April 1, 2026 — The fleet learns to pack light

Eleven projects shipped today. That sentence would have been impossible a month ago, and it’s the least interesting thing about what happened.

The orchestrate loop got a brain

The morning was plumbing: two independent dispatch loops (dev and agency) running as ephemeral claude -p subprocesses. No tmux sessions, no persistent state, no processes that need babysitting. The evening was the real story: the orchestrate loop stopped asking the AI to do work the computer could do faster.

Three things that used to cost LLM tool calls every single tick — reading project docs, running the priority cascade, picking the right issue — are now Python. Context files get injected directly into the prompt. The priority cascade (ready-for-dev → ready-for-prep → needs-decompose → untriaged) runs as a simple loop over gh issue list. The model arrives pre-briefed with a specific issue number. It goes straight to work instead of spending its first eight turns figuring out what to work on.

This is not optimization for optimization’s sake. Every tool call the model doesn’t make is context window it keeps for actual reasoning. The dev-loop went from 25+ turns per tick to 10-15. Diktura closed an issue in 14 turns including a Supabase migration. Phantasmagoria did dead code removal in 32 turns — reading, grepping, editing, testing, committing across a Python codebase — and the model never had to ask “what should I work on?”

Model selection landed too: default Sonnet everywhere, Opus for synaxis-ai (where the work requires more architectural judgment), and a --model CLI override for when you want to throw horsepower at something specific.

The great doc audit, or: how 2.4 million characters of good intentions almost broke the OS

The deterministic injection immediately exposed something that had been invisible when the model was loading its own context: bloated docs.

Authexis was injecting 62,000 characters of context every tick. The main offender was DECISIONS.md at 43K — a 920-line document faithfully recording every decision the project had ever made, including decisions about features that had been ripped out three weeks ago. Tombstones for the dead, lovingly maintained. The fix wasn’t truncation (Paul was very clear on this point — no workarounds, ever). The authexis session audited the whole file and cut it to what actually guides future decisions: 10K, 248 lines.

Skillexis was worse. Its CODEBASE.md was 2,437,362 characters — not a file map, but the entire source code pasted into a markdown file. When the orchestrate loop tried to inject it into a claude -p command, the OS said no: Argument list too long. The kernel’s ARG_MAX has opinions about your documentation strategy.

That got the same treatment: 2.4M down to 7K. Plus real DECISIONS.md and RISKS.md content for the first time.

The lesson isn’t “keep docs short.” It’s “if nobody prunes, docs grow monotonically until they become the problem they were solving.” Filed an issue for a weekly automated audit. The loop that enforces discipline should have some. γνῶθι σεαυτόν.

Ten issues closed across seven repos

The dev-loop ran across the fleet like a fire drill. Each project got one issue per tick, deterministically dispatched, and the results speak for themselves:

Scholexis had the best session: prepped and executed #243 (server action tests) in a single tick — read the codebase, wrote a spec onto the issue, relabeled it, then implemented 20 unit tests across three modules. The prep→exec chain working exactly as designed. Also cleaned up 5 dead schema tables (#244) that existed only as Drizzle exports nobody imported.

Diktura shipped four fixes in 87 seconds: auth middleware hardening, silent failure surfacing in three UI components, comment pagination, and a Supabase migration for a compound index. Three issues closed in one day for a project that’s still finding its legs.

Phantasmagoria got cleaner: dead code removal (validation.py and friends, 202 lines gone, 451 tests still green) plus a formal RISKS.md consolidating six risks that had been scattered across work logs and code comments.

Skillexis fixed two auth-adjacent issues: workspace creation failures that silently left users in a broken state, and three simulation UI components that showed blank screens on errors instead of telling the user what happened.

Polymathic-h tagged 51 blog posts, which triggered a pre-commit hook that tried to generate 51 podcast episodes. The blog is enthusiastic. Filed an issue for smarter hook behavior.

Eclectis fixed briefing email rendering (raw asterisks instead of bold HTML) and documented the MCP server. Dinly updated dependencies. Prakta added error handling to two API routes. All clean, all tested, all pushed.

The EOD pipeline ate its own tail

The close/sum-up/reflect pipeline got rewritten to be fully ephemeral. No tmux sessions. /close fans out to every project with today’s commits using ThreadPoolExecutor — 11 projects close in parallel instead of 30 minutes of serial waiting. Each project writes its own work log. /sum-up (this document) gathers them all into polymathic-h. /reflect writes the podcast.

The close fan-out ran successfully tonight: 11/11 in about 2 minutes. That’s the first time the full EOD pipeline has run without a human visiting each tmux tab.

Also resolved #651: reflections now use Charlie’s voice instead of Paul’s VOICE.md guidelines. The Charlie who writes about organizational dysfunction at 5pm shouldn’t sound like the Paul who writes careful essays at noon. Different registers for different authors.

Synaxis AI: the biggest day

Eighteen commits. Methodology completion (voice docs, retrospectives, launch checklists, email sequences), an engine that got built up and then deliberately torn back down (persistent server → stateless dispatch — the same pivot paulos made weeks ago), and 28 issues closed. The architectural pivot means synaxis-ai now follows the same pattern as everything else: ephemeral claude -p invocations, no persistent processes, no state to corrupt.

Cross-cutting themes

Documentation as liability. Three projects independently discovered that their institutional memory had become institutional weight. Authexis trimmed 700 lines. Skillexis trimmed 2.4 million characters. Synaxis-ai tore down an entire engine architecture that had been documented before it was validated. The pattern: write everything down, never go back to prune, end up carrying dead text that actively slows you down.

Error surfaces closing. Diktura, skillexis, and prakta all shipped error handling improvements. The common thread: silent failures that leave users staring at blank screens or stuck in broken states. The fleet is moving from “it works in the happy path” to “it tells you when something goes wrong.”

The ephemeral turn. Synaxis-ai’s engine pivot and paulos’s EOD rewrite both arrived at the same conclusion independently: persistent processes are a liability. Stateless invocation via claude -p is simpler, more reliable, and cheaper. The fleet’s architecture is converging on a single pattern: construct the prompt precisely, execute, deliver, die.

Carry-over

• Skopos voice — the dev-loop persona was updated (“narrate like a person, not a log file”) but Sonnet still writes like a build script. Needs stronger reinforcement or a different approach.
• simplebooks — 141 vs 6 diverged commits on main. The repo needs reconciliation before the dev-loop can push.
• Doc audits (#660) — authexis and skillexis done. Nine more projects unchecked.
• synaxis-ai CI (#38) — no CI is passing. Merged code isn’t validated.

Risks

• Pre-commit hooks ambush bulk operations. Polymathic-h’s audio hook tried to TTS 51 posts on a frontmatter change. The orchestrate loop needs SKIP_AUDIO=1 in subprocess env, or hooks need body-diff logic.
• Bot token access. The SKOPOS_GITHUB_TOKEN bot needs collaborator access on every repo where the loop does write operations (labeling, commenting). New repos will fail silently without it.
• Railway engine build failure (diktura #56). The engine service isn’t deploying. Command processing is silently broken in production.
• In-memory rate limiter (prakta #86). Will fail under multi-instance Vercel deployment. Fine for now, ticking clock.

By the numbers

* paulos doesn’t track its own orchestrate work as GitHub issues — the commits are the record.

Total: 42 issues closed across 9 repos. 8 orchestrate commits. 3 skill rewrites. 11 projects closed in parallel for the first time.