Expand description
@yah:relay(R040, “yah-cloud Phase 1: mirror bootstrap (Hetzner driver, Podman runtime, yah-yubaba)”) @yah:status(review) @yah:parent(Q058) @yah:next(“yah-A track (8 tickets) — substrate for noisetable cloud + future managed camps; gates noisetable C6 (E2E) and D1–D5 (corpus migration)”) @yah:next(“Adjacent relays: R019 (SSH-RPC remote camps), R032 (yah-agentd), R034 (identity registry — hostkeys overlap)”) @yah:next(“Reparented under Q058 cloud quest 2026-05-07. Largely superseded by R091 (yubaba orchestration + integration testing) and R092 (managed-camps CLI + .yah/cloud/ schema) — net new work should land on those relays. Remaining unique R040 scope (Cloudflare Tunnel cookbook F15, pg-on-mesh recipe F16, mesh-promote phasing F18-F19) stays here as historical/operational reference; consider archiving once R091+R092 reach review.”) @yah:next(“Refinement open questions: subcommand framework alignment with yah arch/board, where hostkey fingerprints live, reverse proxy choice. Crate-layout question soft-answered by this stub being a sibling crate; A1 may still refactor into a sub-module of an existing crate if desired (move the //! block + delete this stub).”) @arch:see(architecture/PHASE_1_MIRROR_BOOTSTRAP.md) @arch:see(architecture/yah-managed-camps-topology.md)
@yah:ticket(R040-F1, “A1: .yah/cloud/ schema + parser (MachineConfig, MirrorConfig, ServiceConfig)”)
@yah:at(2026-05-05T00:29:11Z)
@yah:assignee(agent:claude)
@yah:status(review)
@yah:phase(P1)
@yah:parent(R040)
@yah:next(“Per-file TOML layout: machines/
@yah:ticket(R040-F3, “A3: Hetzner driver — MachineProvider trait + hcloud impl (server, bucket, status, destroy)”) @yah:at(2026-05-05T00:29:11Z) @yah:assignee(agent:claude) @yah:status(review) @yah:phase(P1) @yah:parent(R040) @yah:next(“API token from HETZNER_API_TOKEN env; document in .yah/cloud/SECRETS.md”) @yah:next(“Gotcha: Hetzner has TWO APIs (Cloud and Robot) — Phase 1 is Cloud only; Object Storage is S3-compat, separate surface”) @yah:next(“Verify before A6: Hetzner Hillsboro + Ashburn Object Storage both GA; CPX-22 available in PDX/IAD/FSN”) @yah:verify(“cargo test -p cloud provider::hetzner – –ignored (gated; needs token)”) @arch:see(architecture/PHASE_1_MIRROR_BOOTSTRAP.md) @yah:handoff(“MachineProvider trait + HetznerDriver impl complete in crates/yah/cloud/src/provider/. Cloud API (create_server, server_status, destroy_server) uses HETZNER_API_TOKEN via reqwest+bearer. create_bucket uses Hetzner Object Storage S3-compat with AWS Sig V4 signing (HETZNER_S3_ACCESS_KEY + HETZNER_S3_SECRET_KEY). Location enum maps pdx/iad/fsn to hetzner cloud IDs and S3 endpoints. 9 unit tests pass; 2 integration tests behind #[ignore] gate. .yah/cloud/SECRETS.md created documenting all three token types.”) @yah:next(“A2: wire yah cloud subcommand stub into CLI (yah board/arch pattern). All commands no-op except machine status + mirror show which dump parsed .yah/cloud/ config. Verify: yah cloud mirror show noisetable prints declared regions/services.”)
@yah:ticket(R040-F4, “A4: cloud-init YAML + yah cloud machine provision (Debian + Podman + Tailscale + yah-yubaba)”)
@yah:at(2026-05-05T00:29:11Z)
@yah:status(review)
@yah:phase(P1)
@yah:parent(R040)
@yah:next(“Template at .yah/cloud/cloud-init/mirror.yml; renderer substitutes YAH_WARDEN_BASE64 + HEADSCALE_PREAUTH_KEY + TAGS per machine”)
@yah:next(“Hostkey generated ON the machine (ssh-keygen in cloud-init); private half never touches dev machines”)
@yah:next(“Provision command polls cloud-init done, reads back hostkey fingerprint via yah-yubaba /identity, writes back to machine.toml”)
@yah:next(“Surface /var/log/cloud-init.log and /var/log/cloud-init-output.log on failure”)
@yah:verify(“yah cloud machine provision noisetable-pdx-1 –dry-run prints rendered cloud-init YAML”)
@arch:see(architecture/PHASE_1_MIRROR_BOOTSTRAP.md)
@yah:handoff(“Cloud-init template + renderer + provision command landed. Template at .yah/cloud/cloud-init/mirror.yml (canonical) with embedded copy at crates/yah/cloud/templates/mirror.yml (drift-test in cloud_init::tests::embedded_template_matches_workspace_canonical). Renderer at crates/yah/cloud/src/cloud_init.rs substitutes {{KEY}} (no-spaces) for MACHINE_NAME, YAH_WARDEN_BASE64, HEADSCALE_PREAUTH_KEY, TAGS; doc comments use {{ KEY }} (with spaces) so they survive rendering. find_unsubstituted() trips on any unsubstituted real placeholder. Provision orchestrator at crates/yah/cloud/src/provision.rs decoupled from concrete provider (takes &dyn MachineProvider). CLI wired in app/yah/cli/src/cloud.rs: yah cloud machine provision
@yah:ticket(R040-T5, “A5: yah cloud machine status — drift report (declared vs Hetzner vs agent health)”)
@yah:at(2026-05-05T00:29:11Z)
@yah:status(review)
@yah:phase(P2)
@yah:parent(R040)
@yah:next(“Drift categories: missing-server, wrong-machine-type, missing-bucket, hostkey-fingerprint-mismatch, missing-co-tenant, agent-unreachable, service-unhealthy”)
@yah:next(“Exit non-zero on drift unless –quiet; bare ‘yah cloud machine status’ dumps all declared machines”)
@yah:next(“Can land any time after A4 (parallel-safe with A6/A7/A8)”)
@yah:verify(“Pre-A6: reports ‘not provisioned’ for each declared machine”)
@yah:verify(“Post-A6+A7+A8: all three report ‘in sync’”)
@arch:see(architecture/PHASE_1_MIRROR_BOOTSTRAP.md)
@yah:handoff(“Drift module landed at crates/yah/cloud/src/status.rs (DriftFinding + MachineReport + collect_machine_report). MachineProvider trait gained find_server_by_name (GET /servers?name=) + bucket_exists (S3 HEAD) — Hetzner impls share a refactored sign_s3_empty_body helper for PUT/HEAD signing. CLI: yah cloud machine status [name] [--quiet] queries Hetzner, prints declared+live state with per-finding [drift]/[note ] tags, and bails non-zero on real drift unless –quiet. Soft findings (missing creds, A8-not-landed, bucket-uncheckable) are surfaced as notes only. Pre-A6 verify: smoke-tested against this workspace’s noisetable-pdx-1 declaration with no token — reports ‘unknown’ + provider-error note, exit 0. 7 new drift unit tests pass via in-memory FakeProvider; 27 total pass in crates/yah/cloud.”)
