace_id still wins — the enclosing trace is a fallback, never an
  override — and the probe is fail-open (a raise while labelling a trace is worse than an
  unlabelled trace).
  Known limit: the generic context is a ContextVar, and LangChain copies the context for
  parallel work, so a run started through .batch(), RunnableParallel or an executor thread
  cannot see the enclosing trace. Absent means "unlinked, exactly as before" — never a wrong
  parent.
• decimalai traces import <file>.jsonl answered 422 on every call. POST /api/v1/traces/import
  takes multipart form-data (file + agent_name), and the CLI posted the raw bytes as
  application/x-ndjson — the command the endpoint's own docs sample prints, and the one the
  LangSmith/Braintrust migration guide gives, ended in an httpx traceback for everyone who ran
  it. It now uploads as multipart through a client that drops the SDK's pinned
  Content-Type: application/json (otherwise httpx's boundary never reaches the server),
  posts to the canonical /api/v1/traces/import / /api/v1/traces/import-bulk paths instead
  of the deprecated /api/v1/import/* aliases, gains --agent-name (required for JSONL unless
  every row carries agent_name), and prints the server's detail on a 4xx/5xx instead of a
  bare status code.
• decimalai skills sync could overwrite uncommitted edits with the remote copy. The local
  timestamp preferred git's commit time over the file mtime, which for a SKILL.md with
  uncommitted edits reports the author's newest work as the oldest state on disk. newer_wins
  then answered "remote wins" and — --apply-pulls defaults on — wrote the older body back over
  the edit with ↓ <name> → SKILL.md (v1) as the only notice. The commit time is now used only
  while the file is clean; a dirty or untracked file reports its mtime. A fresh clone is still
  clean, so the checkout protection this existed for is untouched.
• decimalai skills sync counted a frontmatter-only edit as nothing. The backend reports
  reconciled_metadata for a body whose hash matched but whose frontmatter diverged — a real
  write that mints no new version. The summary knew five actions and not that one, so an author
  who edited a description saw every counter read zero. It now prints a metadata count and
  ~ <name> (metadata only — no new version) per skill. Filing it under no_change would have
  said the edit did nothing.
• decimalai skills benchmark crashed before printing the report URL on a run with no
  measured lift. The backend always emits pass_rate and nulls its members when nothing was
  measured (every case's LLM call refused, or every expectation calibration-gated — the normal
  first-draft eval.yaml outcome), so delta_pts >= 0 raised TypeError. Null members now
  render as —, and a run whose verdict is error wears ✗, not the ✓ a passing run gets.

Added

• update_skill(offer_scope=...): set who the trigger router offers a skill to. "workspace" (the default) is every agent in the org; "restricted" is only the agents it has been assigned to. Until now a skill could not be taken off the ambient menu at all, so assigning it to one agent never stopped the router considering it for the others. Metadata only — it does not mint a new version.

Docs

• README: the built-in deterministic scores (completion, has_output, tool_compliance,
  latency, token_efficiency) are attached by the LangChain adapter only — the kill
  switch is instrument(..., builtin_evals=False), not install(...); inject_skill_body is
  documented as tri-state (defaults on for langchain / anthropic / adk, off for
  openai_agents / pydantic_ai), and the delivery and framework tables gain their adk and
  Pydantic AI rows.
• RELEASING.md no longer says releases carry no attestation: every release since 0.10.3 is
  uploaded by publish.yml over Trusted Publishing and carries provenance.

Internal

• The pre-commit ruff hook (scripts/install-git-hooks.sh) lints the staged files under
  decimalai/ rather than the whole tree, so one session's half-finished edit no longer blocks
  another session's commit in a shared worktree. Its flags are literal and pinned against
  ci.yml's Lint job by tests/test_local_lint_hook.py.
• The openai_agents conformance driver's comment stops claiming that seeding the registry
  grades the fabricated-activation mutation; what grades it is
  tests/test_inference_never_writes_the_activation_rung.py.