v0.15.12

[0.15.12] - 2026-09-14

Fixed

• Release preflight no longer requires a3s-deep-research 0.1.5 on crates.io;
  the CLI pins that crate from git, and only registry prerequisites are gated.
• A bubblewrap uid-map denial names the unprivileged user-namespace
  prerequisite and still denies Bash. An unrelated probe failure does not
  invent that repair.
• Help a3s-test 02-composer-help pages down until /sandbox is visible in
  the long reference panel.
• Published-pin scripts fail closed when Cargo.toml path-depends on
  ../code/core, not only when .cargo/config.toml path-patches Core.
  --require-published on a path pin is a local tree, not a shipped pin.
• When published Core refuses isolation with source revision is unknown,
  session create/resume host copy appends the repair (create an initial
  commit) if Core has not already named it. Does not invent Core event text.
• Plan / read-only Code sessions admit absolute paths and Read files[]
  targets that stay inside the workspace; host absolutes and .. escapes
  remain denied (Core surface from 8.5.5+; this release pins 8.5.8).

Changed

• Pin a3s-code-core to published =8.5.8 at git rev
  bcd4efe2fceb50cae9a6a5d40d29e3142143018d (tag v8.5.8). Drop the local
  path = "../code/core" dependency so --require-published can pass.
  Release workflow A3S_CODE_CORE_VERSION / A3S_CODE_CORE_REVISION match
  that pin. Worktree add and remove pass Git the verbatim-prefix-stripped
  path the kernel uses. AgentEvent::UserQuestion on this Core rev does not
  carry allow_free_text; the TUI option picker still works and treats free
  text as off until a later Core publishes that field on the event.

• Discover WorkBuddy AI alongside classic WorkBuddy: macOS accepts
  WorkBuddy AI.app and WorkBuddy.app, account state prefers
  ~/.workbuddy-ai then ~/.workbuddy, and Windows install probes include the
  WorkBuddy AI product directory / executable name.

• Code TUI /goal Slice 2–3 + P3: ACCEPTANCE.md supports kind:command /
  kind:file_exists / kind:manual criteria; the host re-runs checked
  command/file predicates before latching GoalAchieved so checkbox-only
  self-approval cannot close the loop. Host kind:file_exists requires a
  regular file (is_file), matching Core's synthesized test -f (directories
  cannot latch), and the resolved path must remain inside the workspace
  (absolute/../ escapes and outbound symlinks cannot latch). Host also refuses latch when ACCEPTANCE is manual-only — at
  least one kind:command or kind:file_exists criterion must be present and
  re-verified (Core workspace presets alone cannot false-complete a durable
  goal). Verifier iterations keep Plan chrome
  but arm a goal-verify fence (sandboxed bash + writes only under the loop
  dir) and clear Core Plan specialty so verification_reports can form;
  maker iterations keep Auto. Core shell verification matches ACCEPTANCE
  kind:command asserts and synthesizes test -f <path> for
  kind:file_exists under .a3s/loops/*/ACCEPTANCE.md, so Host re-checks
  and Core evidence share the same machine predicates. While a goal run is
  active, session max_parallel_tasks is host-clamped to ≤4 (plan 2–4 wave;
  Ultracode's wider interactive budget does not apply). When the workspace
  declares machine ACCEPTANCE criteria on active loops (running /
  retrying / paused), Core GoalAchieved emission requires a passing
  ACCEPTANCE-derived verification_report for each such loop (workspace
  presets alone cannot authorize emission; completed-loop leftovers and a
  sibling/orphaned active loop's report alone cannot authorize). Starting a
  replacement /goal cancels the prior in-memory run and any other on-disk
  active goal-* loop STATUS so orphans cannot pollute emit. Footer/runtime treat
  ACCEPTANCE criteria coverage as the primary completion meter and demote
  Core plan progress to secondary (not completion). Passing machine criteria
  are fingerprinted into STATE.md ## Verified Evidence as maker skip-hints;
  stale fingerprints cannot skip host latch re-checks. Continue path uses a
  host schedule (schedule_goal_continue): provider_error → retrying +
  backoff sleep; criterion_fail → running + immediate continue. KB home
  hint no longer claims /okf manages packages.

• Code TUI slash IA stage 0/4: every registered slash command carries an
  explicit core / advanced / migrate-out tier; /relay, /fork, and
  /worktree share the Session group with handoff-vs-isolation copy; docs
  lead research with /research (? remains a shortcut). Stage-5 sibling
  deletes stay deferred until hub redirects are stable. Scrubbed residual
  /okf-manages-KB module docs so removed five-pack authoring is not treated
  as live.

Fixed

• DeepResearch planner/report generation no longer nests a one-step
  dynamic_workflow PTC around generate_object. Host-direct structured
  generation with Host-owned retries replaces the opaque
  program script error: path that degraded live runs to no_evidence.
• Pin a3s-deep-research to 0.1.5 (4c113452) so retrieval accepts Core
  staged batch headers natively, skips the multi-minute model selector for
  tiny closed catalogs, and promotes closed-catalog deterministic excerpts
  when model-backed chunk selection fails—acquired text can stay
  claim-eligible instead of collapsing to false no_evidence.