v0.9.8.7 release notes

v0.9.8.7 teaches the team to reuse the research it already owns, to fan a wide gather by context size so no single producer drowns, and to keep assembly calm under those wide fans — driven by four live test runs and a code cadre across the span. This page is the delta on top of v0.9.8.6.

A research library the team reuses, not re-buys

The project folder is now the durable layer: research notes, drafts, and finished products persist across runs while a run’s scratch is transient. Producers are prejudiced toward reuse — before reaching for http_get, a producer mines the team canvas and prior-run artifacts and reuses a grounded, still-fresh note as-is rather than re-fetching the same sources. Notes carry a 30-day freshness TTL: a note past it is re-fetched (and re-stamped), a fresher one is reused. Cross-run reuse is now measured — the engine writes an audit row each time the team-canvas digest is injected, recording how many prior-run files reached the producers. A project that runs the same class of job repeatedly gets cheaper, not just faster.

Fan the work by context size, not a guess

A capable planner only structures a wide research goal into parallel tasks about half the time; when it collapses, one producer inherits the whole scope and rides the compression bands. The engine now binds the fan: an oversized gather task is split into the fewest size-bounded chunks that each fit comfortably inside a producer’s window and together cover the same scope. Size decides whether to cut; the model only picks the cut lines. A task that already fits stays whole — no invented work. The old fixed topic-count guesswork is gone. Tune the cap with MODULATIO_TASK_CONTEXT_CAP_PCT (default: a fifth of the task’s own window).

Assembly stays calm under a wide fan

Joining a multi-piece deliverable is a mechanical operation, so an over-budget assembler now falls back to the engine join instead of recursively decomposing itself into partial assemblies. QC reviews with a 64K window (up from 32K), so the reviewer’s context matches the largest producer tier and a big single deliverable no longer forces a compressed, partial-view judgment. And the content-addressed cheap-verify now resolves every dependency’s output — including the units that write to the drafts fallback convention — so a wide fan’s assembly passes QC by its marks rather than re-reading the whole thing into a blown budget.

Producers can’t leak their scaffolding into the product

A producer sometimes prefixes its reply with a runbook block (Operation: / Definition of Done:) above the first heading. That preamble is now stripped at the assembly join and rejected by a deterministic QC gate — matched by runbook-shaped marker lines, so ordinary prose that merely mentions those terms mid-sentence is left untouched. The engine binds the cleanup that five LLM QC passes had let slip through.

The Leader grounds its size claims

A deliverable’s verify block now carries an engine-measured size (bytes / words / ≈tokens), and the Leader’s rationale must cite it — no more rounding a nine-page draft up to the goal’s “20 pages”. When a brief asks for reputable sources, a user forum or social thread is treated as a lead to a primary source, never a citation.

Reliability, from four live test runs

Cron fails closed when its project was deleted: the job disables itself and opens one ticket in a reserved system project instead of resurrecting an empty shell on a default team.
A wedged model call is diagnosed and surfaced by a CPU-spin watchdog rather than silently hanging the run.
An empty Clay reply (claude -p runtime hiccup) is retried and, if still empty, surfaced as a failure routed to the model fallback — never propagated as a blank message.
The kickoff render-storm at run end is quieted (it could starve the final verdict relay), the finished Leader lane no longer re-animates during post-run learning, and a hidden stream lane re-follows the tail when revealed so an off-screen verdict isn’t invisible.
/reload applies model and config changes to the live services without a TUI restart.
The Leader names each job and captures a named ad-hoc run as a re-kickable, project-local job template; deliverables are named by their title.

See the CHANGELOG for the full delta.