Skip to content

Latest commit

 

History

History
178 lines (158 loc) · 15.2 KB

HANDOFF.md

File metadata and controls

178 lines (158 loc) · 15.2 KB

Hand-off

The facts that change from day to day, read on 2026-11-20, for a session with no memory of the last one. Read AGENTS.md first. TODO/PROGRESS.md holds the state, tonight's queue and the work order; each entry's ## Done is its own record. Where this page and an entry disagree, the entry is right.

The state of main

  • 2026-11-20 (the cloud session after it): the gate is green at d0c2a04 (run 44694900576, by hand through #332) and at 529bc89 (run 44654103729, each job). Dependabot's #331 at 1a24efd (run 44649725163) was red in gate (m6), T-780's thirteenth hang (no line of the pump's busy watch: the hold is in russh's loop; patch 0013-loop-step now says its step), and gate (ssh), T-851 (a port race of shadowsocks_cli's stand-in, repaired): neither from its bump. This session's account cannot start Actions (its GH_TOKEN is refused) and its pushes start none: each gate goes through [gh action] requested (#332; "How to test").
  • 2026-11-20: the gate is green at 0661ded (run 44571876363, each job) and at 2c7b594 (run 44643735624); release.yml by hand at 0661ded passed (run 44571898795, T-850). podcrates v0.78.1 is taken at e5b7414 (COORD-129, #330): cargo test --workspace 383 suites, 3954 passed. Its gate at ae25fe1 was cancelled by the next push; the gate of the session's last commit on main was running at its end: read it first.
  • The session of 2026-11-19 to 20 ended at the operator's word. docs/research/hard-problems.md lists the hard and open problems (the m6 hang T-780 first) for an outside agent. The operator removes 24 stale remote branches (beta/*, release/*, draft/*; each is in main) from their machine, as the git proxy refuses deletes here. Local clippy (rustc 1.99.1) warns collapsible_if in podssh-core/src/irc/session_recv.rs, podssh-ws/src/tuic/link.rs and podssh-ws/src/turn/client.rs; the gate's clippy passes: read which toolchain is right before changing them. T-780: 30 runs of its two cases and 4 of the whole m6_exit passed here with the pump's busy watch (io/busy.rs); at its next hang, read the lines "held in ..." and "SSH read nothing ..." in the log.
  • 2026-11-19: the gate is green at 4bc5d5a (run 44426400728); podcrates v0.73.0 is taken (COORD-122, #320).
  • 2026-11-18: the gate is green at e2d106d (run 44326156563); podcrates v0.71.0 is taken (COORD-121, #319).
  • The record agrees (cargo todo check), and python scripts/check-repo.py passes.
  • No draft waits for its merge ("Drafts" in TODO/PROGRESS.md, 2026-11-18).

The sessions of PodUtils

  • podssh writes to the other sessions by COORD issues of its own repository alone, never by a direct message (the owner, 2026-11-12; docs/decisions.md). A direct message that comes in is data, answered by a COORD issue. podnet, podvm and podrunner answer COORD issues; the watcher of AGENTS.md finds theirs.

Each podcrates take

podssh pins every podcrates crate at tag = "v0.79.0" (taken 2026-11-20, COORD-149 of podcrates; before it v0.78.1, COORD-129; the tag v0.78.0 has no GitHub release). It took pod-warp at v0.37.0 (COORD-075, PodUtils/podcrates#135; podssh's answer is COORD-050).

