Repository navigation
Description
Goal: a fork the owner can trust, so they only make the high-level decisions. Technical work and advice are right, agents do what is expected, each kind of tooling sits in the right place, and the system finds problems nobody thought to ask about (like poor tooling) before they waste time.
Every known gap is listed below with its own issue. Tick an item when its issue closes.
1. Unattended work cannot waste usage, loop or burn CI
- G19 Hard budgets and a loop detector: unattended work stops itself #860 Hard budgets and a loop/stall detector (usage pacing on enforce)
- Capacity rules enforced by the tools, not by the agent's command line #866 Capacity rules enforced by the tools themselves
- Model guard: turn on "only managed hooks" so other hooks cannot change a checked call? #298 Model guard: turn on (owner question)
- Watch every workflow: alert when any workflow is switched off #867 Alert when any workflow is switched off (with Owner question: who switches off fork-tests every 20 minutes? (2-minute security-log look) #517)
- Flaky: Kiro replay turn_interrupt test sees 'active' instead of 'idle' #846 Flaky Kiro replay test (fixed at its cause) (closed)
2. Agents do what is expected
- Check a sub-agent's final report against reality (PR head, CI state, comments, leftovers) instead of trusting it #810 Check a sub-agent's final report against reality
- Root-cause fixes enforced by a check, not only by instructions #864 Root-cause fixes enforced by a check
- Merge gate: a comment that only looks like a lens report must not block a PR #865 Merge gate: a comment that only looks like a lens report must not block
- Orchestrator logic as software, not prose #870 Orchestrator logic as software, not prose
- G8 Agent slot counting: one source (server.agentSlots) #849 One agent slot count
3. Advice you can trust
- Advice you can trust: evidence and an independent check for big recommendations #868 Evidence on claims; an independent check for big recommendations
- Owner-question convention: Issues page Q1: how far should we change T3's Issues page? #828 Issues page Q2: may T3 post your answer to an owner question for you? #829 Issues page Q3: should upstream's issues stay out of our projects? #830 Issues page Q4: may our owner-question wording be the public fork's defaults? #831 (owner answers)
4. Finds what nobody asked about
- G18 Project kickoff tool-fit audit: find poor tooling before work starts #859 Project kickoff tool-fit audit (poor tooling found at the start)
- Blind-spot explorer: a scheduled audit that finds what nobody asked about #869 Scheduled blind-spot explorer
- Waste report: make usage, CI minutes and idle memory visible #871 Waste report (usage, CI minutes, idle memory)
5. Efficient
- G20 Memory reservations: hold only what is used; stop capacity flapping #861 Memory reservations hold only what is used; no capacity flapping
- G21 Faster merges: batched merge queue and shared build caches #862 Batched merge queue and shared build caches
- G7 Merge path: one merge queue and one merge CLI in the fork #848 One merge queue and CLI (does not stop at an unrelated flake)
6. The right tooling in the right place (#834: general tooling lives in the fork)
- G1 GitHub App tokens: fork-kit/gh-apps: call forms and GH_REPO check of the older per-user gh-as #835 (merged)
- G2 Capacity and build slots: heavy facts: per-kind fits from the kernel ledger; kit capacity evaluator and callers #786 (merged), fork-kit/heavy: machine-file holds for heavy facts / t3-capacity (stacked on #786) #843 (merged), heavy capacity: callers without a systemd user manager always get NOT OK (CI runner users) #844 (open)
- G4 Guard hooks kit: fork-kit/agent-guards: the machine's agent guard hooks as a kit (#834 G4) #847 (merged)
- G5 One agent-rules doc (in progress)
- G6 One review skill: fork-kit/review: one general review skill; tiers, failure table, nit-only check and a drift test #845 (merged)
- G9 fork-adopt: starting a new project is not hand work #850 fork-adopt: starting a project is not hand work
- G10 Orchestrator state, handoff and session hooks as a fork feature #851 Orchestrator state, handoff and session hooks
- G11 Guaranteed wake-ups: one mechanism #852 One wake-up mechanism
