Professor' Harvester

Contents

• Why Professor?

• Six things you can watch it do

  • 1. See the fleet

  • 2. Chats that talk to each other

  • 3. Rules that bite

  • 4. Read what the web hides from bots

  • 5. One contract, three runtimes

  • 6. Reload without losing the conversation

• Install

• The discipline layer (templates/)

• The fleet CLI (pfm/)

• Plugins

• Origin

• Help and community

• Contributing

• License

A fleet controller and a discipline layer for Claude Code, Codex, and OpenCode — chats that talk to each other, agents that follow the rules, and a harvester that reads what the web won't show a bot.

Related MCP server: @cyanheads/pubmed-mcp-server

Why Professor?

You already run more than one AI chat. They cannot see each other, they forget the rules the moment context compacts, and a growing share of the web answers them with a 403. Professor is the layer that fixes all three — without touching what your harness is, only what it does.

Six things you can watch it do

Every transcript below is real output from this repository, redacted only of names.

1. See the fleet

pfm ls: every AI chat on the machine — Claude Code, Codex (⬢), and OpenCode — across accounts (🥇🥈), grouped by repo, live (●), resumable (↻), or agent-run (⚙). ⇄ marks chats that talk to other chats; ←here is the one you are sitting in; ✦ opens a new one on any harness. Pick one, attach, or fire it a goal without ever attaching. A chat that scrolled off a closed terminal tab is not gone — it is a resumable transcript, and now somebody can find it.

tab once more and the same fleet is drawn as the sky at the top of this page.

Every project is a star its chats orbit; a spawned chat rises as a moon at its parent's angle, so lineage is visible in the sky itself. When chats talk to each other the sky draws an edge between them, read from a durable comms ledger — an edge is a fact, not a guess. The chronoscope replays the last 24h, and a chat that is dead now still renders as the ghost it was back then.

2. Chats that talk to each other

Two panes, two harnesses. You type one line into the Claude chat on the left; the Codex chat on the right receives it as a signed turn and gets to work.

The footer on the Codex side is the signature: who spoke (sid 0a98b7fe) and the exact command to answer them. A message no sender could be derived for is refused, never delivered anonymously.

pfm chat call is what the Claude chat called — it types a real, signed turn into another chat's pane — under a per-target lock, safe against a busy target (--force-now) and shell-hostile payloads (--file). pfm chat ask waits for the answer, with named exit codes: 0 done · 2 usage · 3 chat dead · 4 no such chat · 5 answer timed out · 6 message not delivered · 7 answered, but another message reached the chat mid-wait. The chat family is registered as chat-professor, with bare MCP tool names such as call and ls, so an agent can spawn, message, read, and retire other chats across all three harnesses.

3. Rules that bite

A subagent tries to run git commit. The git-guard PreToolUse hook answers:

git-guard blocked `git commit -m wip`.
Only gitter writes this: spawn Agent(subagent_type: "gitter") with the repo path and the exact change.

The refusal names the right way. Its twin, pcm-guard, refuses a write to a CLAUDE.md, a .claude/ file or a template by anyone but the pcm agent, whether the Edit tools or a sed -i make it. That is one of the mandatory rules every install ships with: only gitter writes git; framework edits go through the pcm agent; and an error never renders as ABSENCE — "we failed to look" is never shown as "nothing there".

4. Read what the web hides from bots

About one in ten of the world's top 10,000 websites now tells AI crawlers to stay out — among news publishers, more than half (HasData AI Crawler Block Index, July 2026). Harvester fetches the way a reader's browser does and climbs a seven-rung ladder — direct → Chrome-fingerprint TLS → reader proxy → extractor → headless-then-headed browser → Wayback → OCR — until it holds real content:

A block is reported as a block and an app shell is never stored as the page — a failure names every source it checked, never an empty success. DOIs, ISBNs, PMIDs and PMCIDs route through twelve open-access resolvers in parallel; harvester ask -p "…" <sources> feeds the full cached artifacts to a Claude or Codex ask engine, failed sources kept visible as receipts. Harvester is a standalone Go plugin in plugins/harvester/. Its native engine plugins register the harvester MCP server and launch harvester mcp as a stdio proxy to the persistent daemon; the plugin guide lists each engine's measured tool names and install route. The browser rung never solves anything interactive.