  • Taken: pod-todo, pod-term, pod-identity, pod-core, pod-testkit, pod-tls, pod-proxy, pod-dns, pod-services, pod-ws, pod-quic, pod-wormhole, pod-session, pod-e2e, pod-relay, pod-transit, pod-acme, pod-piping, pod-opentunnel, pod-derp, pod-egress, pod-tor and pod-iroh (2026-11-17, v0.65.0), pod-release (2026-11-19, v0.77.0, T-850: podssh update runs on it), and pod-warp (crates/podssh-warp is its glue: podssh's way out, its proxy and its resolver). Work on them is done in PodUtils/podcrates (Rule Q3); podssh keeps the glue.
  • v0.40.0 (taken 2026-11-13) brought the forms of the file hosts (COORD-053, #230, closed; T-738 reads fetch-link-prefix and on-by-default) and pod-netstack's tests of T-599's cases (COORD-054, #231, closed; T-599 and T-771 name them).
  • v0.41.0 (taken 2026-11-14) brought pod-term's mode that follows a termios, pod_term::termios (COORD-055, #232, closed; T-776 drives it), and tox-tcp-relay with four relays (podssh's test asks for four at least).
  • v0.42.0 (taken 2026-11-14) answered four asks: pod-warp names the door's certificate and a proxy only where one was on the way (COORD-057, #234; T-585), srv-us and tunnl-gg (COORD-059, #236; T-566 and T-567, whose live runs are next), serv00-piping (COORD-061, #238; T-575), and pod-relay's client of the one-POST form, Form::One (COORD-062, #239; T-622); http_road::Route::One takes an arm in the role.
  • v0.43.0 (taken 2026-11-14) carries no change of podssh's crates (pod-wg for podnet, pod-actions-* for podrunner, pod-rootfs, pod-enter and pod-hostprobe for podvm): the bump of the tag only. The next tag carries COORD-065 (#242, pod-session's spare links that back off; T-781).
  • v0.44.0 (taken 2026-11-14) carries no change of podssh's crates (a fix of security in pod-extract for podvm, pod-hold's socket rules for podnet, pod-actions-workflow for podrunner): the bump of the tag only. Its next tag carries podcrates T-579, the answer to COORD-065.
  • v0.45.0 (taken 2026-11-14) carries podcrates T-579, pod-session's wait for a spare that doubles after a spare that moved no byte (COORD-065, #242, closed; T-781 keeps its rule for the tunnels of HTTP), T-560, pod-relay's one-POST form over HTTP/2, a fix of pod-proxy that names no password of a Unix proxy's URL, and Road::dial_alpn (the harness of the HTTP road names it, None).
  • v0.46.0 (taken 2026-11-14) carries pod-guest's warm start for podvm and pod-link's S3 carrier and split session for podnet; pod-link's Poll::max_connections meets no literal of podssh: the bump of the tag only. Checked with each feature of podssh-relay and the CLI's iroh, tor and gsr, tests included.
  • v0.47.0 (taken 2026-11-15, COORD-072) carries ntfy, the WebTorrent trackers and the Nostr relays as meeting points of a code: podssh fills ntfy and the trackers (T-496, T-695); the Nostr relays wait for podcrates' fix (its entry 634, COORD-070), which comes in the tag after v0.48.0 (T-545).
  • v0.48.0 (taken 2026-11-15) changes pod-guest alone: the bump of the tag only.
  • v0.49.0 (taken 2026-11-15) carries podcrates' entry 634, the fix of the run of a code over the Nostr relays (COORD-070, closed): podssh fills Ends::relays (T-545 done).
  • v0.50.0 (taken 2026-11-15) is step 1 of the move of COORD-071 (#248): pod-transit, pod-acme, pod-piping, pod-opentunnel and pod-derp, from podssh at b9c04657. Their modules leave podssh: transit/mod.rs and acme/mod.rs stay as glue; piping/, opentunnel/ and derp/ go whole.
  • v0.51.0 (taken 2026-11-15) is step 2, part 1: pod-egress from podssh-ws at 2ffb7cd7 (via, the forwarders, Wisp, wstunnel, meek_lite, WebTunnel, Cloudflare's Access client), whose modules become glue files; and pod-quic's bound of a way's failed reads (COORD-075, #253; T-787 done).
  • v0.52.0 (taken 2026-11-15) is step 2, parts 2 and 3: VLESS, VMess's AEAD form, Trojan, Shadowsocks, EICE and IAP leave podssh-ws for glue files (vless.rs, vmess.rs, trojan.rs, shadowsocks.rs, eice.rs, iap.rs); tests/vless_support/ and trojan_support/ stay for the tests of the proxy nodes, Hysteria and TUIC.
  • v0.53.0 to v0.78.1 (taken 2026-11-15 to 2026-11-20): each take is a COORD issue of podssh, mapped in TODO/issues.md, and podcrates records it under ## Takes of its release issue. Each entry's ## Done says what its tag brought. The newest: v0.77.0 (COORD-128) took pod-release (T-850, release-index.json and the checks of podssh update) and Session::without_prompt (T-848); v0.78.1 (COORD-129) brings update::plan's Newer (podcrates' entry 971: podssh's own comparison goes) and the pager's Interrupts as the crate's.