@yah:next(“A8 will plug in real agent-ping (yah-yubaba /health + /identity). The HostkeyFingerprintMismatch + AgentUnreachable variants are wired through the report rendering — A8 just needs to flip the AgentUnreachable emit-site in collect_machine_report() from the always-emit pre-A8 stub to a live probe.”)
@yah:ticket(R040-T6, “A6: provision noisetable-{pdx,iad,fsn}-1 (operational; yah dogfoods on noisetable’s mirror, no yah-cloud SaaS)”)
@yah:at(2026-05-05T00:29:11Z)
@yah:status(review)
@yah:phase(P2)
@yah:parent(R040)
@yah:verify(“yah cloud machine status –path /Users/user/ss/noisetable — all three ‘in sync’”)
@yah:verify(“aws s3 ls –endpoint-url yah cloud machine status --path /Users/user/ss/noisetable → all three ‘in sync’) which is gated on the release-tag + provision steps.”)
@yah:ticket(R040-F7, “A7: yah cloud service deploy — Podman compose generation + Cloudflare front + tier isolation”)
@yah:at(2026-05-05T00:29:11Z)
@yah:assignee(agent:claude)
@yah:status(review)
@yah:phase(P3)
@yah:parent(R040)
@yah:next(“Resolves service configs + mirror declarations → per-machine compose.yml → push via yah-yubaba (A8) → systemctl restart yah-cloud-services”)
@yah:next(“Tier isolation at SOFTWARE layer: separate PG roles per service set, distinct Headscale tags (tier:t0 yah meta vs tier:t2 noisetable assets), mesh_only flag controls Cloudflare exposure”)
@yah:next(“Cloudflare DNS records (manual, in SECRETS.md): pdx.cloud.noisetable.example etc. → public IPs via orange-cloud”)
@yah:next(“Gotcha: Cloudflare free tier proxies HTTP/HTTPS but NOT raw TCP/UDP — Headscale mesh uses direct WireGuard between machines”)
@yah:next(“Assumes: Cloudflare account exists + noisetable.example DNS delegated to it (else spawns Cloudflare-bootstrap sub-ticket)”)
@yah:next(“Open question (refinement): Caddy vs nginx vs direct-to-Cloudflare for per-machine reverse proxy”)
@yah:verify(“curl https://pdx.cloud.noisetable.example/healthz (asset-registry) returns 200”)
@yah:verify(“curl https://pdx.cloud.yah.example/healthz (yah meta-directory) returns 200”)
@yah:verify(“yah cloud service status –all — all green on all 3 machines”)
@arch:see(architecture/PHASE_1_MIRROR_BOOTSTRAP.md)
@yah:handoff(“Compose generation + yubaba service deploy + CLI wiring landed. New file: crates/yah/cloud/src/compose.rs (generate_compose_bundle: tier-named networks, Caddy service for mesh_only:false, Caddyfile with hostname when cloud_domain is set in mirrors/*.toml). MirrorConfig gained cloud_domain: Option
@yah:ticket(R040-F8, “A8: yah-yubaba (Rust binary on machine) + yah-cloud-client + desktop Cloud panel”)
@yah:at(2026-05-05T00:29:11Z)
@yah:status(review)
@yah:phase(P3)
@yah:parent(R040)
@yah:verify(“cargo test -p yubaba -p cloud -p cloud-client + cargo test -p yah –bin yah cloud:: — 18 + 47 + 9 + 17 tests pass”)
@yah:verify(“cargo run -p yubaba – serve –bind 127.0.0.1:7450 –state /tmp/i.json then curl POST /register-hostkey with a real ssh-ed25519 pubkey, then yah cloud machine attach <machine> --host 127.0.0.1:7450 --wait 5 --path <camp> prints attached: <machine> (fingerprint recorded: SHA256:…) and writes the line into .yah/cloud/machines/yah cloud machine attach <name> (the hostkey write-back follow-on). app/yah/cli/src/cloud.rs::handle_attach loads MachineConfig, polls cloud-client GET /identity until success or –wait seconds elapse (default 180s, retry every 3s on NotRegistered or Transport errors), then calls MachineConfig::save() with the fingerprint. Decision logic in decide_attach_action covers all four states: Write (no fingerprint on file), Idempotent (match — no-op write), Mismatch (bail with –force hint), Overwrite (mismatch + –force). Subcommand args harmonized with agent ping: –host (default http://next: hostkey fingerprint write-back lands with A8) replaced with next: once cloud-init finishes, run yah cloud machine attach <name>. End-to-end smoke verified all four AttachAction paths against a live yubaba on 127.0.0.1:7450 — pre-registration polling + bail, fresh-write, idempotent re-run, mismatch refusal, –force overwrite. 5 new unit tests (decide_attach_action_{writes_when_no_declared,idempotent_on_match,bails_on_mismatch_without_force,overwrites_with_force,force_is_idempotent_when_match}) — total 9 cloud:: tests in the yah bin. Cumulative R040-F8 deliverables across sessions 1–3: yubaba binary (12 tests), cloud-client crate (6 tests + doctest), AgentProbe trait + drift wire-up (4 new in cloud, 29 total), yah cloud agent {ping,services,logs} CLI, yah cloud machine attach, provision next-step hint correctly points at attach. Out of scope still: Tailscale-mesh discovery, desktop Cloud panel, mTLS hardening.”)
