v1.20.0

Minor Changes

• 9a536c4 The Operating Board harness pivot for planr operate (SPEC-004), the openplanr minor
  bump 1.19.0 → 1.20.0 — the agent is the engine; the CLI harnesses, verifies, and records.
  Every landed claim below is backed by a named proof; coordinated sibling work still
  completing under this same version is described as landing with the release, never as a
  phase this repository has already verified.

The mandate replaces the collector as the unit of dispatch. A per-role operating
mandate carries the lens question, declared read boundaries (workspace roots — including
the .planr/ tree — a sensitivity ceiling, and forbidden paths), a required response
schema, and a citation requirement, and it carries no evidence bodies and no evidence
index — structurally forbidden rather than merely omitted. Proven by
tests/unit/operate-adapter-mission-dispatch.test.ts ("prepares a mandate (not a pack)
with declared boundaries and no evidence body ..."), whose fixture fails if any file body
ever leaks into the body-free, index-free mandate.

Evidence is an output, not an input, and citation resolution is the universal gate.
adapter record resolves every citation in a response fail-closed and mints the
evidence-of-record from what was actually cited. A fabricated path, a wrong line range, a
moved revision, or a citation above the role's sensitivity ceiling each becomes a governed
gap — on every dispatch path and every source — and a response resolving zero citations
records its role not_evaluated with a governed gap naming the empty grounding. Proven by
tests/unit/operate-citation-resolution.test.ts ("rejects a fabricated path, a wrong line
range, and a moved revision with distinct reasons and one gap each" and "commits a role
not_evaluated with a governed gap when its citations resolve zero evidence") and by the
above-ceiling refusal in tests/unit/operate-mission-honeytoken-isolation.test.ts
("refuses a read above the sensitivity ceiling inside a granted root").

Hard-blocked secrets in cited content are rejected, not redacted-and-accepted. A
citation whose snapshot contains a hard-blocked secret category is refused as an
unresolvable citation gap instead of being persisted in redacted form. Proven by
tests/unit/operate-citation-resolution.test.ts ("rejects a citation into
HARD-blocked-secret content as unresolvable, never redacted-and-accepted"), with a
soft-secret assignment still redacted-and-accepted so the distinction is exercised both
ways.

A gitignored .planr/ tree is fully citable — by architecture. The dispatched agent
reads the filesystem directly rather than git ls-files, so a project that gitignores its
.planr/ control surface can still ground the three lenses (CPO, CMO, COO) that the
fourth field audit found unsatisfiable when candidates came only from tracked files — the
defect that starved three of six lenses is gone by construction, not by repair. Proven by
tests/unit/operate-mission-honeytoken-isolation.test.ts ("reads a gitignored .planr/
tree — the mission tool walks the filesystem, not git ls-files (finding 2)").

Guided init has no livelock, no dead-end advice, and a revise path. An answer envelope
is accepted on its binding validity (session id, questionnaire digest, project head)
rather than on wall-clock ordering, a transiently stale session un-latches when the tree
is restored, a genuinely terminal rejection names the resumable session id and its exact
--resume command, a previously answered question can be re-answered before apply, and the
questionnaire advertises --answers-file as a stdin-parity transport alternate with
per-question renderability metadata. Proven by
tests/unit/operate-question-session.test.ts ("accepts an answer envelope whose
submittedAt predates the session (livelock regression)" and "un-latches a transiently
stale session when the tree is restored"),
tests/integration/operate-question-resume.test.ts (the resume command surfaced in the
next actions), and tests/unit/operate-question-engine.test.ts ("advertises
--answers-file as a stdin-parity transport alternate with its exact argv" and "carries
repeated-text renderability metadata sufficient to present without improvisation").

Completing the pivot in the same coordinated release. Sibling work landing under this
version retires the now-dead evidence collector — its walks and budgets, role packs,
mission packets, and the pack|mission dispatch-mode split — behind the mandate contract;
renders cycle integrity (citation rejections, boundary refusals, not_evaluated roles) as
a first-class section of the readable tree and a doctor check; makes the persisted
cycles/<id>/report.md a self-contained record with complete registers; corrects the
"commit-safe root" claim to an honest, redaction-based statement; and collapses runtime
classification to mandate-capable-or-unsupported with no silent structured fallback. On the
same schedule, the CLI's own structured-provider advisor path and its --ai planning
surfaces are deprecated (FR4) — functional this release, pointed at the harness flow, and
scheduled for removal, never broken and never silently removed. These items land with this
release rather than ahead of it; their per-task proofs live under
.planr/specs/SPEC-004-operate-agent-harness-architecture/.

Pins the optional planr-pipeline runtime to the exact 0.36.1 — the build that
publishes the additive operating-mandate schema (the same additive pattern as
create-quick-task/create-epic), the regenerated mandate-flow command/skill
instructions, the registry investigation mandates, and the reclassified adapter capability
rows this release dispatches against. Verified:
optionalDependencies["planr-pipeline"] === "0.36.1", the three workflow ref: v0.36.1
pins, and both packed-install e2e assertions.

Unreleased

Deprecations

• CLI-managed Operating Board structured-provider dispatch and the legacy
  --ai planning surfaces remain functional in this release, but now emit one
  shared deprecation notice. Agentic investigation and advisory work belongs in
  the native runtime harness through Protocol v1.3 mandates. Both legacy surfaces
  are scheduled for removal in OpenPlanr 2.0.0; see
  https://openplanr.dev/docs/operate/agent-harness.