Release Notes — v0.8.13.0 “Nothing Invented, Nothing Borrowed”
Date: 2026-09-21 · Cut from: main 39925528e (see tag v0.8.13.0) · Milestone: MVP — server-key abolition complete (#1812), the PM-dogfood standup trust fixes (#1836/#1837), keyless first contact (#1818)
Quality posture: 14,332 tests collected; unit+intent sweep 11,290+ passing at the #1812 landing and 4,875 across every package touched by the standup work; smoke 531; security suite 127; spend-free canonical ratchet 14/14 with membership unchanged; mypy gate at all 24 ceilings. One day of changes since v0.8.12.0 — a fast follow, cut so the weekend’s trust fixes reach alpha before new testers do.
Summary
This release finishes what v0.8.12.0 started and fixes what the first real dogfood session
found. The “server key” is no longer merely refused — the concept is deleted from the
codebase: there is no code path, gated or otherwise, to a product-owned LLM credential.
And the standup flow stops inventing things: accepting the guided-interview offer now
actually starts the interview, the generic placeholder draft (“Made progress on assigned
tasks”) is deleted and unreachable, free-form draft edits genuinely apply, and a keyless
“hi” gets a human acknowledgment instead of a wall of policy.
What’s New
The server-key concept is gone (#1812, closing the BYOC arc)
- Steps 5–6 landed:
PIPER_OPERATOR_SERVER_KEY, the designated-operator principal,
and the transitional “explicit None binding” are all deleted. Every LLM spend is the
acting user’s own key, resolved per request — or an honest refusal. Forgetting to bind
is an error; binding the old operator form raises at bind time.
- The import-time hazard is gone at the root: constructing the LLM client no longer
reads any credential from any store — the server’s long-lived provider clients were
amputated along with their only consumer (the seam itself).
- KG embeddings lose their keychain fallback: document search embeds on your stored
OpenAI key, or refuses honestly.
The standup flow stops fabricating (#1837 + #1836 — from PM’s live alpha transcript)
- Accepting the interview offer starts the interview. You say yes, it asks the first
question — no second “Ready for your standup?” greeting, no mode-fork re-ask.
- The generic placeholder standup is deleted and unreachable. With nothing to build
from, Piper re-enters the interview (the only honest keyless-data path) or says plainly
that there’s nothing to draft. It never presents boilerplate as your day.
- The flow owns its own offers. Ask “didn’t you say you’d do a guided interview?”
and — if it owes you one — it says exactly that and offers to start it, instead of
denying the offer it made three turns earlier.
- Free-form draft edits actually work (#1837 shape 3): “change what I did yesterday —
say I spent the day getting the team back on track” now routes through the
conversational floor on your own key and applies. The old keyword matcher (which fired
on words inside your sentences — the literal cause of the transcript’s wrong turn) is
retired. “Add blocker: …” and “remove …” stay instant and key-free.
- “I’ve updated your standup” is only ever said when something changed (#1836):
the success message is derived from a verified diff. An inapplicable edit gets an
honest “unchanged” answer; a keyless free-form edit gets an honest note that it needs
your key, with the key-free actions named.
- “Start over” now actually starts over (clears the captured answers and re-interviews)
instead of re-rendering the same draft.
- A keyless user’s first message gets a kind-matched acknowledgment — “Hello — good to
meet you.” / “That’s kind — though I haven’t actually done anything yet.” — followed by
the one shared key sentence: “Piper runs on an LLM key of your own, not on anyone
else’s account. Add an OpenAI or Anthropic key in Settings and I’ll be ready when you
are.” Deterministic, zero LLM, zero spend.
- A repeated pleasantry gets a short restatement, not the full copy again; a repeated
substantive request gets the gate’s own refusal.
Deploy identity and credential-store safety
/health now reports the real version, git SHA, and environment (#1839) — alpha
had been reporting a hardcoded environment=staging for months. “What is actually
deployed” is now a curl, not an inference.- Keychain namespace collisions are loud (#1764): a non-default service name over the
DB-backed credential store now raises with a pointer to the issue, instead of silently
colliding credentials across services.
Known limitations (honest)
- These fixes are in the code as of this cut; alpha.pipermorgan.ai runs them only after
this release is deployed to the droplet (deployment is a separate step, same session).
- Free-form standup edits require your own LLM key (they run a model on your account);
the chip actions (add blocker / remove / start over) never do.
- A short polite imperative like “please remove the fluff” can still be mis-read as an
acceptance and finalize a draft — found by this release’s own regression work, tracked
as #1843 (the shared acceptance-vocabulary lane; fix pending its owners’ ruling).
- Gemini has no spend path (its SDK offers no safe per-request credential); OpenAI and
Anthropic keys are the supported pair.
- Known-issues list: see
docs/ALPHA_KNOWN_ISSUES.md (updated with this release).
Version mechanics
- Cut: product content at
39925528e (origin/main, 2026-09-21 morning); the tag
sits on the release commit (version bump + this doc set) directly atop it — lockstep
mode: production fast-forwarded to the tagged commit in this session.
- Tag:
v0.8.13.0
- pyproject.toml and VERSION:
0.8.13.0 (the VERSION file — which #1839’s
/health reads — was caught stale in the same-session docs audit and bumped in a
follow-up commit directly atop the tag; production includes it)
Upgrade instructions
Hosted alpha testers: nothing to do — the update lands when the droplet is redeployed.
Self-hosting: git fetch && git checkout v0.8.13.0, then restart the server
(docker compose up -d --build app or your equivalent). No schema migrations in this
release; alembic upgrade head is a safe no-op.