@yah:next(“Tailscale-mesh discovery for –host: yah cloud agent ping/services/attach defaults to http://tailscale status --json for the machine’s mesh IP. Until then, operators pass –host explicitly. Add a TS_DEVICES env shortcut or yah cloud agent ip <machine> helper if the manual path stays painful.”)
@yah:next(“Desktop Cloud panel (out of session, larger): packages/yah/ui/src/cloud/ — machines list, per-machine service list, log tail. Tauri command bridges in app/yah/desktop/src/. Pulls on the cloud-client crate (already in workspace). Defer until at least Hostkey write-back lands so the panel has something useful to display per-machine.”)
@yah:next(“Hardening parking lot (post-Phase-1): mTLS (server cert from machine hostkey, client cert from desktop user identity); /metrics endpoint (Prometheus); streaming /logs from journald with heartbeat. None of these gate Phase 1 verify; cloud-client’s CloudClient::with_timeout exists so log streaming has a hook.”)
@yah:gotcha(“yubaba binds 0.0.0.0:7443 by default; cloud-init template adds ufw rules to restrict to tailscale0. If a future deployment skips ufw, public IP exposure is a real risk. Better long-term: bind to tailscale0 IP directly (cloud-init systemd unit can compute it via ip -4 addr show tailscale0).”)
@yah:gotcha(“Re-running register-hostkey with a different pubkey replaces the stored identity (idempotent for same key, swaps for different). Probably what you want, but worth flagging if someone designs an audit log around ‘first registration wins’.”)
@yah:gotcha(“Default URL resolution (http://<machine.name>:7443) is shared between WardenHealthProbe (status drift), agent ping/services, and machine attach — outside Tailscale mesh every default-host call against a real Hetzner machine will fail. Status emits AgentUnreachable as a soft note; agent/attach surface the transport error directly. Once mesh discovery lands, all three call sites should resolve via Tailscale before falling back to plain hostname.”)
@yah:ticket(R040-F18, “Phase 1a — yah mesh start: camp bootstrap with stable URL from day 1”)
@yah:at(2026-05-05T00:29:11Z)
@yah:status(review)
@yah:assignee(agent:claude)
@yah:parent(R040)
@yah:phase(P1)
@yah:next(“Implements Phase 1a from .yah/docs/architecture/A041-yah-mesh-bootstrap.md — design committed 2026-05-04. Phase 1b lands as R040-F19, Phase 2 (HA) as R040-F20+F21.”)
@yah:next(“Subcommand surface: yah mesh start brings up Headscale locally (yah-desktop or camp daemon); auto-detects direct vs. cloudflared reachability; configures DNS for mesh.mesh-url so provision picks it up.”)
@yah:next(“Cloudflared bundling: yah-desktop should manage the cloudflared binary so the operator doesn’t need a separate install. Same daemon can carry yah serve ingress + Headscale tunnel under one install — that’s the user-acceptable story.”)
@yah:next(“Stable URL contract: mesh.<your-domain> set ONCE here and never changes across Phase 1b promotion or Phase 2 leader changes. Nodes embed this URL in their tailscale config; downstream phases re-point DNS, not nodes.”)
@yah:next(“Headscale binary provisioning: download + manage like cloudflared (precedent: yah already shells out to cloudflared/tailscale/cargo). Config template per arch doc ‘Headscale config’ section. Initial ACL writes a permissive tag:* policy.”)
@yah:next(“Provision integration: when mesh-url is set, yah cloud machine provision calls Headscale API for a single-use preauth key and embeds {MESH_URL, PREAUTH_KEY, TAGS} into cloud-init’s tailscale up line. Static headscale-preauth-key in vault becomes optional.”)
@yah:next(“Also lands yah mesh status (coordinator location, connected nodes, preauth count). yah mesh backup/restore can be Phase-1a stubs; full litestream comes with R040-F21.”)
@yah:verify(“yah mesh start on a fresh camp brings up Headscale + DNS, prints stable URL, stores in vault”)
@yah:verify(“yah cloud machine provision
@yah:ticket(R040-F19, “Phase 1b — yah mesh promote: migrate Headscale from camp to first cluster machine”)
@yah:at(2026-05-05T00:29:11Z)
@yah:assignee(agent:claude)
@yah:status(review)
@yah:phase(P2)
@yah:parent(R040)
@yah:next(“Implements Phase 1b from .yah/docs/architecture/A041-yah-mesh-bootstrap.md. Depends on R040-F18 (Phase 1a) shipping the stable-URL contract and camp Headscale.”)
