Status: ACCEPTED, 2026-11-16. xian’s word at 06:20 PT, relayed by Exec (mail 01M7JBP51M), verbatim: “1. Yes to ADR-084 and budget 2. Nothing I run personally uses that key, no”. The figures: a seat may spend $10 a run without asking, and the lead-scoring workspace’s $150 monthly limit is the hard cap. Paid runs stay stopped until Step A has put the seat key where a seat can use it and Lead’s check passes. Drafted 2026-11-15 (Arch-authored, from Exec’s docs/internal/operations/seat-test-harness-proposal-2026-11-15.md, draft 2 at 7c571b8fec, with CIO’s, HOST’s and PA’s replies of 19:26 to 19:28).
Related: #2340 (the developer-key binder), #2364 (the provider gate that stopped), #2327 (ops_reader), #2373 (the floored counter), #2374 (reads limited to the installed set), #2381 (the screen atlas), #2382 (stateless MCP server), ADR-080 (code decides permission).
Between 11-10 and 11-15 four rules landed, each sound alone:
fly; production reads and secret changes are xian’s (standing);Every one of them protects xian’s own state. Their sum removed each way a seat had to spend on a model, sign in to beta, or read a production fact. Each loss was patched with “xian does it once”. On 11-15 the patches were a queue of eight manual steps, and xian refused the pattern: “We need to solve that problem, not assign me manual tasks as a workaround for a poorly conceived solution.”
Two facts from that day shape the decision.
scripts/dev_key_binding.py (Arch’s 11-11 ruling) named two key sources: the command-line environment, and the encrypted database store. The store needs ENCRYPTION_MASTER_KEY, which scripts/local_key_setup.py keeps in the worktree’s .env. The binder never loads it. The route was ruled and never run end to end. The fix a seat reached for (a launcher that loads .env) was refused by the classifier, correctly.~/.piper-shared holds files seats made for testing (CIO, 19:26: 17 files, names only read). The rulings of 11-10 to 11-13 and the classifier cut seats off from them. No rule had named them as different from xian’s own.A seat resource is something that exists so agents can run and test Piper. It is:
lead-scoring ($150 a month) is the workspace for seats’ own runs. piper-morgan-test (xian’s $40 limit, which paused the workspace at its limit on 10-29) may be the workspace for keys stored on seat accounts, which pay for served chat reads on beta. The account table’s exception row for web-seed-1386 does not wait on xian: the key, its workspace and the limit are his own setup of 10-20 and 10-29, the helper’s reads are GET only and can’t spend, and the row only stops the helper refusing the account. A seat’s chat turns on that key do wait on his word through Exec (two short answers: nothing of his own uses the workspace; its key on the test account counts as a seat key at $40), since his 11-16 yes named lead-scoring only. If he says something of his own uses the workspace, the row comes out. Answered 2026-11-17 13:31 (xian via Janus, relayed by Exec): to “does anything he runs himself use the workspace”, “No”; to “may seats spend through that key up to its $40 limit”, “Yes, if that’s the right way to do it, sure”. Arch’s answer to his condition, the same day: this conforms to D2 as written here, because the workspace now passes D1’s test and its provider limit is a stop a seat can’t pass (it stopped chat on 10-20 at $10 and on 10-29 at $20, by Anthropic’s own notice and by served refusals). The “one cap enforcer” bullet below is about a seat’s own runs; it does not reach served requests, and nothing here pretends it does. HOST answers the custody half. Not verified by any seat: that the limit reads $40 today (the record is PA’s note of xian’s raise on 10-29); that is his console’s to show. Read by xian on 2026-11-17 (figures as Janus transcribed them from his console, relayed by Exec): the workspace shows $16.19 spent of a $40 monthly limit, resetting 1 December UTC; the organisation shows $49.29 spent of a $250 monthly limit. The organisation limit sits above every workspace limit and is not a seat bound. It is xian’s cap for the whole project, and it counts everything in the organisation, seats’ two workspaces included. The seat bounds stay the workspace limits ($150 and $40, so at most $190 of the $250 in a month). One consequence to keep in view: if a key that serves testers is in the same organisation, seat spend uses the same $250 it draws on, so the room left for served Piper in a month is $250 less what seats spent. Whether that key is in this organisation is not recorded in this repo; it is his console’s to show. A key stored on a seat account spends through served requests, which bind no in-code cap, so for those reads the workspace limit is the whole bound and the whole budget: there is no per-run stop. The seat logs the planned turns before a sitting. An after-cost a seat can read needs a count-only route and is not built.env entry in the seat’s own git-ignored .claude/settings.local.json, under a Piper-specific name: PIPER_SEAT_ANTHROPIC_KEY, PIPER_SEAT_OPENAI_KEY. The file is mode 0600, exists only in the worktrees of seats that run paid calls, and carries a Read deny for itself in those seats’ settings, so the value can’t enter a transcript through the Read tool. A seat checks its settings with a script that prints key names only. The Read deny in a seat’s own settings stops that seat’s Read tool only: CIO’s probe on 11-15 showed a **/name deny is honoured inside the seat’s worktree and does not reach a file outside it. So a second deny, naming the settings files by their path from the home folder, goes in