5. One contract, three runtimes

$ engine-sync codex build .
$ head -1 AGENTS.md
<!-- Generated by engine-sync from CLAUDE.md — edit the source, then run: engine-sync build -->
$ engine-sync codex check .
CODEX CHECK PASS

CLAUDE.md compiles to AGENTS.md; .claude/** compiles to .codex/** and .opencode/**; global agents' .toml roles compile straight into ~/.codex/agents/ as regular files — Codex opens a role with O_NOFOLLOW and refuses a symlink, so a link there is a role that cannot spawn. Every mirror is generated, never tracked — a fresh clone has none of them until its compiler runs. The Stop hook recompiles the mirrors and a drifted mirror fails the check, so the three harnesses can never disagree about the law. Even the verification engine is one TypeScript source compiled for both the Claude Workflow runtime and the Codex SDK.

6. Reload without losing the conversation

Every seat's windows on one tab, confirmed against the provider every two seconds — you see 96% before it becomes 100%:

Account 🥇 is at 96% of its 5-hour window. The usage hook warns the chat mid-task; it finishes the step in flight and moves the conversation to 🥈 itself — same pane, same history:

The red banner is the usage hook — it reaches the model before the limit does. /reload typed by a human goes through a hook too, before the model sees it, so the reboot never spends a turn.

The running chat reboots in place — same pane, same history, new account, model, or effort. --then "prompt" hands the baton unattended; /reload typed by a human runs through a hook without spending a model turn; /handoff --branch carries the full context into a detached successor.

Install

Two independent halves; adopt either without the other.

• pfm, the fleet CLI — one Go binary, touches only your $HOME. Binary or source: INSTALL.md.

• The discipline layer — cloned, then scaffolded into your repo through an interview:

REPO=rezzminator/professor
TAG=$(git ls-remote --tags --sort=-v:refname "https://github.com/${REPO}.git" 'v*' \
  | grep -v '\^{}' | head -1 | sed 's#.*/##')
git clone "https://github.com/${REPO}.git" "$HOME/.professor"
git -C "$HOME/.professor" checkout "$TAG"
cat "$HOME/.professor/docs/SETUP.md"      # the install interview — start here

The checkout is pinned to the latest semantic version tag. A maintainer checkout also runs git config core.hooksPath .githooks so pfm doctor reports pre-push gate=armed. Upgrading? Follow the update workflow: pfm doctor --project-updates reports UPDATED / NEW / GONE-UPSTREAM / LOCAL-DELETED with the upstream diff under each UPDATED row, you carry what applies into your local file, and pin / ignore / drop record your decision — pfm never rewrites a project file after init.

Jumping several versions on one host? UPGRADING.md prints the host fixes in apply order with pfm install --check --plan, and covers single-account gateway hosts.

The discipline layer (templates/)

Clone it into a repo and you get the complete agent, command, hook and script template set; CLAUDE.md is yours to write, and pfm init points you at /quality:claude-md write CLAUDE.md for it. docs/SETUP.md walks an interview that substitutes your project's names into every placeholder; docs/PLACEHOLDERS.md is the substitution law. Every template is the live source file verbatim — never a skeleton.

The single idea underneath it is the honest-looking absence — an instrument that answers "nothing found" both when nothing is there and when the instrument itself is broken. The rules say it out loud:

An empty enumeration is never a verdict.

• One agent writes git. gitter is the only agent that commits; every other agent is read-only.

• Framework edits. A change to .claude/**, a rule file, a CLAUDE.md or templates/** goes through the pcm agent; it loads the quality-prompt contract before its first edit.

• The judge is never the thing being judged. Verdicts are read from disk, never from a brief that asserts green.

• The flight pipeline. Work too large to do directly goes to one foreman, its tier picked by difficulty and ambiguity, never size: flights-foreman-sonnet for a clear path where the pattern is in the code, flights-foreman for design choices, ambiguity, unfamiliar code, an unknown cause, several modules or the toughest work. The foreman reads what the build needs and writes the whole change by its own hand, spawning a child foreman only for an area sharing no file with its own. Once every foreman on the worktree has returned, the main chat sends one flights-lander: it computes its review plan from the diff, spawns cheap hunters, spec walkers and sub-landers in parallel, verifies every candidate on disk, fixes every finding, runs the gate, then gitter commits. /flights:doctor meters the flight from transcripts and git — landed lines per million input tokens — never from what an agent said it did.