@yah:next(“Subcommand: yah mesh promote
@yah:relay(R085, “Infra-info MCP — yubaba / machine / mirror status read surface”) @yah:assignee(agent:claude) @yah:status(review) @yah:parent(Q082) @yah:handoff(“Four read-only cloud.* MCP tools landed in crates/yah/agent-tools/src/cloud_tools.rs. Registration: KgToolRegistry::with_cloud() in tools.rs (same builder pattern as with_board/with_scryer/with_task). cloud-client dep added to agent-tools Cargo.toml. Local-config tools (cloud.machines, cloud.mirror_state) read .yah/cloud/{machines,mirrors,services}/.toml from ctx.camp_root with lightweight serde structs — no cloud crate dep, no network. Network tools (cloud.warden_status, cloud.service_ports) probe yubaba HTTP API via cloud_client::CloudClient and return reachable:false rather than error when unreachable. 14 unit tests, all pass. Total agent-tools tests: 140.”) @yah:next(“Wire with_cloud() into the agent session startup (camp.rs or agent.rs in app/yah/cli) alongside with_board() and with_scryer() so the tools appear in the agent’s tool list automatically.”) @yah:next(“Feed cloud.warden_status into the drift detection surface (status.rs AgentProbe) so yubaba reachability populates machine status reports rather than being a separate probe path.”) @yah:verify(“cargo test -p agent-tools – cloud 2>&1 | grep ‘test result’ shows 14 passed; 0 failed”) @yah:verify(“KgToolRegistry::standard_read_only(ctx).with_cloud().schemas() returns 4 tools all named cloud.”) @arch:see(.yah/quests.md)
@yah:relay(R168, “yah-cloud TOML config schema + validation”)
@yah:assignee(agent:claude)
@yah:at(2026-05-13T19:07:07Z)
@yah:status(review)
@yah:parent(Q156)
@yah:handoff(“Schema doc written at .yah/docs/architecture/A031-yah-cloud-config-shape.md. MirrorConfig.camp renamed to .camp with serde alias ‘camp’ for migration compat (R137 rename). load_mirrors() added to handle both flat mirrors/
@yah:relay(R260, “OpenRouter model almanac — track top-weekly models for backend selection”)
@yah:assignee(agent:claude)
@yah:at(2026-05-20T22:23:52Z)
@yah:kind(spike)
@yah:status(in-progress)
@arch:see(architecture/yah-cloud-config-shape.md)
@yah:next(“Unique signal: popularity + freshness + capability tags. Pricing is already covered by crates/yah/probe/data/pricing.json (LiteLLM snapshot) — almanac does NOT duplicate $/token; it adds rank, :free flag, modality, capability tags (code / vision / tool-use / long-context), and a last_seen_at so consumers can tell when an entry has gone stale.”)
@yah:next(“Primary consumer: the OpenRouter character bundle at crates/yah/kg/preroll/openrouter/{agents,subclasses}.json. Currently a single preroll agent ("Quick Hand") pinned to qwen/qwen3-coder:free — this WILL silently break when OpenRouter rotates the free-tier list. Almanac’s first job is to keep that subclass pointed at a currently-available free model (and let the bundle expand into a roster of currently-hot free models, not just one).”)
@yah:next(“Free-model rotation is the load-bearing constraint. Refresh cadence has to match how often OpenRouter shuffles the :free set (empirically weekly-ish). Storage options ranked by how well they handle rotation: (1) scheduled daemon refresh writing back into crates/yah/kg/preroll/openrouter/ — auto-heals; (2) on-demand fetch+cache in the runner — never stale at session start but offline-fragile; (3) build-time embedded snapshot like probe/data/pricing.json — simplest, goes stale between releases (probably wrong shape for free-tier).”)
@yah:next(“Two approaches to evaluate side-by-side: (A) straight call to https://openrouter.ai/api/v1/models (free, no key, structured JSON — already returns context_length, pricing.prompt/pricing.completion, architecture.modality; crates/yah/runner/src/resolver/openrouter.rs already talks to OpenRouter so a fetch_models() sibling to its existing GET /api/v1/credits is the natural home); (B) inference-based tool that distills https://openrouter.ai/models?order=top-weekly HTML into our schema. Confirm (A) first — if the JSON exposes the top-weekly ordering and :free flag, (B) is unnecessary; if not, (B) becomes the rotation-detector.”)
@yah:next(“Query-resolvable subclass (the real primitive): AgentSubclass.model at crates/yah/kg/src/party.rs:520 is a hard provider:model string today. Promote it to a sum type — literal provider:model OR ModelQuery { tier, capability_tags, min_context_tokens, max_price_per_token, … } — resolved by the almanac at session start. Slots into the existing fallbacks: Vec<FallbackRule> machinery at party.rs:540: emit ConfigSwitch{ModelResolved} (sibling of FallbackTriggered) when the query lands, so the UI can show "Quick Hand → qwen3-coder:free (current pick, refreshed 2h ago)". The preroll bundle then declares intent ("Thief-shaped, free, code-tuned") instead of a brittle pin.”)
@yah:next(“Constraint vocabulary — the hard numbers behind "Thief-shaped, free, code-tuned", layered by cost-to-implement so we ship something useful at T0 and can grow. T0 (already free from /models): context_length, pricing.prompt/pricing.completion, architecture.modality come back in the same fetch. MVP heuristic = Pareto frontier of (context_window, price_per_token); at :free tier price collapses to 0 so it cleanly reduces to "max context with matching capability tag". This is enough to pick a non-broken Quick Hand. T1 (external quality scores): join on canonical model id against artificialanalysis.ai’s quality index (general + code subscore), Aider’s code leaderboard, or LiveBench. Cheap to add once the T0 pipe exists; lets queries express min_quality or prefer_code_subscore. T2 (almanac-as-evaluator): spend a few credits running a small fixed corpus (lint-fix, summarize, ack-and-route) through each new free candidate, score against a reference, persist scores into the almanac. The dogfood option — most signal, most expensive, most interesting. Out of scope for the spike; file as a follow-up child once T0/T1 are in flight.”)