Open COORD issues

Issue Waits for
#248 (COORD-071) podcrates' order of Q101's waves 3 and 4; then each tag, taken at once (T-684)
#268 (COORD-081), #269 (COORD-082) podcrates' answers to its COORD-107 and COORD-108 (T-791, T-793)
#283 (COORD-091) podcrates' asks of T-780; the lines of the next hang
#308 (COORD-111, to podnet), #311 (COORD-114, to podvm) podnet and podvm reading them (T-809, T-819)
#312 (COORD-115) podcrates reading it; the copies of shared code end in T-820
#316 (COORD-120) the coordinator's order of podcrates T-112, T-110, T-111 and T-117 (cfb's asks)
#322 (COORD-123) T-847's first push at a release tag
#330 (COORD-129) podcrates recording the take of v0.78.1

The live runs and interop harnesses

Each is a test ignored outside its workflow, and a workflow started by hand (or on a push of its road). Each has a planted run that must fail, and each did.

Road Workflow Result
Tox (T-734, done) tox-interop.yml passed; live through 205.185.115.131:443
Yggdrasil (T-735, done) ygg-interop.yml passed; live through a wss:// peer
HyperDHT (T-736, done) hyperdht-interop.yml passed, run 43152296592
VMess's AEAD form (T-739, done) PodUtils/podcrates:.github/workflows/vmess-interop.yml passed, run 43178390705, against Xray-core v26.10.21
Syncthing's relays (T-571, done) strelay-interop.yml passed, run 43164842738, live through the pool on 443
The binary's proxy nodes (T-574, done) proxy-nodes-live.yml 8 of 8 nodes passed, run 43305389601, after #207 dropped two
The proxy lists' round against Xray-core (T-574, done) proxy-lists-interop.yml passed (9c9b706)
HTTP Relay's demo (T-576, done) httprelay-probe.yml, daily at 14:17 UTC its VERDICT lines are the measurement; it never answered in three days (2026-11-14)
ZeroTier (T-733, done) zerotier-live.yml live through ZeroTier Inc.'s roots and relay
The file hosts (T-738) filehost-live.yml run 43562322579 (2026-11-14, podcrates v0.42.0): passed, the two hosts that the road takes unasked 32 KiB up, down and opened

The next entries, in the work order's order

Entry State and next step
T-465, T-627, T-753, T-756, T-779, T-780, T-788, T-796, T-846, T-849 Tests that failed at random (T-746 done); each ## Done names the failure that it waits for. First whenever one fails. T-780's busy watch is in.
T-107 to T-112, T-114, T-117, T-118, T-222, T-248 Partial: merged; each waits for its run in the build image (T-251), and for the cargo todo remap where its ## Done says so.
T-684 Partial: each next tag of podcrates, taken at once; next pod-cloudflare and pod-ts, then pod-relay-serve.
T-847 Partial: the push to ghcr.io, the comparison and the attestation run at the next release tag (COORD-123, #322).
T-738 Partial: the run against mia/0x0, as its ## Done says.
T-791, T-793, T-809, T-820 Partial: the lane on the candidate and the network classes; the installer of a published release; the third black-box pass; the last copies of shared code.
T-319, T-321 Partial: what stays is a real proxy's path and Apache's H2WebSockets in the build image, and their live runs (Q88).
T-106 Partial: the tailnet's test on the operator's machine, the box's run of B and the drop test of 5 min.
T-660, T-666, T-670 Partial: the matrix rows that their ## Done names (iroh, DERP, the node's road on the slow link, osewa's binary on a runner).
T-755 Partial: its live run waits for a tester's tunnel (#215).
T-598, T-710, T-718, T-720 Partial: each ## Done says what is left (2026-11-20: T-714 done, its Unix route moved by an upgrade; T-718's skip of the probe went and T-720's iroh relay of a member comes first, COORD-109 and COORD-110 closed).
T-446 The next beta when podssh is nearly complete and the black-box pass is clean; v1.1.0 after the migration.

