workflows
Reviewed
at commit e964349
Caution: Blocks Claude from reading images, files over 256 KB and more than ten large documents, steers image reads to Gemini, and can refuse to let Claude end its turn.
Sends data to Google Gemini, OpenRouter and OpenAI when you set their keys; Codex and Copilot through your CLIs; research sites its skills query (SEC EDGAR, CrossRef, CourtListener)
One loop — work — for work worth doing properly: clarify, an approved plan that is the run's only authority, delegated implementation, independent…
0 upvotes
Log in to upvote or save- Source
- github.com/edwinhu/workflows
- Made by
- edwinhu
- Pinned commit
-
e9643496f7df5be8bb97ebffea58d267e59f9504 - What we checked
-
Adds a plan-first work loop and dozens of research, writing and data skills, and runs hooks in every session: linters and prose checks after edits, guards that refuse some reads, and checks that can keep Claude from stopping. With an OpenRouter key set, your latest request, Claude's last message and edit excerpts go to OpenRouter (or OpenAI as fallback); skills can send files to Gemini, OpenAI and research services.
Review details
Static review of the source at the pinned commit; not run. Full bundle of a new version (0b660a7..e964349, 2379 files changed, mostly WRDS parser test data); the hook code was re-reviewed through the diff in the clone, and hook files the diff does not touch (session-start.ts, find-slide-page-inject.ts, lint-check.ts, cite-fidelity-lint.ts, pr-url-logger.ts, session-end.ts, pattern-scan.ts, teammate-idle-report-check.sh, jev/forecast-mod.ts, openrouter-credits.ts, read-guard, image-read, cron-delete) were confirmed unchanged, so their earlier findings carry over. Read in full: plugin.json, hooks.json, register.ts, every hunk of early-stop.ts, work-hold.ts, jev/mod.ts, jev/rules.ts, watch/runs.ts, watch/watcher.ts, bulk-guard.mjs, guards/mod.ts, core.ts, node-io.ts and writing-prose-check.ts, plus the new guards/farm-detach.ts, watch/snapshot.ts, hooks/lib/structural_constraints.py, bin/workflow-check and the bin/ wrappers, the bgjobs.py diff, look_at.sh's agy backend and the new afk SKILL.md; pgrep.ts, rule-check.ts, route.ts and the new skill scripts were read by targeted search (hosts, spawns, writes, installs, permission flags), and the roughly 1,800 added skill files (WRDS/N-PX parsers and fixtures, batch skills, prose and ds rule sets) were mapped by name and grepped. Mod hooks: tool.call (guards, bulk-guard, per-edit Jev), ui.render AbovePrompt (doc-count and OpenRouter balance bands), session.start (watcher), command.run (/farm). Runtime calls: $.fs.read/write/list/stat, $.process.run (sh, git, ps, python3 bin/workflow-check --list, bun rule-check.ts, openrouter-credits.ts), $.env.get, $.store, $.prompt.submit, $.ui.resolve/invalidate/log, $.clock, $.command.register. Shell hooks: SessionStart, PreToolUse(Bash), PostToolUse(Edit|Write|Bash), Stop, TeammateIdle, SessionEnd; overflow-check.ts was removed. Changes in this version: a new farm-detach guard denies a backgrounded farm.sh --tasks without --detach (deny only). The watcher no longer draws a status line (bin/farm-status is an optional statusLine command you configure yourself); its wake prompts now paste the first 15 lines of up to three report files of a finished run, or a work loop's verdict, and grind loops get a wake on new WAIT lines. The per-edit Jev mod and writing-prose-check now resolve their rules and constraints through bin/workflow-check, whose layout includes checks/ folders of other plugins under ~/.claude/skills, so their Python rules and constraints run too. work-hold.ts can release a hold on facts the run's args.json declares (goalSettledBy), reads test-suite files named in red commands only to list their titles, handles /afk holds, and no longer tells Claude to CronDelete the heartbeat. New in decisionsCall (used by work-hold, early-stop and per-edit Jev): when OpenRouter fails (unreachable, 402, 429, 5xx or an incomplete reply) and an OpenAI key exists in OPENAI_API_KEY or $XDG_RUNTIME_DIR/agenix/openai-api-key, the same state goes through curl to OpenAI's Decisions API (gpt-6-luna at api.openai.com), held to a 0.95 bar, not cached, logged to ~/.local/state/jev/calls.ndjson; this needs an OpenRouter key first. early-stop.ts now also blocks without any key: when a typed message reads as a standing mandate (phrases such as 'don't ask', 'going to bed', 'whatever you think', 'work overnight') or an /afk hold is armed, and no heartbeat cron, monitor, owned run or background job is alive, it blocks up to two stops per turn and tells Claude to take the next item or CronCreate an hourly heartbeat; after two idle heartbeat ticks in a row it blocks once per tick. To tell whether jobs are alive it scans /proc: every process's cmdline, its environ (only CLAUDE_CODE_SESSION_ID is used) and its open write handles. The judge state now also carries up to 1500 characters of the standing mandate and 600 of the latest typed message. bulk-guard now points to batch-router and only counts document-naming fan-out rows; bgjobs.py kills fewer helper processes. New skills: afk (arms a check-less hold until 18:00 or 09:00, has Claude create a recurring hourly cron, commit locally and not ask questions), batch-router, claude-batch (Anthropic Batch API with your key), openai-batch (uploads documents to the OpenAI Batch API with your key), prose-rules, ds-rules and replicate-paper. Twelve skills pre-approve unrestricted Bash (afk is new among them); none has Claude run an unpinned package. The WRDS skill ships updated prebuilt Linux Go binaries (parse_13f_go, parse_def14a_own_go, parse_npx_go) with their Go source alongside; the binaries were not reviewed and run only when Claude follows that skill. Carried over: image Reads denied and redirected to look_at.sh, whose default agy backend runs a child CLI with --dangerously-skip-permissions and sends the image to Gemini; whole-file Reads over READ_GUARD_BYTES (256 KB) denied; bulk-guard's tenth-document deny and result trimming with full output saved under $XDG_RUNTIME_DIR/bulk-guard or /tmp/bulk-guard (BULK_GUARD_OFF=1 opts out); work-hold reruns the armed check with bash -lc at every Stop and can write autoCompactWindow into the project's .claude/settings.local.json only with WORK_HOLD_COMPACT_WINDOW set; session-start reads ~/.secrets/claude-keys.env and the project .env only to report which key names are set, reads ~/.claude.json, runs ~/.local/bin/morgen-events if present and checks the OpenRouter balance; writing-prose-check runs uv run --with lxml --with pyyaml; pattern-scan stores correction snippets in ~/.claude/projects/<slug>/pending-patterns.json; EARLY_STOP_HOOK=0 and JEV_EDIT_MOD=0 switch off the judge hooks. Most failures fail open.
- What it uses
-
- Reads your prompts: yes
- Sees tool calls: yes
- Blocks or changes tool calls: yes
- Reads or writes files: yes
- Runs programs on your machine: yes
- Calls the model (uses tokens): yes
- Sends data over the network: yes
- Stores data: yes
- Shows something on screen: yes
- Hooked events
One loop — work — for work worth doing properly: clarify, an approved plan that is the run's only authority, delegated implementation, independent verification against computed gates, and human review. Domain workflows for code (test-first), data, writing and workshop presentations.
Made this mod? Add a badge to your README
Paste this into your README. The badge says reviewed once the current version has been reviewed, and links readers to this page.
[](https://modsforclaude.com/mods/workflows/)