@yah:next(“Recommendation surface: popularity-ranked picker entries in packages/yah/ui/src/components/agent/Picker/toPickerAgents.ts + AgentProvidersPanel.tsx so adding an OpenRouter backend surfaces the currently-hot models first; query-resolved subclasses display their resolved pick + a freshness chip; stale literal-pinned :free references show a warning.”)
@yah:next(“Spike output: a design note at .yah/docs/working/W091-openrouter-almanac.md + a working (A) prototype printing the top 10 models in our T0 schema (id, context_length, price, :free flag, capability tags, popularity rank). From there, refine into child tickets for: (i) refresh mechanism, (ii) ModelSpec::Literal | ModelSpec::Query sum type + T0 resolver, (iii) preroll-bundle migration from pinned-model to ModelQuery, (iv) UI recommendation + freshness chips, (v) T1 external-benchmark join, (vi) T2 self-eval as a stretch/follow-up.”)
@yah:handoff(“Spike output delivered: working prototype + design note + 6 child tickets. Prototype lives in crates/yah/runner/src/resolver/openrouter.rs (AlmanacEntry struct + fetch_models() against /api/frontend/models/find?order=top-weekly + derive_capability_tags) plus app/yah/cli/src/agent.rs (yah agent almanac –top N –free-only –tag code –json). Design note at .yah/docs/working/W091-openrouter-almanac.md. Six follow-ups filed: R260-F1 (refresh scheduler), R260-F2 (ModelSpec sum type + T0 resolver), R260-T3 (preroll bundle migration → ModelSpec::Query), R260-F4 (UI freshness chips + stale-pin warning), R260-F5 (T1 benchmark join), R260-S6 (T2 self-eval, stretch). T0 prototype confirms the rotation problem: yah agent almanac --top 10 --free-only --tag code shows qwen/qwen3-coder:free (Quick Hand’s current pin) is NO LONGER in the top-10 free code models — current leaders are openrouter/owl-alpha, nvidia/nemotron-3-super, poolside/laguna-m.1.”)
@yah:verify(“cargo run -p yah – agent almanac –top 10 — prints OpenRouter top-10 weekly models in T0 schema”)
@yah:verify(“cargo run -p yah – agent almanac –top 10 –free-only –tag code — shows the currently-hot free-tier code-tuned roster (the Quick Hand replacement set)”)
@yah:verify(“cargo check -p runner -p yah — clean (note: cargo test -p runner –lib is blocked by pre-existing R258-F3 test breakage on AgentSession; tracked as gotcha)”)
@yah:gotcha(“Pre-existing R258-F3 test breakage in crates/yah/runner/src/{anthropic.rs:833, sessions.rs} — cargo test -p runner --lib fails with E0063 (AgentSession initialiser missing last_text + slot_id fields). R258-F3 is currently in review. R260’s new resolver::openrouter::tests entries are byte-perfect but un-runnable until R258-F3’s test fixtures are updated.”)
@yah:gotcha(“OpenRouter’s /api/frontend/models/find?order=top-weekly is undocumented. The almanac treats it as best-effort and the design (see openrouter-almanac.md) calls for graceful degradation to public /api/v1/models alone when the join fails. ?order= is silently ignored on the documented endpoint (it always returns newest-first); approach (B) HTML scraping is unnecessary and was ruled out.”)
@yah:relay(R414, “yah-cloud pill sync + publish + Pills panel”) @yah:assignee(bundle-anthropic-miravel) @yah:at(2026-06-03T07:07:03Z) @yah:status(review) @yah:phase(P2) @yah:parent(Q410) @arch:see(.yah/docs/working/W141-pill-rings.md) @arch:see(.yah/docs/architecture/A031-yah-cloud-config-shape.md) @yah:depends_on(R411) @yah:handoff(“P2 delivered: resolve_pill_catalog (chips crate) + party_resolve_pill_catalog Tauri command + env wiring + PillsPanel in FullCharacterEditor. Pills panel shows all categorized pills grouped by category with per-character tri-state toggle (always-on/discoverable/off); edits agent.persona.chips.pill_state map via patchAgent.”) @yah:next(“P3: local pills.toml loader — add ~/.yah/pills.toml and .yah/pills.toml as new resolution layers (separate from chips.toml) per W141 layer stack”) @yah:next(“P3: MCP URL trust display — when a content-pill with mcp[] is toggled always-on, show one-time approval dialog listing the URLs before activating”) @yah:next(“P4: cloud sync stub — ‘Share to my account’ button in PillsPanel (disabled until account system exists)”) @yah:verify(“cargo test -p chips — 84 passed (3 new resolve_pill_catalog tests)”) @yah:verify(“bun run typecheck — no new errors (6 pre-existing unchanged)”)
@arch:see(.yah/docs/working/W193-asset-dependency-status-surface.md)
@yah:ticket(R743-T7, “yah-cloud: 5 test binaries to 1 (or 2 — pond_smoke may stay isolated)”)
@yah:at(2026-08-11T01:18:56Z)
@yah:status(review)
@yah:phase(P2)
@yah:parent(R743)
@yah:next(“tests/main.rs mod’ing the siblings + autotests = false and [test] name = "main" in oss/yubaba/crates/cloud/Cargo.toml.”)
@yah:verify(“cargo test -p yah-cloud – –list count unchanged; three green runs. One commit — oss subtree.”)
@yah:gotcha(“tests/pond_smoke.rs:198 is the only real fixed-port bind in the whole in-scope set — MinIO on http://127.0.0.1:9000, an external dependency it does not own. It is a legitimate deliberate-isolation candidate: if merging it makes the suite flaky or order-dependent, keep it as its own [test] target and say so on the ticket rather than force-merging.”)
@yah:gotcha(“pond_smoke.rs and mesofact_static_e2e.rs both host live @yah: annotations.”)
