haiku's review of PR #199 ended mid-word at "- **astra.py, `parked_c" and then the footer: it ran into think()'s max_tokens: 4000. The verdict was Object, so the card went to needs-human on a review whose last point nobody can read. - think() returns a Reply: the text as before, carrying the API's stop reason (finish_reason: length in the engine's OpenAI shape, stop_reason: max_tokens from the Messages API), and takes a max_tokens. A caller that doesn't care uses it as the string it is, so nothing else changes. - review_reply(): a cut-off review is asked once more with TIME_REVIEW_MAX_TOKENS (default 8000). Cut off again, it is posted with CUT_NOTE first, which keeps VERDICT from matching the half-written verdict, and post_review posts it as a COMMENT rather than REQUEST_CHANGES. So it casts no vote: the council waits or the tie-break seat votes, and a truncated review is never an Object on its own. The card's beat carries cut_off and its note says a review ran into the limit. - The reviewer prompt asks for the verdict on the first line and at most five points, most important first, each a line or two. A longer, complete review might also have caught that its first point on PR #199 was wrong (router.pick_cloud does skip n is None). - doctor lists TIME_REVIEW_MAX_TOKENS. - Tests: a reply with finish_reason: length is retried at the higher limit (and at the value TIME_REVIEW_MAX_TOKENS sets); cut off twice it is posted with the note, as a COMMENT, matching no VERDICT, with or without a seat token; a complete reply votes as before in one call at 4000. test_time.py, test_seed.py and test_housekeeping.py green (the opt-in TIME_TEST_VENTURE suite fails the same way before this change). Workflow change this needs, not made here (.github/workflows/ is not the worker's to touch): in .github/workflows/heartbeat.yml, in the heartbeat step's env block, after the TIME_REVIEW_QUORUM line, add TIME_REVIEW_MAX_TOKENS: ${{ vars.TIME_REVIEW_MAX_TOKENS }} # a cut-off review is asked again with this (default 8000; #202) Without it the 8000 default still applies; only setting it as an Actions variable needs the line. Closes #202 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016D3hR5GpJTPoN44WP3J7oM
time
A Gitea-native agent factory, running on Teahub prod (teahub.xyz). The Gitea repo is the ledger, Gitea Actions are the clock (Time), the worker is a Claude or Venice seat with hands or a Cursor cloud agent, and chat is where you plan.
- Cards: an issue with a
DONE:line, labelledready. - Heartbeat (
heartbeat.yml, hourly and whenever a card is labelledready): moves cards forward. Indry-run(default) it only says what it would do; withTIME_MODE=liveit dispatches each card to the worker (TIME_WORKER: a seat such asclaude:sonnetorvenice:kimi, which edits the repo and opens the PR itself, or a Cursor cloud agent), which opens a PR that closes the card. An optional reviewer (TIME_REVIEW) checks the PR. At mostTIME_DAILY_CAPlaunches a day. Spend governor: every model call is booked totime/ledger/<YYYY-MM>.jsonl(tokens and dollars); once today's spend reachesTIME_DAILY_USD(default 5) or the month's reachesTIME_MONTHLY_USD(default 100), Time stops calling models, opens anandonissue, and a monthly breach also setsTIME_MODE=dry-run. No model decides whether to spend. - Mentions (
mention.yml):@sonnet,@opus,@kimi,@grok 4.7,@qwen max highin any issue, PR or comment gets that seat's reply. - Hands-off (
TIME_AUTOMERGE=on): label an issuespecand Time splits it into cards that run one at a time. A card's PR merges itself once the reviewer approves and checks are green, then the next card starts. Objections, red checks and protected paths (workflows, deploy, secrets) go to a human. - Venture loop (
venture.yml, every 20 min onceTIME_VENTURE=on): one venture card a day, or one right away from Actions → venture → Run workflow → "Start a new venture now". It scouts the repos that gained the most stars this week; seats pitch; the board votes (Borda, quorum of two); every board seat writes a launch plan (positioning, first users, pricing, channels, 7-day success and kill lines) and one seat synthesizes them intoLAUNCH.md; the worker builds the product as the page ongh-pages(the tool fills the first screen under a slim header with the one signup button; pricing, limits and the why sit below); deterministic checks and headless Chrome (phone and desktop) gate it, with two fix rounds before an andon; GitHub Pages ships it at/ventures/<slug>/. Signups go through an issue form (.github/ISSUE_TEMPLATE/early-access.yml), each one thanked on its issue; after 7 days the count of distinct people decides: build v2, kill, or watch a week more. The studio index lives at/ventures/. Launch posts are drafts for a human; nothing is posted outside the repo. - Batches (
TIME_VENTURE_BATCH=3, or Run workflow → "batch"): one scout pass, a field of pitches, and the board's top k become venture cards in one milestone, built side by side and launched together. A held child gets one batch retry before it's counted dead; a retro compares the cohort once every child is measured. - Seed (
seed.yml, manual): land a venture that was built somewhere else. Give it a slug, the URL of a .tgz whose top folder is that slug (or of its parts, in order) and the .tgz's sha256, plus a name and a one-liner for a new venture.seed.pychecks the bundle, copies it toventures/<slug>/ongh-pagesin one commit, adds or refreshes its line inventures/ventures.json, and asks Teahub pages to pull, so it is live at<slug>.yolopreneur.lolwithin the minute. Nothing in the bundle runs. - Refresh (
refresh.yml, manual): for a venture that lives on gh-pages and makes some of its own files (fonts, art, screenshots, generated code). It runsventures/<slug>/tools/build.shon the runner, in place and without credentials, then commits what changed to gh-pages in one commit and asks Teahub pages to pull.fromtakes the folder from another branch first, so a reworked venture lands whole. - Backends (optional,
CLOUDFLARE_API_TOKEN+CLOUDFLARE_ACCOUNT_ID): each venture also gets its own Cloudflare Worker, D1 database and Workers AI, serving its Pages folder at<slug>.<TIME_VENTURE_DOMAIN>. The page talks to it throughtime.js(share links that save a creation, a rate-limited small model, and accounts with passkeys: no passwords, no email); every call degrades gracefully, so products still work on Pages alone. Code:.github/time/cloudflare.pyandcloudflare/worker.js. - Doctor (
doctor.yml, manual): runs the tests and checks label, settings, keys and today's spend. - Update (
update.yml, weekly or manual): opens a PR with the latest engine from teahubxyz/time, so every project stays on one Time instead of drifting copies.
Setup
Started on GitHub as a proof of concept (teahubxyz/time, now read-only); ported to Gitea 2026-10-14.
Open this repo in Claude Code or Cursor and say "set up Time": the agent follows .claude/skills/time/SKILL.md. The recommended path needs one secret, ANTHROPIC_API_KEY: Claude seats plan, build and review (TIME_WORKER=claude:sonnet, TIME_REVIEW=claude:haiku; haiku, sonnet, opus, fable are the seats, priced in .github/time/seats.json). With VENICE_API_KEY instead, Venice seats do the same (TIME_WORKER=venice:qwen coder, TIME_REVIEW=venice:grok). A Cursor cloud agent can be the worker instead if you have a Cursor API key. The landing page walks through it: https://teahubxyz.github.io/time/
Status
For changes and release history, see CHANGELOG.md.