• On-demand specifications. The user, or the main chat on the user's request, invokes flights-speccer for a spec and token price. It returns to its caller, outside the build/land lifecycle; no lifecycle agent or command invokes it.

• The persona is load-bearing. The Professor prompt replaces the vendor system prompt; the vendor baselines are pinned by sha256 so pfm doctor reports MATCHES / DRIFT / CHECK FAILED / CANNOT CAPTURE — never silence.

Professor ships no GPT agents. claude.gptAgents in the machine config pfm.config.json (default false, disabled) is the switch for the engine-sync GPT-twin feature: it declares the engine-sync capability gpt-agents, and engine-sync installs only agents whose requires that capability satisfies. No agent requires it today. gpt-gateway is a separate model route that serves GPT model IDs over your ChatGPT login.

A legal skill shelf ships for teams that want one.

The fleet CLI (pfm/)

One Go binary with embedded installer assets. Beyond the six moments above:

• Limits, honestly. The Limits tab shows every provider window on the box; a provider it cannot reach never renders as a 0% bar — the panel says no usage source registered instead.

• Crash-safety by construction. reap, archive, heal, and install default to a dry run, and the dry run is the apply's preview. heal backs up the store before it deletes a row.

• Headless exec. pfm headless exec is one scriptable interface for Claude and OpenCode; --engine codex also routes to OpenCode. Prompt, system prompt, schema, and timeout go in; normalized results or native streams come out. See headless README.

• Housekeeping. doctor runs the dependency registry and fleet DB checks; tokens attributes spend per agent from the telemetry ledger and the transcripts (ledger totals sum telemetry observations that can overlap, never a count of unique calls; telemetry with no session is kept apart from every chat, never guessed into one; transcript totals miss background calls such as auto_mode and prompt_suggestion, which only the telemetry ledger sees; and prices default to direct/global Claude list rates, so a Bedrock or Vertex setup sets its billing route in pfm.prices.json); /context-audit rebuilds and prices what each loop loads before its first tool call; statusline renders identity, session and spend; the mirrors' single writer is engine-sync, which the plugin hooks and pfm install run.

• Editor. pfm install --vscode installs the Professor VS Code extension — visible as Professor in the Extensions view and a Professor entry in the terminal + dropdown — and makes the PFM settings profile (its own icon and colour) the default, so each new integrated terminal opens at the fleet picker. To open a Professor terminal, press Ctrl+Shift+Alt+T (macOS: Cmd+Shift+Alt+T), run Professor: New Chat Terminal, or pick Professor from the terminal + dropdown — all three give the next icon and colour; the default + terminal is PFM. On a prompt file (SKILL.md, an agent, a command, CLAUDE.md, AGENTS.md; setting professor.promptLint.globs) the Prompt Lint link above line 1 of the file (it reads "Prompt Lint… Ns · Cancel" while a run goes; the palette's Professor: Prompt Lint does the same) has Claude, on your own subscription through pfm headless exec, check the editor's text against Professor's prompt law (/quality:prompt, built into the extension's system prompt by pfm install); its errors, warnings and info land in the Problems panel, each suggested rewrite a quick fix.

From pfm doctor's own registry: Linux or macOS, amd64 or arm64, plus tmux ≥ 1.8, git, python3 (checked, never installed by pfm install), sh, bash, zsh, and sleep; setsid on Linux, ps/lsof/launchctl/security on macOS. A clone host also needs the engine-sync binary. Go 1.27.1 or newer for source builds and pfm update. The claude and codex CLIs are optional diagnostics. Themes can be skipped with --skip-themes; the Codex probe with --skip-engine codex. Run the dry preview before applying. Harvester's native install routes, configuration and testing are documented in its plugin guide.

Plugins

Claude Code plugins built with Professor, installable on their own. This repo is their marketplace:

claude plugin marketplace add rezzminator/professor
claude plugin install forward-compaction@professor

• forward-compaction: the model writes its own handoff through one tool and the plugin installs it at Claude Code's compaction, with facts recorded from typed events and every later message kept verbatim. It replaces the earlier sub-agent-compact plugin, whose settings it no longer reads (its CHANGELOG maps each one); pfm install disables an enabled sub-agent-compact copy, and fails a claude.plugins entry that still names sub-agent-compact until it is replaced with forward-compaction.

• cache-live-control: a /cache command that switches the prompt-cache TTL (5m, 1h or automatic) for one chat and its sub-agents, live, with no model turn.