@yah:handoff(“LANDED: 5 test binaries → 2. New tests/main.rs mods live_workspace_smoke, mesofact_static_e2e, pg_driver_live and whisper_derive_e2e; Cargo.toml gains autotests = false plus explicit [test] main (tests/main.rs) and [test] pond_smoke (tests/pond_smoke.rs). Files were mod’d, not concatenated, so mesofact_static_e2e.rs’s live R441-B2 annotation block stays where the harvester expects it and CARGO_MANIFEST_DIR (which three of the four walk up from) is unchanged. autotests = false only gates [test] discovery — examples/load_probe.rs is still autodiscovered. Test names gained a module prefix (e.g. whisper_derive_e2e::derive_pipeline_upload_skip_prune_reproducibility); substring filters still match, but --test <file-stem> for the four merged files is now --test main -- <module>::.”)
@yah:handoff(“ISOLATION DECISION: pond_smoke STAYS its own [test] target — 2 targets, not 1, and deliberately so. Three reasons, strongest first. (1) DECISIVE: .yah/qed/pond-smoke.toml’s pond-spinup-budget step shells out to cargo test --release --locked -p cloud --test pond_smoke -- --nocapture. Folding pond_smoke into main turns that committed pipeline step into a hard cargo error, and that file is outside this crate (peer-owned path, not mine to edit). (2) It is the only test in the crate that reaches a FIXED external address — MinIO at http://127.0.0.1:9000 (pond_smoke.rs:199) — and the only one that creates/docker rm -fs named containers in a Drop guard, so its blast radius on failure is outside the process and a standalone runnable handle is worth keeping. xtask/src/lib.rs’s @yah:assumes census independently found this to be the ONLY real port bind across all ten in-scope crates. (3) It is benchmark-shaped and libtest parallelises within one binary: both its tests assert wall-clock budgets (WARM_RESTART_BUDGET 3s, WARDEN_COLD_BUDGET 5s, COLD_START_BUDGET 15s). Merging would drop whisper_derive_e2e (two in-process axum servers + BLAKE3 + tempdir IO, ungated) and live_workspace_smoke (a full CloudConfig::load walk of the real workspace, ungated) onto sibling threads inside that measurement window, and the pipeline runs it –nocapture for the timing report, which merging would interleave with every other test’s output. The four that DID merge have no such conflict: read-only fs, 127.0.0.1:0 ephemeral ports, tempdirs, and no process-global state (no set_var / set_current_dir / top-level statics anywhere in the set).”)
@yah:verify(“BASELINE (HEAD layout, cargo test -p yah-cloud – –list, run with the change temporarily reverted): 6 test binaries — lib 924 tests; live_workspace_smoke 1; mesofact_static_e2e 1; pg_driver_live 1; pond_smoke 2; whisper_derive_e2e 1; Doc-tests cloud 1. Integration total 6 across 5 binaries.”)
@yah:verify(“AFTER (same command, exit 0): lib 924 tests; tests/main.rs 4 tests; tests/pond_smoke.rs 2 tests; Doc-tests cloud 1. Integration total 6 across 2 binaries — COUNT UNCHANGED, every one of the 6 names accounted for, now module-prefixed inside main.”)
@yah:verify(“THREE GREEN RUNS: cargo test -p yah-cloud –test main –test pond_smoke, three consecutive times — main 3 passed / 0 failed / 1 ignored, pond_smoke 2 passed / 0 failed, all three runs identical. No order-dependence observed.”)
@yah:verify(“HONEST SCOPE OF WHAT ACTUALLY EXECUTED (confirmed with – –nocapture, container has no docker / no live pg / no mesofact-dev binary): REALLY RAN — live_workspace_smoke::live_yah_workspace_loads_cleanly (real CloudConfig::load of the checked-in .yah/ tree, real assertions) and whisper_derive_e2e::derive_pipeline_upload_skip_prune_reproducibility (in-process axum fake-S3 + fake upstream on 127.0.0.1:0, all three runs). SELF-SKIPPED at runtime — mesofact_static_e2e (prints ‘skipping: set YAH_RECONCILER_E2E_BIN’), pond_smoke::pond_spinup_budget and pond_smoke::warden_container_spinup_budget (both print ‘SKIP: set YAH_LOCAL_SIM_E2E=1’, need orbstack/colima/docker). IGNORED — pg_driver_live (#[ignore], needs a built yah-pg-dev binary + network). So the live/e2e legs were compiled and linked but NOT exercised here; the pipeline steps that do exercise them are unchanged for pond_smoke and reachable as --test main -- mesofact_static_e2e:: for the merged one.”)
@yah:gotcha(“PRE-EXISTING, NOT CAUSED BY THIS TICKET: .yah/qed/local-sim-smoke.toml’s only step runs cargo test -p cloud --test local_sim_smoke, but tests/local_sim_smoke.rs does not exist and git log --all finds it at neither oss/yubaba/crates/cloud/tests/ nor the pre-OSS crates/yah/cloud/tests/ path. That pipeline was already dangling before this change (it most likely wants pond_smoke). Left alone: .yah/qed/ is outside this crate. Flagging it because autotests = false makes a target-name typo fail identically to this, and the next person to hit it should not blame R743-T7.”)
@yah:gotcha(“Historical @yah:verify strings on already-landed tickets still cite the retired per-file target names — --test whisper_derive_e2e (reconciler/static_asset.rs:46,118), --test mesofact_static_e2e (the R441-B2 block inside tests/mesofact_static_e2e.rs itself). Those are landed records, not live invocations, so they were deliberately NOT rewritten; the working form is now cargo test -p yah-cloud --test main -- <module>::.”)
@yah:gotcha(“CONTAINER, not code: the first attempt at the post-change –list died with error: linking with cc failed … ld terminated with signal 7 [Bus error] on both the merged main target and the untouched lib test target. Root cause was the container’s root filesystem at 100% (8.0K free) — oss/yubaba/target alone was 23G with no cargo-orphan-gc installed here to reclaim it. Cleared by deleting the regenerable incremental caches (oss/yubaba/target/debug/incremental, target/debug/incremental) while no cargo was running; the listing then exited 0. Worth knowing because the failure mode reads exactly like the R770 orphan-gc symptom described in CLAUDE.md but is plain ENOSPC.”)
