2026-04-01

Two-loop architecture shipped

Got the orchestrate system running with two independent dispatch loops:

• Dev loop (paulos orchestrate run -p <project>) — picks GH issues, dispatches claude -p '/dev-loop' as ephemeral subprocess. Five focus projects: authexis, eclectis, simplebooks, diktura, synaxis-ai.
• Agency loop (paulos orchestrate agency -c <client>) — picks Notion deliverables, dispatches agency-triage + agency-loop. One client: Eclectis.
• Scout (paulos orchestrate scout -p <project>) — deep codebase scan, slower cadence.

Key config: focus = true in workspaces.toml controls auto-dispatch. New agency-clients.toml for client list.

Bugs fixed

• Discord 400 on idle cycles: _result_msg("Dev-loop", "idle") returned "Dev-loop" with no “idle” substring, so _notify classified it as success and tried to send a malformed embed. Fixed by passing outcome explicitly.
• Dev-loop idle for synaxis-ai: git remote was still skopos-synaxis/synaxis-ai after folder rename from marketing/. Fixed to pwelty/synaxis-ai. The dev-loop skill relies on gh inferring the repo from the remote.
• Streaming output: _run_cli now streams claude -p stdout line-by-line so the operator can see progress instead of a silent wait.

API key rename

Renamed ANTHROPIC_API_KEY → PAULOS_ANTHROPIC_API_KEY in all project .env files. Claude Code was picking it up and using API credits instead of the Max subscription. Updated llm.py and agent_provider.py to check the new name first with fallback chain.

Docs updated

PRODUCT.md, CODEBASE.md, DECISIONS.md all updated to reflect the two-loop architecture, 12 active skills (up from 9), focus field, agency-clients.toml, and current project portfolio.

Afternoon session: deterministic dispatch and the great doc audit

The orchestrate loop got its brain transplant. Three things that used to burn LLM tokens every tick — reading context files, running the priority cascade, picking the right issue — are now deterministic Python. The model arrives with context pre-loaded and a specific issue number. It goes straight to work.

Deterministic prompt construction

Context files (PRODUCT.md, DECISIONS.md, CODEBASE.md, RISKS.md) are now read in Python and injected into the claude -p prompt. Saves 4+ tool calls per cycle. The priority cascade (ready-for-dev → ready-for-prep → needs-decompose → untriaged) runs in Python too, passing the issue number to the skill as /dev-loop #42.

Model selection

All claude -p runs default to Sonnet. Per-workspace override via model = "opus" in workspaces.toml (synaxis-ai uses Opus). CLI override with --model/-m. Codex CLI integration exists but is shelved — Codex can’t write to .git in sandbox mode, so it can’t commit.

The doc audit

The deterministic injection exposed a problem that was invisible before: bloated docs. Authexis was injecting 62K of context per tick, mostly a 43K DECISIONS.md full of tombstones for features that got ripped out weeks ago. Skillexis was worse — 2.4 million characters in CODEBASE.md (the entire source code pasted into a markdown file). It literally crashed the OS: Argument list too long.

Both got audited and tightened:

• Authexis: 62K → 25K (DECISIONS.md from 43K to 10K)
• Skillexis: 2.4M → 18K (CODEBASE.md from 2.4M to 7K, plus new DECISIONS.md and RISKS.md)

Filed #662 for a weekly automated context doc audit across all workspaces. The loop that enforces discipline should have some.

Notification improvements

Discord notifications now show what the dev-loop actually did (“What happened: removed 4 dead schema exports, 451 tests passing”) instead of just the issue list. Outcome classification fixed — dispatched work returns “success” with a green checkmark, not “idle” with a sleep emoji. Bot token removed from read-only issue queries (was blocking access to repos the bot hadn’t been invited to).

EOD pipeline rewrite

The close/sum-up/reflect pipeline is now fully ephemeral — no tmux sessions needed. /close fans out to all projects with today’s commits in parallel (ThreadPoolExecutor). Each project writes its work log to its own repo. /sum-up then gathers all logs into polymathic-h and synthesizes. /reflect uses Charlie’s voice, not Paul’s VOICE.md (#651).

Fleet dev-loop test runs

Ran the deterministic dispatch across seven different repos. Results:

• scholexis #244 — removed 5 dead schema tables, 103 tests passing
• scholexis #243 — prep→exec chain, added 20 server action tests
• phantasmagoria #263 — removed dead code, 202 lines, 451 tests
• phantasmagoria #264 — created RISKS.md risk register
• prakta #233 — error handling on API routes
• dinly #361 — dependency updates, 327 tests passing
• diktura #53 — four fixes + Supabase migration
• skillexis #199 — auth error handling, 29 tests
• skillexis #200 — error states in 3 UI components
• polymathic-h #242 — tagged 51 blog posts (triggered a pre-commit audio hook ambush, filed #247)

Completed

• #659 (partial) — deterministic prompt construction: context injection and issue selection done
• #651 — Charlie voice in /reflect
• #662 — filed: weekly context doc audit
• #661 — filed: speedy-gonzales provisioning (backlog)
• polymathic-h#247 — filed: pre-commit audio hook on frontmatter changes

Carry-over

• Skopos voice — skill updated but Sonnet still narrates like a build script. Voice reinforcement needs work.
• simplebooks — repo has 141 vs 6 diverged commits. Needs reconciliation before dev-loop can push.
• #660 — doc audit for remaining projects (authexis and skillexis done, others unchecked)

Risks

• Pre-commit hooks can ambush bulk operations — polymathic-h tried to generate 51 podcast episodes when we tagged 51 posts. The orchestrate loop should probably set SKIP_AUDIO=1 in subprocess env, or hooks need smarter change detection.
• Bot token still used for writes — _gh() in orchestrate_loop.py uses SKOPOS_GITHUB_TOKEN for labeling/commenting. The bot needs collaborator access on every repo where the loop does write operations.

Next session

• Run paulos orchestrate close to test the parallel fan-out in production
• Test the full eod pipeline end-to-end (close → sum-up → reflect)
• Address simplebooks repo divergence
• Continue doc audits on remaining projects (#660)
• Improve Skopos narration voice — current skill tweak landed but Sonnet needs stronger reinforcement