the one settings file all PM seats share, and it is xian’s edit. Both denies are tested before any real key is in a file (Lead’s order): a failed test then shows nothing. (A session variable has to come from a file; this is the harness’s own, and every other launch surface is one of xian’s files.) The name differs from ANTHROPIC_API_KEY because Claude Code uses that one and our launch lines strip it. CIO tested the shape with a dummy value on 11-15: the settings env surface reaches the shell, and the command ran unrefused.ANTHROPIC_API_KEY="$PIPER_SEAT_ANTHROPIC_KEY" after the env -u flags, with PIPER_CREDENTIAL_STORE=db. The shell expands the name; the command text never holds the value. No file is read, no keychain, no .env.developer_keys_bound(providers=..., max_usd=...) refuses to start without a cap. The spend chokepoint (#1809, services/llm/clients.py) adds each response’s usage and raises before the cap is passed. Batch submits check the worst-case cost of the rendered requests first (router_batch.py submit --cap-usd, on main at 1e07f0ba74, is that half). The binder half is on main at 857909e2e2 (issue 2383). The in-code cap is per run and in process: a second run starts at zero, so across runs the bound is the workspace’s monthly limit, and the log lines are what make a series of runs visible. A cap refusal ends the run; a run with a refused call is incomplete and its rows are not counted. The binder refuses a cap above the per-run figure unless the run names Exec’s relay of xian’s approval..env is read, and the database is dropped with the run. This is the seat key in a seat’s scratch store, not a new route.scripts/mailbox_bearer_lint.py; its baseline holds paths and lines, never a value). GitHub’s push protection is on, and untested on this key’s shape.env output, a transcript), it writes only the masked form, says so, and Exec has the key rotated. Seats don’t print their environment.scripts/seat_session.py --account <name>, issue 2388, on main at ff8b9d51b7) reads the account’s file, signs in, and saves a session: the access cookie only, in one browser storage state file that serves API reads and Web. The 7-day refresh cookie is never kept, so a session lasts as long as the access token inside the cookie, 30 minutes (JWTService default access_token_expire_minutes=30, no caller overrides it; Web measured the printed expiry on 11-22), and a seat re-runs the helper. (Corrected 2026-11-22 by Arch: this said 24 hours, which is the cookie’s max_age, not the token’s life.) scripts/seat_read.py is the client: GET only, from a closed list of paths in code, each addition a diff HOST reads. It prints the account name and the expiry only. The grant is an exact-command allow for that helper, added by xian. The Read denies on ~/.piper-shared stay as they are; the session file lives there under a name the denies match.
seat_session.py run places the cookie again. The browser holds a seeded account’s session and nothing else: no other sign-in is typed into it and no Connect round trip with a real provider account is done in it.dev/active/seat-grant-dummy-test/README.md at 9e10ac77db: a stand-in helper placed an httpOnly cookie through the debug port with Storage.setCookies, it survived a browser restart on the same profile, and the pinned browser tool attached with --browserUrl read the page signed in; not yet run with beta’s own cookies or in a seat’s live session). Three build points follow from it: the debug port has no authentication, so it binds to 127.0.0.1 only, the browser behind it is a dedicated seat browser and never xian’s own Chrome, it is not started with a wildcard --remote-allow-origins, and its port is read from one place so a stray second browser is noticed (HOST’s bound: what that browser holds is the whole exposure); .mcp.json needs an entry that attaches with --browserUrl for served-read seats; and a LaunchAgent starts that browser and keeps it up (Pard builds it with a cycle-check arm that asserts the port answers; HOST reads it; xian loads it once). The session is two httpOnly cookies (services/auth/session_cookies.py: access cookie max_age 24 hours but the token in it expires after 30 minutes, refresh 7 days, samesite=lax, bound to no address or browser). Page script can’t set them. Today .mcp.json starts the browser tool with --isolated, so every browser restart signs every seat out. The intended shape: each seat that does served reads gets a persistent browser profile outside the repo (--userDataDir, or --browserUrl onto a browser that stays up), and the helper places the access cookie into that profile through the DevTools protocol (the access cookie only, as the helper saves it; the refresh cookie is never kept, so the browser is signed in for at most 30 minutes, the access token’s life, and the seat re-runs the helper). Built as issue 2400: scripts/seat_browser.py proves what listens on the port is that seat’s browser (one process, 127.0.0.1 only, the cohort’s test browser, on that seat’s profile folder, started with --use-mock-keychain and without --remote-allow-origins) before any cookie is sent; the launcher that attaches the browser tool makes the same proof before it attaches (Arch, 2026-11-16). The password and the cookies never pass through a tool call. CIO and Pard own the launch configuration; Web runs the dummy check. Web read its tool surface on 11-15: the browser tool has no set-cookie, storage-state or raw protocol call, so the helper must place the cookies through the browser’s debug port; Web’s part is to open the page and confirm the signed-in name, after a restart and after a promotion.