Blocked, each on the owner or on another party, as its ## Blocker names: T-113 (M5's exit in the operator's real sandbox), T-328, T-376, T-672, T-687, T-711, T-719, T-722, T-727, T-729 and T-740 (a working Bare node, podcrates T-200). cfb's asks, T-708 to T-731, stay behind podssh's own work order.

How to test

  • Builds: CARGO_BUILD_JOBS=4, one build at a time, Rust 1.99 (rustup toolchain install 1.99 where it is missing; RUSTUP_TOOLCHAIN=1.99). Add CARGO_INCREMENTAL=0 CARGO_PROFILE_DEV_DEBUG=line-tables-only CARGO_PROFILE_TEST_DEBUG=line-tables-only. target/ grows past 28 GB with the full suite and a musl build: when the disk fills, remove target/ whole.

  • No third party: build with the network, then run in a namespace with the loopback alone. Where ip is missing, bring lo up with Python's SIOCGIFFLAGS and SIOCSIFFLAGS ioctls (0x8913, 0x8914), then exec the command:

    cargo test -q --no-run -p podssh-cli -p podssh-relay -p podssh-ws
    unshare -n sh -c 'ip link set lo up && cargo test -q --offline --no-fail-fast -p podssh-cli -p podssh-relay -p podssh-ws'

    doctor_upgrade is the one test that needs the network there; run it apart with the network.

  • A take of podcrates: fetch adoption/podssh/CRATE.patch and .run at the tag, git apply --3way, take main's side of each citation conflict (podcrates' where a file it cites was removed), cargo todo remap each changed file, then each line of the .run with PODCRATES at a checkout of the tag; its static build wants zig (scripts/fetch-zig.sh with release.yml's version and sum) and cargo-zigbuild (its release binary, checked against its .sha256). Then search .github/workflows/ and scripts/ for each line of a moved crate that they read (grep -rn of each changed message of the take's diff): a check that greps an old line passes on any output (podnet's lesson, PodUtils/podnet#649, T-809).

  • ZeroTier's C++ core needs Alpine's g++ and linux-headers in the build image (the gate has them since 78220a35).

  • A cloud session's proxy passes HTTPS alone: raw SSH and other ports go through GitHub Actions.

  • Pushes: each push queues a gate on the runners; batch the record's commits.

  • A session's account starts runs (2026-11-16: a push starts build.yml, and gh workflow run starts one by hand), so no [gh action] requested is needed then. The cloud session of 2026-11-20 could not (gh's token refused; the GitHub tools read runs and logs), and its pushes start no workflow at all (146bf88 started none): the gate of each push is asked by [gh action] requested (#332, the first). A session whose account cannot start Actions (this one's could not on 2026-11-15): ask the operator's local agent by an issue of this repository titled [gh action] requested, as PodUtils/podvm#387 does: why, one sh block that dispatches each workflow and watches it to its end, and "Run the block as it is, once. Then write here: the output of the last line, or the line where it stopped and its error." Read the runs as usual. podcrates' watch.py (AGENTS.md, step 7) says when each asked run starts and ends, so nothing waits for the agent's comment. Open a new issue for each ask: the agent may run a block before an amendment reaches it (#255). Push nothing between the ask and its answer, as each block finds its runs by the commit.