• auto-effort: reasoning effort for the main chat and for the sub-agents it launches. Each main-chat turn starts at a planned level, the model raises or lowers it with a set_effort tool, failing tests raise it, and a per-model cache check switches it off where an effort change would break the prompt cache. Start the Agent prompt with [effort: low|medium|high|xhigh|max] to set a sub-agent's effort per spawn; an agent whose frontmatter opts in with effortControl has its effort rise at each failing test run or red sub-task and come home once the trouble resolves. It replaces the earlier auto-effort-control and effort-live-control plugins, carrying their settings over (its README and CHANGELOG state the migration).

• tool-visibility-controller: one visibility field in agent, skill and command frontmatter decides, per agent, which sub-agents, skills, MCP tools and built-in tools it sees and may call — hidden from the listing and refused, never just discouraged.

• rule-control: one visible-to key in a rule file's frontmatter names the loops that receive the rule — the main chat, chosen agent types (globs allowed), or all; a targeted rule leaves the context every loop shares and reaches only those loops, again after each compaction. Claude Code only: engine-sync still carries such a rule to every Codex and OpenCode loop.

• buddy: a tiny ASCII companion above your prompt — Quack the duck, seven more characters or your own — that comments on each turn, suggests your next prompt and answers /buddy questions from what it remembers of the chat.

• callmeter: records every tool call, model request, sub-agent turn and session event into one local SQLite store — counts and sizes, never your prompts or file contents — and answers which files, commands and agents cost the most.

• engine-sync: carries a project's Claude Code agents, commands, skills and MCP servers to Codex and OpenCode — AGENTS.md, .codex/, .opencode/ and the role TOMLs in each Codex home — and keeps them in step after every turn; at pfm install, engine-sync global also carries the always-on global rules of ~/.claude/rules/ to both engines; run it as engine-sync codex ... and engine-sync opencode ....

• cache-keepwarm: keeps the main chat's prompt cache warm while it waits on background agents, shells, Monitors and workflows — one invisible replay before expiry, admitted only while the forecast spend stays under the rewrite it saves; /keepwarm on|off|status controls it.

pfm install ensures the plugins claude.plugins in pfm.config.json lists; unset, that is cache-live-control, forward-compaction, auto-effort, tool-visibility-controller and engine-sync. A bare name is one of ours, installed as {plugin}@professor on every build. A release build installs it from this repo's marketplace, rezzminator/professor. An alpha build of pfm installs it instead from a local checkout: the working tree of {plugin}/plugins/{plugin} under the source clone's plugins/, where each plugin is a git submodule on its develop branch (git submodule update --init), or under the directory claude.pluginCheckoutRoot in pfm.config.json names, which wins when set: it copies each into ~/.local/share/professor-dev/plugins/{plugin}, lists it in that directory's own marketplace, also named professor, and installs {plugin}@professor from there. An alpha build also installs every other plugin the source clone's .claude-plugin/marketplace.json lists the same way, so claude.plugins is the enabled set: a listed plugin is enabled, and one the list does not name is installed from its checkout and disabled. A plugin with no checkout is not installed on an alpha build, never from GitHub, and the line names the root it tried; for a listed one that is a failure. Unset on an alpha build, the list is every plugin the marketplace lists, all enabled; a release build's default stays the five above. The install re-points a professor marketplace Claude Code knows from another source with claude plugin marketplace add, never automatically removing it. A printed manual repair may require claude plugin marketplace remove; it deletes the plugins' saved options and data, so run it only if they may go. A plugin installed under an older id, {plugin}@{plugin} from its own GitHub marketplace or {plugin}@{plugin}-dev from ~/.local/share/{plugin}-dev, moves to {plugin}@professor once that is installed: its pluginConfigs options are copied over, each old install is uninstalled with its data kept (one in a project scope is first installed as {plugin}@professor in that project), the old marketplace is removed once nothing is installed from it, ~/.local/share/{plugin}-dev is deleted, and any old key left in enabledPlugins or pluginConfigs is dropped; pfm doctor names each old copy left. The install builds an enabled callmeter's binary from its checkout into {callmeter home}/bin/{version}/callmeter before refreshing its copy, keeping the binary it replaces as callmeter.prev. Claude Code reads an alpha checkout copy straight from its folder (readFromFolder in claude plugin list --json; claude plugin update answers that there is nothing to update), never from its versioned cache, so a checkout edit, committed or not, reaches a new chat or an open chat's /reload-plugins with no version bump, and a running chat's next hook call can already run newly mirrored files. The install also runs claude plugin update for each plugins/installed_plugins.json entry behind the checkout's plugin.json version. On a release build Claude's versioned cache is what runs: a content change ships with a version bump, and an open chat gets it when it restarts. A listed plugin the install leaves off its checkout copy or disabled is a FAIL line naming the plugin, the account dir and each cause, the summary line counts it as failed=N and the install exits non-zero; pfm doctor reports the same as FAIL, and fails when the callmeter binary the install placed is gone. A full name@marketplace id is installed as named, from Claude Code's own marketplace of that name when it knows one; an entry {"id": "name@marketplace", "source": "owner/repo"} adds the marketplace first (source is what claude plugin marketplace add takes: owner/repo, a git or https URL, a path). An id with neither is reported by name, with its repair, in install and doctor. A plugin neither the list nor, on an alpha build, the marketplace names is never touched in any settings.json.