--use-mock-keychain, so it never asks the login keychain for a storage key. Two mechanisms hold this, both CIO’s, on main at 6f5bfdab94, b31a9b5d54 and b14050a176, read by HOST and by Arch: a PreToolUse Bash hook in project settings (.claude/hooks/guard-browser-launch.sh) refuses a Bash command that starts a Chrome or Chromium program without the flag, and scripts/browser_keychain_check.py, run by every seat at every fire, lists any such browser of this user that is running without it (pid and program path only; it never looks in the keychain). What they are and are not: the hook reads the command’s text, so it is a guard and not a bound. It does not see a launch from inside a script file or another program, Since b14050a176 it fails toward refusing: a command with a browser program among its words is refused unless the flag is there or the command word is on a short list of readers (NOT_LAUNCHERS); the contents of $( ) and backticks are judged as their own commands; a bare program name is refused anywhere in a non-reader command; osascript naming the browser is refused. Arch’s runs at that commit (29 forms, then 19): every wrapper, if, while, eval, !, a leading redirect, bash -c, variables, backticks and command substitution are refused; reads (ls, ps | grep, a commit message, a heredoc going to a file) pass. A command that pipes into a shell loses the reader excuse (f58d8cfd79). Still passing, with no flag: a launch from inside a script file or a program on the reader list (python3 -c; the list holds python3, node, npx, brew and git because seats name the browser to them without starting it). The check sees only what is running when it runs, and it takes a browser under /Applications to be xian’s own. A prose rule in CLAUDE.md also says it; that is not counted as a mechanism. Measured firing in five seats’ sessions of eleven by 09:17 on 11-17 (CIO, PA, Arch, HOST, Exec), six untested. A seat has the hook only after its worktree has merged the commit and its session has re-read settings, which is not instant (Exec’s first test, seconds after the merge, reached the shell); the one-line test in that seat is the only proof.A fact seats need more than once gets a read-only route behind ops_reader. The rules for every such route:
ops_reader is checked on every request. The flag is set only from the admin control, by xian’s hand. No chat or agent path can set it, and it never implies is_admin.The first routes: active keys by provider; the floored-action counter (served); the routing-to-floor tally. A count that would single out a described tester is written in this public repo as “none” or “at least one”.
A one-off fact is asked of xian once and not built.
One database per seat on the shared Postgres server (CREATE DATABASE piper_<seat>), named in the seat’s launch environment so the test tree and alembic pick it up with no per-command flag, and migrated to head at START. No real person’s data is ever copied into one. It is the pattern tests/integration/test_fresh_database_setup.py has used since #2372. Not per-seat containers: the compose file fixes container names and ports. A fresh-clone guide run is the exception and uses a compose override.
A ruling that puts a step on xian’s hands names, in the same text, the standing access that would remove it. A ruling-relay with xian as the actor and no such line is incomplete. Exec checks for the line; CIO may make mail4.py warn on its absence.
Two further items were named in review. One is accepted in shape with its bound; the other is not solved:
workflow_dispatch job in fly-deploy.yml holding a token scoped to the MCP app only. HOST’s bounds: the secret sits in a GitHub environment limited to main, not a plain repository secret (otherwise a branch with an edited workflow could read it); the job takes no free-text input, deploys main’s sha, and prints sha and release only; the token has an expiry if Fly allows one. Unverified: whether a Fly deploy token can also run commands inside the app. If it can, the workflow text on main is the bound, not the token. Lead builds the job. PA’s order, accepted: this job is built first among the items here, and #2382’s MCP deploy is its first run, so that deploy is never a standalone step for xian.Nothing here lets a seat run fly, read the keychain, or pass a classifier refusal.
main (D6), not a plain repository secret. The job is #2384.Adds the GitHub-connected seat account to the #2242 allowlist (one production write), then adds a second connector in his claude.ai settings and signs in to Piper as that account, once.
Rows 4 and 9 include clicks on GitHub’s and claude.ai’s own pages. No seat drives those.
Settled since (11-16): the workspace is lead-scoring. Nothing of xian’s own uses the older lead-scoring-2026-10 key, and he disabled it on the console at about 06:23, so the workspace holds one live key, seat-screens-2026-11. Arch accepts disabled in place of deleted: a disabled key can’t make a call, and only xian can re-enable it. “Operations reader” is ticked on web-seed-1386 (list row 6).
lead-scoring can serve only if the older lead-scoring-2026-10 key is revoked and nothing of xian’s own still spends there. If anything does, the seat key needs its own workspace. A lowered limit also has to allow for what the workspace already spent this month.piper-morgan-test, and a count-only route for the after-cost.env entry: whether a restart is needed (it shows at the real install).max_usd and the chokepoint counter; seat_session.py; the ops_reader routes and their emit-site counters; the closed-list pin; the MCP deploy job; #2382.