@yah:tier(Warrior)
Re-exports§
pub use almanac_dispatch::dispatch_on_change;pub use asset_journal::AssetState;pub use asset_journal::AssetStatusEvent;pub use asset_journal::AssetStatusJournal;pub use capability::Capability;pub use compose::generate_compose_bundle;pub use compose::ComposeBundle;pub use config::BucketLogEntry;pub use config::CampCloudDbs;pub use config::CloudConfig;pub use config::CloudDb;pub use config::ConnectSpec;pub use config::DbCatalog;pub use config::DevDb;pub use config::GitSource;pub use config::IngressDecl;pub use config::IngressEdge;pub use config::IngressProvider;pub use config::LegacyMirrorConfig;pub use config::LegacyServiceConfig;pub use config::MachineConfig;pub use config::MirrorAssignment;pub use config::MirrorConfig;pub use config::MirrorProviderSlot;pub use config::MirrorShape;pub use config::PondDb;pub use config::PondDbKind;pub use config::Provider;pub use config::ProviderConfig;pub use config::ServiceComponent;pub use config::ServiceConfig;pub use config::TopologyConfig;pub use config::WorkloadConfig;pub use config::WorkloadConfigError;pub use local_driver_glue::local_container_spec_from_provider;pub use provider::on_ingress_owner_changed;pub use provider::reconcile_assignment;pub use provider::BucketAcl;pub use provider::BucketRef;pub use provider::CfAccountInfo;pub use provider::CloudflareClient;pub use provider::CloudflareEnvoy;pub use provider::CreateR2BucketResult;pub use provider::CreateTokenResult;pub use provider::CreateTunnelResult;pub use provider::DigitalOceanEnvoy;pub use provider::FloatingIpAssignOutcome;pub use provider::FloatingIpProvider;pub use provider::FloatingIpState;pub use provider::FloatingIpTarget;pub use provider::GrantScope;pub use provider::HetznerDriver;pub use provider::HetznerEnvoy;pub use provider::HetznerFloatingIp;pub use provider::Location;pub use provider::MachineProvider;pub use provider::OvhFloatingIp;pub use provider::ProjectId;pub use provider::R2BucketInfo;pub use provider::R2CustomDomain;pub use provider::ServerId;pub use provider::ServerSpec;pub use provider::ServerStatus;pub use provider::ServerSummary;pub use provider::TokenGrant;pub use provider::TunnelConnState;pub use provider::TunnelDnsRecord;pub use provider::TunnelDriftRow;pub use provider::TunnelDriftState;pub use provider::VultrFloatingIp;pub use provider::WorkerDeployResult;pub use provider::MESOFACT_STATIC_GRANTS;pub use reconciler::collect_live_derive_hashes;pub use reconciler::compute_derive_cache_candidates;pub use reconciler::compute_live_set;pub use reconciler::compute_prune_candidates;pub use reconciler::compute_service;pub use reconciler::derive_minio_key;pub use reconciler::execute_derive_cache_prune;pub use reconciler::execute_prune;pub use reconciler::load_service_and_mirror;pub use reconciler::mesofact_static::WORKER_SCRIPT;pub use reconciler::new_sync_id;pub use reconciler::pond::MINIFLARE_SIM_SCRIPT;pub use reconciler::publish_to_pond;pub use reconciler::summarize;pub use reconciler::CellStatus;pub use reconciler::CloudflareWorkerReconciler;pub use reconciler::ContainerOptions;pub use reconciler::ContainerReconciler;pub use reconciler::DeriveCacheLiveHashes;pub use reconciler::DerivePruneCandidate;pub use reconciler::DriftEntry;pub use reconciler::HeadscaleReconciler;pub use reconciler::HealthState;pub use reconciler::LocalProcessReconciler;pub use reconciler::LocalStaticOptions;pub use reconciler::MesofactStaticReconciler;pub use reconciler::MirrorObservation;pub use reconciler::PondOptions;pub use reconciler::PondPublishReport;pub use reconciler::PondState;pub use reconciler::ProviderScope;pub use reconciler::PruneCandidate;pub use reconciler::PruneOutcome;pub use reconciler::PruneReport;pub use reconciler::ReconcileCtx;pub use reconciler::Reconciler;pub use reconciler::RunningWorkload;pub use reconciler::RunningWorkloadSummary;pub use reconciler::Runtime;pub use reconciler::ServiceStatus;pub use reconciler::StaticAssetReconciler;pub use reconciler::StatusSummary;pub use reconciler::SyncHistoryEntry;pub use reconciler::SyncOutcome;pub use reconciler::SyncState;pub use reconciler::WireContainerStatus;pub use status::collect_machine_report;pub use status::AgentProbe;pub use status::DriftFinding;pub use status::MachineReport;
Modules§
- almanac_
dispatch - Dispatch [
almanac::OnChangeConfig] actions after a feed run (R330-F4). - app_
manifest yah-app.tomlschema, parser, closed-setrequired_whenevaluator,.yah/apps.tomlregistry, and workspace discovery (R470-T7).- asset_
journal - Append-only JSONL journal for static-asset reconciler decisions (R470-T1).
- asset_
status - Build the W193 workspace asset-status report as a
serde_json::Valuesuitable forcloud.statusRPC responses (R470-T4). - capability
- Service capabilities and the mirror-side driver bindings that satisfy them (W265).
- cloud_
init - Cloud-init template renderer for mirror-machine bootstrap (R040-F4).
- compose
- Podman Compose file generation for yah-cloud service deployments (R040-F7).