A plugin option is also an environment variable named {PLUGIN}_{OPTION}: the plugin's name and the option in upper snake case, as in AUTO_EFFORT_FLOOR. The variable beats the option, which beats the default. Two plugins differ: forward-compaction has no options and reads its settings from these variables alone, as in FORWARD_COMPACTION_CEILING; buddy reads none and keeps every setting in its own file, ${XDG_CONFIG_HOME:-~/.config}/buddy/config.json (its README). claude.env in pfm.config.json maps each variable to its string value, and pfm install writes the map into the selected config directory's settings.json env (the primary account by default, normally shared through the accounts' links). Doctor checks the last successful install's recorded target, including --config-dir. pfm owns only the variables it wrote, recorded in ~/.local/share/pfm/install/claude-env-ownership.json: a variable dropped from claude.env is removed unless it was changed by hand since, and one pfm never wrote is left alone. Alpha installs default an unset PLUGIN_LOG_DIR in each managed account directory to the shared ~/.local/state/pfm/plugin-logs folder, or the absolute path set by claude.env.PLUGIN_LOG_DIR; an operator-set value is neither overwritten nor taken into ownership. Release installs and uninstall retire generated settings only while they remain pfm-owned, never deleting the logs. A chat writing that settings.json while pfm edits it keeps its write: pfm reads the file again just before replacing it and plans its edit again on what it finds, at most three times, then fails naming the file — rerun pfm install. Apart from the known PLUGIN_LOG_DIR knob, a variable no installed plugin's name prefixes is a warning. claude.env may not set the launch-owned CACHE_LIVE_CONTROL_MAIN_TTL, CACHE_LIVE_CONTROL_AGENTS_TTL, CALLMETER_HOME or ENGINE_SYNC_HOME; every launch sets the main cache TTL from claude.cache1h. pfm doctor fails a variable missing or different from claude.env, and one left behind after its removal.

pfm install installs no engine-sync, and on any host with a Professor clone (the discipline-layer clone above, every pfm update host) or a Codex home it refuses without the binary, naming the missing path; a binary-only install with neither skips the step. So install the plugin before the first pfm install --yes or pfm update: claude plugin marketplace add rezzminator/professor, then claude plugin install engine-sync@professor, then start a new Claude session. At that session start its wrapper downloads the release binary, checks it against the plugin's SHA256SUMS, and links it at ~/.local/share/engine-sync/bin/engine-sync, the binary the plugin's hooks run; test -x on that path confirms it. Once that binary exists, pfm install --yes links it onto PATH as ~/.local/bin/engine-sync (an existing entry there that is not pfm's link stays untouched, and pfm doctor warns).

Origin

Extracted from a live production monorepo, not designed in the abstract. Every rule here exists because something went wrong without it — the gate that reads disk instead of chat exists because an agent once claimed green; the scoped-commit rule exists because two concurrent commits once swallowed each other's files; the prevention step exists because the same bug class shipped twice. The characters exist because a generic agent wasn't good enough to argue with.

Built by @rezzminator.

Help and community

Questions and ideas go to Discussions, bugs to an issue, vulnerabilities privately through SECURITY.md; SUPPORT.md has the details.

Contributing

Work lands on develop; main moves only by release. CONTRIBUTING.md carries the build, the gates and the rules of a repository whose files ship into other people's agent pipelines. Pull requests are welcome.

License

MIT