v0.9.8.8 release notes

v0.9.8.8 is about watching the run, and trusting the team to finish it. The factory floor’s telemetry rail goes live; the team finishes the job even when a seat dies — the router routes around a crashed model and QC, the producer of last resort, authors the missing pieces; and the seats now tell the truth about themselves when a reasoning model can’t be quieted. Four cadre-cleared arcs, one live kill-test. This page is the delta on top of v0.9.8.7.

Watch the run — live telemetry on the floor

The MOD SQUAD tab’s Run Telemetry rail is no longer a mockup. As a run moves, it paints:

elapsed time and a 10-segment tasks bar with a real percentage,
the QC pass / reject tally (qc ✓ 4 · ✗ 1),
the context load — tokens and compression count, read from the run’s audit log,
and, on the floor below, each producer’s current move rendered in the same icon + verb vocabulary as the stream (◆◆ Randy ▼▼ reading a page). QC steps onto the floor while it reviews, and steps off when the verdict lands.

Everything rests to dashes between runs. The CONSOLE flip is now two real LEADER / MOD SQUAD tabs (each a click target; F4 still cycles), and launching a job no longer yanks you to the floor — you flip to watch when you want to. This is the instrument that made the rest of the release visible: the compression bloat that #16 fixes, the recovery that the self-heal arc drives — you can now see both happen.

The team finishes the job even when a seat dies

A crashed or unloaded local model doesn’t fail cleanly — an OpenAI-compatible endpoint (LM Studio, llama.cpp) reports a dead model as an HTTP 400, which used to slip past the availability classifier. The retry loop would then burn a task’s whole budget in about a second against the dead endpoint, and because a dead seat is always idle, the dispatcher kept feeding it new tasks — a task magnet. All three are fixed:

The crash-shaped 400s are now recognized as availability failures, so the seat’s fallback chain engages (restart the task on a backup model).
Retries back off (2 / 8 / 20s) instead of hammering a dead endpoint, and the wait rides the abort event so the kill switch stays responsive.
The seat is cooled out of the dispatch pool for a cooldown window, so it stops attracting tasks while it’s down; a half-open probe lets it back in when the window expires.

And when the seat stays dead, its exhausted tasks route to the QC backstop instead of dead-ending as blocked.

QC produces the missing pieces

QC has always been the producer of last resort — when a producer can’t clear the bar, QC patches the artifact from its own findings. v0.9.8.8 extends that to a goal’s end: if work is still missing — a task blocked by a failed dependency, or a producer that committed nothing — QC now produces the missing pieces itself, rather than shipping the goal with a hole in it and a “disappointed” verdict while QC sits idle.

The sweep builds each missing artifact from its task’s brief, in dependency order, feeding each build the real content of the pieces it depends on — so an assembler assembles the actual parts, never a fabrication. It stays out of blocks it genuinely can’t fix (a missing linter, a wave path-conflict), spends only QC’s budget (never a producer’s), and runs at most once per goal. The post-run headline is now honest about partial runs, too: a run that landed real deliverables but left a task blocked no longer reads “Nothing usable landed” — it owns both the wins and the reservations.

Seats that tell the truth — thinking-off honesty

A reasoning model in a producer seat bloats its own context with reasoning tokens; the fix is to quiet it. But a request parameter can’t be trusted through a shim: an Ollama or LM Studio /v1 endpoint silently drops (or outright rejects) reasoning_effort, so a reasoning-heavy model behind one can’t actually be quieted that way — even though the engine used to believe it had.

v0.9.8.8 fixes both the mechanism and the honesty:

The engine sends each model family’s in-band toggle where one is proven — Qwen’s /no_think rides the message text, which no shim can drop. A family with no known toggle gets clean messages instead of ignored prose.
When a producer seat can’t be quieted (a reasoning-heavy family behind an opaque OpenAI-compat shim), the engine says so: modulatio doctor grows a Seats section that flags the seat with a remedy, the agent builder warns at seat time, and the run log warns at team build.

The guidance is honest, not a gate — if you want a heavy reasoner’s judgment in a producer seat and you’ll pay the reasoning tax, the engine respects it; the warnings just make sure it’s a decision, not a surprise you find in the compression logs. For producer seats, prefer a non-reasoning model or a toggle-able family, and save the heavy reasoners for the Leader and QC — the judgment seats.

Also in this release

Agents accrue episodic memory from the jobs they run — a per-agent record of what each one worked on, surfaced in the MEMORY tab, where pending QC memory proposals can now be approved (p) or rejected (d).
A SETTINGS tab (CONFIG → SETTINGS) tunes the engine’s knobs — persistent env overrides, retry budgets, per-role context windows — from the TUI, applied without editing a config file.
The activity feed speaks in icons: each tool call and phase renders as a bold phosphor glyph + a plain-language verb, a repeated action coalesces into one line with a counter instead of a wall of identical rows, and /cls (or Ctrl+L) clears the active stream.
Two more Feng-Tui variants — neon phosphor red and purple — join amber / green / cyan on the F2 cycle, and the µ mark (Modulatio’s icon) renders in the active phosphor.

See the CHANGELOG for the full delta.