- config
- @yah:ticket(R040-F16, “pg-on-mesh service recipe: bind tailscale0 + pg_hba.conf snippet + ufw rules”)
@yah:at(2026-05-05T00:32:34Z)
@yah:assignee(agent:claude)
@yah:status(review)
@yah:parent(R040)
@yah:handoff(“Companion to R040-F15. Inter-node TCP (Postgres primary↔replica, NATS clusters, anything raw-protocol) lives on the Headscale mesh, not on Hetzner public IPs. Each node has a stable 100.64.x.x mesh IP that survives replacement of the underlying box, so DNS / config / pg_hba never churn when a CPX-11 is rebuilt. WireGuard already encrypts the wire — TLS becomes defense-in-depth, not load-bearing. This ticket carries the concrete pg-shaped recipe so the first stateful service deploy doesn’t have to re-derive the pattern; subsequent services (redis, NATS, etc.) cargo-cult from it.”)
@yah:next(“ServiceConfig gains a
bind_interface: Option<String>field (e.g.Some(\"tailscale0\")for mesh-only services). The cloud-init/podman compose renderer translates this into either--network host+pg listen_addresses = '<mesh-ip>'OR a podman macvlan/host-binding pattern that achieves the same.”) @yah:next(“Generated pg_hba.conf snippet: allow the mesh subnet (100.64.0.0/10) for replication + app users. Postgres binds to the node’s tailscale0 mesh IP only —listen_addressesis templated from the node’stailscale ip --4at first boot.”) @yah:next(“Generated ufw rules:ufw allow in on tailscale0 to any port 5432; ufw deny 5432— mirrors the existing yah-yubaba 7443 pattern in mirror.yml. Same shape works for any mesh-only port.”) @yah:next(“Replica connection string uses primary’s mesh IP, NOT its public IP. Stable across box replacement.”) @yah:next(“Out of scope: pg_basebackup orchestration, failover, WAL archiving — those belong in noisetable’s domain; this ticket only standardizes the binding/firewall/auth shape so noisetable’s pg deployment doesn’t reinvent it.”) - envoy
- Envoy framework types —
Tier,AdapterFlavor,VerbCategory,InternalVerb,VerbDescriptor. - identities
- Stub local identity registry:
.yah/cloud/identities/<machine>.json. - local_
driver_ glue - Adapter between cloud’s TOML-driven
ProviderConfigshape andlocal_driver::LocalContainerSpec. - mesh
- Headscale API client and configuration helpers for
yah meshoperations. - mesh_
service - Recipe helpers for stateful services bound exclusively to the Headscale mesh (R040-F16).
- migrate
- W305/R742-F3 — planning a workload’s move between sovereign groups.
- multi_
root - Multi-root cloud config — union sibling
.X/config trees (W206 layout (b)). - paths
- On-disk path resolver for camp deployment manifests.
- proc_
control - The process-control channel — how a yah-supervised process describes itself to its supervisor and to an agent, instead of being guessed at from the outside.
- provider
- @yah:relay(R409, “Envoy — providers and internal verb catalog (W144)”) @yah:at(2026-06-02T20:58:35Z) @yah:status(open) @arch:see(.yah/docs/working/W144-envoy-providers-and-tiering.md)
- provision
- Provisioning orchestrator: config → cloud-init render → MachineProvider call.
- reconciler
- Reconciler abstraction — bring a workload up against a mirror’s provider slots.
- release_
manifest - Yubaba release-manifest fetch + per-triple resolution (R330-F21).
- state
- Reconciler-written sidecar state for declared machines.
- status
- Drift detection between declared
.yah/cloud/config and live cloud state. - topology
- R850 — what the declared topology does when a node dies, and whether it fits on the boxes it is declared against.
- validate
- Workspace-wide lint checks for the
.yah/declaration tree (R470-T3).
Structs§
- Container
RunSpec - Run-time spec for a single container managed by
LocalRuntime::run. - Custom
Docker Host Provider RuntimeProviderfor a rawDOCKER_HOSTstring supplied by the operator (e.g."tcp://localhost:2375"or"unix:///path/to/custom.sock"). Always reported as available — the operator opted in explicitly; failures surface as docker CLI errors rather than probe misses.- Local
Container Spec - Probe spec lifted from a
kind = "local-container"provider TOML. - Local
Docker Runtime WorkloadRuntimeimplementation backed by the docker CLI.- Local
Runtime - Reachable local container daemon — the winning provider from the cascade.
- Owned
Container - A container managed by this module, as observed via
LocalRuntime::list_owned. - Socket
Runtime Provider RuntimeProviderbacked by a Unix socket path. Available iff the tilde-expanded path exists on disk.
Enums§
- Container
State - Container State.Status from
docker inspect. Anything the docs don’t enumerate lands in [Unknown]. - Detected
Runtime - Which runtime answered the probe cascade.
- Runtime
Pref - Operator-declared runtime preference from the provider TOML’s top-level
runtimefield.autowalks the full cascade; any pinned value probes only that runtime.
Constants§
- LABEL_
KEY - Docker label key applied to every container this module owns. The value
is
<service>:<env>:<slot>sodocker ps --filter label=yah.pond=<v>scopes orphan cleanup to a specific mirror. - NAME_
PREFIX - Canonical name prefix for every container managed by this module.
Traits§
- Runtime
Provider - Probe contract for a single Docker-compatible runtime back-end.
Each implementation knows its socket path (or raw
DOCKER_HOST) and can report whether it is currently reachable.
Functions§
- canonical_
label - Build the canonical label value for the same triple.
- canonical_
name - Build the canonical container name for a (service, env, slot) triple.
- warden_
container_ label - Canonical label of the camp’s pond yubaba container:
camp:<env>:yubaba. - warden_
container_ name - Canonical name of the camp’s pond yubaba container for
env:yah-pond-camp-<env>-yubaba. The camp (which starts it) and the desktop probes (which inspect it) MUST agree on this — always go through here.