- G12 Runners: one routing, one sandbox, one checkout-heal #853 Runners: one routing, sandbox and checkout-heal
- G13 Upstream sync of forks: keep fork-follow, make it config-driven #854 Upstream sync config-driven
- G14 Cleanup and install drift: one cleaner for every project #855 One cleaner for every project
- G15 Deploy and release with migration and rollback: general flow, project adapters #856 Deploy/release with migration and rollback
- G16 Leases, machine access, scheduled jobs, backups: general kits #857 Leases, machine access, jobs, backups (shared test machines)
- G17 Interaction-limit keeper: one scheduler #858 Interaction-limit keeper
- Fork code: small project leftovers (R1-R3 of the 11-21 audit) #872 Small project leftovers in fork code
7. Ready for very large projects (e.g. a game engine at 144 FPS)
- G22 Graphics/performance lane: what a game-engine-scale project needs #863 GPU/Windows runners, frame-time gate, big assets, engine build caches, a design-decision workflow
8. Planned features (features from the fork's plans; each issue says what is built and what "done" means)
- MCP profiles: finish the feature (Settings pick, T3 tool groups, Codex, optional narrowing) #875 MCP profiles: finish
- Model policy: finish the feature (backstop, Claude sub-agents, Settings panel, handover) #876 Model policy: finish (with Model policy: write the Opus 5.5 / Fable 5.2 rows and check picks on the server (#125-#128) #731 model rows)
- Handoff instead of compaction: finish the feature (agent source, thread menu, roles, later items) #877 Handoff instead of compaction: finish
- Calm sidebar and thread roles: show only the assistant and the project orchestrators #878 Calm sidebar and thread roles
- Personal assistant across projects (setup, mail read-only first, drafts, phone, texting look) #879 Personal assistant across projects
- Armed PR sweep: a red or conflicted auto-merge PR wakes its owning thread within 10 minutes #880 Armed PR sweep
- Account failover: move work to the next account in a set order before one runs out #881 Account failover
- Dynamic agent slots: finish D4-D8 so slot counts come from the machine, not from instructions #882 Dynamic agent slots D4-D8
- Placement: rules decide which machine runs work (P1, P3, P4, P6-P9) #883 Placement by rules
- Machine loss: drain a machine, tell the parent when a machine is lost, move threads #884 Machine loss
- Host tools: agents use the owner's other computers safely (four stages) #885 Host tools (four stages)
- Updates in each machine's own hours, only when quiet; show machines that are behind #886 Updates in each machine's own hours
- Issues page: "Waiting on you" view and Inbox for owner questions (after #828-#831) #887 Issues page: "Waiting on you" and Inbox
- Review rules as a fork feature: finish the judge (C3, C4) and the lens tools (T3-T5) #888 Review rules: judge and lens tools
- Playtest mode F5: thread chip with an End button #889 Playtest thread chip
- Test lanes: a general lane pool and claims rule in fork-ci (R4) #890 Test lanes: general lane pool
- Leak guard: refuse posts and pushes that contain personal details (private pattern file) #892 Leak guard
- Seamless fleet: every prompt from any machine (questions, approvals, secrets, sign-ins) shows up in the thread the owner reads #808 Seamless prompts from every machine
- Agents on your PC: what happens when you play games #222 Agents on a PC while the owner plays games
- head-check: make the required check impossible for a pull request to rewrite (base-pinned job + GitHub App) #34 Required check a PR cannot rewrite
- Model policy: write the Opus 5.5 / Fable 5.2 rows and check picks on the server (#125-#128) #731 Model policy rows
- idea: all the owner's devices reach each other (Tailscale everywhere, one roster, always-on laptop) #891 idea: all devices reach each other
Open owner questions on these: #898 (assistant: chat first, mail later), #900 (homelab kit timing), #903 (device roster or Tailscale only), #904 (computer-use experiments), #102 (fewer tools for a sub-agent).
Suggested order: 1 and 2 first (they make everything else safe to leave running), then 4 (#859 and #869 stop repeats of this audit's misses), then 5 and 6 together, then 7 before such a project starts.
🤖 Generated with Claude Code
Activity